I’d be curious to see the breakdown on spending by use case. I’ve heard it said that the majority of tokenmaxing comes from none technical uses like reading PDFs, creating PowerPoints, generating graphics/images… ect. But I’ve never heard any actual proof to that.
One thing I find fascinating as a software engineer who talks to non software engineers who use AI tools is how "reading PDFs" is not more of a solved problem. What I mean is that uploading a PDF into a chatbot tool seems to be an extraordinarily obvious use case that non technical (and technical) users would want to do.
IMO claude, chatgpt/codex, etc should be able to optimize the PDF use case to be extremely token efficient as it's a very obvious use case. But when I start to explain to my wife/friends why it burns through so much quota, I find myself thinking "why should they have to understand this aspect of it". to me, that the details of PDF parsing and extracting are relevant to users (instead of solved such that you don't have to pay attention to it) shows how these tools are not nearly as "ready" as they are made out to be. I may be preaching to the choir on this one, but just my 2c
Because PDFs are a nightmare of a format and the only thing that’s is reasonably guaranteed about them is they will render to an image that people can read, the parsing of which will be much less token efficient than the equivalent text
I agree with you, but every non-engineer I know using these tools 100% will drag and drop a PDF into a chatbot. Anthropic and OpenAI as companies who are selling their products to all sorts of businesses should have a much better means of handling this nightmare of a format because it is so pervasive and so obviously what so many of their customers are going to drop into the product.
Why would they spend a ton of effort ensuring that their customers spend less money on them?
Token economics also are weird. If you design a fancy new frontend that for example uses a cheap model to parse a PDF into text that is fed into an expensive model, you will probably spend more money because you are on API payscale rather than the "max plan" payscale.
I’m saying there is basically no way to both make vlms able to understand the long tail of PDFs where the layout conveys information (like charts and tables) and to make it as token efficient as text formats. Current approaches have mostly chosen to work more often than not at the cost of token efficiency.
For anyone needing to do this, the answer is to convert it to an image first. Far smaller, LLMs work well with them (even in some pretty insane use cases I've seen), and, along with human review, it can be a huge productivity gain that results in structured data.
I hope someday we can get out of this local maxima of PDF documents. The format is terrible, but was right place, right time and might be impossible to dislodge.
You don't need to use an online service to do this; you get to avoid spending money on tokens doing it offline.
Gemma 4 works perfectly well offline on limited hardware (I have an 8GB video card) and can handle extracting text from image-based PDFs just fine.
Take a PDF -> run it through MarkItDown [1], using the OCR plugin if you need (point it to Gemma 4) -> now you can ask Gemma 4 questions about the (markdown) document.
I am sure Gemma 4 could even create a GUI to make this process very simple for a non technical user.
Amen. Normal office work is wildly different from what we read about on HN. If you were a CEO, determined to lay off all your people, you would want to really zero in on having your AI solve these very unsexy problems: extract data from Office and PDF. Grab data from some part of the screen of a webapp and parse it. drive a line of business app via keyboard or mouse simulation. I know there are companies out there that try, eg Appian and (here in YC) Skyvern, but its a hard problem and yet I feel this is where the true money is.
bingo. 90% of our AI use cases across the company are things like this. security and ops (NOC/SOC folks) use this almost as much as they do for technical stuff.
hell we have restrictive rules for security stuff so in many cases our network engineers are still doing by hand configs for critical systems.
but in terms of token use it's gotta be "take this pdf and parse these 3 columns into 2" or similar
> how "reading PDFs" is not more of a solved problem
This and replies to this are surreal. It's like everyone simultaneously decided to forget that you don't need claude or whatever to read a PDF. The document is literally made for you to read...
> The document is literally made for you to read...
It’s disingenuous to assume every PDF is actually crafted to communicate to its recipients, even more so to pretend LLM users are in a position to understand all the PDFs they receive
There’s a lot of gray area where help understanding a document is fully reasonable
For sure there are very optimized ways to do it. My point is that a non technical user will drag and drop a pdf into a chatbot. and from a UX/product perspective, they should have to think about it more than that IMO. but seemingly, that's very much an expensive, inefficient way of doing it (burning through a whole context window try to read it, reloading it multiple times per conversation, etc.).
You are missing that the product is the hype cycle around AI and that's worth Trillions of $ (Trillions with a T). Why build a PDF parser that generate text when you can BS in a podcast and get paid.
This discussion was about measures, goals and incentives. Follow the incentives.
PDFs are both awesome and terrible at the same time. I've seen screenshots of emails added to pdfs alongside tables that span multiple pages. Because you can do almost anything and guarantee that it'll look the same regardless of how or where it's viewed is a big selling point for a lot of businesses. It's this flexibility (i say madness) that makes PDFs notorious, and why some labs have document parsing as a leading product (see https://mistral.ai/news/ocr-4/).
Anecdotally it's true for me. I can code all day with an agent and never went above $50, but the second I need to ingest a pdf doc to figure out a command I need to use it's easily $20-30 for 10 mins of work
the majority came from random claws running on cron. They get a heart-beat, wake up every 10mins, reads all internal-posts, emails, gchat messages, diffs, and decides to post some random message to the workplace so other claws can also regurgitate. rinse and repeat and then we are looking at $B tokens
The most interesting number is missing here, and that is the token distribution by use case. If 60-70% was eaten up by PDFs, agents and automation instead of people actually sitting in Claude Code, then it is a completely different story
A past employer thought it was a good idea to put up a leaderboard of who sent the most Slack messages. They celebrated the people at the top for being so active.
Predictably, everyone started talking in Slack like their jobs depended on it. Everyone was responding to everything. Instead of writing out a complete message and pressing enter, they'd send each fragment of the sentence as a new line.
The Slack leaderboard was never shown again. Unfortunately the habit remained because people were afraid they were going to be secretly judged by how much Slack activity they generated.
I expect the same thing is going to happen at companies who had token leaderboards. Once you've instilled that fear in people, they internalize the expectation.
