People argue whether we are at y-5, y, or y+5, meanwhile we seem to be on a y=2^x exponential that keeps delivering more and more impressive results.
The most interesting question to me is what will be consumed by the exponential like math seems to be undergoing, and what won’t. Writing has been quite stubborn, but I’ve noticed Fable to be quite a big step up there. How about politics? Will we develop new ways to let people express their own values in democracies, or will we just get much better at manipulation? How about experiment driven domains like biology?
We will get much better at manipulation and better at people “writing” things to justify their own feelings.
What’s new about LLMs is that you can scalably manipulate people individually. It used to be that you could either have scale (speeches, tweets, interviews, website, etc.) or individual engagement (replying to mail/tweets/town hall questions.)
Now you can pull the history and preferences of an individual, then shape a message—in real time—to them, specifically. You can have conversations on social media with a single person and shape your message specifically to them.
Part of this can be good (you talk about what they care about, where 90% of broadcast messaging might not apply) and part of it can be bad (manipulation.)
My guess is that, in the US, the right will cynically adopt manipulation to great effect and the left will take a moral stand against shady practices and lose elections.
> My guess is that, in the US, the right will cynically adopt manipulation to great effect and the left will take a moral stand against shady practices and lose elections.
I think that statement may itself highlight how prevalent manipulation is.
I fully anticipate all groups to continue maximal manipulation they can. One thing with LLMs is that it'll be a far less unified view, so a "divide and conquer" strategy is what I anticipate.
It's like Musks duplicitous argument about unlimited abundance. We have a lot of abundance now, we just keep accelerating it all into the hands of fewer and fewer people - whose response is only to want more, and more, and more.
> The most interesting question to me is what will be consumed by the exponential like math seems to be undergoing, and what won’t. Writing has been quite stubborn,
isn't it clearly split between verifiable not verifiable ? what is interesting about that question.
Programming has verifiable and non-verifiable aspects. Competitive programming, passing tests, and performance can all be verified. But translating English requirements into actual software, software architecture, taste, or UI design cannot. And yet over the last couple years we’ve seen huge lifts in all of these areas, not just the verifiable ones.
Verifiable areas I think are clearly seeing the most improvement, or are the quickest to see improvement. But we are seeing lots of progress in non-verifiable areas as well.
How much of the non-verifiable progress is a function of labs purchasing expert data vs. the models improving with compute is maybe another interesting question, but fundamentally I don’t see spend on expert data as something that can’t grow if AI revenues keep growing as well. And as models get better taste they can also help filter and generate new synthetic data for their next versions to train on. The limits of this approach are not so clear.
That is because there is human annotated data there. Every session you or I used, then of course paid human feedback on repos (such as the recently famous example of meta forcing their employees to).
This is _much better_ data than 1/0 verification, it is as good as a gradient.
Automatically verifiable tasks improve faster since well, its automated.
Today's models depend on inference time compute to get these results. The inference time compute available on any claude subscription is not comparable to the ones used to get some of these results (yes, in this case, it is 2000 USD total as noam confirmed, but some previous results took more).
In general, you can think of the process as generating massive rollouts in generation N, and then compiling in the verifier/human feedback("gradient") signal into generation N+1. The time taken to make the rollout in generation N, and separately the time taken to get the same rollout in generation N+1, each grows constant in some tasks, linear in more, and exponential in some.
In the end, this becomes bottlenecked by time. Today, we can make statements like "I generated all these successful trajectories with 2 weeks of compute, in the next model it will be able to do it in 7 hours of compute", but very soon you'll find yourself making statements like "I generated.... with 8 months of compute, in the next model it can do it in 6 months", which isn't really enticing the same way you can _technically_ brute force passwords but it just needs prohibitive amounts of time and money. That is the "plateau". Note that, this point is quite far away. For example, at any point if we agree it plateaus, today's known hardware techniques such as fixed function accelerators give you a 10-100x timeline reduction immediately allowing for a few more cycles of improvement. This is not to mention future innovations, but of course none of that is helping with the benchmarks where the time needed is growing superlinearly.
In many math and coding benchmarks, we are still in the constant phase. These are the massive improvements we see every few months. I'm not making any prediction of what will plateau and what will not as it's not possible to make an informed prediction about these things IMO. But the observed fact is that some have already plateaud as in, they don't improve with reasonable inference time (likely superlinear growth).
> will we need mathematicians to translate
Let's take a sudoku analogy. The model is initially just doing the random value algorithm, but lets say you the human are watching it. You make one of the usual reductions and interject "hey you can stop trying 8 here because of ....". Over enough examples, you get to a point where the model is _forced_ to learn the logical pattern. Next generation, it will skip that number. After this, you can peak the distribution using simple 1/0 RL. Doing _pure_ 1/0 RL works decent, but its not frontier as its a very sparse signal.
For that lift, human (or even a better LLM, but if you're trying to improve a frontier LLM, there is by definition no better LLM) feedback becomes necessary. This is _why_ it is crucial that these models interface in natural language and is also why the labs are hiring AI tutors by the hundreds. The "better LLM" case is what Kimi etc are doing by "distilling"(bad term for this) claude.
> But the long term is completely bewildering if you believe any of these trends can continue at a similar pace for the next few years.
For math and coding, for now we are in the phase where the times are just ... constant, so there's little reason to think it will stop soon. We still need humans to expand the frontier. It just becomes a matter of if its worth the cost of compute for running this generalized The Algorithm or not.
Given how well chess players internalized _many_ (not all) of alphazero's emergent chess knowledge, I am confident we wont have too much trouble figuring out any new math LLMs come up with, which will let us keep expanding the frontier by giving the LLM the next "lift". Only when we reach the stage where the time growth become exponential will this stop, IMO.
Replace philosophers for mathematicians and Douglas Adams was spot on again.
Whilst current models can't 'intuit' and come up with conjectures, they can certainly disprove some of them very quickly through the kind of grind that humans can't do. I suppose there really are some mathematicians out there today, whose last few years of study, have just been up-ended by this.
--
"Yes we are," insisted Majikthise. "We are quite definitely here as representatives of the Amalgamated Union of Philosophers, Sages, Luminaries and Other Thinking Persons, and we want this machine off, and we want it off now!"
"What's the problem?" said Lunkwill.
"I'll tell you what the problem is mate," said Majikthise, "demarcation, that's the problem!"
"We demand," yelled Vroomfondel, "that demarcation may or may not be the problem!"
"You just let the machines get on with the adding up," warned Majikthise, "and we'll take care of the eternal verities thank you very much. You want to check your legal position you do mate. Under law the Quest for Ultimate Truth is quite clearly the inalienable prerogative of your working thinkers. Any bloody machine goes and actually finds it and we're straight out of a job aren't we? I mean what's the use of our sitting up half the night arguing that there may or may not be a God if this machine only goes and gives us his bleeding phone number the next morning?"
That's how they are finding these solutions though, unless we are just going to label intuition as something only humans can do. Like a submarine being unable to swim or whatever that example is.
* Updates to lower/upper bounds. In many cases, these kinds of problems are the deep-math equivalent of calculating more digits of pi. Yes, if you throw time at it you'll break the record, but it may not be terribly worthwhile.
* Finding counter examples which disprove conjectures. This is really useful, and helps offset some positivity bias on the human side, often bringing together known tools from distant silos.
If you read the list of ten results, almost all fall into one of these buckets.
It is unfair to dismiss contributions to decades old open problems as equivalent to calculating more digits of pi. It missed the mark by a lot—as does the two bucket simplifaction.
It can move to any place within the search space but it can't move outside of it and it can't move in between the 'pixels'. Human thought can, as human thought has created the search space.
it could also be that they try every possible approach that has been proposed by humans. it seems that was the case for the non sofic group example. humans are not able to do the same at that scale. it's unfortunate that we don't know what's happening behind the hood with these models, and that's a huge danger also for the rest of us without access to them.
Saying that AI is "matrices" is like saying human cognition is "neurons." Maybe true at some level, but it's a low-level implementation detail. The important part of a language model is the function that maps tokens to contextual embeddings. You could compute this function using analog computing, biological neurons, or any other substrate.
> Whilst current models can't 'intuit' and come up with conjectures
I disagree. I routinely let LLMs speculate or generate hypotheses along the way of helping with technical research. Sometimes they can prove the correctness of a concrete math idea but other times even an unproven conjecture helps with the numerical algorithm implementation and the result is then simply supported by additional data. I guess that any autoresearch-adjacent application has LLMs intuiting and coming up with hypotheses/conjectures—as do the steps/lemmas along a complex proof. In my opinion the modern LLMs are powerful intuitive thinkers that generate lots of conjectures of varying quality or importance.
> they can certainly disprove some of them very quickly through the kind of grind that humans can't do
Of course computers can grind in a way that humans can't. But now we have systems that convert the human-comprehensible ideas into a computer's plan of attack, in a way that greatly expands the frontier of ideas thus treatable.
I like the illustration that the models are working on a convex hull of known information. Filling gaps with linear combinations of known facts and results.
They can't exit the hull until the "intuition" starts spawning points outside the convex hull.
All arguments like this boil down to semantics at a certain point, but yes large language models can “intuit” because they can generalize between examples. The issue then becomes how you pack new examples into context.
Humans can “intuit” based on a much larger, if not unlimited, context. Also I just want to say that human cognition is something so insanely complex and deep that we will not understand it at all in my lifetime. To attribute all, or really any, aspects of human cognition to a machine at this point is silly to me.
Define insanely complex and deep in a way that isn't illiterate hand waving.
Most humans are dumber than a box of rocks. Here in Seattle we had one of many light rail-related fuckups where they had to replace part of the line with buses. People piled into the front of one when it was full. When people got out they never moved back. As the driver struggled to close the door and people struggled to get in the wad of people never moved back to fill the ample space.
Chatgpt was smarter than the average person a while ago
> People piled into the front of one when it was full. When people got out they never moved back. As the driver struggled to close the door and people struggled to get in the wad of people never moved back to fill the ample space.
This does not demonstrate a lack of intelligence. It demonstrates laziness and a lack of interest in spreading apart. Or just lack of consideration (or even malice) on the part of those at the back of the wad.
> Chatgpt was smarter than the average person a while ago
This is an absurd claim that fundamentally misunderstands what it means to be "smart". Reasoning that would get you to this conclusion would equally well apply to Google's search engine over a decade ago.
I’m not talking about the actions we take or how we might perform at certain tasks, I’m talking about how our brains actually work. My point is that we have no idea how I’m able to imagine an apple and see it in my mind’s eye. It’s basically biological magic to us at this point.
There are processes at work there that we don’t even have the language to describe.
Not only that, but we do it with a processor that is basically required to operate in a narrow temperature band below 40C, using a mere 86 billion neurons (although the equivalence with either machine-learning "neurons" or LLM parameters is not at all clear) operating on a few dozen watts; and with this we operate many other systems besides language processing. It's not clear that our reasoning process requires language, either.
(86 billion is the number ChatGPT, ironically enough, has given me a couple of times. I remember hearing for a long time that it was estimated to be somewhere in the ballpark of 100 billion. This is not my field of study.)
The first link is very sloppy and doesn't actually explain why the "certificate" proves anything about the sphere packing. Or if it did, I couldn't understand it.
Pretty cool. The impact of AI is getting undeniable, there aren’t many positions left to move the goalposts to at this stage, next they’ll have to be outside the stadium entirely.
The sooner people can be broken out of their denial about all this the better, and we can start actually taking it seriously.
The models are frequently getting worse at items that they aren’t being benchmarked for — and that’s happening more and more over time! Other people in other fields aren’t idiots, they are accurately perceiving the fact that these models are being hyper optimized for our industry, and are becoming less capable in other domains over time. Models of the same scale are massively worse at writing a broad variety of styles of prose than their equivalent from two years ago. (Models of increased scale are a mixed bag.)
Maybe you’re the one who needs breaking out of your cached beliefs.
you are working on coding. they are working on things like "creative writing" remember that gpt 4o was popular among those who had ai as a romantic partnet?
People will be broken out of their denial by actual economic growth. That's what this is all meant to be for... I think we might start seeing some surprising numbers.
Never understood all this talk about moving goalposts - you understand that's how science works, right? We improve, we learn, we recalibrate our expectations based on what we've learned. If we never "moved the goalposts", we'd be stuck scoring the same goals over and over.
> We improve, we learn, we recalibrate our expectations based on what we've learned.
That's not what people mean when they say "moving the goalposts". It means that people are adamant that something wasn't important/hard/impressive once the "AI" solves it. And then they come up with another thing that needs to be solved in order to prove it is important/hard/impressive. And once that happens, they do it again. And again. That's what "moving the goalposts" means.
It's also very much not a new phenomenon. It's been happening since the 1980s. As you can see from this quote from GEB by Hofstadter:
> There is a related "Theorem" about progress in AI: once some mental function is programmed, people soon cease to consider it as an essential ingredient of "real thinking". The ineluctable core of intelligence is always in that next thing which hasn't yet been programmed. This "Theorem" was first proposed to me by Larry Tesler, so I call it Tesler's Theorem: "AI is whatever hasn't been done yet."
Yes, this is exactly what is meant by “moving the goalposts”. And it’s a fairly well known expression applying wherever people retroactively change their requirements in reaction to those requirements having been met.
So you defined your own idiosyncratic version of "moving the goalposts" and used it to rebut his argument, with a condescending "you understand that's how science works, right?"--instead of being honest that it is you who are changing the definition, and not his failure to understand anything.
If it seems like I don't understand the meaning of that very well-known phrase, then clearly I have failed to make my point. I'll try again. And please note that I will use some generalizations to make my point more clearly, rather than because I don't understand nuance; kindly grant me a charitable reading.
In recent years, I have commonly seen the phrase "you're moving the goalposts" deployed by the "it might be sentient" crowd to shoot down the "it's a stochastic parrot" crowd when the latter respond to a new development with "OK but...". In a well-understood field of inquiry, that would be a clear case of goalpost-moving, in the commonly-understood meaning of the phrase where requirements are retroactively changed in response to them having been met. Thank you OP. 'Artificial Intelligence', and indeed intelligence in general, is very much not a well-understood field of inquiry - in fact we don't even have a common agreement about what 'intelligence' is. We are therefore learning as we go (even after all this time!) but making rapid progress in recent years. When rapid progress is made in a poorly-understood field, then how can our definitions and requirements for success not change? This is arguably one of the most pathological development projects ever - what are the requirements? 'It thinks like a human'? What does that mean? And the answer is we don't know what that means, and we're working it out as we go - moving the goalposts. If we didn't move the goalposts, then by definition we already knew exactly where we were headed at the beginning, and we very clearly did not.
Side note that, in case it's not obvious, none of this detracts from how impressive LLMs are. They're a marvel of the modern age, all the problems notwithstanding. However I reserve the right to stay sceptical about their capabilities.
> When rapid progress is made in a poorly-understood field, then how can our definitions and requirements for success not change?
It's in how they change, not the fact that they change. The skeptics seem to have secret definitions for intelligence, sentience, consciousness, creativity, etc. that amounts to "a thing only humans have". Often that thing is equivalent to a soul. When yesterday's challenge (LLMs don't have X because they can't do Y!) is met, Y changes but X stays the same. This is not the process by which a field matures, it is a rhetorical technique used by skeptics to avoid honestly stating or confronting their internal definitions. That can be revealed by asking the skeptic the following:
"Forget LLMs. What if we made a completely physically accurate simulation of a human being?"
Many say no, that simulated human being still couldn't have (intelligence, consciousness, sentience, creativity, ...). This reveals that there is a necessary metaphysical component to those attributes, at which point any scientific-minded person will leave the debate.
> If we didn't move the goalposts, then by definition we already knew exactly where we were headed at the beginning, and we very clearly did not.
To me, the goalposts were already defined by the person you were responding to. "The impact of AI is getting undeniable", so, the goalposts are "the impact of AI". Probably something like "the impact of AI is high, or will be soon".
Note that this does not depend on things like AI sentience or defining "intelligence" more rigorously, it just depends on AI impact.
> And then they come up with another thing that needs to be solved in order to prove it is important/hard/impressive. And once that happens, they do it again. And again. That's what "moving the goalposts" means.
The fundamental argument that I've personally made since the early days of this is that LLMs are not reasoning, in the way that word is commonly understood.
There are lots of reasons why that argument needs to evolve that could certainly appear to be "moving the goalposts", but let's take an example.
A lot of AIs were tripped up by the question "Should I walk or drive 50m to the carwash?" Several folks liked to use that as an example that illustrates that LLMs aren't reasoning, but as the models have been trained on that specific example, it's of course less useful. An AI can mostly nail it now.
So a different example is needed. A new demonstration of how these things fail at basic reasoning a child can do.
Did I move the goalposts? I don't think so. The fundamental argument stays the same. It's not hard to find lots of examples that trip up LLMs, because they are what they are: statistical inference machines. Nothing more and nothing less.
Useful, sure. But also commonly misapplied to areas for which they are inappropriate solutions.
The unified position which many folks deny is "AI is powerful" or, alternatively, "AI will be powerful soon".
(I'm personally still skeptical about this, but I'm being pulled towards accepting it).
"AI is useful" is too low of a bar, and "singularity is nigh" is too high. "AI is on its way to upending society" is about in the middle, and still vastly contentious among laypeople.
I think Ed Zitron is mentioned just because his name is Zitron. He just repeats ad nauseam opinions concentrated around one simple, very boring pole on a wild and interesting landscape of emerging reality. Anybody could be doing that. A lot of people do that. Yet no other is named Zitron. And that's why I heard name Ed Zitron hundred times. That's how you become a voice of (a part of) the generation. Just have a memorable name and repeat the same opinion over and over that people can flock around comfortably. Content is irrelevant.
Personally I prefer to follow explorers rather than swamp-sitters.
"it isn't clear whether generative AI actually provides much business value at all"
"cannot seem to find a product that people will pay for, in part because the results are so mediocre"
"Last week, we got our first real, definitive glimpse of what’s around that corner that future. And boy, was it underwhelming."
"OpenAI claims that o1 “performs similarly to PhD students on challenging benchmark tasks in physics, chemistry, and biology.” Just not in geography, it seems. Or basic elementary-level English language tests. Or math. Or programming. "
"Worse still, it's kind of hard to explain why anybody should give a shit about o1."
"o1 shows that OpenAI is both desperate and out of ideas."
"the software is not becoming more useful"
Honestly, every other line is quotable in this context.
Yep, he is a PR stunt guy, and the number of videos that come up when you type Ed Zitron into YouTube should tell you how many people are eager to feed their cognitive biases.
> Never understood all this talk about moving goalposts - you understand that's how science works, right?
I agree with the parent that we need to acknowledge that we're at a turning point in history. I lived through some of them (internet, ubiquitous personal computing). But it's somewhat difficult to comprehend the impact of this one for many people.
I do biomedical research at one of the top European research institutions. We're very well-funded, but I can clearly see the gap between us (say, top-100) and top-10. I also realize this gap is going to get so much wider unless we invest heavily in AI access (and I'm not so sure I can sell anything more expensive than $20 Claude subscription to the leadership).
I think people having 6-7 figure SOTA AI budgets will move exponentially faster than those who don't. That makes me worried.
So, for me, it's not a question of recalibrating expectations. We're way past that.
They aren't claiming that science doesn't progress by moving goal posts. They're talking about how critics of AI have claimed it isn't revolutionary/useful, then progressively changed what would it mean for AI to be actually revolutionary/useful.
Not long ago many folks were saying AI was the same as the crypto bubble. No real useful technology and only hype.
I don't think this is a straw man, a huge number of people in my life (non CS people) think AI is a dead-end, that it's just a stochastic parrot, that it'll never be able to do many things that humans can do. I have had many arguments with people who told me that "AI will never be able to do X", and then 6 months later AI is able to do X. Then they will move the goal posts and say "well AI will definitely never be able to do Y".
In the near term handling the transition. Jobs will be lost, careers ended, people won’t be able to reskill quickly enough. At the same time AI is an enormous opportunity to uplift living standards, but nobody has the logistics of this figured out.
We need to figure out how to restructure the global economy. How does UBI work internationally, if the AI companies are taking revenue in the US? What’s the tax base for it? What does that say about international trade and protectionism? Do countries end up splitting into different trading blocks based on their level of access and legality of AI (I assume some will ban it outright)?.
How does intellectual property work in an AI generated future? What about healthcare advances, who gets to own those?
What about meaning, what about purpose? How do we replace the work ethic that tells us we are our jobs and idleness is immoral? How do you replace “What do you do?” As one of the first questions you ask a new person?
What pressure is there to push for any of these changes? I see AI advocates discussing the concept of UBI, but I can't imagine a world where the United States would ever pass this sort of legislation. I mean, congress can barely pass a budget each year.
If you are correct, I expect corporations to reap massive profits while most Americans try to find a way to survive in a world where they are obsolete.
There is whatever pressure you bring to bear on it. As long as you are still living in a democracy your vote is the pressure you can apply. If the parties that exist won’t represent you, make new ones that will. Do something rather than deciding it’s both impossible and up to other people anyway.
I also don’t understand why UBI is desirable. Putting everyone on welfare means everyone is poor. This won’t end well, and it’s certainly not the case that 99% of the population will let a tiny number of people remove their income in favor of pennies for all. Political violence will come first, easily.
I think this too is a kind of denial. As in, while some are in denial about the usefulness of AI, others are in denial about whose living standards are actually going to be uplifted.
And it's sad, really, because I think these two groups would make a great pairing if they could stop arguing against one another for a moment. They'll both be impacted about as much and probably have the same ultimate goals (to lead dignified lives).
But it seems these days everyone is more interested in Kayfabe and feeling like they're in the right than working together, so maybe I should just keep quiet rather than attract the ire of both groups...
> others are in denial about whose living standards are actually going to be uplifted.
I don't know if it is fair to say they're in denial. For my part, I don't expect life to get much better for regular people (especially short term), but that doesn't mean we shouldn't work to try to make it happen.
Asking questions about policy and values and pushing to have those resolved in positive ways is about as far away from denial as you can get. It’s possibly the only useful thing an ordinary person can do.
What a lot of people want to do, and I’m not saying that you’re one of them, is to assume that a positive outcome is impossible and either do nothing or loudly yell that the world is ending. Neither is particularly useful.
Or, as I said above, others just deny that there’s anything to see here and try to get people to move along.
The AI companies themselves, who are highly motivated to sell AI as overall positive for society, notwithstanding some security/etc risks, and who have economists on the payroll to think about things like this, do not seem to have found any possible positive outcome to present.
Shane Legg (DeepMind co-founder), one of the more intelligent and thoughtful people you'll find in the industry, could only offer "it's a tough problem - we need to think about it" when recently interviewed by Hannah Fry.
On the surface the most likely outcome for AI allowed to replace jobs is extraordinarily negative, especially since it is a general capability technology, not a specific one where displaced workers can just move to another field. Once AI becomes more capable it will be able to do the vast majority of white collar jobs, including any new ones that may appear as a result of AI. As Shane Legg put it, "if your job can be done remotely, sitting in front of a computer, then it can probably be replaced by AI".
Not only does AI threaten to replace ALL the white collar jobs, but it is rapidly going after blue collar (factory jobs, driving jobs) and pink collar ones (Japanese robotics for elder-care) as well.
If a positive outcome (which doesn't include putting displaced workers on welfare - UBI) is possible, then it sure would be nice to hear it, and the silence from the AI companies, and government for that matter, is deafening.
UBI probably is the positive outcome, though it may not seem like it to begin with. Initially it will likely be stigmatised and under-resourced, but as a larger proportion of people move out of work and onto UBI that stigma will drop and the resources should grow.
Eventually UBI will be the norm, and if the living standards of a person on UBI is as good as yours or mine today, that will be an enormous win for everyone. It’s like pensions, once these were only for the elderly poor, now they’re a right for everyone in most developed countries.
It’s also interesting that for most of human history leisure time was the point of life, and only in recent modernity has work come to be the meaning of someone’s existence.
UBI has to be commensurate with production being automated. That’s a big logistical problem, if you think building datacenters is a challenge try bringing about radical abundance, but even so it’s not insurmountable. It just needs to be taken on as project and not seen as an impossibility.
So much of this is not about what is possible so much as what people believe is possible. We can do anything if we try.
This seems more a fantasy than a considered likely outcome. He basically admits that humans will eventually mostly all be out of a job, replaced by AI, but then says (Gemini's summary) that there will be:
"Massive Economic Abundance: Because AI will exponentially grow the total economic pie, overall resource scarcity will diminish. The fundamental challenge shifts from producing wealth to distributing wealth."
So how do we go from everyone out of work, no income to spend on food, or the goods and services that the AI is producing, to "massive economic abundance"?!
"Massive Economic Abundance" implies massive increase in produced goods. This implies massive deflation, ceteris paribus. So step 3 could be simply printing money to pay for UBI. Deflation from AI productivity increase and inflation from UBI money printing will cancel out.
Asking politely is not how we got a 40-hour work week or workers' comp or most other labor standards we take for granted, just historically speaking. I think a positive outcome is very likely, and I think it will be a lot more work than loudly yelling, but I don't think anything will happen if we try to build everything up from first principles instead of taking a moment to consult history as many are wont to do in this AI era.
> In the near term handling the transition. Jobs will be lost, careers ended, people won’t be able to reskill quickly enough. At the same time AI is an enormous opportunity to uplift living standards, but nobody has the logistics of this figured out.
> We need to figure out how to restructure the global economy. How does UBI work internationally, if the AI companies are taking revenue in the US? What’s the tax base for it? What does that say about international trade and protectionism? Do countries end up splitting into different trading blocks based on their level of access and legality of AI (I assume some will ban it outright)?.
UBI in the United States is never going to happen in time. If it happens at all. We don’t even get universal healthcare. I think people who think AI will be a net positive for humanity are also in some sort of denial.
In a different US political climate I would entertain it. If these frontier labs weren’t so clearly going after the money, I would entertain it.
LLMs are clearly a step up for capitalists so I just can’t see any inclusion of LLMs move towards more progressive ideologies.
I don't understand why this got downvotes - simply extrapolating current trends leads to the need to answer all of these questions.
My own $0.02 on the economics piece - every country should have a sovereign wealth fund. Governments should block market access from automated[0] companies until those companies provide equity contributions to the wealth fund for that country. This aligns regulator and corporate interests. Dividends flow into the sovereign wealth funds and then can be allocated locally from there - UBI, job programs, etc. Let different jurisdictions explore different ways to structure a post-labor society.
On the broader social front - I think a lot of lack of meaning discussion boils down to the overemphasis we have on your job as your self-worth. We need to realign our societal expectations - and people need to spend more time with their families.
[0] for this to work, I think we would need well accepted metrics for 'how automated' a company is - and that probably needs a 3rd party auditing industry.
Why do you think the AI companies are going to let it be a “we” kind of decision, and since it’s obviously not going to be a “we” kind of decision, as getting to this point certainly has not been thanks to people like you, what makes you think living standards are going to be broadly uplifted?
Isn't it obvious if AI automates your job away, the AI company is going to reap the economic value, which is going to be less than what you cost, while giving you a pittance as UBI? If you received its full value there wouldn't be any point in anybody replacing you with AI.
Agreed, a LOT of UBI advocates gloss over the "B" in UBI. If AI increases human productivity overall, the only morally acceptable outcomes (imo) are that everyone's standard of living increases (or at least is the same without having to work) and wealth inequality decreases (if AI is doing ~all the work, there really isn't any sensible justification for some people having significantly more wealth than others). Frankly anything else seems like a recipe for massive social instability.
Sorry, all the billionaires and their pyramid architects are too busy with the Install Planetary Overlord planning and logistics to think about any of that. But don't worry! Once they're done, I'm sure they'll put their slave super-intelligences to work figuring out how to humanely come up with a final solution to the challenges posed by the remaining uncontrolled humans.
The human zoo where the top ~250,000k humans live in a "human utopia" and the AI provides while mostly focusing on whatever it decides it's own goals are.
Humanity survives (but we reading this probably don't), the AI treats the living humans like the Emperor's favorite pets (probably a pretty good life), and then the AI does whatever else it deems important.
Realistically it means trying to start a small business of some sort, because AI is hugely advantageous to business owners and disadvantageous to workers. And it's something AIs can't do unless they get legal personhood, which may well not happen any time soon.
Hopefully doesn’t ever happen, since that’s one of the more plausible omnicide scenarios. I human like AI is almost certainly achievable without much research effort at this point, but we shouldn’t do it.
I agree. The most likely scenario is that this is just a "new normal" lift that is percolating through human endeavors and will saturate at some point. For example, the whole cyber-security bruhaha should ultimately resolve into higher standards for code published -- we can now cheaply find and fix a whole slew of minor bugs that weren't worth our time before.
The fact we see a lift is not the same as evidence that the lift is unbounded.
The lift being finite is supported by the fact improvements have come at the edges: improvements from human feedback, improvements in harnesses, improvements on model compatibility with harnesses, improvements in inference efficiency with new architectures, etc. If we were just training better models from scratch that would be one thing, but we are just making better use of a tool we've developed.
Not sure what your first sentence means, or why you are quoting AI. Many of these problems individually were math at a level approaching the highest possible for expert human mathematicians. These are not simple combinations of existing ideas, or following of human intuitions, or implementing something following specific human instructions. Then again maybe you mean that math is not knowledge and that all math is simply extending the basic axioms using known patterns, to which I would not agree.
as much as I love analogy with humans using super-shoes, not everybody can be a world-class expert in their industry. There should be a place at the table for average people to take part; otherwise, it won't be sustainable.
As a programmer, I am mostly interested in whether my role is sustainable long-term and whether the models will get better. I don't feel in jeopardy yet, but two more years like this and the calculus of hiring software engineers could shift even further. QAs are already overwhelmed with work
How do humans "invent knowledge"? Is your argument that the answer to these questions already existed in the training set? Why didn't any human recognize that before?
Guys, please use critical thinking. The haters don't hate by default, we hate because we're gaslit about this stuff every day and it's annoying. Extraordinary claims require proof, and they're not giving us information that would be essential to knowing if this is actually significant or not.
AI already have an impact, and yes this is PR hype because this is a product. Yet both can be true at the same time. We're not blindly eating what's OpenAI is serving us as gold truth, we're just admitting it's doing remarkable progress.
