Norvig (and Chomsky) are talking at cross purposes.
Norvig confuses engineering for science. If an engineer wants something that works (and guarantees that it will work) then you use the most effective tool available to achieve that goal. No problem there. But there's absolutely no reason to think that tells you anything scientific about the phenomenon you're modeling. For example, aeronautical engineering says very little about how birds fly. And - currently relevant - LLMs learn to use and understand language, but they use millions of times more training data, and millions of times more computing power, than a child does. So an LLM may be a great technology, but it's probably not a good model for how humans learn and use language.
Chomsky OTOH doesn't see any use for statistics in science, at least not in linguistics. Fortunately for the field of linguistics, Chomsky's influence in linguistics is waning and linguists are beginning to use statistical tools to do science.
i agree. but it's uncanny that stochastic generation of language yields a form of recognizable intelligence.
in college, i was somewhat involved in cogsci wrt semantic mapping of linguistic concepts based on human surveys (this was well before LLM/AI-word2vec), and it surprises me today how the mathematical properties of relatedness/distance seem now to drive machine intelligence, something which we did not consider at the time at all.
maybe i'm getting in my metaphysical sentimental stage, a turn a lot of adults make after middle age, but i wouldn't fully dismiss the fact that there may be an equivalence.
Indeed. Who'd have thought that mere statistics could do that? Hat's off to them - they made the darn thing work.
There's pretty interesting (I think) research in how effectively it's possible to scale down LLMs. If you're interested, take a look at the BabyLM conferences:
(I have no connection to them.) The work there demonstrates that gap between what we can do now with human-relevant amounts of training text, and what remains to be discovered. Quite a bit, it looks like.
Humans decidedly take much more data than LLMs. Like orders of magnitude. We happen to feed only distilled data to LLMs though. That's why they appear so efficient.
LLMs behave like they're closest, but only during execution. During training, obviously, they don't resemble the human brain at all. No human reads a million books before they begin to speak.
And I'd be hard pressed to identify many cognitive characteristics of LLMs during execution that resemble human thinking, aside from those that are readily parrotable.
Care to back this up? What evidence do we have to suggest that back propagation, self attention, and byte pair encoding are mechanistically similar to human cognition? Similarity of the outputs says nothing about similarity of the machinery underneath. Turboprop engines tell us nothing about how birds fly - doesn't matter that they successfully transport millions of people around the world.
I would say that when it comes to higher level cognition, logic engines are more human like even if they only have a tiny fraction of the practical utility of LLMs.
> Turboprop engines tell us nothing about how birds fly
That is not true.
Turboprop engines and flying birds are different things, but there is a set of common properties.
A blade of a screw propeller functions exactly in the same way as a flapping wing of a bird, except that the wing cannot rotate without limits around the body of the bird, so it must stop, change angles and revert the motion, while the blade can rotate continuously around the axis, functioning on one side like a falling wing and on the other side like a raising wing. The similarities between propeller blades and insect wings are even greater, because insect wings cannot change their shapes so much as those of birds.
For now, our understanding of the networks composed of the neurons of animals is much less advanced than our understanding of wings and propellers, which is why there is a much greater uncertainty about whether there are any similar mechanisms in biological neural networks and in LLMs.
Also, AFAIK, insects rely on their (flat) flapping wings to create lift by creating vortices.
"Nothing" may be a little too strong -- physics is physics, after all -- but (small) birds and airplanes have Reynolds numbers about 1000x different and different underlying material science.
Airplanes are engineered (with a very small number of exceptions) to be stable during flight, whereas bird flight is inherently unstable, requiring a lot of specialized cognition so birds can fly.
I would think a lot of cognitive ornithologists (who I really hope exist!) would be interested in finding out about this.
At this point, I'm at the complete end of my knowledge of bird flight!
We already know that the most basic mechanisms in ANN and actual neurons are very different. Spiking and neurotransmitters are basically everything in brains and don't exist in ANNs.
What reason to we have to believe that LLMs are the closest model we have for how humans learn language.
We considered the hypothesis that humans are a general intelligence that figures language out by exposure. The evidence has consistently come back that that is not how humans actually work. A model where humans are hardwired with a lot of knowledge about language, then fill in the gaps based on exposure still appears to be far more accurate.
The only thing LLMs show is that it is possible to learn language de-novo.
> But the vast majority of people who study interpretation tasks (such as speech recognition) quickly see that interpretation is an inherently probabilistic problem: given a stream of noisy input to my ears, what did the speaker most likely mean? Einstein said to make everything as simple as possible, but no simpler. Many phenomena in science are stochastic, and the simplest model of them is a probabilistic model
Why the "simplest" model? Is a probabilistic model not the only model able to deal with uncertainty? What other kind of models are possible to represent uncertainty and inherent ambiguity that are not probabilistic?