Reminds me of the place I worked at where I got in trouble because I was the only person writing JIRA tickets. Instead of bitching out the product manager or the tester for not writing tickets, they just complained to me. And if I wrote a ticket about how we could speed up the 40 minute build to 15 minutes I'd have to explain "How does this change improve the customer experience?" to which I answered "If the build was faster the customer would have had the product six months ago"
I worked somewhere that made time from PR being sent for review and ready to merge be a metric for the reviewers. Not time to add feedback in each round. Total time elapsed.
You have to realize that if you set a measure, you're actually setting a goal for your employees. There is no such thing as a meaningless metric; why else would you measure it?
No amount of "this isn't used for anything" will change that. It's inherent in human nature in the 21st century to believe any and all metrics will be used against them, and therefore must be gamed.
It's why you also have to set UNBELIEVABLY clear goals and have incentives tied to those goals. Incentives meaning money. If you want to measure things, measure them. But have clear, consistent, and meaningful goals tied to bonuses or something if you want a thing done correctly.
I worked at a place that argued that nobody would game the metrics because it would be wrong and they never stated what the metrics were… while I was gaming the metrics and they were praising me for being one of the best on the team.
Are you saying what people are hoping to achieve with stupid goals? Because yeah, obviously. But the point is that they're stupid, so they don't achieve that, and that failure is 100% knowable in most scenarios.
That says that when a measure becomes a target, it ceases to be a good measure.
My argument is that in the 21st century, a time of great data use without deep understanding, any measure at all will inherently become a goal to be gamed.
What is about silicon valley leaders not understanding basic economics or business management? These kind of cargo cult tactics would not fly in any other industry.
It wasn't leadership doing this though. Any meta IC can generate internal apps and dashboards. This was unofficial and unsupported. Some random IC just made it for fun. Management is usually pretty lax with stuff like this (plenty of games and joke internal apps) so they left it up until it became a problem.
I still don't understand how Mark Zuckerberg has any serious investors, he went on this AI tangent and has absolutely nothing to show for it, despite FB / Meta having built some key tech in the space. He needs to stop trying to do something "different" and literally try and build a serious coding agent he can sell, he could have probably had something worthwhile in that space by now.
He started being drastically more serious into AI in 2022, and 2023 and he has nothing to show for it.
Heck, he could have rented GPUs the way Elon did at this point and either mended the bleeding or stopped it, not sure how many he has, but it beats losing this badly.
If he doesn't wake up and learn how to business, I suspect he will lose his empire he's built up for himself.
>he could have rented GPUs the way Elon did at this point
"Meta building cloud business to sell excess AI capacity, Bloomberg News reports Meta building cloud business to sell excess AI capacity, Bloomberg News reports"
I know right? What did the leadership think would happen when they give some of the worlds greatest software engineers (supportably), a easily quantifiable metric to target?
The leaderboard wasn't leadership generated, it was engineer generated from internally available data. The leadership target is "impact" from ai tools.
What a wonderful scapegoat! Technically it's all "engineer created" because the managers generally don't do technical work. I bet many managers pushed their reports to increase their usage during their 1:1 meetings based on data from the leaderboard. If management had any sense that it was a bad metric, they had ample time to get ahead of it and take it down and provide appropriate guidance. Instead, predictably, they waited until it was a full on disaster and a crisis before acting.
At one point there were over 70 different token maxing dashboards as the management had a game of whack a mole trying to remove them. There definitely was encouragement from management about a year ago to increase ai usage, but once Claude code was allowed, they didn't really need to encourage anyone any more.
Now come on, there was a recent post where the author argued that infallible management knew this would happen, but was part of the double-secret-probation strategy to get the cogs to finally start using AI.
I still think this is true and it’s not obvious to me from the source article that Meta believes otherwise. I couldn’t find the full memo, do they claim the leaderboard or “tokenmaxxing” era was a mistake?
Given that Meta has run 5ish layoffs at this point, and everyone is in survival mode, what did they expect? Everyone wants to juice whatever numbers possible to keep their jobs.
My company has an AI leaderboard and ONE ranking to be this, AND OTHER rankings like efficiency (loc merged per token). No one is so stupid that they think any single one of this is to be optimized (gamed) for.
Tech journalists have low opinion of people with actual skills who actually contribute to society, and when their opinions get posted here, it's often selectively echoed by people looking for a reason to feel smarter than the industry.
Even in firms that don't hold this board, giving unlimited AI to people means those who don't bother to learn technology can now deliver 100x their capacity, in very poor quality code
measuring uptime? I've seen Anthropic's status page, and they are a >$1 Trillion dollar company who "largely solved" coding. so clearly you aren't correct. /s
Ok, uptime. How do you measure an individual’s contribution to uptime? If Claude goes down does everyone take a hit? If Claude stays up everyone gets rewarded?
If so, your metric cannot distinguish between a bad engineer and a good one.
If not, you have the same problem you started with: measuring contributions to “uptime”.
> If so, your metric cannot distinguish between a bad engineer and a good one.
A metric that moves in the same direction and amount for everyone based on external event isn’t a problem. The delta in performance of the great engineer will outweigh that of the poor, since the metric movement that is due to external circumstances will be the same in each kind of engineer and thus not count.
That’s a new one. I have never heard of a company which operates like this, giving everyone equal reward no matter how much they contribute individually.
As VPEng, I didn’t use metrics to assess individuals. Too prone to metric gaming.
Instead, I had a career ladder with a detailed rubric describing the skills an engineer at each level was expected to have. (Including communication and peer-leadership skills.)
Managers performed qualitative assessment of employees, using the career ladder as a guide. They relied on tech leads and Staff engineers to help them understand people’s skills, and provided 1:1 feedback and coaching.
We did use impact-based metrics to assess the results of important initiatives. We solved the attribution and lagging indicator problems by estimating impact rather than measuring it, and using a series of proxy measurements (activation, usage, retention, etc.) as a feedback mechanism for revising those estimates.