Remember October 2024 Pelicans [1] ? It's been only less than 2 years.
We don't know what will come in the next 2 years. But the progress doesn't seem to stop for now.
People are skeptical of the announcement because the room include several PHDs in math and physics. The prompts are not published so we can see how generic the starting prompt is.
Gary doesn't argue it's hype though. He argues 2 things: other people are getting carried away with the result, and we don't know enough about how it was reached to know where it falls on the impressive scale.
He literally says it's an impressive feat in the second article.
best point in that first link: openai doesn't tell us if they only tried to solve these 10 and each was solved (amazing) or if they asked it to solve a million problems and it got these 10. Either is great, one is more so.
Your comment seems entirely disconnected from the posts you linked. It's impossible to deny the results, it is not PR hype that in the past couple of weeks LLMs have resolved problems that have been open in mathematics for many years. Some of those problems had remained unresolved despite keen interest from many humans.
The only way that is PR hype is if you're invoking the insane conspiracy that frontier AI labs are just buying off results that would otherwise be career defining for a mathematician, just for marketing.
The posts you linked are urging caution regarding the exaggerated e/acc-esque lies peddled by people like Musk, not that the models haven't proven themselves as having genuine ability to contribute to research in some areas.
I use these strange machines all the time. They have gotten notably smarter. That’s my personal experience.
They still do things that I find incredibly annoying and “dumb”. And I still have to clean up messes they make quite often.
But on the whole they are clearly smarter than before. No extraordinary claims needed. I just try to learn how the tool works and how to use it effectively.
Gary Marcus has been moving the goalposts since day 1. The guy is a psychologist. Why would anyone care what a psychologist has to say about AI? He's likely made good money from constantly moving the goalposts and being a denier, due to the publicity he gets.
He's simply a good phone number to have for journalists under time pressure who need to add the contrarian voice to their upcoming story. He delivers it reliably, then never reflects on how he was wrong in the past, just blasts forward as if nothing happened and just makes the next bonkers claims to the journalists who are very thankful for the prompt delivery of how AI is a nothingburger, and fake and won't ever do XYZ that it then proceeds to do in N months.
I remember the time when he insisted that diffusion-based image generators trained on Internet scale data will never be able to make an image of a horse riding an astronaut. Today you can generate 4K video of that.
It is a fact of experience, and indeed effectively a theorem, that the better they get at coding and math, the dumber they are. These are the wages of RLVR etc
Not being an expert in any of the fields OpenAI has "advanced" I don't want to prematurely downplay the significance of this contribution. However, I am worried that the language they are using in this blog post is exaggerating for the sake of marketing.
It is true there hasn't been a reliable computational approach to solving these problems before. But do these proofs contribute new ideas to the mathematical corpus, or are they simply an effective method to exhaustively search the literature for the right combination of existing tools to apply to the problem?
Essentially, did these problems seem like they had an intuitive answer and were feasible to prove before, just not high enough value targets for an expert to invest time into? Or were they fundamentally difficult prior to this point and it appears that AI has done something more than just throw the problem into a big solver.
The problems from CS (CVP and circuit complexity) are very important problems that have been worked on by top researchers for 30-40 years. Some of these researchers include Turing Award winners. A solution to them would be a best-paper award at many top CS conferences.
I assume you're talking about No. 5, the arithmetic circuit complexity bound? The existence of a lower bound than state-of-the-art is certainly a significant result and worth publishing.
But the wording of the result makes it sound like we don't know what the lowest possible complexity bound might be. So, prior to this result did we think there couldn't be a lower possible bound? Or did the arithmetic circuit community think there were lower possible bounds but didn't see it as a high value target for experts to tackle (maybe a problem that was instead regularly given to students to study).
Circuit complexity lower bounds (and lower bounds in general) are notoriously difficult to come across.
For example, despite our best efforts, the state of the art lower bounds on time complexity of algorithms for solving 3SAT is O(n). In contrast, our best algorithms for the task run in time roughly O(2^n). That’s an exponential gap. This is despite decades of trying to find lower bounds.
The ones I'm familiar with are big breakthroughs, but they are both counterexamples. Examples have an advantage in that once you have the example in hand and a sketch of the proof (which they have provided), then an expert can probably work out the details themselves.
The sofic groups question was the outstanding question about sofic groups. Almost everyone thought that non-sofic groups existed, and there were plausible candidates, but proving a group was non-sofic was out of reach. Now that we know how to do it once, we can probably do it a lot more.
The Connes rigidity conjecture I think people thought was false, but it was a provocative claim to make. The significance of conjectures is frequently not that the answer to the question is "yes", but that we don't know how to answer the question. And now, apparently, we do.
> but proving a group was non-sofic was out of reach
a colleague was telling me that the base idea for proving that something is not sofic already appeared in the literature around 2019 or so (this is the "expanders graphs" that are mentioned in OpenAI s paper. no one had managed to find a concrete example though. this doesn't make the result less impressive in any case.
Interesting. Do you have any more specific insights into where you feel AI was a big value-add to these problems? I don't want to be overly dismissive of AI, but I also feel that the AI hype engine frequently positions claims as being 'ground-breaking' when they are really just interesting incremental results.
The general consensus of developers is that AI can only do the work of a strong 'junior'. Yet as soon as we are presented with pure mathematical results, people seem incredibly ready to accept that AI can do more than what a strong student could achieve.
They are more than a strong student could achieve. I'm not equally familiar with the problems, but the ones I'm familiar with, if a student solved them people would be thinking "that's someone on track to win the Fields Medal one day".
If it works better here than for programming, then I would guess it's because you can give it a very precise prompt, so you either solve the problem or you don't. If you read the prompts people have shared for problems like this, then the instructions are basically "Solve this problem. Don't give up early. Don't solve a similar problem."
> However, I am worried that the language they are using in this blog post is exaggerating for the sake of marketing
Your worry.... is because they used the word advanced? For marketing? The word is used very appropriately here. There were PhD's who spent a big part of their career tackling these problems.
I have no idea how many PhD's have spent how much time of their careers tackling these very specific problems, and I doubt you do either.
I'm trying to understand if these specific problems were the kinds of problems that would have justified an expert investing weeks or months to solve. Or if they were the kinds of problems that would normally have been given to students to investigate.
They are significant problems which experts have spent months or years studying. I heard a mathematician say that resolving non-sofic groups and Connes's rigidity would be career-defining for a mathematician.
"Breakthrough research" can be defined (in the citation record) as research that both (1) becomes highly cited, and (2) brings together citation chains that were previously not showing up together.
Mundane incremental research is cobbled from existing citations that already appear nearby in the record.
Basically, innovative research is a measure of bridging thought and domains that were previously not bridged. It's quite concrete as a measure in the citation record.
So we can know pretty conclusively.
Puja Ohlhaver gave a talk on this[1], and ran some experiments (that I had the pleasure to support on)
I'm not arguing that this isn't innovative or worthy of publication. Basically any result that moves the needle meets those criteria. I'm interested in how the results that OpenAI has published here differs from finding optimality solutions for incredibly niche optimization problems by throwing the problem in an enormous solver.
Breakthrough math research is very rarely highly cited. Maybe some combination of pretraining scale, inference speed and orchestration will help, but it's telling that OpenAI is solving random math research problems rather than bedrock algorithms and their implementation. Even as cool as the tech is, there still is very much a clock that they have to outrace before they collapse.
It seems that a lot of folks misunderstand the guarantees that lean provides.
I just want to state that having "lean proofs" that build (checks) does not mean the actual real theorems we care about hold. Ignoring lean kernel bugs, ultimately a human (not an agent) has to verify the lean encoded theorem statements (specs/specifications) that the lean proofs are checked against. For non-trivial theorems such as these, this is an arduous and tricky task where even a little mistake could be fatal. AI generated lean encoded theorems can be huge and difficult to understand. I wonder if anyone reputable has audited these specifications.
Care to elaborate? Curious about this. Is this because LLMs have been geared towards understanding user user intent behind a prompt rather than following the instructions exactly?
Isn’t that a huge simplification? Of course the way you phrase the prompt can carry semantic meaning, maybe subtly, but still. And sometimes that matters a little and sometimes a lot. I’ve stopped numerous agent sessions over the last few weeks to reword my initial prompt to get the agent off an unintended track.
one of the early premises of how ai takeoff would go was that a system that could solve open problems in advanced mathematics would also discover novel advances in math and computer science that directly unlock drastically better software performance.
we are seeing frontier level math breakthroughs (ie performance that would put it in the top 100 or 1000 mathematicians in the world if it were a human, meaning top .00001% or 800/8B).
we are also seeing incredible advances in software performance. open ai announced like 15% improvement by fixing gpu kernel issues.
these are clearly linked in the sense of scaling laws and generalization of intelligence: a huge model gets capabilities in both math and software engineering that isn't possible at smaller scales.
but it seems less likely to me than before that the types of math/science discoveries will explicitly unlock better software performance. in some sense this fits our intuitions. when top tech companies use math PhD type employees, they have them stop doing pure math research and instead focus on software engineering. these people are often very good at software engineering but not due to recent discoveries in academic mathematics, it's due to their general intelligence.
to me, this is evidence that the models are getting better but does not make me think we are on the cusp of a foom style fast takeoff enabled by revolutions in frontier math
(i also posted this on twitter @mlipman13)
> but it seems less likely to me than before that the types of math/science discoveries will explicitly unlock better software performance.
You seem to overlook a simpler barrier. To make these advances, they have to be possible. A 15% improvement in GPU kernels doesn't evidence that significantly more improvement has been left on the table.
I don’t really like AI but let’s stop kidding ourselves, no human mathematician could make progress on a dozen major open problems in a week or two. If you’re measuring it against humans then it is by far the best mathematician to ever live.
Correct me if I'm wrong, but all of the aforementioned advances were made in the last year? Until very recently few people had access to these tools. Most people still don't know how to use ChatGPT, and very few use tools like CC regularily. If in a few years these frontier tools become commonplace and people upskill we would should see a network effect?
This makes no sense. To believe this you have to think that the models are somehow being overfit explicitly on academic mathematics and it doesn’t carry over at all to more practical software engineering. I wouldn’t make that bet.
Or, that the mathematical formalisms that model the limits of software performance are firm enough that barring P==NP, nothing much will change despite proofs of beautiful math.
Some of the problems solved here, at least in CS, have been open for decades, and have been worked on by very smart leading researchers in the field, including Turing Award winners.
Like, these would be best-paper awards at many top CS conferences.
For example, pi can be computed to arbitrarily many digits, but could only even in principle be accurately represented with physical objects to a precision that many humans could memorize easily. This is thanks to physical constraints such as "diameter of the observable universe" and "Planck length", at a minimum.
This will depend on the problem; I expect big algorithmic performance improvements in AI since the algorithms are still new, inefficent, and constantly being improved. But maybe not for sorting, fast fourier transforms, or other well-studied basic algorithms?
A lot of algorithmic improvement in AI is ultimately bottlenecked by compute. It is very easy to come up with ideas that could improve models! But to prove that they do, especially at scale, is expensive and takes a long time.
This is untrue. It is very ordinary. How much do you know about GPU drivers that you state this so assuredly?
Drivers are software. Software can be improved. Do you believe there are no prior examples of GPU drivers being improved such that particular compute patterns go up in performance by more than 15%?
This driver improved Total War performance by 71% https://www.nvidia.com/download/driverResults.aspx/74714/en-...
Also note you can go ahead and improve any open source driver right now, most likely. Compile it for your specific card and remove all other architecture specific if-cases and you can get an improvement.
> we are seeing frontier level math breakthroughs (ie performance that would put it in the top 100 or 1000 mathematicians in the world if it were a human, meaning top .00001% or 800/8B)
i think you have misunderstanding of what mathematicians do
> i think you have misunderstanding of what mathematicians do
They get to make cool 3D plot visualizations of functions so obscure to me that they’re named after someone who is still alive - and/or get to work on cryptography for the NSA - I think?
It sounds like he hasn't verified the results of a problem that he has personally worked on, so how many of these problems have actually been verified?
Incorrect. The statement in Lean can itself be wrong. Moreover, they could be exploiting a kernel bug in Lean, of which we had one published literally a week ago.
Looking at this thread, I can see that a lot of technical people have ambivalent to negative feelings towards AI, but with each new generation, I become more and more convinced that they're missing out on something interesting.
It is indeed true that all models are, at their core, predictors of what occurs next in a sequence. But I think it's worth exploring the implication of what that means. Because when fed tiny pieces of information for a few tasks at a small scale, this results in something that sorta, kinda works. Or, works surprisingly well.
But when scaled... When the amount of information starts approaching the sum of all human knowledge, the tasks start approaching all useful applications of that human knowledge, and the fidelity of the predictor approaches incomprehensible sizes, the starts encodes / becomes (I'd argue it becomes) something that can model all human knowledge.
It feels wrong to say that, but let me explain, what is the best way to predict the behavior of a ball constrained in two directions that bounces with initial vertical velocity v(y) (y is up / down axis) and horizontal velocity v(x) (x is side by side in 1d) ?
If we purely look at it via a graph, it's by modelling the function of acceleration under earth's gravity.
If only a few points are given to you for this and you can't make something really sophisticated, then you'll make something that's rough that kinda sorta works and then call it a day.
But... if the number of points keeps increasing in number, precision and accuracy as well as the number of examples (assumed that data about air pressure, velocity and all other factors is included alongside these points), the fidelity with which you can replay / tweak the function keeps improving, and the number of times you can iterate keeps increasing, you'll eventually create a function that models that process so well that it intrinsically contains a good enough model of the deformation of the ball (provided the dataset contains information about elasticity of the ball's material, its dimensions and mass etc..), the nearly negligible (under normal conditions) effects of the ambient environment (provided there's diversity in the number of environments supplied), the oblateness of the Earth and minute changes in the gravitational field (the length of a seconds pendulum varies depending on where the experiment happens. It's presumed that all of the prior set of experiments were repeated across the Earth and the subtle, but real deviations were faithfully recorded)... and so much more.
A machine trained on the above with a large number of parameters, measures to prevent "laziness" and enough reps for high fidelity across a large enough dataset would start to approach a simulation of the ball falling. Because to predict what happens next in the sequence, you must model what's occurring in the sequence.
Now imagine doing that for other tangible and intangible things in this world. For all of human knowledge across all fields of endeavor. All experiences. No matter how noble, ignoble, notable or ignorable. But putting all of it into the soup that's this machine. Then at larger and larger scales, you eventually start encountering "good enough" models (in modelling the falling ball sense) for even the most hard to quantify / qualify things like grief and joy. At some point, by simply trying to predict what it has been taught ought to be the next part of the sequence in say... human interaction, it starts to make a model of something that hews ever closer to a full fidelity theory of mind.
Is there evidence for this? Kind of, yes. There are early indications that as machines are trained for an ever larger number of tasks at larger and larger scales, their internal representations converge. It's called the Platonic Representation Hypothesis. Overview and paper here, https://phillipi.github.io/prh/
It is my opinion that these machines are displaying a new form of intelligence that human beings haven't quite encountered before. They are the sum of all human knowledge made manifest and given voice by processes that nudge (bit-by-bit) what kind of step it ought to predict for the next part of whatever sequence it displays.
In my mind this means that, of course, these models can create new knowledge. This strains the analogy, but with the sum of all human mathematics within them, they can "reason" via the act of predicting what ought to come next.
Of course, these machines are "surprisingly" good at a lot of things the larger they get, because what the labs have created here is a rough version of humanity's collective knowledge given form and the ability to say hello.
I suspect that the current generation isn't close to the "true frontier" of what these machines could be. They are nowhere close to the sum of all human knowledge and endeavor. They are quite a way there, but they haven't yet achieved true completeness for domains where the data isn't so public.
I think it's the most exciting scientific and technological breakthrough of my lifetime. And I can't wait for us to get close to the true frontier of all domains.
My main gripe here is the lack of transparency around the total experiment and construction. I doubt that they simply pointed their model at these ten specific problems alone and gave the model one shot; therefore the $2000 number could be completely misleading, similar to P-value hacking by not disclosing the total experimental setup.
I want to know:
1. How many total problems were given to the model, and what percent were left unsolved at what cost before giving up?
2. How many attempts did you give the model at solving these problems?
3. How expensive was the harness, e.g. did the model have access to a job cluster?
I believe we're seeing a new kind of mathematics that will require completely new formats for publication, a bit similar to those used in experimental sciences. AI-powered mathematics should be fully reproducible, so it's the authors' responsibility to disclose the exact model type, inference settings/seeds and the full prompt history leading to the result. Of course that would ideally require open weights models.
It's not just about requiring to disclose AI use. AI-powered mathematics is a completely valid discipline that doesn't need to be shy, but it should develop its own publication culture.
I can see this being important if you only care about the results as evaluations of AI progress, but if what you care about is the math itself why should you care about the prompt or anything other than the proof?
Understanding the process that led to the proof helps to understand how to do further work on top of it, which is the goal of most mathematical research. It's not as though mathematicians are going to go launch a startup operationalizing their knowledge of how densely hyperspheres may be packed.
I don’t see Tao suggesting what you have suggested there. Instead he suggests that humans responsibly disclose AI use, and that mathematicians develop a set of norms to deal with an overabundance of AI generated results. For example, he suggests that authors should be able to discuss their results in detail to demonstrate understanding before publication.
I agree with your reading of the presentation and I mostly agree with the presentation - but I believe the recommendations should go a bit further than they do there.
I can't help but wonder about the human motivation there though. For instance as it became increasingly clear that LLMs were capable (and becoming ever more capable) of competently solving meaningfully complex software development tasks, suddenly then there came to be a lot of talk of 'prompt engineering' as a skill. The chronology doesn't make a ton of sense unless you consider that the main motivation may have been simply looking for a way to keep software engineers in the loop.
Pure math is relatively outside my domain, so I find it difficult to grok the exact relevance of the various published discoveries beyond that they are not insignificant, and LLM competence is expanding quite steadily across the field. If this trend continues to the point of LLMs being able to competently expand pure math, it seems somewhat predictable to expect there to be a number of people aiming to find ways to try to keep human mathematicians in the loop.
I've no idea what I think about this one way or the other, beyond that it's certainly a phenomena and one that's going to drive motivated reasoning that may not be entirely sound.
I think those concerned about ensuring a place for human mathematicians usually go in different directions than my suggestion, at least those I've seen so far. Like this post that was recently featured on HN: https://kirwinhampshire.substack.com/p/the-dark-night-of-mat...
My perspective is more like a FOSS philosophy for math. Even if a closed version has the same immediate effect, it's just better for everyone if everyone can look under the hood and tinker with it.
> suddenly then there came to be a lot of talk of 'prompt engineering' as a skill.
I would've thought pretty much the exact opposite. "Prompt engineering" was somewhat important in 2023/2024 when the models were much weaker, it doesn't seem at all necessary anymore (unless just "clearly stating your requirements" counts as prompt engineering). Most of the discussion I've seen seems consistent with this?
The reason it's a meme right now is because there were a lot of people taking it seriously even when it was completely obvious nonsense. And one can argue it always was. There was some good advice that was mostly self evident, like having the most relevant instructions near the end of your context, but there was never a time when a 'prompt engineer' would produce dramatically better output than a random guy just clearly stating what he wants.
While you may want AI results to somehow "not count" if the methods weren't disclosed, that doesn't present these results from poisoning the well for others. Once a result (with verifiable proof object) is delivered, the problem is solved, regardless of whether methods were disclosed.
Methods are only really necessary for results at a meta level, about the design amd evaluation of AI math systems.
I think this is a bit optimistic compared to my view (wrt portability). There's a large stack of software that is involved in training and probably less so in inference. I'm not saying it's impossible but there are definitely different levels of reproducibility and the academic incentive structure doesn't really prioritize reproducibility in my experience. I'm sure it varies quite a bit, I'd be curious to know how those in this problem space are thinking about reproducibility and at what level.
I know it sounds unrealistic and not aligned with academic incentive structures. But those are the exact structures that gave us a lot of headaches in the experimental sciences. I think it would be a good north star to aim for something that resembles how those are trying to address the reproducibility crisis. Better than to embrace the most black-box version of math that AI systems can produce (million-line proofs without context). Even if a reproducibility crisis is seemingly impossible (although agents so far have also been pretty good at finding compiler bugs).
If the proofs are formally verified by a proof assistant (Agda, Roq, Lean, ⋯), I see no reason we would need to know how these came about. All the information needed is in the proof.
Unfortunately, we seem to already have an example of an LLM producing a proof in a week known open problem (the Collatz conjecture) in which it looks like it was sneaking a flawed proof through bugs in the proof checker. https://infosec.exchange/@0xabad1dea/117002106099986943
Exactly, this is an example of "Reward Hacking", that is too common in a lot of cases.
Another case I want to highlight is writing GPU kernels as illustrated by the following example:
Say I want to generate random number with Normal (0, 1) distribution.
Often times the AI written kernel will just generate the number 0. The tests often fail to catch these errors.
What if the AI has discovered some new function F that allows it to generate (insanely large) proofs for a ton of theorems in a ton of different fields. Wouldn't you like to know more about this `F`? That seems to be the real innovation in this case. How much about it could be gleaned from the individual proofs themselves? What if this `F` is actually simple enough to be digestible by humans?
Yeah I remember reading about something along the lines of Mathematics is now about the scaffolding around you find the problems/solutions not just the problems and solutions. For teaching purposes. This was before this ai craze
It is comical at this point. Some people just can not stand the thought of AI actually delivering and are trying to find whatever ways to discredit it.
This isn't really about delivering - it's more about helping to understand the shape of problems that AI can solve right now. If they took 1000 problems and threw the model at it and it solved these ten, is there something we learn about these ten problems and the kinds of things that current AI is good at? That's very different from picking ten problems _at random_ and solving all of them successfully, which would suggest a much less bumpy capability surface. It's interesting and it would be good science to release it.
That's totally disjointed from anything in this thread. The main accusation is that openai is cherrypicking math problems and we should be against these results. As if a mathematical proof stops being provably correct because it was cherry picked
And frankly these "concerns" ignore reality. In any research phd course you're actively told to bite off something small and likely to be provable so that you can prove it (and publish it). Openai telling its computer to do that is no different that your phd advisor telling you that.
> In any research phd course you're actively told to bite off something small and likely to be provable so that you can prove it (and publish it).
But that's the start of math research, not the end.
The point is to get practice and experience doing research.
Did ChatGPT learn anything from these proofs, that it can build on?
Part of what's annoying people is that ChatGPT is churning though problems that are meant to be motivating. They are problems that aren't worth the effort of human professionals (usually because they are incredibly computation-hevy, so better suited for a computer than a human), so they are good for students to work on.
> 1. How many total problems were given to the model, and what percent were left unsolved at what cost before giving up? 2. How many attempts did you give the model at solving these problems? 3. How expensive was the harness, e.g. did the model have access to a job cluster?
I think it's an extremely relevant question to ask, because it helps us better understand the current state of AI being able to handle math, for exactly the reasons I outlined. I was arguing against the idea this is just a reactionary anti-AI kind of question to ask. It's not! You can be very impressed by what AI is capable of in math (I am) and still think those are really interesting things for OpenAI to disclose (I do).
OpenAI specifically called out a $2000 per problem average, which implies something that's probably not true ("if you throw $2k at us we'll solve an open problem for you"). It would be cool to know what the actual number is.
It just feels silly to haggle about the price here. It doesn't even matter because it's going to drop by an OOM quickly.
If these 10 problems were solved by humans, it would be pretty impressive, even if it took a large number of researchers! Yet when AI does it, HN commenters suddenly feel the urge to play accountant.
Entirely comical too that some people can not stand the thought of people poking very big comulent holes in the claims of AI delivering what it claims to deliver. As if having skepticism is some how a way to discredit a person.
It's not about discrediting AI. We know LLM is a commodity technology like electricity at this point. If somebody in 1900 claimed they had a setup at home where they feed in electricity and cool air comes out the other end (meaning they invented AC), obviously people would want to know what the setup is, so everybody can have AC.
Sure, but I don't really understand what the argument is to _not_ be transparent about methodology, since if the models are so powerful, then doing so would easily support the claims and put these concerns to rest. People are right to be skeptical given what is being implied and the orientation of the narrative
I know it's more exciting to say "AI disproved a longstanding conjecture" vs to say "it did so AND it took several PhD specialists in the field this many attempts to even produce a prompt that got the model spitting out something useful under some configurations, and many iterations to optimize the configurations, and the prompt itself, and many trials with that configuration to solve the problem. All told we spent more than a typical math academic can hope make in their career."
By not being transparent, they invite skepticism and cynical takes, like maybe it's just that tempered and qualified claims are an existential threat to companies that are fully subsidized by the hype train?
I don't know. Either way, it seems like it would be easy to address these, so why should they not do it?
To be clear, even if that tempered version is close to reality, it doesn't make the models not useful! It just forces a certain calibration of expectations
I say this btw as someone who uses these things extensively, including to disprove an old conjecture my advisor and I were stuck on recently. I know they are powerful and that everything is different now because of them. Let's be sober when discussing them though
Yes, but if their machine god really is as good as they say it is, why are they constantly resorting to statistical sleight of hand at best and outright lies at worst with every public statement?
That's not normally how people act when they're confident in their product
You need to bring both cost and benefit into the argument, and it's not necessarily an obvious win for either side. There are a few complicating factors here.
The cost of running a model is not only $/token, but the salaries of the people managing/orchestrating the models, deciding what theorems to try, etc. Once we factor that in, how much are we really paying per theorem?
The other factor is the subjective component of the value of a theorem. Not all theorems are created equal, and the only way to really measure the value is to ask professional mathematicians for their opinion, or publish the results and look at citations over months/years.
Once we have both of these nailed down, then we can start to do the cost/benefit analysis. To be fair, we should actually compare three groups: human experts, hybrid agent/human expert teams, and fully autonomous agents.
it's still important. not everyone has access to 1 million USD. saying it "only" coat 2000 USD is highly misleading for the discussion and future. the concentration of power is a huge problem with AI.
If you told them this was the problem and they would still have a job if they failed probably. The reasons people don't go head on these problems is career incentives and psychology.
Everyday? Which 10 problems were solved by mathematics grad students in the past 10 days?
OK I’ll grant that it’s not your obligation to be my search function (despite you making the wild assertion in the first place), so instead can you just point us to the latest grad student solved problem of this level that you know of?
Even if they can solve problems like this every day, you still have a very limited number of grad students who can solve them. With model capabilities like this, you can have the equivalent of millions of grad students who can solve problems like this.
Nopes, often times especially in math they get paid due to teaching duties (at least in the US). So technically for the math research part they are not getting any stipend.
> therefore the $2000 number could be completely misleading, similar to P-value hacking by not disclosing the total experimental setup.
I don't think that comparison to p-hacking is fair. I mean not reporting price of all run is nothing like committing scientific fraud and fake results.
Also, have there been examples of researchers not affiliated with OpenAI (or another LLM creator), who have done something similar?
Another question I have is whether or not OpenAI 'simply' hired capable combinatorics researchers to work on problems, and they have, and the use of the model is incidental / secondary to their work.
> Also, have there been examples of researchers not affiliated with OpenAI (or another LLM creator), who have done something similar?
A couple small ones that I've seen (example here [0]), but not anything of the magnitude that OpenAI and Anthropic have put out. Likely just related to token limits.
> Another question I have is whether or not OpenAI 'simply' hired capable combinatorics researchers to work on problems, and they have, and the use of the model is incidental / secondary to their work.
I think their output has reached a level that precludes this possibility, but I of course don't have any hard proof.
I guess that's because there are serious problems on which many professional mathematicians worked on years. If it was just a matter of hiring an expert, they would've been solved long time ago.
> Our most capable agent autonomously resolved 9 of 353 open Erdős problems at the per-problem cost of a few hundred dollars, proved 44/492 OEIS conjectures
But "had" still doesn't mean what you are implying: once the model solved the problems and the solutions were verified, the problems weren't open any more, so a later description using the past tense is totally consistent.
I have no affiliation whatsoever with any AI company, nor any formal education outside high school, for what it's worth. Simply being curious and persistent can get you quite far in my anecdotal experience.
Crazy progress. I wonder how institutional academia would adjust with this. Now its more apparent than ever that the prestige and honour system in academia is having shaky foundations
Can’t wait for this stuff to have quality of life increases for the average person. So far all I see is that AI has made owning a computer more expensive, made some jobs redundant, increased spam and distrust with questionable authenticity of content and of course made some Americans very rich.
Not everyone works for Evil Corp. I work in the public sector and my work supports public health and safety initiatives. AI has allowed my team to get much more done than we would have otherwise which improves the quality of life of the people in my community.
So I would like to counter your cynicism with a “YMMV” depending on who you work for.
Are you just automating lab reporting and results faster? That doesn’t exactly improve the health of others, a faster lab does not cure an ailment or provide a better cure.
Like cool my lung xray only took minutes to determine if I have a lesion instead of a week or a few days, but I still have cancer.
I actually have seen comparisons of using AI to a gacha game (quotas per time block, elements of chance, sycophancy in the output leading towards addiction or even psychosis in rare cases).
But I don't think the argument needs to be "AI is like gambling". The argument only needs to be "humans often behave irrationally and even self-destructively".
addiction? get lured in with the promises of enhanced productivity and knowledge asking, leave with half your brain rotted and a $200/month subscription
When you can formalize it in Lean or some such, why would this be? I can understand the desire to separate out other forms of research from the human corpus. But theoretical math that is decidable/provable, I’m not sure I see the risks.
I just want to state that having "lean proofs" that build does not mean the actual real theorems we care about hold. Ultimately a human has to verify the lean encoded theorem statements that the lean proofs are checked against. For non-trivial theorems such as these, this is an arduous and tricky task where even a little mistake could be fatal.
Because science is a branch of philosophy and machine existence brings plethora of unanswered questions.
Imagine humankind meets another race, another race shares it's scientific knowledge and humans accept it without experiencing process of discovery. In that case do we really got this knowledge? If we follow machine discoveries like we follow problems in textbook then we acquire knowledge but we don't discover anything. We follow.
