I remember at Google at around 2007 - 2009, as Google was massively expanding its data centers, there was a lot of unused capacity, especially during off-hours. Any engineer could run as many jobs as they wanted at zero priority, which means the job would be first in line to be killed if a more important task needed the resource.
I did so many interesting experiments with MapReduces that would run overnight.
For a while, I would even build internal services that were basically "free" because I'd just run them all at priority 0.
Over time those services got less and less reliable as overall usage started to increase, so I was forced to either justify the resources or scale back - but that was a good thing.
I feel like something similar would be a good model for AI token use: big tech companies ought to have their own self-hosted LLM data centers to power their own needs, then let employees use off-hours capacity to experiment.
Outside of experimentation, we should be encouraging token efficiency for everyday tasks. Rather than having a certain number of tokens, engineers should be evaluated based on how much they actually get done.
Using a lot of tokens to automate a process that used to require hours of human labor every week? Good use of tokens, should be encouraged.
Using a lot of tokens to debug an easy frontend bug that could have been fixed by hand, and still took you 4 hours to complete? Waste of tokens, should be discouraged.
"Using a lot of tokens to debug an easy frontend bug that could have been fixed by hand, and still took you 4 hours to complete? Waste of tokens, should be discouraged."
Hahahah good luck with that!
For many of us, what is happening now was super obvious. Telling a new formed crack addict (who you wanted to become addicted) to be more thoughtful about their consumption of crack... yeah not gonna work is it.
Most AI front ends seem to be designed for interactive jobs, so they make it hard to define a job that should be done eventually with zero priority. It makes much more sense to do that with spec-driven development (have work done with the human on the loop rather in the loop), but as far as I know that just isn’t well supported by any front end yet (would be happy to be proven wrong, my experience is with Google front ends).
my money is on: eventually frontier model dev and training becomes basic research funded by governments, and LLM operators become essentially private utilities a la ISPs, competing mostly on data center operational costs and occasionally new chip tech to run models cheaper
and governments will keep running massive data centers with classified frontier models for intelligence and propaganda purposes
Don’t think we’ll see similarly logically behaviour from LLM users tbh. A sizeable portion of the user base seems to insist on through opus at every trivial task
If any company announces that they use token consumption as an employee performance signal, for me that's close to a red flag to stay away from that company.
No company with good engineering leadership should act like this is remotely a good idea.
The where may be the decision makers chasing social media trends. A friend sent me a link to this, this morning, about devs rather than managers, but I suspect it's the same: https://youtu.be/IW3Sbe0Hbgg
You can believe whatever conspiracy theories you want, of course, but the most straightforward explanation is that when you lay off X,000 or XX,000 people, some number of them will be on leave.
And my CTO insists that PR count isn't a performance metric. But guess what number gets used the minute people are forced to stack rank (of course they don't call it that, but... that's basically what it is)?
Do you have any source for this at all? I’ve seen so many different exonerations for Meta’s layoff criteria including claims that engineers using the most AI were laid off because Meta had them build AI tools to replace themselves.
Everyone is oddly confident despite all of the conflicting explanations.
Without any evidence, I would be shocked if performance rating wasn't a factor in the layoffs. But performance rating is not the same thing as AI tool use.
You'd be surprised, I know a few devs in very big tech companies, not faang but you definitely know them, and they all have some kind of token leaderboards, a few told their dev "we don't want you to write a single line of code manually anymore", etc.
I assume the execs perspective is something like: if the top 20% of worker produce 80% of the code with LLMs and the company still works then we can get rid of the bottom 80% of devs and save money
I think there's probably something to token use as some kind of metric. If you aren't using these tools much, you're definitely not going to remain a top contributor. The world is evolving quickly here.
But it's just one signal out of many, and more isn't somehow inherently better beyond a certain point.
But even if the end goal was to lay off 80% of programmers, shouldn't the 20% to keep be the developers delivering the 80% of the code, regardless of whether they spent the most to do it? Like what if the 20% of workers spending the most tokens were actually the bottom 20% in terms of delivery because they were using the worst prompts and having AI constantly implement 5 different versions of everything, then throwing it all out because their prompt was so bad anyway?
Ah, but "who uses the most tokens" is a number, a number generated by a computer no less. Questions like who delivers lots of high quality work require you to do research and make judgements, which is work.