Unfortunately I think we're entering (have entered?) a period of insanity.
The trouble is AI is being sold as an individual engineering accelerant. I suspect at the most AI pilled orgs you'll then see a commensurate push that starts off with measuring usage (tokens), then measuring output (PRs, code reviews), and then a lot of talk about impact while everyone quietly admits that remains as impossible now as it was fifty years ago.
Why? Because leadership is looking to (and selling, both internally and to the market) AI as the solution to all of their problems, which means they have to prove outcomes that justify their sky high AI budgets.
Higher level metrics at the org/division/product/project level aren't satisfying and flashy enough as they're slow moving and attenuated.
And squishy individual or team level assessments that rely on strong management won't show well on a cost-benefit comparison chart to the board.
At bottom I suspect AI pilled leadership wants to turn software into an assembly line and measure accordingly. Your post perfectly captures why it's still not that easy, and that the real problems in software remains the same and are unsolved by AI: building the right thing, at the right time, and then later figuring out what went well, what didn't, and trying to make those successes more repeatable and failures less likely.
> Unfortunately I think we're entering (have entered?) a period of insanity.
It’s been true for a long time. One of the hardest things as a senior leader in software is dealing with people demanding “accountability” (by which they mean making long-term plans with impossibly precise forecasts) and focusing on costs, all while ignoring value and refusing to engage in prioritization. (I swear, if I hear “it’s all important” again…)
People are just… shallow. They operate on feelings and vibes. They follow the herd without thinking critically. Then they get angry when their dreams clash with reality, and they blame the messenger when those dreams turn out to be fantasies.
But you’re right. AI is bringing out the worst in these tendencies. I think it’s because it’s so convincing when you don’t dig deeply, or aren’t an expert in the subject being discussed. On the plus side, it raises the floor, but I think we’re in for some difficult times before the lessons are learned. I don’t think it will take long, though: I suspect that naive use of AI is going to massively speed up the technical debt curve, and where it used to take 5-9 years to destroy a codebase, it will now take closer to 1-2.
> In particular proactively preventing bugs is not rewarded
The main way I think you can proactively prevent bugs in a meaningful way is by crafting and propagating better architecture.
Better (or worse) architecture and adoption of it can be measured through a mix of quantitative and qualitative means so those metrics could be used to evaluate the impact of the engineer driving that architecture.
That's not how managers evaluate engineers at these corporations.
The engineer who haphazardly launched on Friday then promptly saved the team at 3am and worked the weekends gets the promotion, while the one who prevented a bug from happening "didn't get anything done" and gets the PIP.
Not surprising. It seems that the comment section of every coding agent thread has at least one person mentioning they use "tokenmaxxing" to increase their token usage because it was brought up during their quarterly review, at a standup, or some other communique from on high.
Just wonder what happens when more and more companies introduce similar restrictions. Will that lead to devaluations of the LLM companies?
And I still can't exhaust the limits on my Claude Max subscription, despite being more productive than I've ever been in terms of real work (ie, things that actually make money)
For real. I've used 8B tokens in the past month and haven't hit my limits even once. In fact, I can't even get close except for the day I used Fable. I've barely stopped. Claude keeps reminding me to sleep.
Oh gosh, I run out by Wednesday usually. But I'm not really coding with it, per se. Mostly just writing docs and tech manuals and AI generation. I'm in biotech.
Ok I’ll ask since nobody else has — are they not giving their devs a Claude code max or Codex Pro subscription? If so, why is token cost approaching billions? And if not, why not?
The subscriptions are for personal use not enterprise.
i.e. [1] "This article is about paid Max plans for individual consumers. If you're part of an organization looking to use Claude with your team, refer to Team and Enterprise Plans."
Enterprise customers don’t get those plans, at the enterprise level you have to pay by the API rate… so people don’t have limited use, but you’re also not getting the heavily discounted rate the “normal” plans are at.
>Meta plans to spend up to $135 billion on AI infrastructure through 2026 and commits $600 billion to data center buildouts through 2028
And they can't afford a few extra billion that their engineers can utilize right now?
Looks like AI as it develops is intended to be too expensive for regular people in the long run, but if Meta can't even afford it at that rate, who can?
Within a few weeks of telling people at our company that if they don’t use AI they will be replaced by someone who does, they just announced that their allocation with ChatGPT has reset and are now panicking as they blew through their million token allocation for this month in under six hours - you can’t make this shit up.
I can’t tell if you’re complaining that Meta isn’t saving the whales or that their products aren’t good. If it’s the latter, you should double check their financials.
Yes, but a curve trends to zero. A cap is a step function to zero. I think they're saying there's no point in removing those diminishing return gains if there's still positive value to be had.
it's not that difficult to say it confidently if you use any of their services and applications because exactly nothing has changed.
For reference most labor productivity increases for the last 50 years amounted to about 2% per year. If a hypothetical FB engineer had doubled their productivity with their gazillion tokens that would be 30 years of productivity gains in one year. I'd wager the evidence would be quite evident if you opened any of their apps
The only thing I’ve seen Meta release recently are spy glasses, and every employee who has worked on that product should be in prison (with a live 24/7 feed where the world gets to watch them wallow).
The times I’ve been asked to evaluate a prospective candidate and I see that product on their résumé, it’s been an instant veto, in the same category as working at Palantir.
I have never worked there and I am likely very unqualified to ever work there and Zuck has more money than I could dream of so take my comment with that in mind.
Meta sounds like a cluster-F of a place to work. Massive reorgs around wild ideas like the metaverse and everything Ai all the time. Employees terrified of being fired. Incentivizing token spending and then cutting it off. While the overall company may be fine, the dev department sounds rudderless and absolutely miserable.
employees consumed 73.7 trillion tokens in roughly 30 days
, a figure tracked on an internal leaderboard called "Claudeonomics" — a reference to Anthropic's Claude, one of the third-party AI tools widely used inside the company [2]. The leaderboard, which ranked employees and teams by token consumption, inadvertently incentivized usage volume over productive output.