I don't think it's the simplest, actually. Off the top of my mind:
Ternary logic is a super simple model; where true AND unknown equals unknown, and true OR unknown equals true. This deals with uncertainty and ambiguity, just not the probability; notionally you could say everything is 50-50.
Same thing if you have e.g. text-to-speech which returns "the b??? dog" rather than "the (0.8:black, 0.1:blue, 0.1:boy) dog" - no probabilities. Of course, maybe later in the processing you add probabilities back in.
Then there's fuzzy logic, which uses scores that are typically numbers, but those don't express actual probabilities, but more like vibes. They don't necessarily add up to 1, for instance, and you don't apply Bayes' formula to them, or even multiply them, you can use addition too - yolo.
Edit: and also of course backtracking, beam search, that sort of stuff - needn't use probabilities, just heuristics.
Old school search methods allow you to pursue multiple paths, but not necessarily based on probability. At the end (or while pruning) you can use "word knowledge".
Anyway, the probabilistic model doesn't exist. Simple probabilistic models work quite badly on speech recognition (word classification, etc.). A lot of structure needs to be added before they start performing.
My objection is "simpler." What's so "simple" about a probabilistic model? It takes an enormous amount of calculations, so much that you have to predigest incoming information into high-level chunks in order to have any chance of calculating it over small inputs with reasonable effort.
A probabilistic model is the most complicated model. It doesn't even attempt to model the problem, it just looks at the results of an enormous number of tries with different inputs, goes through them all, finds the most similar cases to the current input, and blends them together in a way that carefully avoids any self-contradictory or absurd result.
We see this plainly with LLMs. You train an LLM on a thousand times the (high-level, pre-digested) input that the average human has gotten, and the average human thinks it's a hallucinating moron. It doesn't start to converge on something that passes the smell test to you average human idiot until it has been trained on a million times the input the average human has gotten.
It's not doing what a human does, it's guessing what a human (or any system) might do when confronted with X input. They're not even similar activities, even if they have identical output in the limit. It's like confusing gravity for equations about gravity. Being able to make perfect predictions about the positions of moving, nearby frictionless spherical things in a vacuum at time t doesn't tell us a thing about what gravity is. It tells us how it looks. It should, we distilled it by watching things for a very long time.
> Is a probabilistic model not the only model able to deal with uncertainty?
No, all models deal with uncertainty. The entire point of a model is that it is what you know about a system when you don't know its inputs. The variables in a model are uncertainty. The model is dealing with it.
edit: I think you're interpreting "uncertainty" as a numerical probability. A numerical probability is an artifact of the set of data you're calculating it from, not a thing with independent existence. "Uncertainty" is just a thing that you don't know yet.
If you assume that "uncertainty" is a knowable numeric probability that contributes to some calculation, you're begging the question when you say that the only way you can model it is through probabilistic calculations.
It's so clear that Norvig knows so little about linguistics as a subject, I mean the suggestion that English is "kinda" a pro-drop language like Spanish is just unserious. There's this kind of lazy criticism of generative work where people hear the vague idea of a theory and critique without even seriously reading anything about it. These are facts that generativists are perfectly aware of and have explanations for.
It's sort of like criticising Galileo's putative experiment about dropping two balls from the leaning tower of Pisa by pointing out that a feather indeed falls slower than a iron ball.
It was a terribly misjudged relationship - though Chomsky was in his late 80s and past it by then - but there is not one iota of evidence that Chomsky "raped kids", and I think that highly implausible. Why make such a childish, unfounded slur?
I haven't seen the evidence for "raped kids", but it's certainly true that Chomsky advised and benefitted from Epstein long after any sensible person would have realized that Epstein JUST MIGHT be a bad person.
Unfortunately, Chomsky's always had a problem pulling the ripcord in similar situations. As one of his colleagues once explained, "When most people step in dogshit on the sidewalk, they scrape it off their shoe and keep walking. Chomsky gets stubborn and just keeps stepping in it again and again."
Not that this is relevant to his intellectual contributions, unless it crosses your personal line where involvement becomes complicity.
Norvig (and Chomsky) are talking at cross purposes.
Norvig confuses engineering for science. If an engineer wants something that works (and guarantees that it will work) then you use the most effective tool available to achieve that goal. No problem there. But there's absolutely no reason to think that tells you anything scientific about the phenomenon you're modeling. For example, aeronautical engineering says very little about how birds fly. And - currently relevant - LLMs learn to use and understand language, but they use millions of times more training data, and millions of times more computing power, than a child does. So an LLM may be a great technology, but it's probably not a good model for how humans learn and use language.