As I have snarkily observed at work: if I go $100 over the meal allowance on my business trip, I'll have to have an unpleasant conversation with my manager or finance. If I use $500 in AI tokens unproductively I'll be recognized for being a top AI adopter.
I have seen this type of behavior happen many times in different companies.
For example, at more than one company I've worked for, if you wrote shitty code but got it into "testing" faster than anybody else, you are considered a superior programmer. And then, if you fixed the hundreds of bugs found in your code seen as an extraordinary programmer going above and beyond the call of duty.
The problem is that many companies which had reasonable leadership in the past with the advent of LLM AI started to make rushed (and dubious from my point of view) decisions - using token usage to evaluate an employee performance is just one of them.
There is little new under the big fusion reactor in the sky. I just read a chapter in James Glieck's "The Information" about tokenmaxxing in the telegraphy industry. There used to be a big market for code books to reduce the per-character charges for sending telegrams. Compression was cash in the pocket. The telegraph companies discouraged the practice but were forced to accept it. The telegraph code industry started with the initial commercialization of telegraphy and didn't end until the 1920s.
There was a cost to it though. Codes greatly reduced redundancy, and caused large miscommunications from very small errors. As Glieck explains it, this was the opposite of the African drumming practice of adding redundancy to strengthen the relationship between the rhythm and the language that the drums mimic.
Thanks, that's so odd that I assumed it was about efficiency, which is how I treat tokens. It's hard to imagine a 19th century business man ordering his staff to send as many long winded telegrams as they can.
I don’t think the analogy works too well for one specific reason: you can increase your number of used tokens without ever „sending a telegram“! Run a bunch of Claude sessions, ask them to review various docs sites, create random prototypes, they just throw all of that away. Congrats, you’re a token maxxer
Think of it like this: telegraphs are the hot new thing. The more you send, the more modern and relevant your company. No more Pony Express. You can either have employees sending 1-2 a day. Or, 100 per day. Wow so advanced, so modern, invest now.
This will probably lead to some balancing act like ye olden days of big data etc. Companies want AI native engineers who will use AI to do their work, but don't want AI quality outputs and don't want to drop 200k per year per employee.
AI quality outputs are fine for backoffice work now, but they are awful to read and reason about. Hallucinated features are also difficult to work with.
Isn't that the exact opposite of tokenmaxxing - instead, the telegraphy analogy would be if telegraph operators were ranked by how many hours per day they tied up telegraph lines (highest number of tokens burned/highest $ spend wins) instead of by customers served (programmers delivering features).
What you were describing would be token-minimizing, not maxxing.
I don't like using AI. I don't find it particularly helpful. But my employer insists that we use it and tracks metrics so I make sure to give it pointless busywork daily. That way I show as using it even if it causes more problems than it fixes.
For example in Seattle you pay county fees, and then state fees, and then maybe special fees if you were picked up in the airport.
I took a ride from SEATAC to my hotel in downtown Seattle and besides the ride itself, there were 5 other items on the bill, 4 of which are specific to the place I used Uber.
Then I had the return trip from my hotel to SEATAC, on this one I got EIGHT items on the bill, on top of the ride fare. Some specific to Seattle itself, some specific to the road that the Uber took (a tunnel fee - which is different based on the direction you take it in), etc.
So the real question is what is NOT different between two locations. Less than 15% of the bill.
I also took Uber in India, where you have to share a one-time password with the driver for example, which I've never seen in any other country.
In some other countries the Uber app exists but Uber drivers are actually taxis, so you're actually ordering a taxi via the app.
Essentially every single airport in the world is custom UI and custom walking path guides and pickup instructions, and rules for where pickups/dropoffs/etc can occur can change multiple times in a day, much to everyone's enjoyment. They're almost all private property, and are so valuable that whatever they want is what they get.
And food. Most/~all? major brands get custom integrations.
Hundreds (iirc) of identity verification providers, most or all custom, and constantly weighed against cost and accuracy because it ain't cheap and it ain't good but it is far better than none (both legally and ethically).
No idea how many payment sources they accept, but it's definitely a lot more than anyone who hasn't lived on 5+ continents thinks.
And remember that this is all international. So scale is huge and law changes are constant and frequently conflicting. Darn near every useful feature is illegal somewhere, at some time, for both good and bad reasons.