Meta plans to dismantle the leaderboard and replace it with a centralized monitoring platform called "AI Gateway," which will track usage and spending across teams in real time [2]."
This seems to be an interesting upcoming business, that is:
Helping companies centralize and track their AI usage by employee.
It's stories like this that really dispell the genius/merit theory of successful business. The best you can say about Zuck is he didn't prevent Facebook from becoming huge.
This is the real point. If an average person had access to the same amount of capital and ended up with the same ownership terms, there would have been a more sensible outcome for everyone affected.
As many companies do with all their budgets, down to the trivial and clearly positive EV cost of free coffee. So it goes, cost controls are hard and necessarily imprecise.
that is a fair point. The contrast between Meta and Apple could not be bigger here. Apple has billions of devices and yet they decided to use 3rd party models from OpenAI and later Google to build their AI features rather than building foundational models in house. Yet Meta did opposite: they built models (spending billions of $$$ and firing 10% of the company) for billions of users who rather would not use Meta AI features.
Judging from the decisions and outputs of the last decade or so, the leadership at Meta, including Mark Zuckerberg, have got to be among the most incompetent I have ever seen. They go all in on the worst decisions; not just the worst in hindsight, but also the worst at the time. The only thing keeping them afloat is their monopoly from past purchases. They are a posterchild for why the US is no longer a properly capitalist nation.
Wasn’t this already reported on? FWIW this article links to primary sources from early last month https://www.theinformation.com/articles/tokenminimizing-meta...
Fact is that I can actually read TFA, while your link is paywalled.
I mean… should you be able to? It looks like this is just an AI summarizing a bunch of other paywalled sources. It’s “by MLQ Agent”
I’d be curious to see the breakdown on spending by use case. I’ve heard it said that the majority of tokenmaxing comes from none technical uses like reading PDFs, creating PowerPoints, generating graphics/images… ect. But I’ve never heard any actual proof to that.
One thing I find fascinating as a software engineer who talks to non software engineers who use AI tools is how "reading PDFs" is not more of a solved problem. What I mean is that uploading a PDF into a chatbot tool seems to be an extraordinarily obvious use case that non technical (and technical) users would want to do.
IMO claude, chatgpt/codex, etc should be able to optimize the PDF use case to be extremely token efficient as it's a very obvious use case. But when I start to explain to my wife/friends why it burns through so much quota, I find myself thinking "why should they have to understand this aspect of it". to me, that the details of PDF parsing and extracting are relevant to users (instead of solved such that you don't have to pay attention to it) shows how these tools are not nearly as "ready" as they are made out to be. I may be preaching to the choir on this one, but just my 2c
Because PDFs are a nightmare of a format and the only thing that’s is reasonably guaranteed about them is they will render to an image that people can read, the parsing of which will be much less token efficient than the equivalent text
I agree with you, but every non-engineer I know using these tools 100% will drag and drop a PDF into a chatbot. Anthropic and OpenAI as companies who are selling their products to all sorts of businesses should have a much better means of handling this nightmare of a format because it is so pervasive and so obviously what so many of their customers are going to drop into the product.
Why would they spend a ton of effort ensuring that their customers spend less money on them?
Token economics also are weird. If you design a fancy new frontend that for example uses a cheap model to parse a PDF into text that is fed into an expensive model, you will probably spend more money because you are on API payscale rather than the "max plan" payscale.
> I agree with you, but every non-engineer I know using these tools 100% will drag and drop a PDF into a chatbot
I'm an engineer and use my coding agent to deal with PDFs all the time. It can reach for unix tools if it needs them.
I don't think I understand why this is a problem - it uses tokens, but it removes drudgery. This is the entire promise of the technology.
I’m saying there is basically no way to both make vlms able to understand the long tail of PDFs where the layout conveys information (like charts and tables) and to make it as token efficient as text formats. Current approaches have mostly chosen to work more often than not at the cost of token efficiency.
I think they’ve just decided that vision gives the best results and the token issue will take care of itself.
Once you pay full price for tokens plus margin, it is better for the company to burn as many tokens as possible.
For the same reason as why the oil companies want everyone to use large cars.
For anyone needing to do this, the answer is to convert it to an image first. Far smaller, LLMs work well with them (even in some pretty insane use cases I've seen), and, along with human review, it can be a huge productivity gain that results in structured data.
I agree with your recomendation, but converting a pdf to an image is by no means smaller. PDFs are much closer to SVGs then to jpegs.
Why can't I just take a screenshot of the PDF and feed that into the llm?
Been building various LLM+PDF pipelines at work. As soon as you need to e.g. parse tables etc. it becomes a lot of hard work!
I hope someday we can get out of this local maxima of PDF documents. The format is terrible, but was right place, right time and might be impossible to dislodge.
The problem is that for 99% of people in 99% of cases they work fine. It's hard for people to understand that they're trash.
Source; my last job working with accessibility and that nightmare.
You don't need to use an online service to do this; you get to avoid spending money on tokens doing it offline.
Gemma 4 works perfectly well offline on limited hardware (I have an 8GB video card) and can handle extracting text from image-based PDFs just fine.
Take a PDF -> run it through MarkItDown [1], using the OCR plugin if you need (point it to Gemma 4) -> now you can ask Gemma 4 questions about the (markdown) document.
I am sure Gemma 4 could even create a GUI to make this process very simple for a non technical user.
[1] https://github.com/microsoft/markitdown
Amen. Normal office work is wildly different from what we read about on HN. If you were a CEO, determined to lay off all your people, you would want to really zero in on having your AI solve these very unsexy problems: extract data from Office and PDF. Grab data from some part of the screen of a webapp and parse it. drive a line of business app via keyboard or mouse simulation. I know there are companies out there that try, eg Appian and (here in YC) Skyvern, but its a hard problem and yet I feel this is where the true money is.
bingo. 90% of our AI use cases across the company are things like this. security and ops (NOC/SOC folks) use this almost as much as they do for technical stuff.
hell we have restrictive rules for security stuff so in many cases our network engineers are still doing by hand configs for critical systems.
but in terms of token use it's gotta be "take this pdf and parse these 3 columns into 2" or similar
> how "reading PDFs" is not more of a solved problem
This and replies to this are surreal. It's like everyone simultaneously decided to forget that you don't need claude or whatever to read a PDF. The document is literally made for you to read...