Chomsky OTOH doesn't see any use for statistics in science, at least not in linguistics. Fortunately for the field of linguistics, Chomsky's influence in linguistics is waning and linguists are beginning to use statistical tools to do science.
i agree. but it's uncanny that stochastic generation of language yields a form of recognizable intelligence.
in college, i was somewhat involved in cogsci wrt semantic mapping of linguistic concepts based on human surveys (this was well before LLM/AI-word2vec), and it surprises me today how the mathematical properties of relatedness/distance seem now to drive machine intelligence, something which we did not consider at the time at all.
maybe i'm getting in my metaphysical sentimental stage, a turn a lot of adults make after middle age, but i wouldn't fully dismiss the fact that there may be an equivalence.
Indeed. Who'd have thought that mere statistics could do that? Hat's off to them - they made the darn thing work.
There's pretty interesting (I think) research in how effectively it's possible to scale down LLMs. If you're interested, take a look at the BabyLM conferences:
https://babylm.github.io/
(I have no connection to them.) The work there demonstrates that gap between what we can do now with human-relevant amounts of training text, and what remains to be discovered. Quite a bit, it looks like.
Humans decidedly take much more data than LLMs. Like orders of magnitude. We happen to feed only distilled data to LLMs though. That's why they appear so efficient.
> LLM may be a great technology, but it's probably not a good model for how humans learn and use language.
but from all models, it's the closest, "all models are wrong, but some are useful"
LLMs behave like they're closest, but only during execution. During training, obviously, they don't resemble the human brain at all. No human reads a million books before they begin to speak.
And I'd be hard pressed to identify many cognitive characteristics of LLMs during execution that resemble human thinking, aside from those that are readily parrotable.
As a model of human cognition, LLMs deep suck.
> but from all models, it's the closest
Care to back this up? What evidence do we have to suggest that back propagation, self attention, and byte pair encoding are mechanistically similar to human cognition? Similarity of the outputs says nothing about similarity of the machinery underneath. Turboprop engines tell us nothing about how birds fly - doesn't matter that they successfully transport millions of people around the world.
I would say that when it comes to higher level cognition, logic engines are more human like even if they only have a tiny fraction of the practical utility of LLMs.
> Turboprop engines tell us nothing about how birds fly
That is not true.
Turboprop engines and flying birds are different things, but there is a set of common properties.
A blade of a screw propeller functions exactly in the same way as a flapping wing of a bird, except that the wing cannot rotate without limits around the body of the bird, so it must stop, change angles and revert the motion, while the blade can rotate continuously around the axis, functioning on one side like a falling wing and on the other side like a raising wing. The similarities between propeller blades and insect wings are even greater, because insect wings cannot change their shapes so much as those of birds.
For now, our understanding of the networks composed of the neurons of animals is much less advanced than our understanding of wings and propellers, which is why there is a much greater uncertainty about whether there are any similar mechanisms in biological neural networks and in LLMs.
Also, AFAIK, insects rely on their (flat) flapping wings to create lift by creating vortices.
"Nothing" may be a little too strong -- physics is physics, after all -- but (small) birds and airplanes have Reynolds numbers about 1000x different and different underlying material science.
Airplanes are engineered (with a very small number of exceptions) to be stable during flight, whereas bird flight is inherently unstable, requiring a lot of specialized cognition so birds can fly.
I would think a lot of cognitive ornithologists (who I really hope exist!) would be interested in finding out about this.
At this point, I'm at the complete end of my knowledge of bird flight!
We already know that the most basic mechanisms in ANN and actual neurons are very different. Spiking and neurotransmitters are basically everything in brains and don't exist in ANNs.
What reason to we have to believe that LLMs are the closest model we have for how humans learn language.
We considered the hypothesis that humans are a general intelligence that figures language out by exposure. The evidence has consistently come back that that is not how humans actually work. A model where humans are hardwired with a lot of knowledge about language, then fill in the gaps based on exposure still appears to be far more accurate.
The only thing LLMs show is that it is possible to learn language de-novo.
> The only thing LLMs show is that it is possible to learn language de-novo.
I agree, but would add, "But not under conditions that humans learn language."
> But the vast majority of people who study interpretation tasks (such as speech recognition) quickly see that interpretation is an inherently probabilistic problem: given a stream of noisy input to my ears, what did the speaker most likely mean? Einstein said to make everything as simple as possible, but no simpler. Many phenomena in science are stochastic, and the simplest model of them is a probabilistic model
Why the "simplest" model? Is a probabilistic model not the only model able to deal with uncertainty? What other kind of models are possible to represent uncertainty and inherent ambiguity that are not probabilistic?