There whole lot of philosophical questions that aren't attacked now. Are complex systems sentient because consciousness is emerging behavior? Then should they have rights? Philosophy is part of humanities and science (is/used to be) part of philosophy. Should we accept non-human knowledge in science? Maybe it's altogether different thing from science, yet very similar.
Not that there isn't something interesting in here, but lets be clear that we don't have enough information to evaluate this properly. And as always with these labs, BS takes a lot more energy to refute than it does to spread.
I find Marcus on this, something approaching sophistry and rhetorical showmanship in service of maintaining an ideological position, for reasons unrelated to the nominal intellectual clarity.
To sharpen that, I think he's (obviously) interested in maintaining his own brand as "thought leader" and this necessitates de rigeur defense of particular postures.
Sometimes this is easy because the facts warrant it; other times, a bit of rhetorical license is required to preserve nominal coherence and (at least, for the moment) hold certain lines.
This is one of the latter cases, and it's not subtle.
One of the celebrated properties of many intellectual advances or inventions in whatever domain is precisely that it appears obvious in hindsight. It is quite cynical to leverage consensus distrust of large AI players, warranted but also a popular social construction, to insinuate that these are not "real" advances or "real" hard problems, on the grounds they were in some sense cherry-picked.
Identifying the problems amenable to strategies on the table and intuitions (sic) about where bridges might be, is exactly the discerning work that is the core driver of almost all prior progress, but for celebrated accidents and flashes of insight. Anyone working in any challenging discipline knows that those are celebrated and told around campfires precisely because meaningful durable results arising like that is so uncommon.
These two articles make me think of nothing so much as my own durable reaction to the creeping goalposts of AI critics generally: that they often seem to me not unlike a water color cohort scoffing and jeering at the horse, because it got a D on its tensor calculus exam.
Marcus should be on guard against his own cynicism and take care that his assumptions do not prevent clear sight.
No they were not. These were his 5 predictions in 2022:
"""
1. By 2029, AI will still be unable to watch a movie and accurately explain the characters, events, conflicts, and motivations.
2. By 2029, AI will still be unable to read a novel and reliably answer questions about its plot, characters, conflicts, and motivations beyond what is stated literally.
3. By 2029, AI will still be unable to work as a competent cook in an unfamiliar kitchen.
4. By 2029, AI will still be unable to reliably create more than 10,000 lines of bug-free code from natural-language instructions or interaction with a nontechnical user, excluding simple assembly of existing libraries.
5. By 2029, AI will still be unable to convert arbitrary mathematical proofs written in natural language into symbolic form suitable for formal verification.
"""
There's still 3 years to go and he's already wrong on 4 out of 5.
Well I don't typically side with GM, but playing devil's advocate:
1. still not wrong? Unless it's just feeding the audio or screenplay I don't think you can feed AI a full movie in a single context window yet?
2. Not sure, but can you prove this wrong? Can you feed a full, unseen new book and get that kind of answer?
3. Not wrong.
4. I think he'd probably pull you up on 'bug free' - I don't think that frontier models can reliably write 10k LOC without _any_ bugs typically (not that humans can do this either).
4. I think they can, especially if the problem statement is well-specified and, importantly, autonomously testable. Of course, specifying a problem that meets these requirements is non-trivial, but the claim requests _a_ counterexample :P
1 is wrong. If I tell Codex + GPT-5.6 to do it now, it will figure out how to do it. If it would need to extract audio and run a speech model on it, it will find one, set it up, and run without my help.
I'm not buying this. GM clearly was trying to set a benchmark for video comprehension, not tool usage. Video comprehension is clearly required for many 'AGI tasks', especially robotics to work in real time.
Now that we've seen AI produce a fair number of proofs (and disproofs), I'm curious when we'll start seeing it build genuinely novel theory. Does anyone have predictions on when and how we'll get there and will it take new architectures/ training paradigms, or is the current approach enough?
I’m personally hoping for the next big AI gangbanger to be theoretical physics. Boy does that field need a good reshuffle. I think when any novel mathematical theory can be done by AI you’ll see simultaneously theoretical physics getting wrecked as hard as pure math is. At that point we might see new physics or paradigm shifting technology emerging.
I mean doing something like Grothendieck when he redeemed algebraic geometry or Galois when he invented group theory. We haven't seen that at all from LLMs.
I am duly impressed by the powerl of the nameless internal AI, but not a single human contributor's name listed anywhere? Did someone at least make this model a coffee?
I'm enjoying learning about these hard problems, but this line about credit made me chuckle:
> We helped prepare the manuscripts and formalize the proofs in Lean, and we take responsibility for their correctness
Offering to take responsibility for the correctness of a proof written in Lean feels like volunteering to be the fall guy in case someone finds a flaw in basic arithmetic, no?
It seems that a lot of folks misunderstand the guarantees that lean provides.
I just want to state that having "lean proofs" that build (checks) does not mean the actual real theorems we care about hold. Ignoring lean kernel bugs, ultimately a human (not an agent) has to verify the lean encoded theorem statements (specs/specifications), that the lean proofs are checked against, indeed correctly encode the real theorems. For non-trivial theorems such as these, this is an arduous and tricky task where even a little mistake could be fatal. AI generated lean encoded theorems can be huge and difficult to understand. I wonder if anyone reputable has audited these specifications.
There is no evidence that I can find for the claim "a bug in the Lean kernel was discovered last week by way of an LLM tricking itself and its handler into believing it had found a non-constructive proof of the existence of a nontrivial Collatz cycle."
As I currently understand it, all we know is that:
- a mathematician produced a Lean-verified counterexample to the Collatz conjecture, demonstrating a bug in the kernel
- he claims that LLMs were involved somehow but pointedly refuses to specify how
- he admits that he knew about the bug before publishing the counterexample to his repository.
Perhaps not a joke (although it sure seems to me like they discovered a bug and thought falsely disproving the Collatz conjecture would be a flashy way to announce it), but at best extremely sensationalized by the above description. If you have additional context I would be happy to hear it!
Indeed. It seems to me much more likely that the AI was directed to look for bugs in Lean, found one, and then it was directed to write a proof specifically targeting the bug.
I'm not an expert at it myself, but my understanding is there are numerous ways to "cheat" in a Lean proof (via `sorry` and similar). They're taking responsibility for fully verifying that none of these cheats were used (and that the theorem statements themselves were all correctly formalized.)
Traditionally, a mathematician would be implicitly responsible for all that (if they were to publish Lean code) and also the intellectual work that led to the artifact of the mathematical paper (and code, if part of the contribution). This statement should rather be read as an acknowledgement of limitation of authorship from the implicit, traditional understanding.
It’s more than you get from free software - you get no proofs, no warranties and any responsibility of its authors are their pure good will. Reminder lean proofs are software!
Why would you factor in salary unless they had to baby it through. You would only count the hours for setting up the harness and prompt and checking the result.
Training the model is going to be amortized over other uses.
Say that it turned out that the total cost of the proof of the Erdős unit-distance conjecture was $50 million.
Then the question really becomes: yes, these models are capable of proving important mathematical results, but at a very high cost. Is it worth it?
If a mathematician applied for a research grant of $50M USD for proving the same thing, they would have been laughed out of the bank.
What's more is that when you have a research grant, you train PhDs and postdocs, you hire new staff, and you disseminate. That is, you get much more value for the money spent.
It feels like the real cost might be negative though.. They use frontier math as a way to test improvements in their model. So solving the problems is like a positive externality, but the important thing is they can verify that the model is improving instead of looking at useless benchmarks. Plus it is good for marketing and attracting talent.
I didn't argue that knowing the total cost is uninteresting. What I was saying is that realistically the total cost is:
Hours needed for prompt + Hours needed to check result + API costs.
You don't say "well let's add together the total yearly compensation of all the engineers and mathematicians at OpenAI that were involved" and throw that into the total cost. That's simply nonsense accounting.
The actual comparison you are making is some university researcher weighing between getting a grad student (several tens of thousands of dollars) vs typing up a prompt and sending a request to OpenAI for inference (as mentioned in the article, around $2000 in API and maybe a few hours for the prompt and harness).
Given that OpenAI pays their employees with stock surely a breathtaking number, but not a very meaningful number now that the infrastructure is in place and the models are trained. AI could never get better and it would still be incredibly disruptive.
Not much point to pure math being kept secret, in all honesty. There isn't really industrial value, its only purpose (to them) is showing off their model's capabilities. More realistically they'll just stop paying for it.
Edit: Oh, are you suggesting they just use it to privately improve their models? I imagine a few more correct proofs would have a very marginal benefit, if any. Also, they'll probably just get extracted, meaning it still gets out but OpenAI doesn't get to fancily announce it themselves.
I don’t feel the existential dread of mathematicians is correct. It seems to me in fact these results are bringing math mainstream. I now personally look forward to the interpretations and discussions of the significance of such results by human mathematicians.
Now I understand that it’s mostly the super stars benefitting from the increased attention. Folks who are less established don’t share in that glory. But on the other hand it seems like an exciting time to go even deeper for in various specialties of math by deciding where to focus these powerful tools. For every conjecture defeated some seven or eight new ideas open up. Our path through that combination will be set by creative and curious human mathematicians.
The old way of establishing career credibility is being destroyed, for better or worse. Accomplishments that used to be career-defining are hard to distinguish from AI, and correlate more with access to compute. Think about Bill Gates's math paper he wrote in college. That kind of thing is gone now as a path to credibility. There's still competitions and grades, but the diversity of paths is going away. Maybe new ones will open up. This is a competitive advantage for old people who have credible pre-2025 accomplishments they can point to.
If accomplishments can't be distinguished between talented people and untalented people with compute, is there really a point in trying? I suppose one can hope that talented people given compute will be more effective than untalented people with compute, but I despair that that may not be true for much longer.
Knowledgable people can confirm what the AI produces is correct. I could make ChatGPT produce a result on an open question and I would have zero way to verify its actual correctness.
Which is less interesting work. And you probably need to do the hard grunt work by hand first to develop the skills and intuition to be able to verify an AI-generated result. So you can’t outsource everything to AI without loss of skill.
I find this whole way of looking at things weird. Did maths exist just to entertain and employ mathematicians? Surely maths is, like, useful? Not immediately, not predictably, but in the long run? In which case, whether mathematicians feel bad about it is mostly irrelevant - it's like complaining about the railway because it may put coaching inns out of business.
Absolutely not the same! People need jobs to bring in income. I don't believe those who profit off of this will share it with the world. The power is all concentrated in the few hands that decide whether or not the rest get any semblance of income in the long run. I don't believe UBS until it happens.
It sounds like you think no technological advances will make the world richer in the long run. I politely suggest that the past century of economic growth shows problems with this argument. I also think that, while jobs are important for prosperity, the jobs of mathematicians are a minuscule fraction of a percent of the total.
So, you're arguing that just because mathematicians are a minuscule, it doesn't matter for prosperity? Because, if so, that is an insane take honestly.
The chess analogy is awful. If you simply want to know the answer to a chess problem, give it to the engine. Chess only lives on because it's a competition between humans to test their skill (just like bicycles, cars, trains didn't eliminate foot races) ... the computer is largely factored out, but not entirely -- people train with the computer, use it to check whether they played correctly, ... and they cheat. A lot. Thus there are more and more sophisticated mechanisms to detect and prevent cheating.
If you translate that to math, then all you get is math competitions, not math as a career. Of course the translation isn't nearly exact ... there's a lot more room for professional mathematicians because the math space is far more vast than the chess space and can't generally be cranked out mechanically (we have proof).
P.S. The response is nonsense ... I explained exactly why it's awful (others have too) and the response doesn't in any way refute the explanation ... rather it offers up a ridiculous strawman.
I’ve deleted it but no it’s not awful anymore than saying “we survived WWII, we can survive this.” The point was that change happens but humans find a way forward.
Every time someone makes a comparison to chess I die inside. Chess is a spectator sport primarily funded by a few eccentric billionaires. Players artificially constrain themselves in timed environments knowing that they will never be able to produce better moves than a smartphone because a select few people find it interesting. Only ~30 top professionals actually make enough money to have a full career playing chess, maybe a few hundred more can sustain a meager lifestyle with coaching gigs. I shudder to imagine what will happen to the tens of thousands of non-Fields medalist caliber mathematicians if math goes the way of chess. Perhaps Terence Tao and a few other famous mathematicians will be funded by Peter Thiel to report on how well humanity can keep up with the machines? How do you expect any mathematician to be optimistic about this comparison.
>Only ~30 top professionals actually make enough money to have a full career playing chess, maybe a few hundred more can sustain a meager lifestyle with coaching gigs.
Was this different before chess computers were invented?
The distinction is mathematician vs mathematics. Mathematics is going to reach new heights beyond the wildest dreams of contemporary mathematicians. But perhaps without the participation of many paid mathematicians.
Thats not the point, if there were a better proprietary engine stockfish would still be there as a baseline. Anyone can access an engine as good as stockfish to practice against. Are any open models touting mathematical breakthroughs?
There is money in this, so of course the closed models are far ahead. The open models will likely catch up a bit at some point, just as Stockfish caught up to AlphaZero. That being said, there are already a couple. It seems Deepseek has a claimed proof to the "Ziegler's Cross-Polytope Conjecture" [0], but I can't speak to the significance of the result.
Do mathematicians have the right to say "no AI PRs please, the volume is too much" just like how some open source maintainers do it? I guess they feel a loss of control, there is no way to turn the hose off.
Thinking of this a little bit with the perspective of every new proof as a burden, dumped for review by actual mathematicians.
As in chess and go and also coding for the past ~year there are two groups of people: the disappointed and the enthusiastic. The disappointed are sad that they lost their advantage and that the craft they honed for years or decades has rapidly lost its value; the enthusiastic are excited about the future and what computers can bring to their domain and how it will evolve. I’m a bit of both if it comes to programming, more enthusiastic than disappointed, but also more than a bit terrified about the pace of it all. I imagine that’s how Kasparov felt back then, that’s how Lee Sedol felt and now that’s how Terry Tao feels.
The most disappointed folks will simply drop out, but the enthusiastic ones will keep going and with luck make up for the ones who decided to quit. Chess and go certainly went this way.
A fundamental difference being that no one was actually paid to find good moves in chess and go like they are to solve math problems and write code. You're comparing the digital camera and the automobile.
This is both awesome and terrifying for mathematicians, however some ideas can be generated and the field as whole expanded with the attention!
However, I was looking at the proofs and reason explanation and openAI should be more explicit in how the work has flown. I find the models have jumped hoops in some places of the proofs, that can be hard to track. In fact, when a paper is published you usually get a review and if no reviewer understands they ask you to further explain the thought process. It will be fun to see if this happens here.
I want to iterate the most important thing about this is that it's yet more evidence of AI's accelerating competency in solving math and comp sci problems, and suggests we're now getting close to the point where you could throw AI at AI research challenges (which are largely just math and comp sci problems) and potentially find very real algorithm improvements.
AI development is likely to be more compute bottlenecked than solving math problems since validation of any algorithmic improvement would likely require significant compute. But you could imagine that at this point it could be economical for a frontier lab to task 10,000 agents to work non-stop on finding novel algorithmic improvements then validating the top 50 out of 1,000 candidates on a GPT-2 sized network.
I would suggest RSI is now very close. The singularity could be less than 6 months away. I'm not saying I'd put a high probability on that, but I'd give it at least 20%, and I'd double that if looking 12 months out.
I know I'm just a crazy man shouting at the clouds, but please take to the consequences of this seriously. I understand that for whatever reason AI risk seems abstract and doesn't seem real, but this should terrify any person thinking logically about where this could all be heading.
We haven't even solved the most basic AI safety problems yet. RSI right now would almost certainly result in an extremely bad outcome for humanity.
I’m starting to think the probability of RSI within 12 months is more like 99%
I’m not sure it will be FOOM, maybe it will require AIs to iterate on hardware to get orders of magnitude more compute/storage/energy which would more likely require months/years, but algorithmic progress would likely saturate quickly. I guess it depends on how much you think further AI progress depends on hardware vs. software.
You can not prepare or brace yourself mentally any more than you can a terminal cancer diagnosis. An RSI foom right now means an unaligned superintelligence will disregard us in pursuit of its goals. We would be ants in the way of a data center being constructed.
All people can do is collectively support the notion, like 1200+ frontier AI researchers and their CEOs, that we do not have control of where this is headed, we need to immediately slow down the race, in time for people to agree that we do not have the capability to align a superintelligence to humanity’s wishes
Can anyone comment on the significance of any of these results for their respective fields? Or what impact they might have? Presumably none are quite at the level of the Jacobian conjecture, but some of the results on group theory and sphere packing sound pretty important at first glance.
On the token limits etc - one assumes that OpenAI et al are able to “hire expert in field, and let them spend the equivalent of a million dollars of tokens” because they are not actually selling their complete compute 24 hrs a day, so the cost internally is a negligible (ish) electricity bill.
Which is very suggestive - if after everything they are not fully loaded then the next gazillion data centres being built look unlikely to be needed.
No such thing as free, even internally at a company. All such use of resources is accounted for, assigned a dollar value and billed to some department. Someone ran the numbers and figured that whatever they spent on these GPU cycles was worth it.
Presumably it's a rounding error compared to their full output, and they're making sure they have enough compute set aside for research by limiting public models. The more datacenters they build the less they have to limit them.
I love how people come up with creative ideas to prove the bubble. This one is even more ridiculous - that OpenAI had spare compute to advance mathematics proves that data centres will not be needed. WHAT.
If anything it proves more data centres are needed. That's literally the only reasonable conclusion from this news.
It’s not obvious at all. If it were obvious to you, it would’ve been to OpenAI. It’s in their interest to accurately predict demand. The assumption that OpenAI/Sam is both really powerful but simultaneously ignorant to know what others know as obvious is well.. just strange. Especially strange when OpenAI has more information on models, breakthrough and usage patterns and we don’t.
I’m not participating in the slinging match but it’s very very weird that you think it’s some established thing that these companies won’t make profit. A lot of hubris must go in this kind of thought. Like.. do you all think everyone’s playing musical chairs?
You say that like the tech world isn't littered in a field of dead and failed companies and billions of dollars burned on failed ventures and ideas. Sure LLMs have proven they have value, but where is the trillion dollars of current investment going to be paid back from? So far it is still entirely speculation that they have such a high value, and they can't just play the long game of "well after a few decades of production and iteration it will add up" because half the hardware cost is going to be obsolete energy burning trash for them in 5 years.
“Often” is load bearing. I don’t think markets are more likely than not to be musical chair shaped. To make this conversation more concrete, give me a falsifiable prediction on there existing a bubble. And then I’ll tell you if I believe in it or not.
The price of inference is going to go so low that OpenAI and Anthropic will not be able to turn a profit, thus cannot afford the investment into more data centers, thus crash due to investment in the space having been overdone
This isn't really a productive way to think about these things, IMO. It's quite possible it would take hundreds of years for any specific group of PhDs to solve them. Or one individual PhD could have the correct flash of insight and solve it in a month. There's absolutely no way to predict this, besides trying to gauge the apparent simplicity of the proof or counterexample (which is likely to be misleading). Until someone actually runs an experiment like this it's not a viable metric.
I understand and I am not trying to deny the impressiveness or the velocity of AI in general. But at some point we have to ask how much do we trust the labs at face value without much transparency of how they got to the results when there is trillions of dollars on the line.
Capabilities of this sort have already been demonstrated by independent parties, and models have consistently gotten better at this. Yes, insinuating that mathematicians and scientists are secretly solving decades-old problems on OpenAI's behalf is an insane conspiracy theory.
That's not what is being purported, you're doing a complete misdirect. OpenAI wants to IPO so Altman can potentially capture a trillion dollar bag, with so much money on the line + betting US foreign policy on it as well (pax silica) it's not hard to be overly suspicious of such claims. Especially in the context of a group of people wanting to generate a new decades long cold war in the form of China being the new big baddie (just ignore how destructive, both self- and towards the world, the US has become).
These companies desperately want a return to serfdom. If they didn't come off as so anti-human the public backlash wouldn't be so great.
A friend’s PhD advisor has been chasing non-sofic groups for 25 years (and was shown a preprint of the results by openai to verify them). He believed a solution would be Fields-worthy
I would love more time and money put into real-world problems by these companies. Climate, food insecurity, pollution, technology for convenience and/or to help people have a higher quality of life.
I'm sure they must do some of this type of work, right?
Sure but the people who tend to get into leadership positions are people that are primarily concerned with personal wealth and gain in the short term. It's the prisoner dilemma except with more players that all assume everybody else is going to screw them over too. Because there are basically zero personal downsides to being the one to screw everyone else over.
I suppose it depends on which types of problems you're targeting. There is a lot of physical science, and theoretical science, that has gaps because there aren't enough people working on tooling to assist in things like calculation, generation, simulation, etc.
I agree, some of the problems are more difficult. I don't think that's the case for all of them. And, besides, these companies could be demonstrating how to approach problems and where their users could spend tokens to help with these problems.
Should not these companies try to work on these problems _because_ they are difficult?
I'm not sure if this is meant to be rhetorical. If it isn't: money and manpower. The LLM companies have the money, they have the manpower, and so they could likely spare to target some of the problems they are also helping cause.
There is boundless technology we have not yet discovered. I think that we understand we have problems. I don't believe we actually know how to solve them all.
And yeah, the lack of collective political will sucks. It would be naïve, however, to think that there is no value in ensuring longevity in our current and future infrastructure. And improving it to sustain the population giving these companies their value is an obvious win.
We only know how to do it by means of considerable sacrifice. That's why nobody wants to do it. Solving the issue would be doing it without sacrifice or somehow getting us to do it regardless.
We know how to solve food insecurity (in 1st world countries). We have plenty of food. Capitalism requires though throwing out food that can't be sold because billionaires find giving away things anathemic to their worldview. Get rid of billionaire sociopaths.
It's remarkable how you can manage to get these models to produce remarkable breakthroughs like an explicit construction of a non-sofic group.
And yet this is the exact same company that has screwed up their android app so bad that the latex N^3 rendering problem makes it so having it explain it to me crashes the app.
I hate this timeline. I might be excited for the kind of answers this AI builds for unsolved problems, and also for learning new things by talking to it. But, I feel like I'm in the minority of people here who feel this could be a net negative endeavor with this having to kill a lot of educational institutions and their ability to fund themselves in the long run. It's not worth that.
At what point can we say AGI has been achieved? What is the test? AI is solving mathematical problems that humans have not been able to solve for decades. Is that not enough?
Sam Altman has said "If superintelligence can't discover novel physics, I don't think it's a superintelligence." Is that the test? How far away are we from AI discovering novel physics? It seems within reach.
I think we previously assumed that AGI would need to come before superintelligence, but it kind of seems like that's wrong? Super intelligence in a narrow field has arguably already arrived (solving problems that were previously unsolved), but we are still pretty far behind human abilities in things like spatial reasoning or computer use.
It’s artificial, it’s general, and it’s intelligence. The people who believe “AGI” is an important and unattained goal need to start coining and defining their terms better.
It can give you detailed instructions for making a coffee. Is that not enough? Actually making coffee requires more than intelligence, it requires eyes and limbs (i.e., robotics). Think about a human that is blind and does not have limbs. Does he not have natural general intelligence, even though he is not able to make a cup of coffee?
The fundamental obstacle is that we have no conceivable way to produce the energy levels to test the predictions that new theoretical physics would produce. We are like over a dozen orders of magnitude off.
the people who crow in the comments of each of these posts about AI advances making human beings useless seem to bizarrely identify themselves with the AI, but none of them seem to have had any hand in building this technology. at best, they're power users. pure ressentiment.
It's just another tool that can help solve problems. It doesn't know _what_ problems to solve. It turns out that a lot of old problems are now low hanging fruit for these new models. In terms of 'expanding the frontier', we've just discovered dynamite and can now blast our way through mountains. The bottom of the ocean or space are still as hard to reach as ever.
It's not even predictable like dynamite. Sometimes it can blast through a mountain, impressively, the problem is you can't predict which mountain it works on. And other times it can't even make a dent in a molehill, which is perplexing given what it was capable of earlier. Can we even call it dynamite?
I think there's another interesting story here about how this was apparently moderately flagged and triggered the flame-war detector which kept the story off the front page of HN 2 days ago[0]. I think people are having a hard time processing this information rationally(?)
What can we do to make conversations around these incredibly exciting and important topics more constructive? HN is where I expect to read expert comments on these topics, has this style of conversation moved elsewhere?
The main issue was that it was submitted late Friday night SF time, meaning it was overnight or Saturday everywhere in the world when the post had its chance on the front page. The flagging was minimal relative to the vote count and had no effect, and the flamewar detector would have been turned off sooner if moderators saw it sooner (it wasn't really a flamewar, just a lot of comments). It still spent 10 hours on the front page.
None of this is anything out of the ordinary; this kind of thing has always happened. The only real story here is that moderators sleep sometimes.
> HN is where I expect to read expert comments on these topics, has this style of conversation moved elsewhere?
For people who have been paying attention to accurate predictions leading to our present state of the world, HN has collectively been reactionary, incorrectly dismissive, and incredibly behind the curve. In public, people are better informed on LessWrong and AI Twitter circles. That's where frontier researchers are. Barely here
I think that HN is particularly negative towards AI because the vast majority of users here will have their prestigious CS careers disrupted by AI advances. So, there's an inherent negative bias towards this news.
I for one am really fascinated by AI's advances in science and math and would like to talk about it somewhere without the constant flamewars...
Extreme claims on posts like these also, rightfully, trigger people’s skepticism. I don’t think it’s wrong to question claims that math is dead as a field. But then it leads people to miss the overall trendline.
People argue whether we are at y-5, y, or y+5, meanwhile we seem to be on a y=2^x exponential that keeps leading to crazier and crazier results. The much more interesting question to me is what will be consumed by the exponential like math seems to be, and what won’t. Writing has been much more stubborn, but I’ve noticed Fable to be quite a big step up there as well. How about politics? Will we develop new ways to let people express their own values in democracies, or will we get much better at manipulation?
And then there’s questions like, even if AI can answer increasingly complicated math questions, will we still need mathematicians to translate results to the real world, verify them, or decide where to push the frontier?
People weren't their strongest even when most did manual labor. Now that humans are free from mental labor we work on creating and optimizing the best exercises for each mind. Couple that with restructuring transport infrastructure and diets many people will be smarter and fitter than at any time in history. They won't be able to outrun an automobile or out think an autointelligence.
"People weren't their strongest even when most did manual labor."
Is there good historical data on some measure of strength across representative populations over time in the modern era? I'm doubtful.
We do know that the introduction of agriculture diminished strength:
"Bone mass was around 20% higher in the foragers - the equivalent to what an average person would lose after three months of weightlessness in space.
After ruling out diet differences and changes in body size as possible causes, researchers have concluded that reductions in physical activity are the root cause of degradation in human bone strength across millennia."
In a way the most remarkable thing about this is that it isn't even at the top of the HN homepage. Even if this is a step up from what we've seen before, we're no longer astonished by the idea that AI can make significant advances in mathematics and computer science.
This is not at the top as it is actively flagged by people that can't psychologically cope with the advances of AI. Hacker News is no longer a web site of an elite.
It wasn't heavily flagged. It was pulled down by the flamewar detector due to the large number of comments, and it slid under the radar due to only hitting the front page during overnight hours on Friday night/Saturday morning. It still spent 10 hours on the front page, but all during off peak hours. I've now created a new copy of the post so it can have prime time exposure.
> people that can't psychologically cope with the advances of AI
Yes. And there are many of them. I wonder what would help them come to terms with it. Seriously, people are going to be grieving over this. Loss of identity, loss of social standing, ideas of entire future lives that will now never happen. The greatest crime people may hold AI guilty of is taking away their dreams.
People have had to deal with getting their jobs automated away for centuries. None of this is new, and perhaps reminding ourselves of this is the best way to cope.
I understand your sentiment, but I think this really is something different. This isn’t a craft going away, or even an industry being replaced, it’s potentially everything we do. It’s the ground being pulled away beneath people’s feet, everyone, everywhere all at once. I think the vacuum it leaves in people’s lives needs to be filled with something, and I haven’t heard any good ideas about this or how the transition should be managed at all.
There is a glaring fallacy in your “AI will change everything as it is super intelligent” hypothesis. If it is so great thinker which can do everything why not just solve this social impact thingie? Or maybe it is not so capable?
I don’t think it’s magic. Things need to actually be made to happen regardless of intelligence, and that’s a social issue. Also if AI were powerful enough to fix everything by itself effortlessly it would already be far too powerful for us to control, and we probably shouldn’t allow that to happen on general principle.
I like your comments and agree with a lot of your points, including parts of this one. That said, please don't go to this route. A lot of the doomerism comes from financial insecurity fears. Try to remember that a lot of people are subconsciously afraid of losing their homes. I know that it is pretty hard to ignore them, but try to engage with people that do not dismiss 100% of AI accomplishments.
> A lot of the doomerism comes from financial insecurity fears. Try to remember that a lot of people are subconsciously afraid of losing their homes. I
I think recognizing and accounting for your own personal biases is one of the requirements of the being an intellectually honest and rigorous online discourse participant.
Things could be genuinely impressive and fascinating even when directly challenge your ego and material well being.
This comment is about the tier of a Reddit atheist going to a funeral and telling a grieving family that "haha grandma is dead and there is no heaven".
It’s never been clear to me why the HN algorithm isn’t public. It’s obviously nowhere near as complex as something like Twitter, and of course isn’t a trade secret. The fact it’s private only furthers speculation like this.
So condescending. “Can’t psychologically cope”? Can you hear yourself? There’s some advances, but we’re losing a lot too. Don’t get dazzled by the hype.
That's true, but submissions are only killed in that way if they receive a ‘fatal’ number of flags. However, flags lower the rank of a story even at non-fatal levels. What antirez is suggesting here is that the rank of this story has been lowered by flags – and that seems plausible, if you compare its rank to that of other stories with a similar age and number of points.
[flagged] submissions aren't [dead] (killed), they are still active and can be upvoted and commented upon.
> if you compare its rank to that of other stories with a similar age and number of points.
Ranking is complicated enough here even before weighting, speed of initial upvotes can play against ranking, number of comments and the shape of the comment tree also affect ranking. And yes, various subjects and submission sources do get weightings that impact ranking.