> The document is literally made for you to read...
It’s disingenuous to assume every PDF is actually crafted to communicate to its recipients, even more so to pretend LLM users are in a position to understand all the PDFs they receive
There’s a lot of gray area where help understanding a document is fully reasonable
If you can't understand it how can you be sure that the thing that helps you "understand" it does it right?
The best way to parse pdfs is to convert them to images and feed them into the llm.
This workflow is highly optimized.
For sure there are very optimized ways to do it. My point is that a non technical user will drag and drop a pdf into a chatbot. and from a UX/product perspective, they should have to think about it more than that IMO. but seemingly, that's very much an expensive, inefficient way of doing it (burning through a whole context window try to read it, reloading it multiple times per conversation, etc.).
Absolutely this. Never try to parse a native PDF document with any expectation of coherence or consistency.
You are missing that the product is the hype cycle around AI and that's worth Trillions of $ (Trillions with a T). Why build a PDF parser that generate text when you can BS in a podcast and get paid.
This discussion was about measures, goals and incentives. Follow the incentives.
PDFs are both awesome and terrible at the same time. I've seen screenshots of emails added to pdfs alongside tables that span multiple pages. Because you can do almost anything and guarantee that it'll look the same regardless of how or where it's viewed is a big selling point for a lot of businesses. It's this flexibility (i say madness) that makes PDFs notorious, and why some labs have document parsing as a leading product (see https://mistral.ai/news/ocr-4/).
Anecdotally it's true for me. I can code all day with an agent and never went above $50, but the second I need to ingest a pdf doc to figure out a command I need to use it's easily $20-30 for 10 mins of work
the majority came from random claws running on cron. They get a heart-beat, wake up every 10mins, reads all internal-posts, emails, gchat messages, diffs, and decides to post some random message to the workplace so other claws can also regurgitate. rinse and repeat and then we are looking at $B tokens
I'd guess through LLM embedded PoC projects.
You can rack up token consumption extremely quickly when you embed LLMs into automated processes or products.
I'd be very surprised if these numbers are just typical coding usage with no scripting/pipeline/automation stuff
One thing we use it for is for forking tools internally because of politics.
Tokenn maxing also comes from lesser technical guys who had no moat pre AI
Using AI to suddenly deliver massive amounts of code without questioning the requirements
The most interesting number is missing here, and that is the token distribution by use case. If 60-70% was eaten up by PDFs, agents and automation instead of people actually sitting in Claude Code, then it is a completely different story
"The leaderboard, which ranked employees and teams by token consumption, inadvertently incentivized usage volume over productive output."
Who could possibly have predicted that happening?
A past employer thought it was a good idea to put up a leaderboard of who sent the most Slack messages. They celebrated the people at the top for being so active.
Predictably, everyone started talking in Slack like their jobs depended on it. Everyone was responding to everything. Instead of writing out a complete message and pressing enter, they'd send each fragment of the sentence as a new line.
The Slack leaderboard was never shown again. Unfortunately the habit remained because people were afraid they were going to be secretly judged by how much Slack activity they generated.
I expect the same thing is going to happen at companies who had token leaderboards. Once you've instilled that fear in people, they internalize the expectation.
Reminds me of the place I worked at where I got in trouble because I was the only person writing JIRA tickets. Instead of bitching out the product manager or the tester for not writing tickets, they just complained to me. And if I wrote a ticket about how we could speed up the 40 minute build to 15 minutes I'd have to explain "How does this change improve the customer experience?" to which I answered "If the build was faster the customer would have had the product six months ago"
I worked somewhere that made time from PR being sent for review and ready to merge be a metric for the reviewers. Not time to add feedback in each round. Total time elapsed.
Insanity
You have to realize that if you set a measure, you're actually setting a goal for your employees. There is no such thing as a meaningless metric; why else would you measure it?
No amount of "this isn't used for anything" will change that. It's inherent in human nature in the 21st century to believe any and all metrics will be used against them, and therefore must be gamed.
It's why you also have to set UNBELIEVABLY clear goals and have incentives tied to those goals. Incentives meaning money. If you want to measure things, measure them. But have clear, consistent, and meaningful goals tied to bonuses or something if you want a thing done correctly.
I worked at a place that argued that nobody would game the metrics because it would be wrong and they never stated what the metrics were… while I was gaming the metrics and they were praising me for being one of the best on the team.
It was an unreal experience.
Kinda.
The answer is simpler on the surface: focus.
Generally the problem is the larger the firm’s operations, the harder it is to focus.
Apple is the only firm that has done well on this consistently and doesn’t have a huge grave yard of failures to show for it.
Are you saying what people are hoping to achieve with stupid goals? Because yeah, obviously. But the point is that they're stupid, so they don't achieve that, and that failure is 100% knowable in most scenarios.
PUBLIC graveyard
I wonder if this should be codified as a rule of thumb, or an unofficial "law"? One perhaps we can reference easily among our peers.
You mean like Goodhart's Law? (Unless you're being sarcastic)
That's close but not quite.
That says that when a measure becomes a target, it ceases to be a good measure.
My argument is that in the 21st century, a time of great data use without deep understanding, any measure at all will inherently become a goal to be gamed.
This will inevitably be allocated like other budgets, and from talking to Meta folks about GPU budgets, it is going to be brutal.
"when a measure becomes a goal, it stops being a measure"
It's surprising how often this principle is applicable.
What is about silicon valley leaders not understanding basic economics or business management? These kind of cargo cult tactics would not fly in any other industry.