---
This is not at all to say I think Uber is efficient, clearly it is not. Not by an enormous margin. But there is a legitimate need for truly absurd complexity, because the world is not consistent. You see similar things happen anywhere [thing] tightly interacts with humans.
Vegas: ordering a tax "to a hotel" - hotels have different entrances, pickup / dropoff there during crazy times is hard. Uber UI for Vegas is unique / some features are designed to make it easier for driver and passanger to find each other
Airports: different regulations, different rules for pickup/dropoff. Also scammers who pretend to be in a car, walk with their phones around pick-up ares in airport and do bait-and-switch (saw that in Istanbul SAW and in Dubai Al Maktoum)
There's an excellent HN thread that talks about this very question (that comes up on HN every now and then - what _does_ company X do that needs so many engineering resources?): https://news.ycombinator.com/item?id=25375921
TL;DR: Managing a taxi service (that's what Uber is in my mind, not whatever "ride share" means) that spans cities and states, never mind countries, is extremely complicated. To their credit, Uber manages to make it look simple to the end user, prompting such comments as "meh it's just a few screens how hard could it be", which is triumph of product engineering as far as I am concerned.
There’s been a bug when ordering an Uber in my (quite small, circa 20k pop) town recently in that they think our annual festival (which is in July) is on now and try to force a stupid pickup point which in my case is about the same distance from my local pub as my house is but in the opposite direction. I guess things like this need some sort of maintenance (which apparently they’re not getting!).
The New Delhi airport has a single pickup location for all Uber rides (which moved recently). Also the Indian government requires separate reporting from Uber depending on whether the rider is using an Indian or foreign credit card to pay. They also have to report to the government anyone who hailed the cab on WiFi rather than LTE (for "security" reasons they can never quite articulate).
Every different country Uber operates in is a moving legal and regulatory target
Sure, but custom integrations seem unlikely to explain the majority of Uber's technical headcount. Let's say they average a dedicated engineer for each of their 1000 largest markets/locations. Let's assume another 200 across the countless smaller markets. Let's assume 50% overhead atop this for things like infra, tools, and management. These all seem like exceedingly generous estimates to me.
They actually had 5,000 engineers in the tokenmaxxing blog post. That's a lot of engineers for the rest of Uber's business activities.
There are always newer technologies and techniques to be implemented. Better algorithms. Larger deployments. Better reliability. There are also almost always bugs to fix. So, so many bugs.
Weren’t they trying to do their own self-driving thing?
I think this is partly a problem with companies that have had heavy investment. Uber’s value isn’t based on what they are doing, it is based on the idea that they are going to render ideas like owning your own car or taking public transit obsolete (I mean that’s an exaggeration but less of one than it ought to be).
That's true, I guess possibly they figured out that humans using their own vehicles turn out to be cheaper than contracting or owning self-driving vehicles.
Uber is regularly offering 50¢/mile trips now (which they charge significantly more than that to the rider for). There's no way an autonomous vehicle is going to get operating costs that low anytime soon.
I think you’re missing how complex international operations and optimization are.
Each country has their own laws around what uber is and isn’t allowed to do. This needs to be formalized in code. For example you actually call a taxi, though the uber app, and the amount you pay is per mile, not a fixed fare decided ahead of time. To add to this complexity, some cities will have their own laws. What happens if you take an uber from town a to b, where each one has different laws ? A lawyer probably has an answer but the app needs to adhere to that.
On top of that laws change all the time.
Optimization, well you can always optimize something. speed, costs, paths etc.
In a way this never ends.
I think the part we interact with as consumers is a tiny sliver of the complexity those services have to build and operate.
Uber is at a large enough scale that this analysis doesn't work. You and I do not care even a tiny bit about "Eats for the Way", one of their planned features this year (https://www.uber.com/us/en/newsroom/go-get-2026/) that lets Uber Black passengers specify that their car should arrive with their Starbucks coffee order. But if 0.01% of users order 1 additional ride a month because of this, that's about 200k rides a year, which may well be sufficient to justify the development costs.
shiny new tools but people only want to use them on the same old problems. how can we innovate the development of crud apps even more?! that was what plagued the web dev landscape for some time. Constantly seeking newer lazier means of producing the same old product. I admit it has an allure but if companies are no longer constrained by dev effort / labour then they can only ponder their own reflection as the source of their failures.