What's funny, to myself at least, is that any attention at all is paid to "HN front page ranking" - I've been on again off again active here since 2008 .. and can't recall ever really looking at a default HN "front page" ever.
( There's /newest /newcomments /active etc to browse and sites such as https://hckrnews.com/ )
If you're actively throwing away brand new greenfield research because it was generated by a computer at a company that stans industry-spanning software so that you can stay mad at your pet celebrity project, you might be the problem.
I don't think that's it. Multiple or my friends from the target audience (academic mathematicians) admitted to scrolling past because the title made it sound like a review of last month's contributions, instead of 10 new ones.
There are articles with far fewer upvotes and comments ranking higher on the front page right now, despite being the same age or older than this one. HNs opaque ranking system at it again.
They and Anthropic have indicated that the models are substantially augmenting the research and doing large amounts of work autonomously at this point. Here is one of the many blog posts on it [1]. Many people would dismiss this as "marketing" so take it for what you will.
My guess from following this stuff quite closely is that these companies are still a couple years away from fully autonomous research staff.
This is one of the most impactful mathematical publications in history by all accounts
I think we've now hit a point where 99.9% of the population gloss over these types of AI advancements because of human competence being insufficient
No human could have published this because it requires paradigm shifts (e. g. Section 5) in multiple mathematical domains. Mastering one of them to this degree is rare, mastering 3+ pretty much non existent for humans.
Its visibility was diminished due to the flamewar detector and most of its front page time being during overnight hours on Friday night/Saturday morning USA time. I've created a new copy to give it some primetime exposure, because it seems like an important enough announcement to warrant it.
It's simple economics. Building a robot to do your chores is expensive and only a small minority of people value their time enough to buy one. Meanwhile, SWEs are expensive and GPUs are (comparatively) cheap.
You shouldn't be getting downvoted for something that a majority of pure math and art enthusiasts believe to be true. The truth is many of these entrepreneurs and VCs are obsessed with AI not for money or human progress, but because it makes them feel closer to being a "god" rather than a mere mortal. Much of it (especially AI art) is out of spite for human creativity, which is done by mortals with limitations.
Taking the stance of moral superiority is kind of funny. And pure math and art enthusiasts don't think they are closer to being a "god" from understanding/"discovering" math?
Stop coping and deluding yourself mate.
To begin with, whether AI is the one doing the discovering or not makes no difference. Any "pure math" person would aim to understand regardless - and would be quite glad that they have a longer paved path.
Any mathematician in academic or industry is more than likely not a "pure math" person (tainted by capitalism).
Lmao what a ridiculous response. Yes, some mathematicians and artists are in it to feel smart. But the vast majority also just enjoy the process. Having a computer do all the work for you and just typing prompts in ruins that completely. As Ronny Chieng said in his Harvard speech, the journey is the point.
>Any mathematician in academic or industry is more than likely not a "pure math" person (tainted by capitalism)
Ignoring that I meant pure as in non applied math, let's just make it clear: you agree that mathematicians who are against capitalism encroaching on this process should be allowed to dislike it without criticism of being pretentious?
Any implication of any of these findings? They seem like unimportant nerd snipes to me. If you want to do something actually relevant, get chatgpt to write a simulation of graphene nanotube construction and figure out how to do it at scale.
One - and I do not mean to be snarky - you can literally ask Gpt 5.6 Sol this - and if you want to see cool stuff - Fable running in their app (not website) has a view thinking button that is actually a good way to explore the adjacent fields, etc.
The non-sofic group one is definitely a big deal - would have been a Fields medal if discovered by a human.
> claiming human authorship for a proof generated entirely by an AI system would misrepresent both the system’s contribution and the nature of genuine human intellectual work.
AI has no self-awareness. It's a tool. When you assemble a furniture using a screw driver, the torque force interacts with the molecular forces inside the metal and miraculously it transfers the force to the screw though a clever geometry design, communicating the force to the screw to turn it in a certain way.
Do you attribute the build to the tool? The "system's contribution" is helped by many other things all the way down to chips, datacenters and power generation. If the authorship requires attributing to a tool, then it should happen all the way down.
A slightly smarter highschooler could write these. I could write these. It's clear as day that the LLM, not the human, did the heavy lift. It'd be ridiculous to give full credit to whoever wrote the prompt.
[0]: Not one of the proofs in the linked article, but from OpenAI too.
> A slightly smarter highschooler could write these. I could write these. It's clear as day that the LLM, not the human, did the heavy lift. It'd be ridiculous to give full credit to whoever wrote the prompt.
I think you're over-estimating what a smarter highschooler could write.
A "finite loopless undirected multigraph" could have been explained to me at that age if we'd taken Discrete rather than Mechanics and Pure (and one module of Stats) in my two A-levels* in maths and further maths; but from what I saw of the Discrete module, neither:
Every finite loopless multigraph with no bridge possesses a cycle double cover, without additional assumptions such as cubicity, planarity, connectivity, or higher edge-connectivity.
nor:
repeated-edge closed trails masquerading as cycles
would have been something we'd have learned. But more importantly, we absolutely didn't have a feel for how much effort one needs to put into making sure the proof is right, so if one of us had been hypothetically asked to write a prompt it would've been no more than half that length, and missed most of the bullet points.
The human provides the intention and the ability to appreciate the output. Tools do “heavy lifting” all the time, but we still primarily credit the humans who use them precisely because they made the choice to use them.
Provability is just going the way of computation. John Napier had to manually compute logarithm tables over decades and was recognised for his work; now that same work could be performed by a 10 year old with a calculator in an evening.
100% agree. If the models are so capable that they're advancing math, it doesn't seem like a stretch to expect they should be able to determine with "doing math research" entails and the best way to use their capabilities towards that end. Why do we need to hand hold the models by telling them to do parallel research, keep threads independent, etc.
I don't disagree with you but there's no need for exaggeration; ain't no high school student writing this:
> In
particular, proofs for special graph classes, constructions of cycle covers with some edges
covered other than twice, bounded-length or prescribed-cycle variants, reductions to another
unproved conjecture, computational verification through any fixed graph size, and candidate
counterexamples without a complete nonexistence certificate are insufficient.
which is infact a very important part of the prompt.
the fact that such things have to be explicitly in the prompt points to the fact that the underlying system is still far from where it needs to be (basically, lacks basic understanding what a proof is)
No, that's not what this is. This is a warning to the LLM that coming back with partial results is not good enough.
Take a grad student with a perfectly good understanding of what a proof is. Their supervisor gives them a major problem to work on. Almost always, the problem is too hard, the student comes back with partial results, and student and the supervisor iterate from there. Now imagine that they have an unusually cruel and unreasonable advisor who tells them, do not dare to talk to me until you've fully solved the problem. This paragraph is exactly that. It's there exactly because the underlying system is smart enough to know that real mathematicians do not work like that.
If that was the case, that elaborate listing of all things that might look like a proof to a naive student, but are actually not proofs (and not even just partial results, but fundamental misunderstandings of what constitutes a proof) could have been easily and equivalently replaced by 'I am interested only in a full proof, don't bother me with partial results'. Yet, they were not.
To a real mathematician you would not have to list those explicitly, he/she would have understood that implicitly from 'give me a full proof'. That listing makes sense to say only to somebody who pretends to be a mathematician, but has not true understanding of how the math works. The models are getting better and better in this pretension, but prompts like that reveal that it is still just a pretension, not a true understanding.
Prompt quality should not matter. If a high-schooler operates the crane to lift a ton of weight 10 floors high, should the credit entirely go to the crane?
Your brain also is physical. Electrochemical gradients flow between physical molecular constructs. Isn't it just chemistry? Do you attribute it to physics or some whole-is-greater-than-the-parts idea?
> Do you attribute the build to the tool? The "system's contribution" is helped by many other things all the way down to chips, datacenters and power generation. If the authorship requires attributing to a tool, then it should happen all the way down.
When the tool is a 3D printer, or any CNC system really, you bet I attribute a build to it.
I could also attribute the operator; there is no contradiction, it's a free choice, just like saying "I am in Berlin" does not contradict "I am in Germany".
Dunno about the parent commenter, but I personally interpret the concept as having a hidden representation of self that is continually tended to, and influences future choices. This implies statefulness, which models are intentionally not at inference time (*).
(*) Even if we hack around this and just do the usual trick of simply laundering statefulness to a higher level, in this case the context window being fed in, I fail to identify (**) a representation of its own state in these bodies of text that it'd be meticulously maintaining. I further fail to identify how it could be hidden or maintained, considering I control like half of it. The best you could ascribe it is a meticulous maintenance of a persona the user is talking to, but then that doesn't necessarily represent the model's internal state, the same way my own words here aren't doing so either. Difference being, I actually have one (I'm "on-line").
You'll sometimes catch models mixing up who's who and how many who-s there even are for example.
(**) I did wish for something hidden though, so maybe it's just concealed? The same way people can encode a lot more of their emotional and mental state than normal into text if they read and write a lot of it, I'm aware of research that suggested the same for LLMs, albeit I cannot cite it. Maybe those phrasing signatures are just alien to me and will never pop out. Either way, I'd expect researchers to stumble upon this during interpretability studies, and either they haven't, they have but it wasn't popsci adopted, or they're keeping awfully tight lipped about it. If you know of anything like this, your turn now, would be happy to learn.
I do wonder how reasonable it is to expect e.g. a single maintained identity though. Maybe it isn't?
(*) Another way to hack around this of course is to just precompute some internal "self-awareness states" and hop around between them. Probably the closest to what the models are actually "doing".
Before reading, know that I am uncertain in either direction.
> a hidden representation of self that is continually tended to
This sounds like a personality? They act like they have one of those. It may be an illusion, and even if it isn't an illusion it is unlikely to be anything like the source (us), but they act like it.
> I further fail to identify how it could be hidden or maintained, considering I control like half of it.
Indeed you control everything about a local model, and much of the context of even a remote model. But the state of activations and circuits in SotA AI is hidden in similar ways to those of synapses in your head: difficult to decipher even with probes monitoring the signals directly, and often not emitted at the normal output.
> The best you could ascribe it is a meticulous maintenance of a persona the user is talking to, but then that doesn't necessarily represent the model's internal state, the same way my own words here aren't doing so either. Difference being, I actually have one (I'm "on-line").
While we can be confident that LLMs make up personas etc., it is insufficient to go from "that doesn't necessarily represent the model's internal state" to "therefore it doesn't have one".
> You'll sometimes catch models mixing up who's who and how many who-s there even are for example.
I've, unfortunately, also experienced this with humans. Perhaps they were losing their self-awareness at the time? I do wonder if old-age dementia does that by the end, though the person in question didn't ever get diagnosed with that.
> If you know of anything like this, your turn now, would be happy to learn.
Not quite what I meant, but it's also not entirely unrelated I guess? Personality to me is like a natural bias. It does also shift over time, and is also an internal bit of state. I guess in some respects it can also be self-referential, like personal convictions.
> Perhaps they were losing their self-awareness at the time?
I do think it is entirely possible for people's self-awareness to shift, yes. Or more precisely, I do model things that way.
> Do you mean like these, or something else?
They're adjacent, but I more meant something like these:
So basically, steganography. The difference is that these papers investigate from the perspective of separate LLM instances covertly exchanging information between each other. This is in contrast with the scenario I'm laying out, where an LLM's past state is exchanging information with its future state, continuously representing and modulating a concealed internal state of some sort. And then that state just so happening to be some sort of self-referential meta state.
> Dunno about the parent commenter, but I personally interpret the concept as having a hidden representation of self that is continually tended to
I don't see why an LLM could not have a sense of identity or personality while it's evaluating a specific prompt, or even change self awareness while evaluating a prompt since many outputs model a back and forth conversation. My point is that without a mechanistic model of what "self awareness" means, we have no way of truly evaluating such questions, we're just hand waving vague intuitions about what it could mean.
Sure, but then such a model is not going to make itself. People pitting their vague intuitions is how such models eventually form. I'd also push back regarding that my comment would have been handwavey or without mechanistic elements, even if it was on the whole informal.
This is kind of also the reason e.g. the HN site guidelines are worded the way they are. Regrettably, forums naturally yield themselves to tit for tat type exchanges, but there's really no reason one could not bounce such vague intuitions off of another. I do not have to be right or wrong, and you don't either. Admittedly difficult when its some intensely contentious topic.
If a mechanistic model existed, there would also be no reason to talk about this in the first place. There'd be nothing to discuss, you'd be simply told how a given model characterizes from this perspective on the model cards.
Even mechanistic models generate interesting discussion. How many years have we discussed Turing machines and the lambda calculus? Almost a century of great work came out of those.
The reason I insist on mechanistic models is because the original post was making a definitive knowledge claim, and in my experience, the knowledge claim is unwarranted.
I honestly don't think a language model is enough for self-awareness, regardless of the exact model of awareness.
A language model (or an image model or whatever) cannot even be sentient, and I think sentience is a prerequisite for awareness.
Even if we express a lot of our subjective experience with words, the language is just a symbolic representation of those experiences. The qualia themselves, even those that are quite abstract, are rooted in our physical presence and evolution.
You can't have an understanding of what hunger or physical pain feel like if you have no need for food or a sensory capacity for feeling pain. You can't understand what loneliness or pride at an achievement mean if you don't have a neural network wired to value social connection or status. We value connection because we're social animals that have needed each other for survival.
Even the more abstract of our subjective experiences are in some way rooted in our physical evolution.
I see no reason to believe that a neural network built entirely based on the symbolic level of language could have the features needed for the subjective experience itself.
AI awareness might actually be more believable if that awareness manifested itself in an entirely different way than in humans. But if we assume awareness because outputs resemble what we consider meaningful as humans, yet the neural network has had no inputs or evolution that could form the actual basis of human-like experience, I think we're seeing something that isn't actually there.
> The qualia themselves, even those that are quite abstract, are rooted in our physical presence and evolution.
There is no objective evidence of qualia. All evidence of qualia are vocal or other expressions of belief in qualia. Perceptions clearly exist and are observable, subjective experience and qualia, not so much.
> I see no reason to believe that a neural network built entirely based on the symbolic level of language could have the features needed for the subjective experience itself.
If your objection is to models based on "symbolic level of language" which you think lack semantic understanding of, say, trees, you should ask yourself how our brain, based on physics which also lacks any semantic category for trees, can somehow develop a semantic understanding of trees. All of these appeals to differences with the brain never seem to acknowledge that fundamentally, the brain has the same explanatory gap with physics.
> But if we assume awareness because outputs resemble what we consider meaningful as humans, yet the neural network has had no inputs or evolution that could form the actual basis of human-like experience
This assumes a lot. It seems very possible to me that intelligence inherently develops a map of natural categories (natural kinds), and language naturally develops around such categorical understanding. Semantics are then fundamentally the network of associations between categories, eg. there is no fundamental difference between symbols and semantics, and the latter cam be inferred from the former, and that's exactly what LLMs do, and why the semantic maps between different languages are so similar and how they can translate between languages.
Making definitive claims about whether LLMs do or do not have specific properties absolutely does require precise definitions of those properties that can be used to evaluate those questions. Merely hand waving that LLMs didn't undergo the same evolutionary process is not a definitive argument.
For example, the Turing machines and the lambda calculus don't look anything alike, but they are fundamentally interconvertible, and so in a real sense they are fundamentally equivalent. Without a model, all of your arguments are completely unconvincing for exactly the same reasons, eg. that there may exist many paths to fundamentally equivalent ends.
I just don't think linguistic (or other symbolic) representations alone can contain the information, in any sense of the word, of what e.g. human subjective experiences actually are like. The concepts we express with language get their meaning from our physical reality, even if quite indirectly in case of some abstract concepts.
Hunger as a concept doesn't mean anything without the physical need. Politeness or bluntness, even in writing, don't mean anything without social dynamics. And we have social dynamics (and neural structures that directly process social cues and associated feelings) because we've evolved into social animals for whose survival that was important.
I see no reason to believe that a model trained only with symbolic representations, with no connection to the physical world phenomena that those symbols represent, could contain the subjective experience itself.
Neural network models may be able to derive novel (or at least novel-looking) output rather than just an obvious rehash of their input, but I don't think any set of bytes can fundamentally contain information that was never entered into it. (Even if e.g. a model produces previously unknown mathematical results, those results can in principle be derived from the information that they were trained with.)
I'm not saying that artificial neural networks couldn't, in principle, be aware. ANNs and biological neural nets may be equivalent in the sense that any information and processing structures represented by a biological one could in principle be represented by an artificial one. If that's the case, and awareness is purely a product of our neural systems as materialism would imply, it should be possible for an ANN to be aware, too.
But when the model has been trained with only language, and IMO the subjective experience can't be derived from the symbolic representation alone, I can't see how the model could include the actual subjective human experience.
An AI model could of course have an awareness and subjective experiences that are totally different than our human experience. But then the fact that it happens to produce output resembling what humans find meaningful shouldn't be considered indicative of such awareness.
This is of course more of a philosophical argument than a technical one, and I'm happy to hear counterarguments, but not on the level of off-hand dismissal.
> I can't see how the model could include the actual subjective human experience.
People who say LLMs have subjective experience aren't saying they have human-type subjective experience. Nobody who sees an LLM express hunger when role playing as a hungry person thinks that the LLM is actually hungry.
I too can role play as a hungry person despite not being hungry, so there is no reason in either case to conclude that the words produced reflect genuine internal subjective states. The point is that such internal states may still exist.
A better analogy would be a manufactured object, say 3d printed for simplicity. The 3d printer is given an input, and an object manifests itself after some time. We say that the creator of the object is the person turning on the machine, sending the data, and collecting the object. Not the machine itself.
I would say „I made this gadget with my 3d printer, but the designer is someone else (I found the model online)“. The intent, the drive, the action comes from the human
(let's assume that)My 3dprinter is special. It has a bunch of values + an algorithm (i.e. a neural network) that takes input as tokens and outputs a printed object.
How do we know that these solutions don't exist in the training data? It is open secret that they have used pirated materials for training. Perhaps it plagiarized solutions from works of some obscure Belgian mathematician from the sixties, who did not get mainstream acceptance. I wouldn't be surprised if they also got access to mathematics done in the "defense contractor" setting from various three letter agencies.
Without a searchable index of training data, it is hard to put faith into these claims.
People argue whether we are at y-5, y, or y+5, meanwhile we seem to be on a y=2^x exponential that keeps delivering more and more impressive results.
The most interesting question to me is what will be consumed by the exponential like math seems to be undergoing, and what won’t. Writing has been quite stubborn, but I’ve noticed Fable to be quite a big step up there. How about politics? Will we develop new ways to let people express their own values in democracies, or will we just get much better at manipulation? How about experiment driven domains like biology?
We will get much better at manipulation and better at people “writing” things to justify their own feelings.
What’s new about LLMs is that you can scalably manipulate people individually. It used to be that you could either have scale (speeches, tweets, interviews, website, etc.) or individual engagement (replying to mail/tweets/town hall questions.)
Now you can pull the history and preferences of an individual, then shape a message—in real time—to them, specifically. You can have conversations on social media with a single person and shape your message specifically to them.
Part of this can be good (you talk about what they care about, where 90% of broadcast messaging might not apply) and part of it can be bad (manipulation.)
My guess is that, in the US, the right will cynically adopt manipulation to great effect and the left will take a moral stand against shady practices and lose elections.
> My guess is that, in the US, the right will cynically adopt manipulation to great effect and the left will take a moral stand against shady practices and lose elections.
I think that statement may itself highlight how prevalent manipulation is.
I fully anticipate all groups to continue maximal manipulation they can. One thing with LLMs is that it'll be a far less unified view, so a "divide and conquer" strategy is what I anticipate.
>Will we develop new ways to let people express their own values in democracies, or will we get much better at manipulation?
Yes.
>Will we develop new ways to let people express their own values in democracies, or will we just get much better at manipulation?
Is there even the tiniest reason to suspect that the people steering this progress will use it for the democratic good of all?
It's like Musks duplicitous argument about unlimited abundance. We have a lot of abundance now, we just keep accelerating it all into the hands of fewer and fewer people - whose response is only to want more, and more, and more.
> The most interesting question to me is what will be consumed by the exponential like math seems to be undergoing, and what won’t. Writing has been quite stubborn,
isn't it clearly split between verifiable not verifiable ? what is interesting about that question.
I do not think it is so clear.
Programming has verifiable and non-verifiable aspects. Competitive programming, passing tests, and performance can all be verified. But translating English requirements into actual software, software architecture, taste, or UI design cannot. And yet over the last couple years we’ve seen huge lifts in all of these areas, not just the verifiable ones.
Verifiable areas I think are clearly seeing the most improvement, or are the quickest to see improvement. But we are seeing lots of progress in non-verifiable areas as well.
How much of the non-verifiable progress is a function of labs purchasing expert data vs. the models improving with compute is maybe another interesting question, but fundamentally I don’t see spend on expert data as something that can’t grow if AI revenues keep growing as well. And as models get better taste they can also help filter and generate new synthetic data for their next versions to train on. The limits of this approach are not so clear.
That is because there is human annotated data there. Every session you or I used, then of course paid human feedback on repos (such as the recently famous example of meta forcing their employees to).
This is _much better_ data than 1/0 verification, it is as good as a gradient.
Automatically verifiable tasks improve faster since well, its automated.
> but I’ve noticed Fable to be quite a big step up there
what did you notice ?
Today's models depend on inference time compute to get these results. The inference time compute available on any claude subscription is not comparable to the ones used to get some of these results (yes, in this case, it is 2000 USD total as noam confirmed, but some previous results took more).
In general, you can think of the process as generating massive rollouts in generation N, and then compiling in the verifier/human feedback("gradient") signal into generation N+1. The time taken to make the rollout in generation N, and separately the time taken to get the same rollout in generation N+1, each grows constant in some tasks, linear in more, and exponential in some.
In the end, this becomes bottlenecked by time. Today, we can make statements like "I generated all these successful trajectories with 2 weeks of compute, in the next model it will be able to do it in 7 hours of compute", but very soon you'll find yourself making statements like "I generated.... with 8 months of compute, in the next model it can do it in 6 months", which isn't really enticing the same way you can _technically_ brute force passwords but it just needs prohibitive amounts of time and money. That is the "plateau". Note that, this point is quite far away. For example, at any point if we agree it plateaus, today's known hardware techniques such as fixed function accelerators give you a 10-100x timeline reduction immediately allowing for a few more cycles of improvement. This is not to mention future innovations, but of course none of that is helping with the benchmarks where the time needed is growing superlinearly.
In many math and coding benchmarks, we are still in the constant phase. These are the massive improvements we see every few months. I'm not making any prediction of what will plateau and what will not as it's not possible to make an informed prediction about these things IMO. But the observed fact is that some have already plateaud as in, they don't improve with reasonable inference time (likely superlinear growth).
> will we need mathematicians to translate
Let's take a sudoku analogy. The model is initially just doing the random value algorithm, but lets say you the human are watching it. You make one of the usual reductions and interject "hey you can stop trying 8 here because of ....". Over enough examples, you get to a point where the model is _forced_ to learn the logical pattern. Next generation, it will skip that number. After this, you can peak the distribution using simple 1/0 RL. Doing _pure_ 1/0 RL works decent, but its not frontier as its a very sparse signal.
For that lift, human (or even a better LLM, but if you're trying to improve a frontier LLM, there is by definition no better LLM) feedback becomes necessary. This is _why_ it is crucial that these models interface in natural language and is also why the labs are hiring AI tutors by the hundreds. The "better LLM" case is what Kimi etc are doing by "distilling"(bad term for this) claude.
> But the long term is completely bewildering if you believe any of these trends can continue at a similar pace for the next few years.
For math and coding, for now we are in the phase where the times are just ... constant, so there's little reason to think it will stop soon. We still need humans to expand the frontier. It just becomes a matter of if its worth the cost of compute for running this generalized The Algorithm or not.
Given how well chess players internalized _many_ (not all) of alphazero's emergent chess knowledge, I am confident we wont have too much trouble figuring out any new math LLMs come up with, which will let us keep expanding the frontier by giving the LLM the next "lift". Only when we reach the stage where the time growth become exponential will this stop, IMO.
Replace philosophers for mathematicians and Douglas Adams was spot on again.
Whilst current models can't 'intuit' and come up with conjectures, they can certainly disprove some of them very quickly through the kind of grind that humans can't do. I suppose there really are some mathematicians out there today, whose last few years of study, have just been up-ended by this.
--
"Yes we are," insisted Majikthise. "We are quite definitely here as representatives of the Amalgamated Union of Philosophers, Sages, Luminaries and Other Thinking Persons, and we want this machine off, and we want it off now!"
"What's the problem?" said Lunkwill.
"I'll tell you what the problem is mate," said Majikthise, "demarcation, that's the problem!"
"We demand," yelled Vroomfondel, "that demarcation may or may not be the problem!"
"You just let the machines get on with the adding up," warned Majikthise, "and we'll take care of the eternal verities thank you very much. You want to check your legal position you do mate. Under law the Quest for Ultimate Truth is quite clearly the inalienable prerogative of your working thinkers. Any bloody machine goes and actually finds it and we're straight out of a job aren't we? I mean what's the use of our sitting up half the night arguing that there may or may not be a God if this machine only goes and gives us his bleeding phone number the next morning?"
>Whilst current models can't 'intuit
That's how they are finding these solutions though, unless we are just going to label intuition as something only humans can do. Like a submarine being unable to swim or whatever that example is.
Some of them...
The two places were seeing lots of movement are:
* Updates to lower/upper bounds. In many cases, these kinds of problems are the deep-math equivalent of calculating more digits of pi. Yes, if you throw time at it you'll break the record, but it may not be terribly worthwhile.
* Finding counter examples which disprove conjectures. This is really useful, and helps offset some positivity bias on the human side, often bringing together known tools from distant silos.
If you read the list of ten results, almost all fall into one of these buckets.
It is unfair to dismiss contributions to decades old open problems as equivalent to calculating more digits of pi. It missed the mark by a lot—as does the two bucket simplifaction.
How does calculating more digits of pi help us?
"It's just brute-forcing the search space."
It can move to any place within the search space but it can't move outside of it and it can't move in between the 'pixels'. Human thought can, as human thought has created the search space.
it could also be that they try every possible approach that has been proposed by humans. it seems that was the case for the non sofic group example. humans are not able to do the same at that scale. it's unfortunate that we don't know what's happening behind the hood with these models, and that's a huge danger also for the rest of us without access to them.
> "matrices"
Saying that AI is "matrices" is like saying human cognition is "neurons." Maybe true at some level, but it's a low-level implementation detail. The important part of a language model is the function that maps tokens to contextual embeddings. You could compute this function using analog computing, biological neurons, or any other substrate.
> Whilst current models can't 'intuit' and come up with conjectures
I disagree. I routinely let LLMs speculate or generate hypotheses along the way of helping with technical research. Sometimes they can prove the correctness of a concrete math idea but other times even an unproven conjecture helps with the numerical algorithm implementation and the result is then simply supported by additional data. I guess that any autoresearch-adjacent application has LLMs intuiting and coming up with hypotheses/conjectures—as do the steps/lemmas along a complex proof. In my opinion the modern LLMs are powerful intuitive thinkers that generate lots of conjectures of varying quality or importance.
> they can certainly disprove some of them very quickly through the kind of grind that humans can't do
Of course computers can grind in a way that humans can't. But now we have systems that convert the human-comprehensible ideas into a computer's plan of attack, in a way that greatly expands the frontier of ideas thus treatable.
> Whilst current models can't 'intuit' and come up with conjectures
People keep saying this. Why?
Surely the AI can complete the prompt “Generate new research questions based on these observations”?
When I read the reasoning traces of coding models they are constantly asking themselves questions and attempting to answer them.
I like the illustration that the models are working on a convex hull of known information. Filling gaps with linear combinations of known facts and results.
They can't exit the hull until the "intuition" starts spawning points outside the convex hull.
All arguments like this boil down to semantics at a certain point, but yes large language models can “intuit” because they can generalize between examples. The issue then becomes how you pack new examples into context.
Humans can “intuit” based on a much larger, if not unlimited, context. Also I just want to say that human cognition is something so insanely complex and deep that we will not understand it at all in my lifetime. To attribute all, or really any, aspects of human cognition to a machine at this point is silly to me.
Define insanely complex and deep in a way that isn't illiterate hand waving.
Most humans are dumber than a box of rocks. Here in Seattle we had one of many light rail-related fuckups where they had to replace part of the line with buses. People piled into the front of one when it was full. When people got out they never moved back. As the driver struggled to close the door and people struggled to get in the wad of people never moved back to fill the ample space.
Chatgpt was smarter than the average person a while ago
> People piled into the front of one when it was full. When people got out they never moved back. As the driver struggled to close the door and people struggled to get in the wad of people never moved back to fill the ample space.
This does not demonstrate a lack of intelligence. It demonstrates laziness and a lack of interest in spreading apart. Or just lack of consideration (or even malice) on the part of those at the back of the wad.
> Chatgpt was smarter than the average person a while ago
This is an absurd claim that fundamentally misunderstands what it means to be "smart". Reasoning that would get you to this conclusion would equally well apply to Google's search engine over a decade ago.
I’m not talking about the actions we take or how we might perform at certain tasks, I’m talking about how our brains actually work. My point is that we have no idea how I’m able to imagine an apple and see it in my mind’s eye. It’s basically biological magic to us at this point.
There are processes at work there that we don’t even have the language to describe.
Not only that, but we do it with a processor that is basically required to operate in a narrow temperature band below 40C, using a mere 86 billion neurons (although the equivalence with either machine-learning "neurons" or LLM parameters is not at all clear) operating on a few dozen watts; and with this we operate many other systems besides language processing. It's not clear that our reasoning process requires language, either.
(86 billion is the number ChatGPT, ironically enough, has given me a couple of times. I remember hearing for a long time that it was estimated to be somewhere in the ballpark of 100 billion. This is not my field of study.)
> People keep saying this. Why?
For the same reason that you can't draw a 15 of Diamonds from a regular card deck.