I don't think it's the simplest, actually. Off the top of my mind:
Ternary logic is a super simple model; where true AND unknown equals unknown, and true OR unknown equals true. This deals with uncertainty and ambiguity, just not the probability; notionally you could say everything is 50-50.
Same thing if you have e.g. text-to-speech which returns "the b??? dog" rather than "the (0.8:black, 0.1:blue, 0.1:boy) dog" - no probabilities. Of course, maybe later in the processing you add probabilities back in.
Then there's fuzzy logic, which uses scores that are typically numbers, but those don't express actual probabilities, but more like vibes. They don't necessarily add up to 1, for instance, and you don't apply Bayes' formula to them, or even multiply them, you can use addition too - yolo.
Edit: and also of course backtracking, beam search, that sort of stuff - needn't use probabilities, just heuristics.
Old school search methods allow you to pursue multiple paths, but not necessarily based on probability. At the end (or while pruning) you can use "word knowledge".
Anyway, the probabilistic model doesn't exist. Simple probabilistic models work quite badly on speech recognition (word classification, etc.). A lot of structure needs to be added before they start performing.
My objection is "simpler." What's so "simple" about a probabilistic model? It takes an enormous amount of calculations, so much that you have to predigest incoming information into high-level chunks in order to have any chance of calculating it over small inputs with reasonable effort.
A probabilistic model is the most complicated model. It doesn't even attempt to model the problem, it just looks at the results of an enormous number of tries with different inputs, goes through them all, finds the most similar cases to the current input, and blends them together in a way that carefully avoids any self-contradictory or absurd result.
We see this plainly with LLMs. You train an LLM on a thousand times the (high-level, pre-digested) input that the average human has gotten, and the average human thinks it's a hallucinating moron. It doesn't start to converge on something that passes the smell test to you average human idiot until it has been trained on a million times the input the average human has gotten.
It's not doing what a human does, it's guessing what a human (or any system) might do when confronted with X input. They're not even similar activities, even if they have identical output in the limit. It's like confusing gravity for equations about gravity. Being able to make perfect predictions about the positions of moving, nearby frictionless spherical things in a vacuum at time t doesn't tell us a thing about what gravity is. It tells us how it looks. It should, we distilled it by watching things for a very long time.
> Is a probabilistic model not the only model able to deal with uncertainty?
No, all models deal with uncertainty. The entire point of a model is that it is what you know about a system when you don't know its inputs. The variables in a model are uncertainty. The model is dealing with it.
edit: I think you're interpreting "uncertainty" as a numerical probability. A numerical probability is an artifact of the set of data you're calculating it from, not a thing with independent existence. "Uncertainty" is just a thing that you don't know yet.
If you assume that "uncertainty" is a knowable numeric probability that contributes to some calculation, you're begging the question when you say that the only way you can model it is through probabilistic calculations.
This is an old, old post by Norvig fyi.
It's so clear that Norvig knows so little about linguistics as a subject, I mean the suggestion that English is "kinda" a pro-drop language like Spanish is just unserious. There's this kind of lazy criticism of generative work where people hear the vague idea of a theory and critique without even seriously reading anything about it. These are facts that generativists are perfectly aware of and have explanations for.
It's sort of like criticising Galileo's putative experiment about dropping two balls from the leaning tower of Pisa by pointing out that a feather indeed falls slower than a iron ball.
Really great post, though a bit old, still relevant. Peter Norvig is one of the better thinkers of our generation.
Have to say that I confronted myself with the fact that I can't read longer posts anymore due to overuse of AI. Guess a detox is required...
chomsky was also a close friend of epstein and raped kids
It was a terribly misjudged relationship - though Chomsky was in his late 80s and past it by then - but there is not one iota of evidence that Chomsky "raped kids", and I think that highly implausible. Why make such a childish, unfounded slur?
There is absolutely no evidence or even any allegations whatsoever that Chomsky "raped kids".
I haven't seen the evidence for "raped kids", but it's certainly true that Chomsky advised and benefitted from Epstein long after any sensible person would have realized that Epstein JUST MIGHT be a bad person.
Unfortunately, Chomsky's always had a problem pulling the ripcord in similar situations. As one of his colleagues once explained, "When most people step in dogshit on the sidewalk, they scrape it off their shoe and keep walking. Chomsky gets stubborn and just keeps stepping in it again and again."
Not that this is relevant to his intellectual contributions, unless it crosses your personal line where involvement becomes complicity.
His Wikipedia article [0] gives a good summary of the connection they had.
[0] https://en.wikipedia.org/wiki/Noam_Chomsky#Friendship_with_J...