These kind of cargo cult tactics show up in all kinds of businesses.
But yeah, it's like they've never actually met human beings...
Give an example.
> Who could possibly have predicted that happening?
Charles Goodhart :-)
https://en.wikipedia.org/wiki/Goodhart%27s_law
It’s funny how many times the same thing happens at each large company. I think people’s thought process is this:
> Oh wow! If I paid for this myself I would have spent a lot of money! Are other people spending as much as me? I’m going to create a leaderboard!
> Oh no, my misinformed manager is using the leaderboard as a slight of hand for work. I need to game this now.
Then the leaderboard is banned… I can’t see how this ever really goes up the chain beyond director.
There is zero chance that this is how the leaderboards came to be.
What makes you assert that?
> Who could possibly have predicted that happening?
Everyone except the executives who get paid millions to predict exactly that.
Not a problem. There are thousands of employees standing by, willing to sacrifice their jobs for their vision.
It's a hard job, someone has to not pay consequences for bad decisions.
presumably they covered this in the boilerplate MBA courses they took, non?
It wasn't leadership doing this though. Any meta IC can generate internal apps and dashboards. This was unofficial and unsupported. Some random IC just made it for fun. Management is usually pretty lax with stuff like this (plenty of games and joke internal apps) so they left it up until it became a problem.
I still don't understand how Mark Zuckerberg has any serious investors, he went on this AI tangent and has absolutely nothing to show for it, despite FB / Meta having built some key tech in the space. He needs to stop trying to do something "different" and literally try and build a serious coding agent he can sell, he could have probably had something worthwhile in that space by now.
He started being drastically more serious into AI in 2022, and 2023 and he has nothing to show for it.
Heck, he could have rented GPUs the way Elon did at this point and either mended the bleeding or stopped it, not sure how many he has, but it beats losing this badly.
If he doesn't wake up and learn how to business, I suspect he will lose his empire he's built up for himself.
>he could have rented GPUs the way Elon did at this point
"Meta building cloud business to sell excess AI capacity, Bloomberg News reports Meta building cloud business to sell excess AI capacity, Bloomberg News reports"
https://www.reuters.com/business/meta-sell-excess-ai-computi...
That's fair, I guess still having that sort of leaderboard is a foolish way to burn a billion dollars.
Haven’t got numbers so I might be wrong, but I suspect it is dwarfed by the present size and future potential of Meta’s ads business.
I know right? What did the leadership think would happen when they give some of the worlds greatest software engineers (supportably), a easily quantifiable metric to target?
The leaderboard wasn't leadership generated, it was engineer generated from internally available data. The leadership target is "impact" from ai tools.
What a wonderful scapegoat! Technically it's all "engineer created" because the managers generally don't do technical work. I bet many managers pushed their reports to increase their usage during their 1:1 meetings based on data from the leaderboard. If management had any sense that it was a bad metric, they had ample time to get ahead of it and take it down and provide appropriate guidance. Instead, predictably, they waited until it was a full on disaster and a crisis before acting.
At one point there were over 70 different token maxing dashboards as the management had a game of whack a mole trying to remove them. There definitely was encouragement from management about a year ago to increase ai usage, but once Claude code was allowed, they didn't really need to encourage anyone any more.
Budget impact is technically impact.
Now come on, there was a recent post where the author argued that infallible management knew this would happen, but was part of the double-secret-probation strategy to get the cogs to finally start using AI.
I still think this is true and it’s not obvious to me from the source article that Meta believes otherwise. I couldn’t find the full memo, do they claim the leaderboard or “tokenmaxxing” era was a mistake?
Would they admit it if it was? Or would they try to find a plausible rationalisation for wasting billions without any return?
Given that Meta has run 5ish layoffs at this point, and everyone is in survival mode, what did they expect? Everyone wants to juice whatever numbers possible to keep their jobs.
My company has an AI leaderboard and ONE ranking to be this, AND OTHER rankings like efficiency (loc merged per token). No one is so stupid that they think any single one of this is to be optimized (gamed) for.
Tech journalists have low opinion of people with actual skills who actually contribute to society, and when their opinions get posted here, it's often selectively echoed by people looking for a reason to feel smarter than the industry.
Even in firms that don't hold this board, giving unlimited AI to people means those who don't bother to learn technology can now deliver 100x their capacity, in very poor quality code
unfortunately at big tech, this shit will keep happening.
people who make it to managers tend to have bozo tendencies & are yes men.
before it was lines of code, Jira tickets closed. Now it's tokens spent.
How dare you question the most effective allocators of capital.
measure outcomes (impact), not effort (token usage, lines of code, code coverage, hours worked, etc.)
The whole phenomenon of metric based Eng evaluations is because leadership does not trust line managers to evaluate individual engineers.
What outcomes though? The ones I’ve seen posted are still nonsensical metrics that a publicly traded firm absolutely doesn’t care about.
It wants to see faster R&D, higher revenues from existing assets, greater operating margins, higher sales to invested capital ratio and so on…
The best way to measure that for a software firm is up-time of services, usage and project completion duration
measuring uptime? I've seen Anthropic's status page, and they are a >$1 Trillion dollar company who "largely solved" coding. so clearly you aren't correct. /s
Ok, uptime. How do you measure an individual’s contribution to uptime? If Claude goes down does everyone take a hit? If Claude stays up everyone gets rewarded?
If so, your metric cannot distinguish between a bad engineer and a good one.
If not, you have the same problem you started with: measuring contributions to “uptime”.
> If so, your metric cannot distinguish between a bad engineer and a good one.
A metric that moves in the same direction and amount for everyone based on external event isn’t a problem. The delta in performance of the great engineer will outweigh that of the poor, since the metric movement that is due to external circumstances will be the same in each kind of engineer and thus not count.
That’s a new one. I have never heard of a company which operates like this, giving everyone equal reward no matter how much they contribute individually.
Unfortunately that is a group metric, we need individual metrics
Okay.