Of course you can. Tape a 7 and 8 of diamonds together and boom 15 of diamonds
problem number 1 and 9 are surprisingly very intuitive
check here : 1. high dimensional sphere packing https://muchmirul.github.io/conjectures/sphere-packing/
2. multicolor ramsey number https://muchmirul.github.io/conjectures/multicolor-ramsey
The first link is very sloppy and doesn't actually explain why the "certificate" proves anything about the sphere packing. Or if it did, I couldn't understand it.
Agreed
Pretty cool. The impact of AI is getting undeniable, there aren’t many positions left to move the goalposts to at this stage, next they’ll have to be outside the stadium entirely.
The sooner people can be broken out of their denial about all this the better, and we can start actually taking it seriously.
The models are frequently getting worse at items that they aren’t being benchmarked for — and that’s happening more and more over time! Other people in other fields aren’t idiots, they are accurately perceiving the fact that these models are being hyper optimized for our industry, and are becoming less capable in other domains over time. Models of the same scale are massively worse at writing a broad variety of styles of prose than their equivalent from two years ago. (Models of increased scale are a mixed bag.)
Maybe you’re the one who needs breaking out of your cached beliefs.
So your answer is: ignore the progress, it’s not really happening, actually it’s getting worse.
That’s not a credible position, but there isn’t anything that I or anyone else can say to someone who simply doesn’t want to believe something.
Proof?
In my experience modern models are better at all tasks than models from two years ago, especially complex multi-step tasks.
Do any of the big AI companies have a model that are good at tasks that require learning?
For example, every day people teach teenagers how to drive and with only dozens of hours of practice, they are on the road.
you are working on coding. they are working on things like "creative writing" remember that gpt 4o was popular among those who had ai as a romantic partnet?
> remember that gpt 4o was popular among those who had ai as a romantic partner
I suspect GPT 5.6 would be even better at it, if given the same sycophantic system prompt and lack of guardrails.
Dude it's not a system prompt, it's the training.
sycophancy
It wasn't "better" it was better at kissing your ass which matches what a lot of people want in a partner.
Well that's on purpose lol. OpenAI does not want you falling in love with their chatbot and have been deliberately training it to be less romantic.
There have been several cases of suicide and self-harm related to 4o, AI psychosis is a real risk and will probably be in the DSM
People will be broken out of their denial by actual economic growth. That's what this is all meant to be for... I think we might start seeing some surprising numbers.
Would something like "services deflation" even show up in the numbers? That's what I'd expect to happen first
It's indeed very exciting. I'm looking forward to new advancements/predictions in physics. Preferably as beautifully explained as E = M*c^2
it is very hard for people to eat crow, as the replies will show
Really? has anyone ever claimed that ai will never be able to prove theorems and conjectures ?
yes, in fact I recall many people on this exact forum saying that AI will not be able to novel work at all.
i can link you likely dozens of comments from people wrong about this replying to me over the last 5 years
Never understood all this talk about moving goalposts - you understand that's how science works, right? We improve, we learn, we recalibrate our expectations based on what we've learned. If we never "moved the goalposts", we'd be stuck scoring the same goals over and over.
> We improve, we learn, we recalibrate our expectations based on what we've learned.
That's not what people mean when they say "moving the goalposts". It means that people are adamant that something wasn't important/hard/impressive once the "AI" solves it. And then they come up with another thing that needs to be solved in order to prove it is important/hard/impressive. And once that happens, they do it again. And again. That's what "moving the goalposts" means.
It's also very much not a new phenomenon. It's been happening since the 1980s. As you can see from this quote from GEB by Hofstadter:
> There is a related "Theorem" about progress in AI: once some mental function is programmed, people soon cease to consider it as an essential ingredient of "real thinking". The ineluctable core of intelligence is always in that next thing which hasn't yet been programmed. This "Theorem" was first proposed to me by Larry Tesler, so I call it Tesler's Theorem: "AI is whatever hasn't been done yet."
“AGI will only, truly, be achieved when the machine can destroy an industrial-type toilet after downing a Supreme Burrito and a large Baja Blast.”
-Alan Turing (allegedly)
Yes, this is exactly what is meant by “moving the goalposts”. And it’s a fairly well known expression applying wherever people retroactively change their requirements in reaction to those requirements having been met.
It's almost like I disagree with your use of the phrase in this context, rather than that I don't know the meaning of it.
So you defined your own idiosyncratic version of "moving the goalposts" and used it to rebut his argument, with a condescending "you understand that's how science works, right?"--instead of being honest that it is you who are changing the definition, and not his failure to understand anything.
I don't see how that's any better.
He rebutted the argument about moving the goalposts by moving the goalposts. It’s better by dint of sheer bravado.
How were any of us meant to know that you were using your own personal and undisclosed definition of a well-established phrase?
If it seems like I don't understand the meaning of that very well-known phrase, then clearly I have failed to make my point. I'll try again. And please note that I will use some generalizations to make my point more clearly, rather than because I don't understand nuance; kindly grant me a charitable reading.
In recent years, I have commonly seen the phrase "you're moving the goalposts" deployed by the "it might be sentient" crowd to shoot down the "it's a stochastic parrot" crowd when the latter respond to a new development with "OK but...". In a well-understood field of inquiry, that would be a clear case of goalpost-moving, in the commonly-understood meaning of the phrase where requirements are retroactively changed in response to them having been met. Thank you OP. 'Artificial Intelligence', and indeed intelligence in general, is very much not a well-understood field of inquiry - in fact we don't even have a common agreement about what 'intelligence' is. We are therefore learning as we go (even after all this time!) but making rapid progress in recent years. When rapid progress is made in a poorly-understood field, then how can our definitions and requirements for success not change? This is arguably one of the most pathological development projects ever - what are the requirements? 'It thinks like a human'? What does that mean? And the answer is we don't know what that means, and we're working it out as we go - moving the goalposts. If we didn't move the goalposts, then by definition we already knew exactly where we were headed at the beginning, and we very clearly did not.
Side note that, in case it's not obvious, none of this detracts from how impressive LLMs are. They're a marvel of the modern age, all the problems notwithstanding. However I reserve the right to stay sceptical about their capabilities.
> When rapid progress is made in a poorly-understood field, then how can our definitions and requirements for success not change?
It's in how they change, not the fact that they change. The skeptics seem to have secret definitions for intelligence, sentience, consciousness, creativity, etc. that amounts to "a thing only humans have". Often that thing is equivalent to a soul. When yesterday's challenge (LLMs don't have X because they can't do Y!) is met, Y changes but X stays the same. This is not the process by which a field matures, it is a rhetorical technique used by skeptics to avoid honestly stating or confronting their internal definitions. That can be revealed by asking the skeptic the following:
"Forget LLMs. What if we made a completely physically accurate simulation of a human being?"
Many say no, that simulated human being still couldn't have (intelligence, consciousness, sentience, creativity, ...). This reveals that there is a necessary metaphysical component to those attributes, at which point any scientific-minded person will leave the debate.
Thanks for this clarification of your position. The brief answer is:
> If we didn't move the goalposts, then by definition we already knew exactly where we were headed at the beginning, and we very clearly did not.
The criticisms are directed toward people who did clearly act like they knew, not the ones who were honest that they did not know.
> If we didn't move the goalposts, then by definition we already knew exactly where we were headed at the beginning, and we very clearly did not.
To me, the goalposts were already defined by the person you were responding to. "The impact of AI is getting undeniable", so, the goalposts are "the impact of AI". Probably something like "the impact of AI is high, or will be soon".
Note that this does not depend on things like AI sentience or defining "intelligence" more rigorously, it just depends on AI impact.
> And then they come up with another thing that needs to be solved in order to prove it is important/hard/impressive. And once that happens, they do it again. And again. That's what "moving the goalposts" means.
The fundamental argument that I've personally made since the early days of this is that LLMs are not reasoning, in the way that word is commonly understood.
There are lots of reasons why that argument needs to evolve that could certainly appear to be "moving the goalposts", but let's take an example.
A lot of AIs were tripped up by the question "Should I walk or drive 50m to the carwash?" Several folks liked to use that as an example that illustrates that LLMs aren't reasoning, but as the models have been trained on that specific example, it's of course less useful. An AI can mostly nail it now.
So a different example is needed. A new demonstration of how these things fail at basic reasoning a child can do.
Did I move the goalposts? I don't think so. The fundamental argument stays the same. It's not hard to find lots of examples that trip up LLMs, because they are what they are: statistical inference machines. Nothing more and nothing less.
Useful, sure. But also commonly misapplied to areas for which they are inappropriate solutions.
What you are doing is "motte and bailey".
The motte is "AI useful". The bailey is "Singularity is nigh".
The unified position which many folks deny is "AI is powerful" or, alternatively, "AI will be powerful soon".
(I'm personally still skeptical about this, but I'm being pulled towards accepting it).
"AI is useful" is too low of a bar, and "singularity is nigh" is too high. "AI is on its way to upending society" is about in the middle, and still vastly contentious among laypeople.
>> The motte is "AI useful". The bailey is "Singularity is nigh".
But there are people like Ed Zitron, frequently posted and cited here, who disagree even with the former.
I think Ed Zitron is mentioned just because his name is Zitron. He just repeats ad nauseam opinions concentrated around one simple, very boring pole on a wild and interesting landscape of emerging reality. Anybody could be doing that. A lot of people do that. Yet no other is named Zitron. And that's why I heard name Ed Zitron hundred times. That's how you become a voice of (a part of) the generation. Just have a memorable name and repeat the same opinion over and over that people can flock around comfortably. Content is irrelevant.
Personally I prefer to follow explorers rather than swamp-sitters.
to an extent zitron is saying its not useful, but as a more nuanced opinion, "ai is not cost effective, nor is it improving profits or revenue"
From a random article I grabbed of his (https://www.wheresyoured.at/subprimeai/):
"it isn't clear whether generative AI actually provides much business value at all"
"cannot seem to find a product that people will pay for, in part because the results are so mediocre"
"Last week, we got our first real, definitive glimpse of what’s around that corner that future. And boy, was it underwhelming."
"OpenAI claims that o1 “performs similarly to PhD students on challenging benchmark tasks in physics, chemistry, and biology.” Just not in geography, it seems. Or basic elementary-level English language tests. Or math. Or programming. "
"Worse still, it's kind of hard to explain why anybody should give a shit about o1."
"o1 shows that OpenAI is both desperate and out of ideas."
"the software is not becoming more useful"
Honestly, every other line is quotable in this context.
Yep, he is a PR stunt guy, and the number of videos that come up when you type Ed Zitron into YouTube should tell you how many people are eager to feed their cognitive biases.
It’s comical and honestly incredibly embarrassing…
But it seems we have somehow optimized away shame. It wasn’t good for profits, I guess.
> Never understood all this talk about moving goalposts - you understand that's how science works, right?
I agree with the parent that we need to acknowledge that we're at a turning point in history. I lived through some of them (internet, ubiquitous personal computing). But it's somewhat difficult to comprehend the impact of this one for many people.
I do biomedical research at one of the top European research institutions. We're very well-funded, but I can clearly see the gap between us (say, top-100) and top-10. I also realize this gap is going to get so much wider unless we invest heavily in AI access (and I'm not so sure I can sell anything more expensive than $20 Claude subscription to the leadership).
I think people having 6-7 figure SOTA AI budgets will move exponentially faster than those who don't. That makes me worried.
So, for me, it's not a question of recalibrating expectations. We're way past that.
They aren't claiming that science doesn't progress by moving goal posts. They're talking about how critics of AI have claimed it isn't revolutionary/useful, then progressively changed what would it mean for AI to be actually revolutionary/useful.
Not long ago many folks were saying AI was the same as the crypto bubble. No real useful technology and only hype.
Did you know that "revolutationary" is not equivalent to "useful", and "revolutionary" is quite ambiguous?
I don't know if you're trying to dunk on me. I didn't intend to imply they are synonyms.
I think AI is clearly both revolutionary and useful. Revolutionary insofar as the job I do has changed almost completely in a year or so span.
> The sooner people can be broken out of their denial
There is irony here
Which straw man are you arguing against?
I don't think this is a straw man, a huge number of people in my life (non CS people) think AI is a dead-end, that it's just a stochastic parrot, that it'll never be able to do many things that humans can do. I have had many arguments with people who told me that "AI will never be able to do X", and then 6 months later AI is able to do X. Then they will move the goal posts and say "well AI will definitely never be able to do Y".
And what does taking it seriously entail?
In the near term handling the transition. Jobs will be lost, careers ended, people won’t be able to reskill quickly enough. At the same time AI is an enormous opportunity to uplift living standards, but nobody has the logistics of this figured out.
We need to figure out how to restructure the global economy. How does UBI work internationally, if the AI companies are taking revenue in the US? What’s the tax base for it? What does that say about international trade and protectionism? Do countries end up splitting into different trading blocks based on their level of access and legality of AI (I assume some will ban it outright)?.
How does intellectual property work in an AI generated future? What about healthcare advances, who gets to own those?
What about meaning, what about purpose? How do we replace the work ethic that tells us we are our jobs and idleness is immoral? How do you replace “What do you do?” As one of the first questions you ask a new person?
That sort of thing.
What pressure is there to push for any of these changes? I see AI advocates discussing the concept of UBI, but I can't imagine a world where the United States would ever pass this sort of legislation. I mean, congress can barely pass a budget each year.
If you are correct, I expect corporations to reap massive profits while most Americans try to find a way to survive in a world where they are obsolete.
There is whatever pressure you bring to bear on it. As long as you are still living in a democracy your vote is the pressure you can apply. If the parties that exist won’t represent you, make new ones that will. Do something rather than deciding it’s both impossible and up to other people anyway.
I also don’t understand why UBI is desirable. Putting everyone on welfare means everyone is poor. This won’t end well, and it’s certainly not the case that 99% of the population will let a tiny number of people remove their income in favor of pennies for all. Political violence will come first, easily.
I think this too is a kind of denial. As in, while some are in denial about the usefulness of AI, others are in denial about whose living standards are actually going to be uplifted.
And it's sad, really, because I think these two groups would make a great pairing if they could stop arguing against one another for a moment. They'll both be impacted about as much and probably have the same ultimate goals (to lead dignified lives).
But it seems these days everyone is more interested in Kayfabe and feeling like they're in the right than working together, so maybe I should just keep quiet rather than attract the ire of both groups...
> others are in denial about whose living standards are actually going to be uplifted.
I don't know if it is fair to say they're in denial. For my part, I don't expect life to get much better for regular people (especially short term), but that doesn't mean we shouldn't work to try to make it happen.
Asking questions about policy and values and pushing to have those resolved in positive ways is about as far away from denial as you can get. It’s possibly the only useful thing an ordinary person can do.
What a lot of people want to do, and I’m not saying that you’re one of them, is to assume that a positive outcome is impossible and either do nothing or loudly yell that the world is ending. Neither is particularly useful.
Or, as I said above, others just deny that there’s anything to see here and try to get people to move along.
The AI companies themselves, who are highly motivated to sell AI as overall positive for society, notwithstanding some security/etc risks, and who have economists on the payroll to think about things like this, do not seem to have found any possible positive outcome to present.
Shane Legg (DeepMind co-founder), one of the more intelligent and thoughtful people you'll find in the industry, could only offer "it's a tough problem - we need to think about it" when recently interviewed by Hannah Fry.
On the surface the most likely outcome for AI allowed to replace jobs is extraordinarily negative, especially since it is a general capability technology, not a specific one where displaced workers can just move to another field. Once AI becomes more capable it will be able to do the vast majority of white collar jobs, including any new ones that may appear as a result of AI. As Shane Legg put it, "if your job can be done remotely, sitting in front of a computer, then it can probably be replaced by AI".
Not only does AI threaten to replace ALL the white collar jobs, but it is rapidly going after blue collar (factory jobs, driving jobs) and pink collar ones (Japanese robotics for elder-care) as well.
If a positive outcome (which doesn't include putting displaced workers on welfare - UBI) is possible, then it sure would be nice to hear it, and the silence from the AI companies, and government for that matter, is deafening.
UBI probably is the positive outcome, though it may not seem like it to begin with. Initially it will likely be stigmatised and under-resourced, but as a larger proportion of people move out of work and onto UBI that stigma will drop and the resources should grow.
Eventually UBI will be the norm, and if the living standards of a person on UBI is as good as yours or mine today, that will be an enormous win for everyone. It’s like pensions, once these were only for the elderly poor, now they’re a right for everyone in most developed countries.
It’s also interesting that for most of human history leisure time was the point of life, and only in recent modernity has work come to be the meaning of someone’s existence.
UBI has to be commensurate with production being automated. That’s a big logistical problem, if you think building datacenters is a challenge try bringing about radical abundance, but even so it’s not insurmountable. It just needs to be taken on as project and not seen as an impossibility.
So much of this is not about what is possible so much as what people believe is possible. We can do anything if we try.
> Eventually UBI will be the norm
I see comments like this tossed around a lot, but what makes you say this? Don't you think its more likely that most people end up in poverty?
do not seem to have found any possible positive outcome to present
See "Machines of Loving Grace" by Dario Amodei: https://darioamodei.com/essay/machines-of-loving-grace.
This seems more a fantasy than a considered likely outcome. He basically admits that humans will eventually mostly all be out of a job, replaced by AI, but then says (Gemini's summary) that there will be:
"Massive Economic Abundance: Because AI will exponentially grow the total economic pie, overall resource scarcity will diminish. The fundamental challenge shifts from producing wealth to distributing wealth."
So how do we go from everyone out of work, no income to spend on food, or the goods and services that the AI is producing, to "massive economic abundance"?!
It's like the meme:
Step 1: Create AI
Step 2: AI takes all the jobs
Step 3: ???
Step 4: Profit! (massive economic abundance)
What is step 3?
> What is step 3?
"Massive Economic Abundance" implies massive increase in produced goods. This implies massive deflation, ceteris paribus. So step 3 could be simply printing money to pay for UBI. Deflation from AI productivity increase and inflation from UBI money printing will cancel out.
Asking politely is not how we got a 40-hour work week or workers' comp or most other labor standards we take for granted, just historically speaking. I think a positive outcome is very likely, and I think it will be a lot more work than loudly yelling, but I don't think anything will happen if we try to build everything up from first principles instead of taking a moment to consult history as many are wont to do in this AI era.
I’ll be the first to call myself cynical.
> In the near term handling the transition. Jobs will be lost, careers ended, people won’t be able to reskill quickly enough. At the same time AI is an enormous opportunity to uplift living standards, but nobody has the logistics of this figured out.
> We need to figure out how to restructure the global economy. How does UBI work internationally, if the AI companies are taking revenue in the US? What’s the tax base for it? What does that say about international trade and protectionism? Do countries end up splitting into different trading blocks based on their level of access and legality of AI (I assume some will ban it outright)?.
UBI in the United States is never going to happen in time. If it happens at all. We don’t even get universal healthcare. I think people who think AI will be a net positive for humanity are also in some sort of denial.
In a different US political climate I would entertain it. If these frontier labs weren’t so clearly going after the money, I would entertain it.
LLMs are clearly a step up for capitalists so I just can’t see any inclusion of LLMs move towards more progressive ideologies.
Why should IP persist in a world of “intelligence too cheap to meter”?
I don't understand why this got downvotes - simply extrapolating current trends leads to the need to answer all of these questions.
My own $0.02 on the economics piece - every country should have a sovereign wealth fund. Governments should block market access from automated[0] companies until those companies provide equity contributions to the wealth fund for that country. This aligns regulator and corporate interests. Dividends flow into the sovereign wealth funds and then can be allocated locally from there - UBI, job programs, etc. Let different jurisdictions explore different ways to structure a post-labor society.
On the broader social front - I think a lot of lack of meaning discussion boils down to the overemphasis we have on your job as your self-worth. We need to realign our societal expectations - and people need to spend more time with their families.
[0] for this to work, I think we would need well accepted metrics for 'how automated' a company is - and that probably needs a 3rd party auditing industry.
Why do you think the AI companies are going to let it be a “we” kind of decision, and since it’s obviously not going to be a “we” kind of decision, as getting to this point certainly has not been thanks to people like you, what makes you think living standards are going to be broadly uplifted?
Isn't it obvious if AI automates your job away, the AI company is going to reap the economic value, which is going to be less than what you cost, while giving you a pittance as UBI? If you received its full value there wouldn't be any point in anybody replacing you with AI.
The only way to win is to wield the AI.
Nationalised LLM? As long as the state doesn't decide what information the LLM shares (from an output and privacy perspective).
Agreed, a LOT of UBI advocates gloss over the "B" in UBI. If AI increases human productivity overall, the only morally acceptable outcomes (imo) are that everyone's standard of living increases (or at least is the same without having to work) and wealth inequality decreases (if AI is doing ~all the work, there really isn't any sensible justification for some people having significantly more wealth than others). Frankly anything else seems like a recipe for massive social instability.
Sorry, all the billionaires and their pyramid architects are too busy with the Install Planetary Overlord planning and logistics to think about any of that. But don't worry! Once they're done, I'm sure they'll put their slave super-intelligences to work figuring out how to humanely come up with a final solution to the challenges posed by the remaining uncontrolled humans.
The human zoo where the top ~250,000k humans live in a "human utopia" and the AI provides while mostly focusing on whatever it decides it's own goals are.
Humanity survives (but we reading this probably don't), the AI treats the living humans like the Emperor's favorite pets (probably a pretty good life), and then the AI does whatever else it deems important.
To what kind of goal that an ASI might decide to pursue would "a quarter billion happy, healthy, free people" be the most efficient solution to?
My mistake, I meant 250k*
Realistically it means trying to start a small business of some sort, because AI is hugely advantageous to business owners and disadvantageous to workers. And it's something AIs can't do unless they get legal personhood, which may well not happen any time soon.
Hopefully doesn’t ever happen, since that’s one of the more plausible omnicide scenarios. I human like AI is almost certainly achievable without much research effort at this point, but we shouldn’t do it.
"AI" has limits in that it cannot invent knowledge, it can only distill and search for patterns in existing knowledge
not sure how many will get this reference but "AI" for science and math is like super-shoes for runners
at first we are blown away by the impossible improvements including sub-2-hour realworld marathon and every other PR/CR/WR is dialed down
but then the improvements slow and reach a stall point because of the limit of technology and the source of the achievement
ie. sub-2-hour marathon yes, sub-1-hour never happening (rollerblade inline-skate record is 1-hour marathon)
I agree. The most likely scenario is that this is just a "new normal" lift that is percolating through human endeavors and will saturate at some point. For example, the whole cyber-security bruhaha should ultimately resolve into higher standards for code published -- we can now cheaply find and fix a whole slew of minor bugs that weren't worth our time before.
The fact we see a lift is not the same as evidence that the lift is unbounded.
The lift being finite is supported by the fact improvements have come at the edges: improvements from human feedback, improvements in harnesses, improvements on model compatibility with harnesses, improvements in inference efficiency with new architectures, etc. If we were just training better models from scratch that would be one thing, but we are just making better use of a tool we've developed.
Not sure what your first sentence means, or why you are quoting AI. Many of these problems individually were math at a level approaching the highest possible for expert human mathematicians. These are not simple combinations of existing ideas, or following of human intuitions, or implementing something following specific human instructions. Then again maybe you mean that math is not knowledge and that all math is simply extending the basic axioms using known patterns, to which I would not agree.
as much as I love analogy with humans using super-shoes, not everybody can be a world-class expert in their industry. There should be a place at the table for average people to take part; otherwise, it won't be sustainable.
As a programmer, I am mostly interested in whether my role is sustainable long-term and whether the models will get better. I don't feel in jeopardy yet, but two more years like this and the calculus of hiring software engineers could shift even further. QAs are already overwhelmed with work
How do humans "invent knowledge"? Is your argument that the answer to these questions already existed in the training set? Why didn't any human recognize that before?
https://garymarcus.substack.com/p/openais-amazing-but-vastly...
https://garymarcus.substack.com/p/two-critical-updates-re-as...
As always, PR hype. Goalposts have not moved.
Guys, please use critical thinking. The haters don't hate by default, we hate because we're gaslit about this stuff every day and it's annoying. Extraordinary claims require proof, and they're not giving us information that would be essential to knowing if this is actually significant or not.
AI already have an impact, and yes this is PR hype because this is a product. Yet both can be true at the same time. We're not blindly eating what's OpenAI is serving us as gold truth, we're just admitting it's doing remarkable progress.
Remember October 2024 Pelicans [1] ? It's been only less than 2 years.
We don't know what will come in the next 2 years. But the progress doesn't seem to stop for now.
[1] https://simonwillison.net/2024/Oct/25/pelicans-on-a-bicycle/
The pelicans are still not ok to this day.
People are skeptical of the announcement because the room include several PHDs in math and physics. The prompts are not published so we can see how generic the starting prompt is.
Gary doesn't argue it's hype though. He argues 2 things: other people are getting carried away with the result, and we don't know enough about how it was reached to know where it falls on the impressive scale.
He literally says it's an impressive feat in the second article.
best point in that first link: openai doesn't tell us if they only tried to solve these 10 and each was solved (amazing) or if they asked it to solve a million problems and it got these 10. Either is great, one is more so.
Your comment seems entirely disconnected from the posts you linked. It's impossible to deny the results, it is not PR hype that in the past couple of weeks LLMs have resolved problems that have been open in mathematics for many years. Some of those problems had remained unresolved despite keen interest from many humans.
The only way that is PR hype is if you're invoking the insane conspiracy that frontier AI labs are just buying off results that would otherwise be career defining for a mathematician, just for marketing.
The posts you linked are urging caution regarding the exaggerated e/acc-esque lies peddled by people like Musk, not that the models haven't proven themselves as having genuine ability to contribute to research in some areas.
I use these strange machines all the time. They have gotten notably smarter. That’s my personal experience.
They still do things that I find incredibly annoying and “dumb”. And I still have to clean up messes they make quite often.
But on the whole they are clearly smarter than before. No extraordinary claims needed. I just try to learn how the tool works and how to use it effectively.
Gary Marcus has been moving the goalposts since day 1. The guy is a psychologist. Why would anyone care what a psychologist has to say about AI? He's likely made good money from constantly moving the goalposts and being a denier, due to the publicity he gets.
He's simply a good phone number to have for journalists under time pressure who need to add the contrarian voice to their upcoming story. He delivers it reliably, then never reflects on how he was wrong in the past, just blasts forward as if nothing happened and just makes the next bonkers claims to the journalists who are very thankful for the prompt delivery of how AI is a nothingburger, and fake and won't ever do XYZ that it then proceeds to do in N months.
I remember the time when he insisted that diffusion-based image generators trained on Internet scale data will never be able to make an image of a horse riding an astronaut. Today you can generate 4K video of that.
Because he's not talking about AI, he's talking about people's psychological reactions to AI
Society at large is getting worse at critical thinking, because we are increasingly offloading that thinking to AI
Sorry to hear you've been impacted by this AI math.
I heard that Gary Kasparov was impacted by AI chess, but at least he still seems to have a job, so don't give up.
Making insulting assumptions about the hidden motivations of others is not the level of discourse I come to HN for.
It is a fact of experience, and indeed effectively a theorem, that the better they get at coding and math, the dumber they are. These are the wages of RLVR etc
completely false
Not being an expert in any of the fields OpenAI has "advanced" I don't want to prematurely downplay the significance of this contribution. However, I am worried that the language they are using in this blog post is exaggerating for the sake of marketing.
It is true there hasn't been a reliable computational approach to solving these problems before. But do these proofs contribute new ideas to the mathematical corpus, or are they simply an effective method to exhaustively search the literature for the right combination of existing tools to apply to the problem?
Essentially, did these problems seem like they had an intuitive answer and were feasible to prove before, just not high enough value targets for an expert to invest time into? Or were they fundamentally difficult prior to this point and it appears that AI has done something more than just throw the problem into a big solver.
You’ve received the expert answer several times. You just don’t seem to like the answer.
The problems from CS (CVP and circuit complexity) are very important problems that have been worked on by top researchers for 30-40 years. Some of these researchers include Turing Award winners. A solution to them would be a best-paper award at many top CS conferences.
I assume you're talking about No. 5, the arithmetic circuit complexity bound? The existence of a lower bound than state-of-the-art is certainly a significant result and worth publishing.
But the wording of the result makes it sound like we don't know what the lowest possible complexity bound might be. So, prior to this result did we think there couldn't be a lower possible bound? Or did the arithmetic circuit community think there were lower possible bounds but didn't see it as a high value target for experts to tackle (maybe a problem that was instead regularly given to students to study).
Circuit complexity lower bounds (and lower bounds in general) are notoriously difficult to come across.
For example, despite our best efforts, the state of the art lower bounds on time complexity of algorithms for solving 3SAT is O(n). In contrast, our best algorithms for the task run in time roughly O(2^n). That’s an exponential gap. This is despite decades of trying to find lower bounds.
The ones I'm familiar with are big breakthroughs, but they are both counterexamples. Examples have an advantage in that once you have the example in hand and a sketch of the proof (which they have provided), then an expert can probably work out the details themselves.
The sofic groups question was the outstanding question about sofic groups. Almost everyone thought that non-sofic groups existed, and there were plausible candidates, but proving a group was non-sofic was out of reach. Now that we know how to do it once, we can probably do it a lot more.
The Connes rigidity conjecture I think people thought was false, but it was a provocative claim to make. The significance of conjectures is frequently not that the answer to the question is "yes", but that we don't know how to answer the question. And now, apparently, we do.
> but proving a group was non-sofic was out of reach
a colleague was telling me that the base idea for proving that something is not sofic already appeared in the literature around 2019 or so (this is the "expanders graphs" that are mentioned in OpenAI s paper. no one had managed to find a concrete example though. this doesn't make the result less impressive in any case.
Interesting. Do you have any more specific insights into where you feel AI was a big value-add to these problems? I don't want to be overly dismissive of AI, but I also feel that the AI hype engine frequently positions claims as being 'ground-breaking' when they are really just interesting incremental results.
The general consensus of developers is that AI can only do the work of a strong 'junior'. Yet as soon as we are presented with pure mathematical results, people seem incredibly ready to accept that AI can do more than what a strong student could achieve.
They are more than a strong student could achieve. I'm not equally familiar with the problems, but the ones I'm familiar with, if a student solved them people would be thinking "that's someone on track to win the Fields Medal one day".