Well there is a lot of ongoing maintenance cost. There is probably still some marginal gains possible on the matching side. There are new products to launch. So while one specific software can mostly be finished, the total software of a company is always changing.
> When will Uber (or your favourite company) be 'done'? They've been writing software for 16 years
I suppose it becomes easier once the browsers, Android and iOS have been frozen for a little longer than 16 years. Nevermind the changing regulatory field and new products (when was Uber Eats launched?).
In that 16-year period, Covid-19 emerged, as did viable self-driving and partnership with Waymo. A networked, people-facing app can't ever be "done", unless you have perfect prescience. Internal tech-stacks are a living thing: keeping a service that on the outside appears to be unchanging is a lot of work! Scaling is a lot of work! Scaling services and maintenance feed off each other.
It’s death for a tech company to be “done,” since that means no more growth. So they will all bloat indefinitely until they implode or get absorbed. It’s simply the fate of all VC-fueled startups.
Tokenmaxxing makes no sense, it is akin to write extremely inefficient SQL / Spark Jobs, full of cartesian joins, ultra skewed datasets, etc, just for the sake of using as much compute / memory / IO as possible.
This always happens when the metric becomes the goal, companies should nurture and foster an environment where AI is used in the most efficient way possible, first asking "do we really need an agent for this" and if so, what kind of agent is needed, what model, reasoning level, etc.
They should also promote projects that aim at saving tokens, increasing cache hits, codifying the information in ways such they use as less context as possible (graphs of knowledge are pretty good for this!)
It's toddler-level logic. "You can achieve positive outcomes by using X. Therefore, we need to use as much X as possible to maximize positive outcomes."
It's like trying to win a race by setting a gas station on fire.
The argument in favor of "tokenmaxxing" has always been that it's creating space for employees to freely explore the broad and novel space of AI-enabled workflows. I've seen a number of use cases where I'm skeptical any value is being produced, but a number of others where some team or another has finally solved a long-standing problem of theirs with an agentic workflow that would have been hard to justify to a cost review committee.
> They should also promote projects that aim at saving tokens, increasing cache hits, codifying the information in ways such they use as less context as possible (graphs of knowledge are pretty good for this!)
My understanding is that most big "tokenmaxxing" companies do have teams who are working on this in the background.
+1 I find the general disdain for C-suite or senior engineering leadership on HN so silly. These people didn't get promoted or hired because of nepotism. A lot of them moved up the engineering ladder and are familiar with how software engineering works and the incentives involved. Yes, some of them are sheep and will blindly copy what is fashionable but so do a large swath of ICs.
If you want incredibly fast adoption of AI within a company, the best thing you can do is to signal from the top that tokenmaxxing will be rewarded (or at least not be punished for it).
1. It forces everyone including the lazy ones who normally wouldn't invest their time in learning anything new to actually install codex/claude and learn to use them.
2. It prevents any middle manager from putting up blockers for adoption/experimentation ("this is new, I don't trust this, let's do it the old familiar way", "this might be expensive, we care about efficiency here", etc). Once the C-suite dictates tokenmaxxing is allowed, every middle manager will fall in line instantly.
3. Tokenmaxxing is not choice you have to live with the rest of your life. A year or two from now, once C-suite is satisfied with the rate of AI adoption within their org/company, they can just as easily switch the focus to efficiency. Teams will be asked to justify their token spend and start to optimize.
>These people didn't get promoted or hired because of nepotism. A lot of them moved up the engineering ladder and are familiar with how software engineering works and the incentives involved.
I would argue that you have an unreasonably optimistic view about corporate culture. There is a substantial amount of adverse selection and political maneuvering going on all the way up to the top. Tokenmaxxing just goes to show this.
That's part of the reason why this website is hosted under the YCombinator name, after all. Hackers are strongly meritocratic, which is not something you will find in a big company.
> These people didn't get promoted or hired because of nepotism. A lot of them moved up the engineering ladder and are familiar with how software engineering works and the incentives involved.
Tokenmaxxing exists because executives think employees are resistant to change. Thats it, a way to incentivize/force every employee to experiment with a new technology. Obviously once they think everyone is utilizing AI the tokkenmaxxing stuff will end.