How?
This is an org pushing thousands of PRs a day. How do you solve the attribution problem for any one engineer's work given some set of impact metrics?
And keep in mind, most common impact metrics are trailing indicators, often over relative long time horizons.
As VPEng, I didn’t use metrics to assess individuals. Too prone to metric gaming.
Instead, I had a career ladder with a detailed rubric describing the skills an engineer at each level was expected to have. (Including communication and peer-leadership skills.)
Managers performed qualitative assessment of employees, using the career ladder as a guide. They relied on tech leads and Staff engineers to help them understand people’s skills, and provided 1:1 feedback and coaching.
We did use impact-based metrics to assess the results of important initiatives. We solved the attribution and lagging indicator problems by estimating impact rather than measuring it, and using a series of proxy measurements (activation, usage, retention, etc.) as a feedback mechanism for revising those estimates.
Right, so you're sane. :D
Unfortunately I think we're entering (have entered?) a period of insanity.
The trouble is AI is being sold as an individual engineering accelerant. I suspect at the most AI pilled orgs you'll then see a commensurate push that starts off with measuring usage (tokens), then measuring output (PRs, code reviews), and then a lot of talk about impact while everyone quietly admits that remains as impossible now as it was fifty years ago.
Why? Because leadership is looking to (and selling, both internally and to the market) AI as the solution to all of their problems, which means they have to prove outcomes that justify their sky high AI budgets.
Higher level metrics at the org/division/product/project level aren't satisfying and flashy enough as they're slow moving and attenuated.
And squishy individual or team level assessments that rely on strong management won't show well on a cost-benefit comparison chart to the board.
At bottom I suspect AI pilled leadership wants to turn software into an assembly line and measure accordingly. Your post perfectly captures why it's still not that easy, and that the real problems in software remains the same and are unsolved by AI: building the right thing, at the right time, and then later figuring out what went well, what didn't, and trying to make those successes more repeatable and failures less likely.
> Unfortunately I think we're entering (have entered?) a period of insanity.
It’s been true for a long time. One of the hardest things as a senior leader in software is dealing with people demanding “accountability” (by which they mean making long-term plans with impossibly precise forecasts) and focusing on costs, all while ignoring value and refusing to engage in prioritization. (I swear, if I hear “it’s all important” again…)
People are just… shallow. They operate on feelings and vibes. They follow the herd without thinking critically. Then they get angry when their dreams clash with reality, and they blame the messenger when those dreams turn out to be fantasies.
But you’re right. AI is bringing out the worst in these tendencies. I think it’s because it’s so convincing when you don’t dig deeply, or aren’t an expert in the subject being discussed. On the plus side, it raises the floor, but I think we’re in for some difficult times before the lessons are learned. I don’t think it will take long, though: I suspect that naive use of AI is going to massively speed up the technical debt curve, and where it used to take 5-9 years to destroy a codebase, it will now take closer to 1-2.
> measure outcomes (impact)
This is also not easy. In particular proactively preventing bugs is not rewarded
> In particular proactively preventing bugs is not rewarded
The main way I think you can proactively prevent bugs in a meaningful way is by crafting and propagating better architecture.
Better (or worse) architecture and adoption of it can be measured through a mix of quantitative and qualitative means so those metrics could be used to evaluate the impact of the engineer driving that architecture.
That's not how managers evaluate engineers at these corporations.
The engineer who haphazardly launched on Friday then promptly saved the team at 3am and worked the weekends gets the promotion, while the one who prevented a bug from happening "didn't get anything done" and gets the PIP.
It's not flashy.
When shit just works for months or years no one is going to come and praise you for stuff you did a while back.
You are better off breaking stuff and then fixing them to show how useful you are.
Not surprising. It seems that the comment section of every coding agent thread has at least one person mentioning they use "tokenmaxxing" to increase their token usage because it was brought up during their quarterly review, at a standup, or some other communique from on high.
Just wonder what happens when more and more companies introduce similar restrictions. Will that lead to devaluations of the LLM companies?
Not sure if I missed it but I couldn't find any information in the article to explain where the "approaching billions" estimate is coming from.
I could believe it, but I'd want to see something a little more concrete.
And I still can't exhaust the limits on my Claude Max subscription, despite being more productive than I've ever been in terms of real work (ie, things that actually make money)
Because that’s heavily subsidised, whereas companies have to pay something closer to the actual price.
Enjoy it while you can, because it won’t last forever. Per-token billing is quite eye opening in terms of how much it can cost
For real. I've used 8B tokens in the past month and haven't hit my limits even once. In fact, I can't even get close except for the day I used Fable. I've barely stopped. Claude keeps reminding me to sleep.
Oh gosh, I run out by Wednesday usually. But I'm not really coding with it, per se. Mostly just writing docs and tech manuals and AI generation. I'm in biotech.
Jesus. 8B tokens in a _month_? Do you even think at all anymore?
Ok I’ll ask since nobody else has — are they not giving their devs a Claude code max or Codex Pro subscription? If so, why is token cost approaching billions? And if not, why not?
They can't.
The subscriptions are for personal use not enterprise.
i.e. [1] "This article is about paid Max plans for individual consumers. If you're part of an organization looking to use Claude with your team, refer to Team and Enterprise Plans."
[1]: https://support.claude.com/en/articles/11049741-what-is-the-...
Enterprise customers don’t get those plans, at the enterprise level you have to pay by the API rate… so people don’t have limited use, but you’re also not getting the heavily discounted rate the “normal” plans are at.
>Meta plans to spend up to $135 billion on AI infrastructure through 2026 and commits $600 billion to data center buildouts through 2028
And they can't afford a few extra billion that their engineers can utilize right now?
Looks like AI as it develops is intended to be too expensive for regular people in the long run, but if Meta can't even afford it at that rate, who can?
Big enterprises don't get to have those subscriptions. OpenAI or Anthropic simply won't sell them to you if you need a couple thousand of those.