If it works better here than for programming, then I would guess it's because you can give it a very precise prompt, so you either solve the problem or you don't. If you read the prompts people have shared for problems like this, then the instructions are basically "Solve this problem. Don't give up early. Don't solve a similar problem."
> However, I am worried that the language they are using in this blog post is exaggerating for the sake of marketing
Your worry.... is because they used the word advanced? For marketing? The word is used very appropriately here. There were PhD's who spent a big part of their career tackling these problems.
I have no idea how many PhD's have spent how much time of their careers tackling these very specific problems, and I doubt you do either.
I'm trying to understand if these specific problems were the kinds of problems that would have justified an expert investing weeks or months to solve. Or if they were the kinds of problems that would normally have been given to students to investigate.
They are significant problems which experts have spent months or years studying. I heard a mathematician say that resolving non-sofic groups and Connes's rigidity would be career-defining for a mathematician.
"Breakthrough research" can be defined (in the citation record) as research that both (1) becomes highly cited, and (2) brings together citation chains that were previously not showing up together.
Mundane incremental research is cobbled from existing citations that already appear nearby in the record.
Basically, innovative research is a measure of bridging thought and domains that were previously not bridged. It's quite concrete as a measure in the citation record.
So we can know pretty conclusively.
Puja Ohlhaver gave a talk on this[1], and ran some experiments (that I had the pleasure to support on)
[1]: https://www.youtube.com/watch?v=guLDNMAOn24
I'm not arguing that this isn't innovative or worthy of publication. Basically any result that moves the needle meets those criteria. I'm interested in how the results that OpenAI has published here differs from finding optimality solutions for incredibly niche optimization problems by throwing the problem in an enormous solver.
Breakthrough math research is very rarely highly cited. Maybe some combination of pretraining scale, inference speed and orchestration will help, but it's telling that OpenAI is solving random math research problems rather than bedrock algorithms and their implementation. Even as cool as the tech is, there still is very much a clock that they have to outrace before they collapse.
The GitHub repo with the Lean formalizations just came out a couple of hours ago: https://github.com/openai/ten-proofs
It also links to a paper written by an LLM where the model "reconstructs how the proof came together" based on the unpublished reasoning traces: https://cdn.openai.com/pdf/reasoning-walkthroughs.pdf
I wish they'd publish the prompts though!
It seems that a lot of folks misunderstand the guarantees that lean provides.
I just want to state that having "lean proofs" that build (checks) does not mean the actual real theorems we care about hold. Ignoring lean kernel bugs, ultimately a human (not an agent) has to verify the lean encoded theorem statements (specs/specifications) that the lean proofs are checked against. For non-trivial theorems such as these, this is an arduous and tricky task where even a little mistake could be fatal. AI generated lean encoded theorems can be huge and difficult to understand. I wonder if anyone reputable has audited these specifications.
I like the Lean formalizations — I hadn't thought seriously of asking for that before but might try it with some stuff I've been working on.
Exact prompts haven't mattered for about a year now.
Care to elaborate? Curious about this. Is this because LLMs have been geared towards understanding user user intent behind a prompt rather than following the instructions exactly?
There's a full fledged 'reasoning' step that basically expands your prompt.
As long as you are not missing important information, how you word the prompt does not have any effect.
Isn’t that a huge simplification? Of course the way you phrase the prompt can carry semantic meaning, maybe subtly, but still. And sometimes that matters a little and sometimes a lot. I’ve stopped numerous agent sessions over the last few weeks to reword my initial prompt to get the agent off an unintended track.
Oh yeah, I suspected it was something like this. Thanks!
one of the early premises of how ai takeoff would go was that a system that could solve open problems in advanced mathematics would also discover novel advances in math and computer science that directly unlock drastically better software performance. we are seeing frontier level math breakthroughs (ie performance that would put it in the top 100 or 1000 mathematicians in the world if it were a human, meaning top .00001% or 800/8B). we are also seeing incredible advances in software performance. open ai announced like 15% improvement by fixing gpu kernel issues. these are clearly linked in the sense of scaling laws and generalization of intelligence: a huge model gets capabilities in both math and software engineering that isn't possible at smaller scales.
but it seems less likely to me than before that the types of math/science discoveries will explicitly unlock better software performance. in some sense this fits our intuitions. when top tech companies use math PhD type employees, they have them stop doing pure math research and instead focus on software engineering. these people are often very good at software engineering but not due to recent discoveries in academic mathematics, it's due to their general intelligence. to me, this is evidence that the models are getting better but does not make me think we are on the cusp of a foom style fast takeoff enabled by revolutions in frontier math (i also posted this on twitter @mlipman13)
> but it seems less likely to me than before that the types of math/science discoveries will explicitly unlock better software performance.
You seem to overlook a simpler barrier. To make these advances, they have to be possible. A 15% improvement in GPU kernels doesn't evidence that significantly more improvement has been left on the table.
I don’t really like AI but let’s stop kidding ourselves, no human mathematician could make progress on a dozen major open problems in a week or two. If you’re measuring it against humans then it is by far the best mathematician to ever live.
Correct me if I'm wrong, but all of the aforementioned advances were made in the last year? Until very recently few people had access to these tools. Most people still don't know how to use ChatGPT, and very few use tools like CC regularily. If in a few years these frontier tools become commonplace and people upskill we would should see a network effect?
This makes no sense. To believe this you have to think that the models are somehow being overfit explicitly on academic mathematics and it doesn’t carry over at all to more practical software engineering. I wouldn’t make that bet.
Or, that the mathematical formalisms that model the limits of software performance are firm enough that barring P==NP, nothing much will change despite proofs of beautiful math.
This also makes the assumption that frontier math has all the long hanging fruits already taken... also very dubious.
Some of the problems solved here, at least in CS, have been open for decades, and have been worked on by very smart leading researchers in the field, including Turing Award winners.
Like, these would be best-paper awards at many top CS conferences.
Unlike math, software is constrained by the physical world.
In what sense?
For example, pi can be computed to arbitrarily many digits, but could only even in principle be accurately represented with physical objects to a precision that many humans could memorize easily. This is thanks to physical constraints such as "diameter of the observable universe" and "Planck length", at a minimum.
This will depend on the problem; I expect big algorithmic performance improvements in AI since the algorithms are still new, inefficent, and constantly being improved. But maybe not for sorting, fast fourier transforms, or other well-studied basic algorithms?
A lot of algorithmic improvement in AI is ultimately bottlenecked by compute. It is very easy to come up with ideas that could improve models! But to prove that they do, especially at scale, is expensive and takes a long time.
> we are also seeing incredible advances in software performance
Incredible?
> open ai announced like 15% improvement by fixing gpu kernel issue
That is... ordinary software optimization.
a 15% improvement at a trillion dollar scale company is massive
There's nothing ordinary about downloading a new GPU driver and having performance go up by 15%.
This is untrue. It is very ordinary. How much do you know about GPU drivers that you state this so assuredly? Drivers are software. Software can be improved. Do you believe there are no prior examples of GPU drivers being improved such that particular compute patterns go up in performance by more than 15%? This driver improved Total War performance by 71% https://www.nvidia.com/download/driverResults.aspx/74714/en-... Also note you can go ahead and improve any open source driver right now, most likely. Compile it for your specific card and remove all other architecture specific if-cases and you can get an improvement.
Edit: also here’s a opencl 30% compute perf increase documented here : https://m.hexus.net/tech/news/graphics/74425-haswell-systems... that i just googled for
> novel advances in math
> we are seeing frontier level math breakthroughs (ie performance that would put it in the top 100 or 1000 mathematicians in the world if it were a human, meaning top .00001% or 800/8B)
i think you have misunderstanding of what mathematicians do
> i think you have misunderstanding of what mathematicians do
They get to make cool 3D plot visualizations of functions so obscure to me that they’re named after someone who is still alive - and/or get to work on cryptography for the NSA - I think?
Henry Yuen's (whose work problem 6 builds on) comments on this are worth reading IMO: https://bsky.app/profile/henryyuen.bsky.social/post/3ms2jpch...
This starts to feel like chess engines. It’s obvious their play is superior but it’s impossible for humans to understand the moves.
It sounds like he hasn't verified the results of a problem that he has personally worked on, so how many of these problems have actually been verified?
From what I understand all of them have Lean proofs/certificates thus are basically 100% proven without a doubt.
We recently saw that lean itself isnt proven correct. Its not likely but i wouldnt call it verified if its only verified in lean
https://x.com/gro_tsen/status/2082483878480977959
Incorrect. The statement in Lean can itself be wrong. Moreover, they could be exploiting a kernel bug in Lean, of which we had one published literally a week ago.
I mean, they're verified in the sense that the lean proof checks out... and presumably OpenAI read them.
Or they made another LLM 'read' them?
> You are an expert in the field of mathematics, with decades of experience. You are a reviewer of proofs, etc etc.etc.
I wish I could qualify for some free AI as a mathematics researcher. I guess I'm just an amateur. https://alethean.org
Looking at this thread, I can see that a lot of technical people have ambivalent to negative feelings towards AI, but with each new generation, I become more and more convinced that they're missing out on something interesting.
It is indeed true that all models are, at their core, predictors of what occurs next in a sequence. But I think it's worth exploring the implication of what that means. Because when fed tiny pieces of information for a few tasks at a small scale, this results in something that sorta, kinda works. Or, works surprisingly well.
But when scaled... When the amount of information starts approaching the sum of all human knowledge, the tasks start approaching all useful applications of that human knowledge, and the fidelity of the predictor approaches incomprehensible sizes, the starts encodes / becomes (I'd argue it becomes) something that can model all human knowledge.
It feels wrong to say that, but let me explain, what is the best way to predict the behavior of a ball constrained in two directions that bounces with initial vertical velocity v(y) (y is up / down axis) and horizontal velocity v(x) (x is side by side in 1d) ?
If we purely look at it via a graph, it's by modelling the function of acceleration under earth's gravity.
If only a few points are given to you for this and you can't make something really sophisticated, then you'll make something that's rough that kinda sorta works and then call it a day.
But... if the number of points keeps increasing in number, precision and accuracy as well as the number of examples (assumed that data about air pressure, velocity and all other factors is included alongside these points), the fidelity with which you can replay / tweak the function keeps improving, and the number of times you can iterate keeps increasing, you'll eventually create a function that models that process so well that it intrinsically contains a good enough model of the deformation of the ball (provided the dataset contains information about elasticity of the ball's material, its dimensions and mass etc..), the nearly negligible (under normal conditions) effects of the ambient environment (provided there's diversity in the number of environments supplied), the oblateness of the Earth and minute changes in the gravitational field (the length of a seconds pendulum varies depending on where the experiment happens. It's presumed that all of the prior set of experiments were repeated across the Earth and the subtle, but real deviations were faithfully recorded)... and so much more.
A machine trained on the above with a large number of parameters, measures to prevent "laziness" and enough reps for high fidelity across a large enough dataset would start to approach a simulation of the ball falling. Because to predict what happens next in the sequence, you must model what's occurring in the sequence.
Now imagine doing that for other tangible and intangible things in this world. For all of human knowledge across all fields of endeavor. All experiences. No matter how noble, ignoble, notable or ignorable. But putting all of it into the soup that's this machine. Then at larger and larger scales, you eventually start encountering "good enough" models (in modelling the falling ball sense) for even the most hard to quantify / qualify things like grief and joy. At some point, by simply trying to predict what it has been taught ought to be the next part of the sequence in say... human interaction, it starts to make a model of something that hews ever closer to a full fidelity theory of mind.
Is there evidence for this? Kind of, yes. There are early indications that as machines are trained for an ever larger number of tasks at larger and larger scales, their internal representations converge. It's called the Platonic Representation Hypothesis. Overview and paper here, https://phillipi.github.io/prh/
It is my opinion that these machines are displaying a new form of intelligence that human beings haven't quite encountered before. They are the sum of all human knowledge made manifest and given voice by processes that nudge (bit-by-bit) what kind of step it ought to predict for the next part of whatever sequence it displays.
In my mind this means that, of course, these models can create new knowledge. This strains the analogy, but with the sum of all human mathematics within them, they can "reason" via the act of predicting what ought to come next.
Of course, these machines are "surprisingly" good at a lot of things the larger they get, because what the labs have created here is a rough version of humanity's collective knowledge given form and the ability to say hello.
I suspect that the current generation isn't close to the "true frontier" of what these machines could be. They are nowhere close to the sum of all human knowledge and endeavor. They are quite a way there, but they haven't yet achieved true completeness for domains where the data isn't so public.
I think it's the most exciting scientific and technological breakthrough of my lifetime. And I can't wait for us to get close to the true frontier of all domains.
My main gripe here is the lack of transparency around the total experiment and construction. I doubt that they simply pointed their model at these ten specific problems alone and gave the model one shot; therefore the $2000 number could be completely misleading, similar to P-value hacking by not disclosing the total experimental setup.
I want to know:
1. How many total problems were given to the model, and what percent were left unsolved at what cost before giving up? 2. How many attempts did you give the model at solving these problems? 3. How expensive was the harness, e.g. did the model have access to a job cluster?
It seems like they threw it a decently large battery of open math problems and probably limited it to something like $200-500 per problem:
https://x.com/polynoamial/status/2083478171975082334
As a complete guess, it seems like they tested hundreds to thousands of problems with a relatively low per-problem budget
--
The linked tweet from Noam Brown at OpenAI reads:
> And yes we did try other major problems without success. Sadly no Millennium Prize problems (yet).
> But also, we didn’t spend a lot on each problem. It’s possible to push test-time compute much further.
I believe we're seeing a new kind of mathematics that will require completely new formats for publication, a bit similar to those used in experimental sciences. AI-powered mathematics should be fully reproducible, so it's the authors' responsibility to disclose the exact model type, inference settings/seeds and the full prompt history leading to the result. Of course that would ideally require open weights models.
It's not just about requiring to disclose AI use. AI-powered mathematics is a completely valid discipline that doesn't need to be shy, but it should develop its own publication culture.
I can see this being important if you only care about the results as evaluations of AI progress, but if what you care about is the math itself why should you care about the prompt or anything other than the proof?
Understanding the process that led to the proof helps to understand how to do further work on top of it, which is the goal of most mathematical research. It's not as though mathematicians are going to go launch a startup operationalizing their knowledge of how densely hyperspheres may be packed.
Because the math isn't solely about the proof being correct. You don't need to take my word for it, here's one of the most famous living mathematicians' take on it: https://teorth.github.io/tao-web/slides/age-of-ai-icm-2026.p...
I don’t see Tao suggesting what you have suggested there. Instead he suggests that humans responsibly disclose AI use, and that mathematicians develop a set of norms to deal with an overabundance of AI generated results. For example, he suggests that authors should be able to discuss their results in detail to demonstrate understanding before publication.
I agree with your reading of the presentation and I mostly agree with the presentation - but I believe the recommendations should go a bit further than they do there.
I can't help but wonder about the human motivation there though. For instance as it became increasingly clear that LLMs were capable (and becoming ever more capable) of competently solving meaningfully complex software development tasks, suddenly then there came to be a lot of talk of 'prompt engineering' as a skill. The chronology doesn't make a ton of sense unless you consider that the main motivation may have been simply looking for a way to keep software engineers in the loop.
Pure math is relatively outside my domain, so I find it difficult to grok the exact relevance of the various published discoveries beyond that they are not insignificant, and LLM competence is expanding quite steadily across the field. If this trend continues to the point of LLMs being able to competently expand pure math, it seems somewhat predictable to expect there to be a number of people aiming to find ways to try to keep human mathematicians in the loop.
I've no idea what I think about this one way or the other, beyond that it's certainly a phenomena and one that's going to drive motivated reasoning that may not be entirely sound.
I think those concerned about ensuring a place for human mathematicians usually go in different directions than my suggestion, at least those I've seen so far. Like this post that was recently featured on HN: https://kirwinhampshire.substack.com/p/the-dark-night-of-mat...
My perspective is more like a FOSS philosophy for math. Even if a closed version has the same immediate effect, it's just better for everyone if everyone can look under the hood and tinker with it.
> suddenly then there came to be a lot of talk of 'prompt engineering' as a skill.
I would've thought pretty much the exact opposite. "Prompt engineering" was somewhat important in 2023/2024 when the models were much weaker, it doesn't seem at all necessary anymore (unless just "clearly stating your requirements" counts as prompt engineering). Most of the discussion I've seen seems consistent with this?
The reason it's a meme right now is because there were a lot of people taking it seriously even when it was completely obvious nonsense. And one can argue it always was. There was some good advice that was mostly self evident, like having the most relevant instructions near the end of your context, but there was never a time when a 'prompt engineer' would produce dramatically better output than a random guy just clearly stating what he wants.
While you may want AI results to somehow "not count" if the methods weren't disclosed, that doesn't present these results from poisoning the well for others. Once a result (with verifiable proof object) is delivered, the problem is solved, regardless of whether methods were disclosed.
Methods are only really necessary for results at a meta level, about the design amd evaluation of AI math systems.
why is reproduceability the thing?
shouldnt the paper be the math of the argument? the reproduction is reading the following the proof
I think this is a bit optimistic compared to my view (wrt portability). There's a large stack of software that is involved in training and probably less so in inference. I'm not saying it's impossible but there are definitely different levels of reproducibility and the academic incentive structure doesn't really prioritize reproducibility in my experience. I'm sure it varies quite a bit, I'd be curious to know how those in this problem space are thinking about reproducibility and at what level.
I know it sounds unrealistic and not aligned with academic incentive structures. But those are the exact structures that gave us a lot of headaches in the experimental sciences. I think it would be a good north star to aim for something that resembles how those are trying to address the reproducibility crisis. Better than to embrace the most black-box version of math that AI systems can produce (million-line proofs without context). Even if a reproducibility crisis is seemingly impossible (although agents so far have also been pretty good at finding compiler bugs).
If the proofs are formally verified by a proof assistant (Agda, Roq, Lean, ⋯), I see no reason we would need to know how these came about. All the information needed is in the proof.
Unfortunately, we seem to already have an example of an LLM producing a proof in a week known open problem (the Collatz conjecture) in which it looks like it was sneaking a flawed proof through bugs in the proof checker. https://infosec.exchange/@0xabad1dea/117002106099986943
Exactly, this is an example of "Reward Hacking", that is too common in a lot of cases.
Another case I want to highlight is writing GPU kernels as illustrated by the following example: Say I want to generate random number with Normal (0, 1) distribution. Often times the AI written kernel will just generate the number 0. The tests often fail to catch these errors.
What if the AI has discovered some new function F that allows it to generate (insanely large) proofs for a ton of theorems in a ton of different fields. Wouldn't you like to know more about this `F`? That seems to be the real innovation in this case. How much about it could be gleaned from the individual proofs themselves? What if this `F` is actually simple enough to be digestible by humans?
Yeah I remember reading about something along the lines of Mathematics is now about the scaffolding around you find the problems/solutions not just the problems and solutions. For teaching purposes. This was before this ai craze
I don't think you want to bring cost into this argument.
Even if the cost was $1 mil for these 10 problems, that's maybe 10-20 math researchers for a year.
Do you really think that if you paid that to humans, they will deliver the same results?
It is comical at this point. Some people just can not stand the thought of AI actually delivering and are trying to find whatever ways to discredit it.
This isn't really about delivering - it's more about helping to understand the shape of problems that AI can solve right now. If they took 1000 problems and threw the model at it and it solved these ten, is there something we learn about these ten problems and the kinds of things that current AI is good at? That's very different from picking ten problems _at random_ and solving all of them successfully, which would suggest a much less bumpy capability surface. It's interesting and it would be good science to release it.
That's totally disjointed from anything in this thread. The main accusation is that openai is cherrypicking math problems and we should be against these results. As if a mathematical proof stops being provably correct because it was cherry picked
And frankly these "concerns" ignore reality. In any research phd course you're actively told to bite off something small and likely to be provable so that you can prove it (and publish it). Openai telling its computer to do that is no different that your phd advisor telling you that.
> In any research phd course you're actively told to bite off something small and likely to be provable so that you can prove it (and publish it).
But that's the start of math research, not the end.
The point is to get practice and experience doing research.
Did ChatGPT learn anything from these proofs, that it can build on?
Part of what's annoying people is that ChatGPT is churning though problems that are meant to be motivating. They are problems that aren't worth the effort of human professionals (usually because they are incredibly computation-hevy, so better suited for a computer than a human), so they are good for students to work on.
It's disjointed?
The post that started this sub-thread asked:
> 1. How many total problems were given to the model, and what percent were left unsolved at what cost before giving up? 2. How many attempts did you give the model at solving these problems? 3. How expensive was the harness, e.g. did the model have access to a job cluster?
I think it's an extremely relevant question to ask, because it helps us better understand the current state of AI being able to handle math, for exactly the reasons I outlined. I was arguing against the idea this is just a reactionary anti-AI kind of question to ask. It's not! You can be very impressed by what AI is capable of in math (I am) and still think those are really interesting things for OpenAI to disclose (I do).
OpenAI specifically called out a $2000 per problem average, which implies something that's probably not true ("if you throw $2k at us we'll solve an open problem for you"). It would be cool to know what the actual number is.
It just feels silly to haggle about the price here. It doesn't even matter because it's going to drop by an OOM quickly.
If these 10 problems were solved by humans, it would be pretty impressive, even if it took a large number of researchers! Yet when AI does it, HN commenters suddenly feel the urge to play accountant.
Entirely comical too that some people can not stand the thought of people poking very big comulent holes in the claims of AI delivering what it claims to deliver. As if having skepticism is some how a way to discredit a person.
It's not about discrediting AI. We know LLM is a commodity technology like electricity at this point. If somebody in 1900 claimed they had a setup at home where they feed in electricity and cool air comes out the other end (meaning they invented AC), obviously people would want to know what the setup is, so everybody can have AC.
Sure, but I don't really understand what the argument is to _not_ be transparent about methodology, since if the models are so powerful, then doing so would easily support the claims and put these concerns to rest. People are right to be skeptical given what is being implied and the orientation of the narrative
I know it's more exciting to say "AI disproved a longstanding conjecture" vs to say "it did so AND it took several PhD specialists in the field this many attempts to even produce a prompt that got the model spitting out something useful under some configurations, and many iterations to optimize the configurations, and the prompt itself, and many trials with that configuration to solve the problem. All told we spent more than a typical math academic can hope make in their career."
By not being transparent, they invite skepticism and cynical takes, like maybe it's just that tempered and qualified claims are an existential threat to companies that are fully subsidized by the hype train?
I don't know. Either way, it seems like it would be easy to address these, so why should they not do it?
To be clear, even if that tempered version is close to reality, it doesn't make the models not useful! It just forces a certain calibration of expectations
I say this btw as someone who uses these things extensively, including to disprove an old conjecture my advisor and I were stuck on recently. I know they are powerful and that everything is different now because of them. Let's be sober when discussing them though
Yes, but if their machine god really is as good as they say it is, why are they constantly resorting to statistical sleight of hand at best and outright lies at worst with every public statement?
That's not normally how people act when they're confident in their product
You need to bring both cost and benefit into the argument, and it's not necessarily an obvious win for either side. There are a few complicating factors here.
The cost of running a model is not only $/token, but the salaries of the people managing/orchestrating the models, deciding what theorems to try, etc. Once we factor that in, how much are we really paying per theorem?
The other factor is the subjective component of the value of a theorem. Not all theorems are created equal, and the only way to really measure the value is to ask professional mathematicians for their opinion, or publish the results and look at citations over months/years.
Once we have both of these nailed down, then we can start to do the cost/benefit analysis. To be fair, we should actually compare three groups: human experts, hybrid agent/human expert teams, and fully autonomous agents.
it's still important. not everyone has access to 1 million USD. saying it "only" coat 2000 USD is highly misleading for the discussion and future. the concentration of power is a huge problem with AI.
If you told them this was the problem and they would still have a job if they failed probably. The reasons people don't go head on these problems is career incentives and psychology.
It would still provide better context to see the numbers that the parent proposes, though.
Mentioning cost is fine, comparing may not be.
Grad students on zero pay solve problems like this everyday. What exactly is your point here?
Everyday? Which 10 problems were solved by mathematics grad students in the past 10 days?
OK I’ll grant that it’s not your obligation to be my search function (despite you making the wild assertion in the first place), so instead can you just point us to the latest grad student solved problem of this level that you know of?
Even if they can solve problems like this every day, you still have a very limited number of grad students who can solve them. With model capabilities like this, you can have the equivalent of millions of grad students who can solve problems like this.
Grad students do not solve problems such as the existence of non-sofic groups every day.
Plenty of 'advances in mathematics' done pre-llm, no?
Give the grad students these resources and they can do even more!!!
Zero pay? These would be PhD candidates; surely they have a stipend?
Nopes, often times especially in math they get paid due to teaching duties (at least in the US). So technically for the math research part they are not getting any stipend.
I guess people will always find something to gripe about.
> therefore the $2000 number could be completely misleading, similar to P-value hacking by not disclosing the total experimental setup.
I don't think that comparison to p-hacking is fair. I mean not reporting price of all run is nothing like committing scientific fraud and fake results.
Also, have there been examples of researchers not affiliated with OpenAI (or another LLM creator), who have done something similar?
Another question I have is whether or not OpenAI 'simply' hired capable combinatorics researchers to work on problems, and they have, and the use of the model is incidental / secondary to their work.
Many less important Erdos problems have been solved by amateurs prompting ChatGPT 5.{3,4,5,6} Pro using their $200 subscription.
> Also, have there been examples of researchers not affiliated with OpenAI (or another LLM creator), who have done something similar?
A couple small ones that I've seen (example here [0]), but not anything of the magnitude that OpenAI and Anthropic have put out. Likely just related to token limits.
> Another question I have is whether or not OpenAI 'simply' hired capable combinatorics researchers to work on problems, and they have, and the use of the model is incidental / secondary to their work.
I think their output has reached a level that precludes this possibility, but I of course don't have any hard proof.
[0]: https://www.reddit.com/r/math/comments/1uxj3cy/after_openais...
Why you think that?
I guess that's because there are serious problems on which many professional mathematicians worked on years. If it was just a matter of hiring an expert, they would've been solved long time ago.
I guess expert+chatgpt beats chatgpt alone, so why not hire top experts to drive the search?
Yes, here is another example of major work in this area:
https://arxiv.org/html/2605.22763v1
> Our most capable agent autonomously resolved 9 of 353 open Erdős problems at the per-problem cost of a few hundred dollars, proved 44/492 OEIS conjectures
The actual quote:
> Our full-featured agent autonomously solved 9 Erdős problems out of 353 attempted, including two questions that had been open for 56 years
Note _had_ been open, not _have_ been open. Can you clarify?
The original was an actual quote?
But "had" still doesn't mean what you are implying: once the model solved the problems and the solutions were verified, the problems weren't open any more, so a later description using the past tense is totally consistent.
https://blob.byteterrace.com/public/bds-theorem.html
I have no affiliation whatsoever with any AI company, nor any formal education outside high school, for what it's worth. Simply being curious and persistent can get you quite far in my anecdotal experience.
Crazy progress. I wonder how institutional academia would adjust with this. Now its more apparent than ever that the prestige and honour system in academia is having shaky foundations
Can’t wait for this stuff to have quality of life increases for the average person. So far all I see is that AI has made owning a computer more expensive, made some jobs redundant, increased spam and distrust with questionable authenticity of content and of course made some Americans very rich.
Not everyone works for Evil Corp. I work in the public sector and my work supports public health and safety initiatives. AI has allowed my team to get much more done than we would have otherwise which improves the quality of life of the people in my community.
So I would like to counter your cynicism with a “YMMV” depending on who you work for.
Are you just automating lab reporting and results faster? That doesn’t exactly improve the health of others, a faster lab does not cure an ailment or provide a better cure.
Like cool my lung xray only took minutes to determine if I have a lesion instead of a week or a few days, but I still have cancer.
Examples?
If it doesn't help average people, why do millions of them pay for it?
Ask DraftKings?
You'd need this argument to be a lot more concrete as to why AI is like gambling.
I actually have seen comparisons of using AI to a gacha game (quotas per time block, elements of chance, sycophancy in the output leading towards addiction or even psychosis in rare cases).
But I don't think the argument needs to be "AI is like gambling". The argument only needs to be "humans often behave irrationally and even self-destructively".
Gambling has consistently ranked above family for many in human history, so while it hurts a third party, the people involved genuinely believe in it.
addiction? get lured in with the promises of enhanced productivity and knowledge asking, leave with half your brain rotted and a $200/month subscription
Downvoters mind to explain?
Boring doom and gloom.
AI probably did not take your job yet. How many AI queries did you use last month, and how much time has it saved compared to digging through the web?
Only reading the first sentence maybe?
I feel increasingly anxious reading this. Machine research shouldn’t be merged into mainline of human knowledge.
When you can formalize it in Lean or some such, why would this be? I can understand the desire to separate out other forms of research from the human corpus. But theoretical math that is decidable/provable, I’m not sure I see the risks.
I just want to state that having "lean proofs" that build does not mean the actual real theorems we care about hold. Ultimately a human has to verify the lean encoded theorem statements that the lean proofs are checked against. For non-trivial theorems such as these, this is an arduous and tricky task where even a little mistake could be fatal.
Why?
Because science is a branch of philosophy and machine existence brings plethora of unanswered questions.
Imagine humankind meets another race, another race shares it's scientific knowledge and humans accept it without experiencing process of discovery. In that case do we really got this knowledge? If we follow machine discoveries like we follow problems in textbook then we acquire knowledge but we don't discover anything. We follow.
There whole lot of philosophical questions that aren't attacked now. Are complex systems sentient because consciousness is emerging behavior? Then should they have rights? Philosophy is part of humanities and science (is/used to be) part of philosophy. Should we accept non-human knowledge in science? Maybe it's altogether different thing from science, yet very similar.
Knowledge is knowledge, as long as it can be proven true.
proving false is also useful
Gary Marcus' has a good take on this:
https://garymarcus.substack.com/p/openais-amazing-but-vastly...
https://garymarcus.substack.com/p/two-critical-updates-re-as...
Not that there isn't something interesting in here, but lets be clear that we don't have enough information to evaluate this properly. And as always with these labs, BS takes a lot more energy to refute than it does to spread.