I am certain that the max sustainable boost from AI use -- with code review and otherwise all-in -- is approximately 20% with the appropriately skilled senior engineering talent, and the token budget for any engineer should not exceed that.
I do not believe that engineers who are tokenmaxxing are truely productive and I have not seen any evidence whatsoever (perhaps the opposite).
I've personally found that with the right flow and codebase knowledge, that's achievable with sustainable levels of effort.
I think it's because not many know how to measure it properly.
I can output 5 useless/bad features in a day with Claude or I can output 1 useful feature per 2 day period. Which one has better impact on ROI?
In this example, it might seem like it's an easy answer. But, in the real world, it is a lot more nuanced and much more difficult to measure and so not many are bothering to do it and are opting in for the simple solution of following the hype.
AI for engineering productivity seems to be widely misunderstood to be a magic button that produces the same result, but faster and more cheaply. And based on that reasoning, you should want to force employees to tokenmax, because, why wouldn't you want to get more results but faster and cheaper?
A more nuanced view would be something like:
* AI lets you achieve your roadmap somewhat faster, but:
* You incur tech debt that's similar to if you hired a dev temporarily for the features. You don't necessarily have someone on the team that understands the new code.
* Similarly, you aren't upskilling your junior team members. So you aren't getting skill/wage arbitrage as much as before.
* You will complicate the product. P2 features are P2 for a reason, but AI can cause them to be included and complicate the product for lower marginal gain.
I find it shocking that anyone ever thought tokenmaxxing was a good idea.
AI maximalists like to compare the technology to electricity. Imagine if in the early days of electrification, a CEO had rewarded staff for increasing the amount of electricity they consumed rather than finding ways to use it for business impact. Institutionalizing people who showed signs of mental illness was popular in those days, and I suspect that would have been the outcome.
Regularly experimenting with AI tools as they improve and relying on them where they provide an advantage is a good idea at both individual and institutional levels. Maximizing usage for its own sake is not.
It's amazing that it took months to figure this out. "Well we thought that if engineers are told to maximize costs through AI use, to consume as much as possible of a resource that costs us money, then obviously good things will happen. Imagine my surprise when it didn't turn out that way."
Imagine if engineers were ranked based on their AWS spend. People allocate VMs and fill databases with terabytes of random bits, to get to the top of the AWS leaderboard. If you don't do this, you're ranked at the bottom, and good luck at the next review cycle. Who could have expected that this is not the road to success?
You say "amazing that it took months to figure this out" as if the answer to the question is obvious.
But it's not. Some FAANGs are doing amazing things with unlimited tokens. Other companies have no clue what to do with tokens, they've just told their engineers to max them.
It really depends on how you're using the tokens. If you're just using them for Codex and Claude Code - yeah, tokenmaxxing is incredibly dumb.
> Some FAANGs are doing amazing things with unlimited tokens. Others have no clue what to do with tokens.
Unlimited tokens is different from “use AI a lot or we will fire you, and we are counting token consumption as usage”. Obviously the latter is stupid and yet it was done in many places.
I'm not convinced it actually was done in many places, although I understand why in a bad job market people don't trust that it isn't happening in secret. Every time I've heard of a token leaderboard or such it's come with a denial that the company is using it as an employee performance metric.
> But it's not. Some FAANGs are doing amazing things with unlimited tokens
Giving someone unlimited access to a resources is not the same as directing or incentivizing them to use it for the sake of using it which is what the parent comment criticized.
As for the other FAANGs, Meta and Google have (not good but still) frontier models of their own, so they are very different from a company paying API costs per token.
Show me some fang that have made nice outwards facing products through a fully embraced AI workflow?
AI is an accelerator that engineers should know and have access to, but it's not something that should have mandated usage and quotas around. It's also absolutely dangerous for young engineers and the like - it fundamentally denies you of the "learning" aspect. I'm now seeing in interviews young graduates being given AI tasks to complete and they come back with a correct solution and no concept of how it is working.
You learn and reinforce learning by DOING and reading in depth. High level summaries don't teach anything and are the kinds of things only VPs care about. So, unless the intention in the future is for everyone to be a VP using AI to do the work, we need some middle ground here and some real thought around implementation of these tools or there's going to be a generational canyon gap of knowledge between being able to "say" and being able to "do".