Within a few weeks of telling people at our company that if they don’t use AI they will be replaced by someone who does, they just announced that their allocation with ChatGPT has reset and are now panicking as they blew through their million token allocation for this month in under six hours - you can’t make this shit up.
A million tokens is like $15 with SOTA models… that’s their allocation?
It was more like some kind of credit thing, not tokens.
All those billions spent on tokens by Meta, and not a single iota of value generated by any of it
I can’t tell if you’re complaining that Meta isn’t saving the whales or that their products aren’t good. If it’s the latter, you should double check their financials.
I guess maybe they can crank out more ads in their dystopian ad space of a social network site.
I love how confidently you say this, with no evidence provided (and I doubt you have any.)
Just a pristine comment section yap.
If there was a positive return on token spend they wouldn’t be capping it now would they?
Note that the person I'm replying to didn't say "positive return." They said "one iota of value."
In addition, as other commenters have pointed out, diminishing returns is a real phenomenon.
It’s possible for something to have diminishing returns.
Having a speed limit does not imply the utility of driving is zero.
That does not follow. Many things have diminishing returns curves.
Yes, but a curve trends to zero. A cap is a step function to zero. I think they're saying there's no point in removing those diminishing return gains if there's still positive value to be had.
Was there a new product released by Meta that we are not aware of? The last thing I read about was the Instagram account take-over AI-bug.
>I love how confidently you say this,
it's not that difficult to say it confidently if you use any of their services and applications because exactly nothing has changed.
For reference most labor productivity increases for the last 50 years amounted to about 2% per year. If a hypothetical FB engineer had doubled their productivity with their gazillion tokens that would be 30 years of productivity gains in one year. I'd wager the evidence would be quite evident if you opened any of their apps
The only thing I’ve seen Meta release recently are spy glasses, and every employee who has worked on that product should be in prison (with a live 24/7 feed where the world gets to watch them wallow).
The times I’ve been asked to evaluate a prospective candidate and I see that product on their résumé, it’s been an instant veto, in the same category as working at Palantir.
Uber, Microsoft and now Meta. All tokenmaxing to the max on Claude
This is what you get when token consumption becomes a KPI...
I have never worked there and I am likely very unqualified to ever work there and Zuck has more money than I could dream of so take my comment with that in mind.
Meta sounds like a cluster-F of a place to work. Massive reorgs around wild ideas like the metaverse and everything Ai all the time. Employees terrified of being fired. Incentivizing token spending and then cutting it off. While the overall company may be fine, the dev department sounds rudderless and absolutely miserable.
>"The internal memo disclosed that Meta
employees consumed 73.7 trillion tokens in roughly 30 days
, a figure tracked on an internal leaderboard called "Claudeonomics" — a reference to Anthropic's Claude, one of the third-party AI tools widely used inside the company [2]. The leaderboard, which ranked employees and teams by token consumption, inadvertently incentivized usage volume over productive output.
Meta plans to dismantle the leaderboard and replace it with a centralized monitoring platform called "AI Gateway," which will track usage and spending across teams in real time [2]."
This seems to be an interesting upcoming business, that is:
Helping companies centralize and track their AI usage by employee.
Anyway, great article!
It's stories like this that really dispell the genius/merit theory of successful business. The best you can say about Zuck is he didn't prevent Facebook from becoming huge.
> The best you can say about Zuck is he didn't prevent Facebook from becoming huge.
Doesn't the legend go Zuck didn't even see the big picture until Sean Parker spelled it out for him?
This is the real point. If an average person had access to the same amount of capital and ended up with the same ownership terms, there would have been a more sensible outcome for everyone affected.
That is insane. I'm sure companies will learn the absolute wrong lesson from this, and attempt to centralize and kneecap token usage.
As many companies do with all their budgets, down to the trivial and clearly positive EV cost of free coffee. So it goes, cost controls are hard and necessarily imprecise.
Tokens are less valuable than the eyeball metric of the Dotcom era. At least the eyeballs were real then.
I'd argue most of the AI value is related to how 'Dead' the internet is.
This talk of tokens is wasteful.
Ultimately the spend on tokens has to benefit the firm financially or it won’t continue spending on it.
in an old-timey cartoon voice
"Now he tells me!"
bonk
Clearly no one is using Meta’s customer facing AI products. Why aren’t they using their own gpu/compute for development?
Because Muse isn't good enough and why use Muse if they'll let you use Opus for free?
that is a fair point. The contrast between Meta and Apple could not be bigger here. Apple has billions of devices and yet they decided to use 3rd party models from OpenAI and later Google to build their AI features rather than building foundational models in house. Yet Meta did opposite: they built models (spending billions of $$$ and firing 10% of the company) for billions of users who rather would not use Meta AI features.
2 week old news OP;
Various discussions:
Meta’s chaotic AI strategy
https://news.ycombinator.com/item?id=48523271
Companies rein in AI usage as costs strain budgets
https://news.ycombinator.com/item?id=48602571
Meta CTO Andrew Bosworth Admits the Company's AI Reorg Was 'Atrocious'
https://news.ycombinator.com/item?id=48548461
Tokenmaxxing is dead, long live tokenmaxxing
https://news.ycombinator.com/item?id=48708795
Had a similar comment but what’s even weirder and I seemed to have missed entirely at first glance is that this is an AI news aggregator agent?
The article is “by MLQ Agent.”
Judging from the decisions and outputs of the last decade or so, the leadership at Meta, including Mark Zuckerberg, have got to be among the most incompetent I have ever seen. They go all in on the worst decisions; not just the worst in hindsight, but also the worst at the time. The only thing keeping them afloat is their monopoly from past purchases. They are a posterchild for why the US is no longer a properly capitalist nation.
I don't understand though. How will all the AI users replace all the non-AI users if they can't spend money that isn't theirs to win by default?
Don't worry, once we achieve post-scarcity they will have more tokens than they could ever dream for spending.
How soon until this becomes part of the "no one wants to work anymore" argument