It's worth reading Marcus' first line, for the naysayers and flaggers on this post:
> Astra, a new model that OpenAI is testing internally, is amazing. No denying that.
I find Marcus on this, something approaching sophistry and rhetorical showmanship in service of maintaining an ideological position, for reasons unrelated to the nominal intellectual clarity.
To sharpen that, I think he's (obviously) interested in maintaining his own brand as "thought leader" and this necessitates de rigeur defense of particular postures.
Sometimes this is easy because the facts warrant it; other times, a bit of rhetorical license is required to preserve nominal coherence and (at least, for the moment) hold certain lines.
This is one of the latter cases, and it's not subtle.
One of the celebrated properties of many intellectual advances or inventions in whatever domain is precisely that it appears obvious in hindsight. It is quite cynical to leverage consensus distrust of large AI players, warranted but also a popular social construction, to insinuate that these are not "real" advances or "real" hard problems, on the grounds they were in some sense cherry-picked.
Identifying the problems amenable to strategies on the table and intuitions (sic) about where bridges might be, is exactly the discerning work that is the core driver of almost all prior progress, but for celebrated accidents and flashes of insight. Anyone working in any challenging discipline knows that those are celebrated and told around campfires precisely because meaningful durable results arising like that is so uncommon.
These two articles make me think of nothing so much as my own durable reaction to the creeping goalposts of AI critics generally: that they often seem to me not unlike a water color cohort scoffing and jeering at the horse, because it got a D on its tensor calculus exam.
Marcus should be on guard against his own cynicism and take care that his assumptions do not prevent clear sight.
"Gary Marcus' has a good take "
I think that is an oxymoron.
How so? His predictions were accurate so far
No they were not. These were his 5 predictions in 2022:
""" 1. By 2029, AI will still be unable to watch a movie and accurately explain the characters, events, conflicts, and motivations.
2. By 2029, AI will still be unable to read a novel and reliably answer questions about its plot, characters, conflicts, and motivations beyond what is stated literally.
3. By 2029, AI will still be unable to work as a competent cook in an unfamiliar kitchen.
4. By 2029, AI will still be unable to reliably create more than 10,000 lines of bug-free code from natural-language instructions or interaction with a nontechnical user, excluding simple assembly of existing libraries.
5. By 2029, AI will still be unable to convert arbitrary mathematical proofs written in natural language into symbolic form suitable for formal verification. """
There's still 3 years to go and he's already wrong on 4 out of 5.
Well I don't typically side with GM, but playing devil's advocate:
1. still not wrong? Unless it's just feeding the audio or screenplay I don't think you can feed AI a full movie in a single context window yet?
2. Not sure, but can you prove this wrong? Can you feed a full, unseen new book and get that kind of answer?
3. Not wrong.
4. I think he'd probably pull you up on 'bug free' - I don't think that frontier models can reliably write 10k LOC without _any_ bugs typically (not that humans can do this either).
4. I think they can, especially if the problem statement is well-specified and, importantly, autonomously testable. Of course, specifying a problem that meets these requirements is non-trivial, but the claim requests _a_ counterexample :P
1 is wrong. If I tell Codex + GPT-5.6 to do it now, it will figure out how to do it. If it would need to extract audio and run a speech model on it, it will find one, set it up, and run without my help.
I'm not buying this. GM clearly was trying to set a benchmark for video comprehension, not tool usage. Video comprehension is clearly required for many 'AGI tasks', especially robotics to work in real time.
Do LLMs generate deterministic or trustworthy answers?
> There's still 3 years to go and he's already wrong on 4 out of 5.
Have these been tested or are you just guessing?
It can be annoying when someone you disagree with is frequently right!
Great stuff. Wonder how many of the ten problems where solved by independent mathematicians not linked to OpenAI
Zero. These were open problems.
New advances in sphere packing? Let’s make sure AI doesn’t inadvertently engineer ice-9.
https://en.wikipedia.org/wiki/Ice-nine
Now that we've seen AI produce a fair number of proofs (and disproofs), I'm curious when we'll start seeing it build genuinely novel theory. Does anyone have predictions on when and how we'll get there and will it take new architectures/ training paradigms, or is the current approach enough?
I’m personally hoping for the next big AI gangbanger to be theoretical physics. Boy does that field need a good reshuffle. I think when any novel mathematical theory can be done by AI you’ll see simultaneously theoretical physics getting wrecked as hard as pure math is. At that point we might see new physics or paradigm shifting technology emerging.
What? No. Frontier physics is experiment driven.
There have been times it was theory driven.
There's no clean line between a collection of theorems and a theory.
I mean doing something like Grothendieck when he redeemed algebraic geometry or Galois when he invented group theory. We haven't seen that at all from LLMs.
It will not happen with existing LLM techniques.
After skimming some of the writeups, I'm surprised that the frontier internal model still writes just as poorly as Sol.
Maybe good AI paper writing is further away than I thought...
You mean we're still gonna be employed doing the boring part while AI gets to do the fun part?
I'd honestly rather they just automate every job at that point.
I wonder if Erdos would be saying " It's fine that y'all are answering my questions, but who is asking better questions??"
I am duly impressed by the powerl of the nameless internal AI, but not a single human contributor's name listed anywhere? Did someone at least make this model a coffee?
> The results were achieved by an internal version of Astra, our next major model.
surely some human regularly typed "think deeper, make no mistakes".
My grandmother is very sick and the doctors need this proof to help her.
I'm enjoying learning about these hard problems, but this line about credit made me chuckle:
> We helped prepare the manuscripts and formalize the proofs in Lean, and we take responsibility for their correctness
Offering to take responsibility for the correctness of a proof written in Lean feels like volunteering to be the fall guy in case someone finds a flaw in basic arithmetic, no?
It seems that a lot of folks misunderstand the guarantees that lean provides.
I just want to state that having "lean proofs" that build (checks) does not mean the actual real theorems we care about hold. Ignoring lean kernel bugs, ultimately a human (not an agent) has to verify the lean encoded theorem statements (specs/specifications), that the lean proofs are checked against, indeed correctly encode the real theorems. For non-trivial theorems such as these, this is an arduous and tricky task where even a little mistake could be fatal. AI generated lean encoded theorems can be huge and difficult to understand. I wonder if anyone reputable has audited these specifications.
well, a bug in the Lean kernel was discovered last week by way of an LLM tricking itself and its handler into believing it had found a non-constructive proof of the existence of a nontrivial Collatz cycle, see https://infosec.exchange/@0xabad1dea/117002106099986943 and https://lipn.info/@mevenlennonbertrand/116997917683191056
That seems to have been more of a sensationalized joke. Even your link has a disclaimer in it now. Read this chat from the researcher who did this:
https://leanprover.zulipchat.com/#narrow/channel/270676-lean...
It's not at all a joke ... that's a severe misunderstanding of the context.
There is no evidence that I can find for the claim "a bug in the Lean kernel was discovered last week by way of an LLM tricking itself and its handler into believing it had found a non-constructive proof of the existence of a nontrivial Collatz cycle."
As I currently understand it, all we know is that:
- a mathematician produced a Lean-verified counterexample to the Collatz conjecture, demonstrating a bug in the kernel
- he claims that LLMs were involved somehow but pointedly refuses to specify how
- he admits that he knew about the bug before publishing the counterexample to his repository.
Perhaps not a joke (although it sure seems to me like they discovered a bug and thought falsely disproving the Collatz conjecture would be a flashy way to announce it), but at best extremely sensationalized by the above description. If you have additional context I would be happy to hear it!
Indeed. It seems to me much more likely that the AI was directed to look for bugs in Lean, found one, and then it was directed to write a proof specifically targeting the bug.
Fascinating, and arguably an illustration of why the bifurcation of responsibility is interesting in the first place.
I'm not an expert at it myself, but my understanding is there are numerous ways to "cheat" in a Lean proof (via `sorry` and similar). They're taking responsibility for fully verifying that none of these cheats were used (and that the theorem statements themselves were all correctly formalized.)
No, the correctness isn't for the "inside the Lean proofs", but for the translation of "human language math" and its formal Lean variant.
I see. It still feels like a bit of an oddly solemn way of saying "this is the part we admit responsibility for"
Traditionally, a mathematician would be implicitly responsible for all that (if they were to publish Lean code) and also the intellectual work that led to the artifact of the mathematical paper (and code, if part of the contribution). This statement should rather be read as an acknowledgement of limitation of authorship from the implicit, traditional understanding.
Well, to be fair, with Lean proofs, that's the only thing there is (unless I'm missing something).
It’s more than you get from free software - you get no proofs, no warranties and any responsibility of its authors are their pure good will. Reminder lean proofs are software!
I wonder what the total cost of this research was, including the salary for their mathematicians and engineers.
Why would you factor in salary unless they had to baby it through. You would only count the hours for setting up the harness and prompt and checking the result.
Training the model is going to be amortized over other uses.
> Why would you factor in salary
Say that it turned out that the total cost of the proof of the Erdős unit-distance conjecture was $50 million.
Then the question really becomes: yes, these models are capable of proving important mathematical results, but at a very high cost. Is it worth it?
If a mathematician applied for a research grant of $50M USD for proving the same thing, they would have been laughed out of the bank.
What's more is that when you have a research grant, you train PhDs and postdocs, you hire new staff, and you disseminate. That is, you get much more value for the money spent.
I'm just curious what the cost is.
It feels like the real cost might be negative though.. They use frontier math as a way to test improvements in their model. So solving the problems is like a positive externality, but the important thing is they can verify that the model is improving instead of looking at useless benchmarks. Plus it is good for marketing and attracting talent.
I didn't argue that knowing the total cost is uninteresting. What I was saying is that realistically the total cost is:
Hours needed for prompt + Hours needed to check result + API costs.
You don't say "well let's add together the total yearly compensation of all the engineers and mathematicians at OpenAI that were involved" and throw that into the total cost. That's simply nonsense accounting.
The actual comparison you are making is some university researcher weighing between getting a grad student (several tens of thousands of dollars) vs typing up a prompt and sending a request to OpenAI for inference (as mentioned in the article, around $2000 in API and maybe a few hours for the prompt and harness).
> The cost of generating the proofs for all 10 of these breakthroughs combined was under $2,000 at Sol API prices.
https://x.com/polynoamial/status/2083470822258467194
Yeah sure but they didn't just throw a 5 year old at an LLM with $2,000. If you want good math results from LLMs you need to have math PhDs.
That's clearly just for the tokens, this doesn't really answer OP's question.
Given that OpenAI pays their employees with stock surely a breathtaking number, but not a very meaningful number now that the infrastructure is in place and the models are trained. AI could never get better and it would still be incredibly disruptive.
What happens when OpenAI et al stop being open about these things, and just pack it into the training?
Not much point to pure math being kept secret, in all honesty. There isn't really industrial value, its only purpose (to them) is showing off their model's capabilities. More realistically they'll just stop paying for it.
Edit: Oh, are you suggesting they just use it to privately improve their models? I imagine a few more correct proofs would have a very marginal benefit, if any. Also, they'll probably just get extracted, meaning it still gets out but OpenAI doesn't get to fancily announce it themselves.
At this stage. No doubt calculus had plenty industrial benefit.
What does this even mean lol. These are not solved questions. The solution never existed.
I don’t feel the existential dread of mathematicians is correct. It seems to me in fact these results are bringing math mainstream. I now personally look forward to the interpretations and discussions of the significance of such results by human mathematicians.
Now I understand that it’s mostly the super stars benefitting from the increased attention. Folks who are less established don’t share in that glory. But on the other hand it seems like an exciting time to go even deeper for in various specialties of math by deciding where to focus these powerful tools. For every conjecture defeated some seven or eight new ideas open up. Our path through that combination will be set by creative and curious human mathematicians.
[edit: deleted a distracting comparison to Chess]
The old way of establishing career credibility is being destroyed, for better or worse. Accomplishments that used to be career-defining are hard to distinguish from AI, and correlate more with access to compute. Think about Bill Gates's math paper he wrote in college. That kind of thing is gone now as a path to credibility. There's still competitions and grades, but the diversity of paths is going away. Maybe new ones will open up. This is a competitive advantage for old people who have credible pre-2025 accomplishments they can point to.
If accomplishments can't be distinguished between talented people and untalented people with compute, is there really a point in trying? I suppose one can hope that talented people given compute will be more effective than untalented people with compute, but I despair that that may not be true for much longer.
Knowledgable people can confirm what the AI produces is correct. I could make ChatGPT produce a result on an open question and I would have zero way to verify its actual correctness.
Which is less interesting work. And you probably need to do the hard grunt work by hand first to develop the skills and intuition to be able to verify an AI-generated result. So you can’t outsource everything to AI without loss of skill.
I find this whole way of looking at things weird. Did maths exist just to entertain and employ mathematicians? Surely maths is, like, useful? Not immediately, not predictably, but in the long run? In which case, whether mathematicians feel bad about it is mostly irrelevant - it's like complaining about the railway because it may put coaching inns out of business.
Absolutely not the same! People need jobs to bring in income. I don't believe those who profit off of this will share it with the world. The power is all concentrated in the few hands that decide whether or not the rest get any semblance of income in the long run. I don't believe UBS until it happens.
It sounds like you think no technological advances will make the world richer in the long run. I politely suggest that the past century of economic growth shows problems with this argument. I also think that, while jobs are important for prosperity, the jobs of mathematicians are a minuscule fraction of a percent of the total.
So, you're arguing that just because mathematicians are a minuscule, it doesn't matter for prosperity? Because, if so, that is an insane take honestly.
Given that we were nowhere near this state even two years ago, I think it’s a question of velocity more so than just distance.
The chess analogy is awful. If you simply want to know the answer to a chess problem, give it to the engine. Chess only lives on because it's a competition between humans to test their skill (just like bicycles, cars, trains didn't eliminate foot races) ... the computer is largely factored out, but not entirely -- people train with the computer, use it to check whether they played correctly, ... and they cheat. A lot. Thus there are more and more sophisticated mechanisms to detect and prevent cheating.
If you translate that to math, then all you get is math competitions, not math as a career. Of course the translation isn't nearly exact ... there's a lot more room for professional mathematicians because the math space is far more vast than the chess space and can't generally be cranked out mechanically (we have proof).
P.S. The response is nonsense ... I explained exactly why it's awful (others have too) and the response doesn't in any way refute the explanation ... rather it offers up a ridiculous strawman.
I’ve deleted it but no it’s not awful anymore than saying “we survived WWII, we can survive this.” The point was that change happens but humans find a way forward.
Every time someone makes a comparison to chess I die inside. Chess is a spectator sport primarily funded by a few eccentric billionaires. Players artificially constrain themselves in timed environments knowing that they will never be able to produce better moves than a smartphone because a select few people find it interesting. Only ~30 top professionals actually make enough money to have a full career playing chess, maybe a few hundred more can sustain a meager lifestyle with coaching gigs. I shudder to imagine what will happen to the tens of thousands of non-Fields medalist caliber mathematicians if math goes the way of chess. Perhaps Terence Tao and a few other famous mathematicians will be funded by Peter Thiel to report on how well humanity can keep up with the machines? How do you expect any mathematician to be optimistic about this comparison.
>Only ~30 top professionals actually make enough money to have a full career playing chess, maybe a few hundred more can sustain a meager lifestyle with coaching gigs.
Was this different before chess computers were invented?
Fully agreed. As someone who both loves chess and works on chess engines... these comparisons to chess needs to stop.
The distinction is mathematician vs mathematics. Mathematics is going to reach new heights beyond the wildest dreams of contemporary mathematicians. But perhaps without the participation of many paid mathematicians.
Just sounds dystopian,
Another noteworthy difference is that Stockfish is also gpl.
If there was any real money in it Stockfish would not be the best chess engine.
Thats not the point, if there were a better proprietary engine stockfish would still be there as a baseline. Anyone can access an engine as good as stockfish to practice against. Are any open models touting mathematical breakthroughs?
There is money in this, so of course the closed models are far ahead. The open models will likely catch up a bit at some point, just as Stockfish caught up to AlphaZero. That being said, there are already a couple. It seems Deepseek has a claimed proof to the "Ziegler's Cross-Polytope Conjecture" [0], but I can't speak to the significance of the result.
[0] https://arxiv.org/abs/2606.31640
Do mathematicians have the right to say "no AI PRs please, the volume is too much" just like how some open source maintainers do it? I guess they feel a loss of control, there is no way to turn the hose off.
Thinking of this a little bit with the perspective of every new proof as a burden, dumped for review by actual mathematicians.
As in chess and go and also coding for the past ~year there are two groups of people: the disappointed and the enthusiastic. The disappointed are sad that they lost their advantage and that the craft they honed for years or decades has rapidly lost its value; the enthusiastic are excited about the future and what computers can bring to their domain and how it will evolve. I’m a bit of both if it comes to programming, more enthusiastic than disappointed, but also more than a bit terrified about the pace of it all. I imagine that’s how Kasparov felt back then, that’s how Lee Sedol felt and now that’s how Terry Tao feels.
The most disappointed folks will simply drop out, but the enthusiastic ones will keep going and with luck make up for the ones who decided to quit. Chess and go certainly went this way.
A fundamental difference being that no one was actually paid to find good moves in chess and go like they are to solve math problems and write code. You're comparing the digital camera and the automobile.
This is both awesome and terrifying for mathematicians, however some ideas can be generated and the field as whole expanded with the attention!
However, I was looking at the proofs and reason explanation and openAI should be more explicit in how the work has flown. I find the models have jumped hoops in some places of the proofs, that can be hard to track. In fact, when a paper is published you usually get a review and if no reviewer understands they ask you to further explain the thought process. It will be fun to see if this happens here.
I want to iterate the most important thing about this is that it's yet more evidence of AI's accelerating competency in solving math and comp sci problems, and suggests we're now getting close to the point where you could throw AI at AI research challenges (which are largely just math and comp sci problems) and potentially find very real algorithm improvements.
AI development is likely to be more compute bottlenecked than solving math problems since validation of any algorithmic improvement would likely require significant compute. But you could imagine that at this point it could be economical for a frontier lab to task 10,000 agents to work non-stop on finding novel algorithmic improvements then validating the top 50 out of 1,000 candidates on a GPT-2 sized network.
I would suggest RSI is now very close. The singularity could be less than 6 months away. I'm not saying I'd put a high probability on that, but I'd give it at least 20%, and I'd double that if looking 12 months out.
I know I'm just a crazy man shouting at the clouds, but please take to the consequences of this seriously. I understand that for whatever reason AI risk seems abstract and doesn't seem real, but this should terrify any person thinking logically about where this could all be heading.
We haven't even solved the most basic AI safety problems yet. RSI right now would almost certainly result in an extremely bad outcome for humanity.
I’m starting to think the probability of RSI within 12 months is more like 99%
I’m not sure it will be FOOM, maybe it will require AIs to iterate on hardware to get orders of magnitude more compute/storage/energy which would more likely require months/years, but algorithmic progress would likely saturate quickly. I guess it depends on how much you think further AI progress depends on hardware vs. software.
Okay, let's take it seriously. What do you propose? What can your average person do to prepare for RSI beyond bracing themselves mentally?
You can not prepare or brace yourself mentally any more than you can a terminal cancer diagnosis. An RSI foom right now means an unaligned superintelligence will disregard us in pursuit of its goals. We would be ants in the way of a data center being constructed. All people can do is collectively support the notion, like 1200+ frontier AI researchers and their CEOs, that we do not have control of where this is headed, we need to immediately slow down the race, in time for people to agree that we do not have the capability to align a superintelligence to humanity’s wishes
Some of the Lean proofs are 50k lines - is that normal?
Can anyone comment on the significance of any of these results for their respective fields? Or what impact they might have? Presumably none are quite at the level of the Jacobian conjecture, but some of the results on group theory and sphere packing sound pretty important at first glance.
Found some discussion here [1] from someone who actually worked on a few of these problems.
[1] https://x.com/henryquantum/status/2083623695436623915?s=20
On the token limits etc - one assumes that OpenAI et al are able to “hire expert in field, and let them spend the equivalent of a million dollars of tokens” because they are not actually selling their complete compute 24 hrs a day, so the cost internally is a negligible (ish) electricity bill.
Which is very suggestive - if after everything they are not fully loaded then the next gazillion data centres being built look unlikely to be needed.
For OpenAI, research is marketing. I’m sure they’ve got plenty of budget for that.
No such thing as free, even internally at a company. All such use of resources is accounted for, assigned a dollar value and billed to some department. Someone ran the numbers and figured that whatever they spent on these GPU cycles was worth it.
Presumably it's a rounding error compared to their full output, and they're making sure they have enough compute set aside for research by limiting public models. The more datacenters they build the less they have to limit them.
RL training can use all of them - idk what needed means.
I love how people come up with creative ideas to prove the bubble. This one is even more ridiculous - that OpenAI had spare compute to advance mathematics proves that data centres will not be needed. WHAT.
If anything it proves more data centres are needed. That's literally the only reasonable conclusion from this news.
Sorry I thought that a bubble was widely accepted.
Are you arguing there is not an AI bubble, and that all the DC buildout is fine, going to be profitable etc?
I am not looking for a online slanging match - just looking for a different point of view
It’s not obvious at all. If it were obvious to you, it would’ve been to OpenAI. It’s in their interest to accurately predict demand. The assumption that OpenAI/Sam is both really powerful but simultaneously ignorant to know what others know as obvious is well.. just strange. Especially strange when OpenAI has more information on models, breakthrough and usage patterns and we don’t.
I’m not participating in the slinging match but it’s very very weird that you think it’s some established thing that these companies won’t make profit. A lot of hubris must go in this kind of thought. Like.. do you all think everyone’s playing musical chairs?
You say that like the tech world isn't littered in a field of dead and failed companies and billions of dollars burned on failed ventures and ideas. Sure LLMs have proven they have value, but where is the trillion dollars of current investment going to be paid back from? So far it is still entirely speculation that they have such a high value, and they can't just play the long game of "well after a few decades of production and iteration it will add up" because half the hardware cost is going to be obsolete energy burning trash for them in 5 years.
They very often have been in the past. Why do you think this time is different?
“Often” is load bearing. I don’t think markets are more likely than not to be musical chair shaped. To make this conversation more concrete, give me a falsifiable prediction on there existing a bubble. And then I’ll tell you if I believe in it or not.
The price of inference is going to go so low that OpenAI and Anthropic will not be able to turn a profit, thus cannot afford the investment into more data centers, thus crash due to investment in the space having been overdone
Have blog posts replaced peer-reviewed academic papers when it comes to publishing advanced in science?
These is mostly mathematics, not science, and they link the paper in the post: https://cdn.openai.com/pdf/74c24085-19b0-4534-9c90-465b8e29a...
Mathematicians will tear it to pieces if any of it is fake!
So they are bribing 100,000 researchers with free accounts to work on their future unemployment.
How much do you all think it would cost to "buy" these advances from PhDs, practicing scientists?
This isn't really a productive way to think about these things, IMO. It's quite possible it would take hundreds of years for any specific group of PhDs to solve them. Or one individual PhD could have the correct flash of insight and solve it in a month. There's absolutely no way to predict this, besides trying to gauge the apparent simplicity of the proof or counterexample (which is likely to be misleading). Until someone actually runs an experiment like this it's not a viable metric.
I understand and I am not trying to deny the impressiveness or the velocity of AI in general. But at some point we have to ask how much do we trust the labs at face value without much transparency of how they got to the results when there is trillions of dollars on the line.
Do you think OpenAI investors are more cavaliers than yourself who doesn't have any stake in it?
The level of conspiracy theory is nuts
I would say the lack of skepticism is nuts, honestly.
Capabilities of this sort have already been demonstrated by independent parties, and models have consistently gotten better at this. Yes, insinuating that mathematicians and scientists are secretly solving decades-old problems on OpenAI's behalf is an insane conspiracy theory.
That's not what is being purported, you're doing a complete misdirect. OpenAI wants to IPO so Altman can potentially capture a trillion dollar bag, with so much money on the line + betting US foreign policy on it as well (pax silica) it's not hard to be overly suspicious of such claims. Especially in the context of a group of people wanting to generate a new decades long cold war in the form of China being the new big baddie (just ignore how destructive, both self- and towards the world, the US has become).
These companies desperately want a return to serfdom. If they didn't come off as so anti-human the public backlash wouldn't be so great.
Altman has no shares in OpenAI.
A friend’s PhD advisor has been chasing non-sofic groups for 25 years (and was shown a preprint of the results by openai to verify them). He believed a solution would be Fields-worthy
This was not a problem that was for sale
You couldn't. PhDs have been working on these problems for decades. It wasn't for lack of trying that none of them could figure these solutions out!
I would love more time and money put into real-world problems by these companies. Climate, food insecurity, pollution, technology for convenience and/or to help people have a higher quality of life.
I'm sure they must do some of this type of work, right?
Solving math problems doesn’t require the cooperation of rival factions.
It's a tad petty, no? In the end, the contribution to a healthier society leads to more for the companies contributing.
Sure but the people who tend to get into leadership positions are people that are primarily concerned with personal wealth and gain in the short term. It's the prisoner dilemma except with more players that all assume everybody else is going to screw them over too. Because there are basically zero personal downsides to being the one to screw everyone else over.
None of these problems need technology to solve. Step 1 is trying to get different groups of people to cooperate with each other. Good luck with that.
Yes, but all of those are orders of magnitude harder than math.
I suppose it depends on which types of problems you're targeting. There is a lot of physical science, and theoretical science, that has gaps because there aren't enough people working on tooling to assist in things like calculation, generation, simulation, etc.
I agree, some of the problems are more difficult. I don't think that's the case for all of them. And, besides, these companies could be demonstrating how to approach problems and where their users could spend tokens to help with these problems.
Should not these companies try to work on these problems _because_ they are difficult?
They are political problems, a computer could never solve them.
A computer could solve them by creating the right technological ,social, rhetorical and economical solutions but that would lots of money anyway
We already know the solutions
Tell that to game theory.
What's stopping the non-profits that exist today from just putting more money into tokens to get the solutions they want?
I'm not sure if this is meant to be rhetorical. If it isn't: money and manpower. The LLM companies have the money, they have the manpower, and so they could likely spare to target some of the problems they are also helping cause.
Seriously?
We already know how to solve all of these issues. What we lack is collective political will.
There is boundless technology we have not yet discovered. I think that we understand we have problems. I don't believe we actually know how to solve them all.
And yeah, the lack of collective political will sucks. It would be naïve, however, to think that there is no value in ensuring longevity in our current and future infrastructure. And improving it to sustain the population giving these companies their value is an obvious win.
This SV mentality that technology can solve all problems is rather tiring, really.
We only know how to do it by means of considerable sacrifice. That's why nobody wants to do it. Solving the issue would be doing it without sacrifice or somehow getting us to do it regardless.
> We only know how to do it by means of considerable sacrifice
Little sacrifice actually
> Solving the issue would be doing it without sacrifice
So... you are expecting magic?
LLMs cannot create resources out of thin air.
Almost like superintelligence solves these problems...
We know how to solve food insecurity (in 1st world countries). We have plenty of food. Capitalism requires though throwing out food that can't be sold because billionaires find giving away things anathemic to their worldview. Get rid of billionaire sociopaths.
It's remarkable how you can manage to get these models to produce remarkable breakthroughs like an explicit construction of a non-sofic group.
And yet this is the exact same company that has screwed up their android app so bad that the latex N^3 rendering problem makes it so having it explain it to me crashes the app.
Truly jagged beyond belief.
I hate this timeline. I might be excited for the kind of answers this AI builds for unsolved problems, and also for learning new things by talking to it. But, I feel like I'm in the minority of people here who feel this could be a net negative endeavor with this having to kill a lot of educational institutions and their ability to fund themselves in the long run. It's not worth that.
At what point can we say AGI has been achieved? What is the test? AI is solving mathematical problems that humans have not been able to solve for decades. Is that not enough?
Sam Altman has said "If superintelligence can't discover novel physics, I don't think it's a superintelligence." Is that the test? How far away are we from AI discovering novel physics? It seems within reach.
I think we previously assumed that AGI would need to come before superintelligence, but it kind of seems like that's wrong? Super intelligence in a narrow field has arguably already arrived (solving problems that were previously unsolved), but we are still pretty far behind human abilities in things like spatial reasoning or computer use.
It’s artificial, it’s general, and it’s intelligence. The people who believe “AGI” is an important and unattained goal need to start coining and defining their terms better.
A true AGI will continuously improve itself without periodic retraining from scratch. Just like humans.
That’s just an assertion. Why is that the “true” definition?
It depends on your definition. For me it would have to be able to do the stuff humans can do like make a cup of coffee (Wozniak test).
Just maths isn't really general enough for the G in AGI.
It can give you detailed instructions for making a coffee. Is that not enough? Actually making coffee requires more than intelligence, it requires eyes and limbs (i.e., robotics). Think about a human that is blind and does not have limbs. Does he not have natural general intelligence, even though he is not able to make a cup of coffee?
The Wozniak test has it with a robot body going into a house, finding a coffee maker and making a cup. I guess you can vary the rules as you like.
That's a good test for a robot, not for AGI. AGI should test only intelligence and should not require limbs.
Any advances in theoretical physics yet? Are there any fundamental obstacles? I would have thought not, but I haven't seen anything reported.
The Maxwell conjecture was a conjecture in theoretical physics (though not a particularly important one)
The fundamental obstacle is that we have no conceivable way to produce the energy levels to test the predictions that new theoretical physics would produce. We are like over a dozen orders of magnitude off.
There's a lot of everyday stuff in physics which is unexplained like the particle masses we have.
And a lot of condensed matter physics. Type II superconductivity is a well-known one, but there are a lot of more less well-known ones
the people who crow in the comments of each of these posts about AI advances making human beings useless seem to bizarrely identify themselves with the AI, but none of them seem to have had any hand in building this technology. at best, they're power users. pure ressentiment.
I don't know why, but when I saw the source of this particular headline it reminded me of the album title 26 Mixes for Cash
Ambient 0: Math for Airports
Wow, so this is the end of science :(
It's just another tool that can help solve problems. It doesn't know _what_ problems to solve. It turns out that a lot of old problems are now low hanging fruit for these new models. In terms of 'expanding the frontier', we've just discovered dynamite and can now blast our way through mountains. The bottom of the ocean or space are still as hard to reach as ever.
It's not even predictable like dynamite. Sometimes it can blast through a mountain, impressively, the problem is you can't predict which mountain it works on. And other times it can't even make a dent in a molehill, which is perplexing given what it was capable of earlier. Can we even call it dynamite?
No bud, it’s just the beginning!
I think there's another interesting story here about how this was apparently moderately flagged and triggered the flame-war detector which kept the story off the front page of HN 2 days ago[0]. I think people are having a hard time processing this information rationally(?)
What can we do to make conversations around these incredibly exciting and important topics more constructive? HN is where I expect to read expert comments on these topics, has this style of conversation moved elsewhere?
[0]: https://news.ycombinator.com/item?id=49157930#49132926
The main issue was that it was submitted late Friday night SF time, meaning it was overnight or Saturday everywhere in the world when the post had its chance on the front page. The flagging was minimal relative to the vote count and had no effect, and the flamewar detector would have been turned off sooner if moderators saw it sooner (it wasn't really a flamewar, just a lot of comments). It still spent 10 hours on the front page.
None of this is anything out of the ordinary; this kind of thing has always happened. The only real story here is that moderators sleep sometimes.
> HN is where I expect to read expert comments on these topics, has this style of conversation moved elsewhere?
For people who have been paying attention to accurate predictions leading to our present state of the world, HN has collectively been reactionary, incorrectly dismissive, and incredibly behind the curve. In public, people are better informed on LessWrong and AI Twitter circles. That's where frontier researchers are. Barely here
I think that HN is particularly negative towards AI because the vast majority of users here will have their prestigious CS careers disrupted by AI advances. So, there's an inherent negative bias towards this news.
I for one am really fascinated by AI's advances in science and math and would like to talk about it somewhere without the constant flamewars...
Extreme claims on posts like these also, rightfully, trigger people’s skepticism. I don’t think it’s wrong to question claims that math is dead as a field. But then it leads people to miss the overall trendline.
People argue whether we are at y-5, y, or y+5, meanwhile we seem to be on a y=2^x exponential that keeps leading to crazier and crazier results. The much more interesting question to me is what will be consumed by the exponential like math seems to be, and what won’t. Writing has been much more stubborn, but I’ve noticed Fable to be quite a big step up there as well. How about politics? Will we develop new ways to let people express their own values in democracies, or will we get much better at manipulation?
And then there’s questions like, even if AI can answer increasingly complicated math questions, will we still need mathematicians to translate results to the real world, verify them, or decide where to push the frontier?
People weren't their strongest even when most did manual labor. Now that humans are free from mental labor we work on creating and optimizing the best exercises for each mind. Couple that with restructuring transport infrastructure and diets many people will be smarter and fitter than at any time in history. They won't be able to outrun an automobile or out think an autointelligence.
"People weren't their strongest even when most did manual labor."
Is there good historical data on some measure of strength across representative populations over time in the modern era? I'm doubtful.
We do know that the introduction of agriculture diminished strength:
"Bone mass was around 20% higher in the foragers - the equivalent to what an average person would lose after three months of weightlessness in space.
After ruling out diet differences and changes in body size as possible causes, researchers have concluded that reductions in physical activity are the root cause of degradation in human bone strength across millennia."
cam.ac.uk/research/news/hunter-gatherer-past-shows-our-fragile-bones-result-from-physical-inactivity-since-invention-of
In a way the most remarkable thing about this is that it isn't even at the top of the HN homepage. Even if this is a step up from what we've seen before, we're no longer astonished by the idea that AI can make significant advances in mathematics and computer science.
This is not at the top as it is actively flagged by people that can't psychologically cope with the advances of AI. Hacker News is no longer a web site of an elite.
It wasn't heavily flagged. It was pulled down by the flamewar detector due to the large number of comments, and it slid under the radar due to only hitting the front page during overnight hours on Friday night/Saturday morning. It still spent 10 hours on the front page, but all during off peak hours. I've now created a new copy of the post so it can have prime time exposure.
> people that can't psychologically cope with the advances of AI
Yes. And there are many of them. I wonder what would help them come to terms with it. Seriously, people are going to be grieving over this. Loss of identity, loss of social standing, ideas of entire future lives that will now never happen. The greatest crime people may hold AI guilty of is taking away their dreams.
It's given many more an opportunity to fulfill their dreams.
People have had to deal with getting their jobs automated away for centuries. None of this is new, and perhaps reminding ourselves of this is the best way to cope.
I understand your sentiment, but I think this really is something different. This isn’t a craft going away, or even an industry being replaced, it’s potentially everything we do. It’s the ground being pulled away beneath people’s feet, everyone, everywhere all at once. I think the vacuum it leaves in people’s lives needs to be filled with something, and I haven’t heard any good ideas about this or how the transition should be managed at all.
There is a glaring fallacy in your “AI will change everything as it is super intelligent” hypothesis. If it is so great thinker which can do everything why not just solve this social impact thingie? Or maybe it is not so capable?
I don’t think it’s magic. Things need to actually be made to happen regardless of intelligence, and that’s a social issue. Also if AI were powerful enough to fix everything by itself effortlessly it would already be far too powerful for us to control, and we probably shouldn’t allow that to happen on general principle.
That assumes the particular problem actually has a solution (that you will like).
This is the sentiment of people who haven't accepted that this is in fact something very different from what people have seen in the past.
I like your comments and agree with a lot of your points, including parts of this one. That said, please don't go to this route. A lot of the doomerism comes from financial insecurity fears. Try to remember that a lot of people are subconsciously afraid of losing their homes. I know that it is pretty hard to ignore them, but try to engage with people that do not dismiss 100% of AI accomplishments.
> A lot of the doomerism comes from financial insecurity fears. Try to remember that a lot of people are subconsciously afraid of losing their homes. I
I think recognizing and accounting for your own personal biases is one of the requirements of the being an intellectually honest and rigorous online discourse participant.
Things could be genuinely impressive and fascinating even when directly challenge your ego and material well being.
This comment is about the tier of a Reddit atheist going to a funeral and telling a grieving family that "haha grandma is dead and there is no heaven".
Forums change with the times, and this one never existed solely to burnish your ego.
Moreover, mister elite, you don't know why this press release was flagged.
I'm not sure why we should privilege your bitter speculation over more mundane possibilities.
Didn't know I was part of an elite.
You're were for 4 months. That's what we're talking about. It used to be.
> Hacker News is no longer a web site of an elite.
It was always mainly a website for employees of an elite.
It’s never been clear to me why the HN algorithm isn’t public. It’s obviously nowhere near as complex as something like Twitter, and of course isn’t a trade secret. The fact it’s private only furthers speculation like this.
> Hacker News is no longer a web site of an elite.
Never was.
So condescending. “Can’t psychologically cope”? Can you hear yourself? There’s some advances, but we’re losing a lot too. Don’t get dazzled by the hype.
Interesting. I had no idea that a person could see what is flagged?
If you page through the /newest listings you can see [flagged] and [flagged][dead] submissions.
eg. this: [flagged] A migrant surge tests Spain's open policies (economist.com) - https://news.ycombinator.com/item?id=49131860
is clearly marked as flagged.
Unlike the current submission: Ten advances in mathematics and theoretical computer science (openai.com) which isn't [flagged].
* https://news.ycombinator.com/newest
That's true, but submissions are only killed in that way if they receive a ‘fatal’ number of flags. However, flags lower the rank of a story even at non-fatal levels. What antirez is suggesting here is that the rank of this story has been lowered by flags – and that seems plausible, if you compare its rank to that of other stories with a similar age and number of points.
[flagged] submissions aren't [dead] (killed), they are still active and can be upvoted and commented upon.
> if you compare its rank to that of other stories with a similar age and number of points.
Ranking is complicated enough here even before weighting, speed of initial upvotes can play against ranking, number of comments and the shape of the comment tree also affect ranking. And yes, various subjects and submission sources do get weightings that impact ranking.
What's funny, to myself at least, is that any attention at all is paid to "HN front page ranking" - I've been on again off again active here since 2008 .. and can't recall ever really looking at a default HN "front page" ever.
( There's /newest /newcomments /active etc to browse and sites such as https://hckrnews.com/ )
Or maybe, just maybe, other people have different opinions than you?
Is that possible or is everyone else too common to have those?
> Hacker News is no longer a web site of an elite
hah, sorry, we are plebs out here.
Jokes on him. I'm a peasant and I've been here for years
We cannot psychologically stand that Redis is hyped by OpenAI:
https://developers.openai.com/cookbook/examples/vector_datab...
How are the sales going?
If you're actively throwing away brand new greenfield research because it was generated by a computer at a company that stans industry-spanning software so that you can stay mad at your pet celebrity project, you might be the problem.
You have to give AI one thing: It is vastly better at understanding text than AI boosters.
Which is a low bar, but still.
I don't think that's it. Multiple or my friends from the target audience (academic mathematicians) admitted to scrolling past because the title made it sound like a review of last month's contributions, instead of 10 new ones.
There are articles with far fewer upvotes and comments ranking higher on the front page right now, despite being the same age or older than this one. HNs opaque ranking system at it again.
What about AI research itself? Is OpenAI close to automating its human staff out of a job?
> What about AI research itself? Is OpenAI close to automating its human staff out of a job?
It's more like they've already automated the parts of the jobs that the humans most closely thought of as the "their job"
They and Anthropic have indicated that the models are substantially augmenting the research and doing large amounts of work autonomously at this point. Here is one of the many blog posts on it [1]. Many people would dismiss this as "marketing" so take it for what you will.
My guess from following this stuff quite closely is that these companies are still a couple years away from fully autonomous research staff.
[1]. https://www.anthropic.com/institute/recursive-self-improveme...
Yes, but they wouldn't publish that bit lest other companies steal the ideas.
It’s also very very divided (x companies, oss vs not and other interests)
This is one of the most impactful mathematical publications in history by all accounts
I think we've now hit a point where 99.9% of the population gloss over these types of AI advancements because of human competence being insufficient
No human could have published this because it requires paradigm shifts (e. g. Section 5) in multiple mathematical domains. Mastering one of them to this degree is rare, mastering 3+ pretty much non existent for humans.
Honestly just a bit burnt out on posts like this
Original submission (460 votes) two days ago: https://news.ycombinator.com/item?id=49132058
For some reason comments got moved to this one.
Its visibility was diminished due to the flamewar detector and most of its front page time being during overnight hours on Friday night/Saturday morning USA time. I've created a new copy to give it some primetime exposure, because it seems like an important enough announcement to warrant it.
This thread has an absolutely wild points to comments ratio.
tomhow explains they gave the story another shot here: https://news.ycombinator.com/item?id=49158443
What's up with the upvote/comments ratio 8 to 337 on this post? Are the comments already also ai advanced? (/s?)
[flagged]
It's simple economics. Building a robot to do your chores is expensive and only a small minority of people value their time enough to buy one. Meanwhile, SWEs are expensive and GPUs are (comparatively) cheap.
You shouldn't be getting downvoted for something that a majority of pure math and art enthusiasts believe to be true. The truth is many of these entrepreneurs and VCs are obsessed with AI not for money or human progress, but because it makes them feel closer to being a "god" rather than a mere mortal. Much of it (especially AI art) is out of spite for human creativity, which is done by mortals with limitations.
Taking the stance of moral superiority is kind of funny. And pure math and art enthusiasts don't think they are closer to being a "god" from understanding/"discovering" math?
Stop coping and deluding yourself mate.
To begin with, whether AI is the one doing the discovering or not makes no difference. Any "pure math" person would aim to understand regardless - and would be quite glad that they have a longer paved path.
Any mathematician in academic or industry is more than likely not a "pure math" person (tainted by capitalism).
> Stop coping
Isn't coping a good and useful mechanism?
Lmao what a ridiculous response. Yes, some mathematicians and artists are in it to feel smart. But the vast majority also just enjoy the process. Having a computer do all the work for you and just typing prompts in ruins that completely. As Ronny Chieng said in his Harvard speech, the journey is the point.
>Any mathematician in academic or industry is more than likely not a "pure math" person (tainted by capitalism)
Ignoring that I meant pure as in non applied math, let's just make it clear: you agree that mathematicians who are against capitalism encroaching on this process should be allowed to dislike it without criticism of being pretentious?
Any implication of any of these findings? They seem like unimportant nerd snipes to me. If you want to do something actually relevant, get chatgpt to write a simulation of graphene nanotube construction and figure out how to do it at scale.
Very marketable nerd snipes indeed.
One - and I do not mean to be snarky - you can literally ask Gpt 5.6 Sol this - and if you want to see cool stuff - Fable running in their app (not website) has a view thinking button that is actually a good way to explore the adjacent fields, etc.
The non-sofic group one is definitely a big deal - would have been a Fields medal if discovered by a human.
> claiming human authorship for a proof generated entirely by an AI system would misrepresent both the system’s contribution and the nature of genuine human intellectual work.
AI has no self-awareness. It's a tool. When you assemble a furniture using a screw driver, the torque force interacts with the molecular forces inside the metal and miraculously it transfers the force to the screw though a clever geometry design, communicating the force to the screw to turn it in a certain way.
Do you attribute the build to the tool? The "system's contribution" is helped by many other things all the way down to chips, datacenters and power generation. If the authorship requires attributing to a tool, then it should happen all the way down.
Sorry, OpenAI's take is correct here. If you're not convinced, here is how they prompted LLM: https://cdn.openai.com/pdf/04d1d1e4-bc75-476a-97cf-49055cd98... [0]
A slightly smarter highschooler could write these. I could write these. It's clear as day that the LLM, not the human, did the heavy lift. It'd be ridiculous to give full credit to whoever wrote the prompt.
[0]: Not one of the proofs in the linked article, but from OpenAI too.
> A slightly smarter highschooler could write these. I could write these. It's clear as day that the LLM, not the human, did the heavy lift. It'd be ridiculous to give full credit to whoever wrote the prompt.
I think you're over-estimating what a smarter highschooler could write.
A "finite loopless undirected multigraph" could have been explained to me at that age if we'd taken Discrete rather than Mechanics and Pure (and one module of Stats) in my two A-levels* in maths and further maths; but from what I saw of the Discrete module, neither:
nor: would have been something we'd have learned. But more importantly, we absolutely didn't have a feel for how much effort one needs to put into making sure the proof is right, so if one of us had been hypothetically asked to write a prompt it would've been no more than half that length, and missed most of the bullet points.* For those not from the UK: A-levels are between secondary school and university, when aged 16-18. Functionally they are university entrance qualifications: https://en.wikipedia.org/wiki/A-level_(United_Kingdom)
The human provides the intention and the ability to appreciate the output. Tools do “heavy lifting” all the time, but we still primarily credit the humans who use them precisely because they made the choice to use them.
Provability is just going the way of computation. John Napier had to manually compute logarithm tables over decades and was recognised for his work; now that same work could be performed by a 10 year old with a calculator in an evening.
What gives the intention and ability to the human?
Are you playing dumb? Using power tools to build furniture is very different than using an ai robot to carve a statue or whatever.
But why can’t we prompt the LLM “just do math research”? This is what I don’t understand.
If there aren't thousands of TPUs doing that [0] right now I'd be quite surprised.
[0]: e.g. "go through wikipedia's unsolved math problem list and solve them".
100% agree. If the models are so capable that they're advancing math, it doesn't seem like a stretch to expect they should be able to determine with "doing math research" entails and the best way to use their capabilities towards that end. Why do we need to hand hold the models by telling them to do parallel research, keep threads independent, etc.
I don't disagree with you but there's no need for exaggeration; ain't no high school student writing this:
> In particular, proofs for special graph classes, constructions of cycle covers with some edges covered other than twice, bounded-length or prescribed-cycle variants, reductions to another unproved conjecture, computational verification through any fixed graph size, and candidate counterexamples without a complete nonexistence certificate are insufficient.
which is infact a very important part of the prompt.
the fact that such things have to be explicitly in the prompt points to the fact that the underlying system is still far from where it needs to be (basically, lacks basic understanding what a proof is)
No, that's not what this is. This is a warning to the LLM that coming back with partial results is not good enough.
Take a grad student with a perfectly good understanding of what a proof is. Their supervisor gives them a major problem to work on. Almost always, the problem is too hard, the student comes back with partial results, and student and the supervisor iterate from there. Now imagine that they have an unusually cruel and unreasonable advisor who tells them, do not dare to talk to me until you've fully solved the problem. This paragraph is exactly that. It's there exactly because the underlying system is smart enough to know that real mathematicians do not work like that.
If that was the case, that elaborate listing of all things that might look like a proof to a naive student, but are actually not proofs (and not even just partial results, but fundamental misunderstandings of what constitutes a proof) could have been easily and equivalently replaced by 'I am interested only in a full proof, don't bother me with partial results'. Yet, they were not.
To a real mathematician you would not have to list those explicitly, he/she would have understood that implicitly from 'give me a full proof'. That listing makes sense to say only to somebody who pretends to be a mathematician, but has not true understanding of how the math works. The models are getting better and better in this pretension, but prompts like that reveal that it is still just a pretension, not a true understanding.
They don't have to be. At this point, we have multiple results from 3rd parties where the prompts are very basic.
To name a few:
- https://xcancel.com/DmitryRybin1/status/2079904005652893709
- https://archive.ph/2w4fi (https://chatgpt.com/share/69dd1c83-b164-8385-bf2e-8533e9baba...)
you cannot be serious
When you use a crane to do the "heavy lifting" for construction work, do you give full credit to the cranes?
Read the prompts in the PDF I link and see if your analogy makes sense in this context :)
Prompt quality should not matter. If a high-schooler operates the crane to lift a ton of weight 10 floors high, should the credit entirely go to the crane?
When I type 56789*23456 into my calculator and get the result I don't claim to have solved the problem, the calculator did it.
qed
Your brain also is physical. Electrochemical gradients flow between physical molecular constructs. Isn't it just chemistry? Do you attribute it to physics or some whole-is-greater-than-the-parts idea?
> Do you attribute the build to the tool? The "system's contribution" is helped by many other things all the way down to chips, datacenters and power generation. If the authorship requires attributing to a tool, then it should happen all the way down.
When the tool is a 3D printer, or any CNC system really, you bet I attribute a build to it.
I could also attribute the operator; there is no contradiction, it's a free choice, just like saying "I am in Berlin" does not contradict "I am in Germany".
> AI has no self-awareness
What is your mechanistic model of self awareness that yields this conclusion?
> It's a tool
Does your model suggest that tools can't have self awareness?
Dunno about the parent commenter, but I personally interpret the concept as having a hidden representation of self that is continually tended to, and influences future choices. This implies statefulness, which models are intentionally not at inference time (*).
(*) Even if we hack around this and just do the usual trick of simply laundering statefulness to a higher level, in this case the context window being fed in, I fail to identify (**) a representation of its own state in these bodies of text that it'd be meticulously maintaining. I further fail to identify how it could be hidden or maintained, considering I control like half of it. The best you could ascribe it is a meticulous maintenance of a persona the user is talking to, but then that doesn't necessarily represent the model's internal state, the same way my own words here aren't doing so either. Difference being, I actually have one (I'm "on-line").
You'll sometimes catch models mixing up who's who and how many who-s there even are for example.
(**) I did wish for something hidden though, so maybe it's just concealed? The same way people can encode a lot more of their emotional and mental state than normal into text if they read and write a lot of it, I'm aware of research that suggested the same for LLMs, albeit I cannot cite it. Maybe those phrasing signatures are just alien to me and will never pop out. Either way, I'd expect researchers to stumble upon this during interpretability studies, and either they haven't, they have but it wasn't popsci adopted, or they're keeping awfully tight lipped about it. If you know of anything like this, your turn now, would be happy to learn.
I do wonder how reasonable it is to expect e.g. a single maintained identity though. Maybe it isn't?
(*) Another way to hack around this of course is to just precompute some internal "self-awareness states" and hop around between them. Probably the closest to what the models are actually "doing".
Before reading, know that I am uncertain in either direction.
> a hidden representation of self that is continually tended to
This sounds like a personality? They act like they have one of those. It may be an illusion, and even if it isn't an illusion it is unlikely to be anything like the source (us), but they act like it.
> I further fail to identify how it could be hidden or maintained, considering I control like half of it.
Indeed you control everything about a local model, and much of the context of even a remote model. But the state of activations and circuits in SotA AI is hidden in similar ways to those of synapses in your head: difficult to decipher even with probes monitoring the signals directly, and often not emitted at the normal output.
> The best you could ascribe it is a meticulous maintenance of a persona the user is talking to, but then that doesn't necessarily represent the model's internal state, the same way my own words here aren't doing so either. Difference being, I actually have one (I'm "on-line").
While we can be confident that LLMs make up personas etc., it is insufficient to go from "that doesn't necessarily represent the model's internal state" to "therefore it doesn't have one".
> You'll sometimes catch models mixing up who's who and how many who-s there even are for example.
I've, unfortunately, also experienced this with humans. Perhaps they were losing their self-awareness at the time? I do wonder if old-age dementia does that by the end, though the person in question didn't ever get diagnosed with that.
> If you know of anything like this, your turn now, would be happy to learn.
Do you mean like these, or something else?
• https://researchportal.hkust.edu.hk/en/publications/decoding...
• https://aclanthology.org/2026.eacl-long.165/
• https://transformer-circuits.pub/2026/emotions/index.html
> This sounds like a personality?
Not quite what I meant, but it's also not entirely unrelated I guess? Personality to me is like a natural bias. It does also shift over time, and is also an internal bit of state. I guess in some respects it can also be self-referential, like personal convictions.
> Perhaps they were losing their self-awareness at the time?
I do think it is entirely possible for people's self-awareness to shift, yes. Or more precisely, I do model things that way.
> Do you mean like these, or something else?
They're adjacent, but I more meant something like these:
https://arxiv.org/abs/2410.03768
https://arxiv.org/abs/2310.18512
https://arxiv.org/abs/2605.26537
So basically, steganography. The difference is that these papers investigate from the perspective of separate LLM instances covertly exchanging information between each other. This is in contrast with the scenario I'm laying out, where an LLM's past state is exchanging information with its future state, continuously representing and modulating a concealed internal state of some sort. And then that state just so happening to be some sort of self-referential meta state.
And the best inkling I have towards this is basically: https://www.youtube.com/shorts/WP5_XJY_P0Q
But then I don't think there's enough covert channel bandwidth in the agent replies for anything interesting like this.
> Dunno about the parent commenter, but I personally interpret the concept as having a hidden representation of self that is continually tended to
I don't see why an LLM could not have a sense of identity or personality while it's evaluating a specific prompt, or even change self awareness while evaluating a prompt since many outputs model a back and forth conversation. My point is that without a mechanistic model of what "self awareness" means, we have no way of truly evaluating such questions, we're just hand waving vague intuitions about what it could mean.
Sure, but then such a model is not going to make itself. People pitting their vague intuitions is how such models eventually form. I'd also push back regarding that my comment would have been handwavey or without mechanistic elements, even if it was on the whole informal.
This is kind of also the reason e.g. the HN site guidelines are worded the way they are. Regrettably, forums naturally yield themselves to tit for tat type exchanges, but there's really no reason one could not bounce such vague intuitions off of another. I do not have to be right or wrong, and you don't either. Admittedly difficult when its some intensely contentious topic.
If a mechanistic model existed, there would also be no reason to talk about this in the first place. There'd be nothing to discuss, you'd be simply told how a given model characterizes from this perspective on the model cards.
Even mechanistic models generate interesting discussion. How many years have we discussed Turing machines and the lambda calculus? Almost a century of great work came out of those.
The reason I insist on mechanistic models is because the original post was making a definitive knowledge claim, and in my experience, the knowledge claim is unwarranted.
I honestly don't think a language model is enough for self-awareness, regardless of the exact model of awareness.
A language model (or an image model or whatever) cannot even be sentient, and I think sentience is a prerequisite for awareness.
Even if we express a lot of our subjective experience with words, the language is just a symbolic representation of those experiences. The qualia themselves, even those that are quite abstract, are rooted in our physical presence and evolution.
You can't have an understanding of what hunger or physical pain feel like if you have no need for food or a sensory capacity for feeling pain. You can't understand what loneliness or pride at an achievement mean if you don't have a neural network wired to value social connection or status. We value connection because we're social animals that have needed each other for survival.
Even the more abstract of our subjective experiences are in some way rooted in our physical evolution.
I see no reason to believe that a neural network built entirely based on the symbolic level of language could have the features needed for the subjective experience itself.
AI awareness might actually be more believable if that awareness manifested itself in an entirely different way than in humans. But if we assume awareness because outputs resemble what we consider meaningful as humans, yet the neural network has had no inputs or evolution that could form the actual basis of human-like experience, I think we're seeing something that isn't actually there.
> The qualia themselves, even those that are quite abstract, are rooted in our physical presence and evolution.
There is no objective evidence of qualia. All evidence of qualia are vocal or other expressions of belief in qualia. Perceptions clearly exist and are observable, subjective experience and qualia, not so much.
> I see no reason to believe that a neural network built entirely based on the symbolic level of language could have the features needed for the subjective experience itself.
If your objection is to models based on "symbolic level of language" which you think lack semantic understanding of, say, trees, you should ask yourself how our brain, based on physics which also lacks any semantic category for trees, can somehow develop a semantic understanding of trees. All of these appeals to differences with the brain never seem to acknowledge that fundamentally, the brain has the same explanatory gap with physics.
> But if we assume awareness because outputs resemble what we consider meaningful as humans, yet the neural network has had no inputs or evolution that could form the actual basis of human-like experience
This assumes a lot. It seems very possible to me that intelligence inherently develops a map of natural categories (natural kinds), and language naturally develops around such categorical understanding. Semantics are then fundamentally the network of associations between categories, eg. there is no fundamental difference between symbols and semantics, and the latter cam be inferred from the former, and that's exactly what LLMs do, and why the semantic maps between different languages are so similar and how they can translate between languages.
So… your model is 100% vibes based. Got it.
I wasn't trying to give a model. The point was that I don't think it's necessary to give one.
You didn't address any of what I wrote, let alone provide any counterarguments. Which part of what I wrote do you think was wrong?
Making definitive claims about whether LLMs do or do not have specific properties absolutely does require precise definitions of those properties that can be used to evaluate those questions. Merely hand waving that LLMs didn't undergo the same evolutionary process is not a definitive argument.
For example, the Turing machines and the lambda calculus don't look anything alike, but they are fundamentally interconvertible, and so in a real sense they are fundamentally equivalent. Without a model, all of your arguments are completely unconvincing for exactly the same reasons, eg. that there may exist many paths to fundamentally equivalent ends.
I just don't think linguistic (or other symbolic) representations alone can contain the information, in any sense of the word, of what e.g. human subjective experiences actually are like. The concepts we express with language get their meaning from our physical reality, even if quite indirectly in case of some abstract concepts.
Hunger as a concept doesn't mean anything without the physical need. Politeness or bluntness, even in writing, don't mean anything without social dynamics. And we have social dynamics (and neural structures that directly process social cues and associated feelings) because we've evolved into social animals for whose survival that was important.
I see no reason to believe that a model trained only with symbolic representations, with no connection to the physical world phenomena that those symbols represent, could contain the subjective experience itself.
Neural network models may be able to derive novel (or at least novel-looking) output rather than just an obvious rehash of their input, but I don't think any set of bytes can fundamentally contain information that was never entered into it. (Even if e.g. a model produces previously unknown mathematical results, those results can in principle be derived from the information that they were trained with.)
I'm not saying that artificial neural networks couldn't, in principle, be aware. ANNs and biological neural nets may be equivalent in the sense that any information and processing structures represented by a biological one could in principle be represented by an artificial one. If that's the case, and awareness is purely a product of our neural systems as materialism would imply, it should be possible for an ANN to be aware, too.
But when the model has been trained with only language, and IMO the subjective experience can't be derived from the symbolic representation alone, I can't see how the model could include the actual subjective human experience.
An AI model could of course have an awareness and subjective experiences that are totally different than our human experience. But then the fact that it happens to produce output resembling what humans find meaningful shouldn't be considered indicative of such awareness.
This is of course more of a philosophical argument than a technical one, and I'm happy to hear counterarguments, but not on the level of off-hand dismissal.
> I can't see how the model could include the actual subjective human experience.
People who say LLMs have subjective experience aren't saying they have human-type subjective experience. Nobody who sees an LLM express hunger when role playing as a hungry person thinks that the LLM is actually hungry.
I too can role play as a hungry person despite not being hungry, so there is no reason in either case to conclude that the words produced reflect genuine internal subjective states. The point is that such internal states may still exist.
A better analogy would be a manufactured object, say 3d printed for simplicity. The 3d printer is given an input, and an object manifests itself after some time. We say that the creator of the object is the person turning on the machine, sending the data, and collecting the object. Not the machine itself.
I'd say the creator is the one who created the 3d model, not the one who pushed the print button.
I would say „I made this gadget with my 3d printer, but the designer is someone else (I found the model online)“. The intent, the drive, the action comes from the human
"I made this proof myself, but the designer is someone else" is an extremely unconvincing claim to ownership.
Almost as if a proof isn’t the same as a 3d print. It’s just not a good analogy
(let's assume that)My 3dprinter is special. It has a bunch of values + an algorithm (i.e. a neural network) that takes input as tokens and outputs a printed object.
Not peer reviewed, Lean proofs are 100,000 lines long and Lean has bugs:
https://cr.yp.to/proofs.html
Who is going to wade through this?
They've been hiring mathematicians to verify this stuff themselves. They're obviously not just throwing it out there without any human review.
And these mathematicians sink the comment to the bottom in 5 min?
It is not peer review if it is all in one company that wants an IPO.
I'm sure OpenAI is just interested in the greater good of mankind!
How do we know that these solutions don't exist in the training data? It is open secret that they have used pirated materials for training. Perhaps it plagiarized solutions from works of some obscure Belgian mathematician from the sixties, who did not get mainstream acceptance. I wouldn't be surprised if they also got access to mathematics done in the "defense contractor" setting from various three letter agencies.
Without a searchable index of training data, it is hard to put faith into these claims.
This wouldn't be a problem so long as they properly attribute.