OP (solenoid0937) is an unfounded AI-hype peddler and an Anthropic shill (check their comment history), do not expect them to provide an actual example of their wild claims.
You could have easily disproved my claim by linking your comments with a more balanced or nuanced opinions on the matter, except you cannot, because there is only even more outlandish and wild stuff you say about AI and LLMs.
It‘s perfectly ok to share opinions that aren’t nuanced or balanced. You seem to have something strongly against that specific user, the fact that you felt the need to go through so much of their history, post a massive list of their „suspicious“ comments, and mention in multiple places how they are a shill is pretty concerning imho and doesn’t make you look good at all. Their activity looks fine, they seem to be enthusiastic and optimistic about the technology, but that’s pretty much it.
And now you’re asking them to somehow disprove they aren’t a shill? How would that even work. You seem unnecessarily antagonistic towards that user
I did not have to go through "so much of their history", this is just the last 9 days. There are considerably wilder claims from them earlier, when the account was solely focused on propping up LLM-hype and defending Anthropic.
We are living through a period of time with one of the potentially most disruptive technologies ever being developed. A lot (A LOT) of money is invested into it, a lot of livelihoods are and might be affected by it, and some people stand to gain A LOT from it. So there are significant interests to sway public opinion in favour of LLMs and AI, to hype it up to unreasonable extent, to muddle the waters of a reasonable discourse. Accounts such as solenoid0937 are unleashed on public forums to achieve that, and because of that we have to take everything they write with a huge grain of salt, or even ignore completely, as there is just no true information in their comments.
You yourself got baited by them by considering what they wrote seriously regarding "amazing things with unlimited tokens". Now the idea that "LLMs 1) are used in one of FAANGs massively and 2) are used to produce amazing things" is planted. Will you remember later that they did not actually provide any evidence of that? The account has been doing this trick multiple times over the last few weeks.
For me, as someone who is actually using LLMs in their work, a single balanced comment on the matter would have been more than enough to consider them not being a shill. Unfortunately, instead they have claimed recently that they went completely full-on with agentic coding (https://news.ycombinator.com/item?id=48245721) skipping reviews and pushing to prod directly (https://news.ycombinator.com/item?id=48243651) at a FAANG, no less. And they claimed it in such a manner that this is objectively the only proper way to do the work, and all other approaches are doomed. How is this anything else than peddling unfounded LLM-hype, I do not know.
The example above might seem like a singular episode, but they have been doing it over and over for the last year and a half, so this is now a pattern. No actual evidence for any of their claims is provided, so the only thing left is AI-hype, and pretty wild at that. So why would a reasonable person, with ostensibly enough money to retire (https://news.ycombinator.com/item?id=48252297), ostensibly working at a FAANG, spend all their days spreading unfounded AI-hype in a degrading manner on HN and defending Anthropic? Given the vested financial interest in the technology, the most plausible answer here is that they are paid to do so.
I have explicitly stated more than once, beforehand, what would have given you a benefit of doubt from my perspective. Even after reading all that you opted for answer evasion, rather than providing any substance, as you do with all the questions addressed to you. I do understand why you had to do it here though, because the claims regarding AI and LLM you have made before are even more outlandish than what has been posted over the last 10 days and similarly without a single shred of proof.
In other words, people who are productive get more done when you scale up what they're already doing, and people who aren't productive will not magically become productive when you scale up what they're already doing. That's incredibly obvious, because we've seen how this plays out repeatedly in so many different ways (lines of code, commits, tickets closed, etc.), and it has nothing to do with tokens or even programming, but just how trying to manage people works.
The inability of leaders to understand Goodhart’s Law is always a sight to behold. They see a number go up and pat themselves on the back for how well their employees are making it go up without ever wondering if the thing they care about is happening.
That’s an even worse characterization of them isn’t it? They don’t even care about the end result just the metric. That would take them from clueless to malicious.
I wouldn't describe it as malice. If your job is to make the line go up, you make the line go up, and are rewarded for doing so, then you have done your job.
The point of this was always to explore what is possible with AI as quickly as possible. Obviously, there is going to be a lot of waste, but the 5-10% of employees who are truly thinking about it and discovering novel applications are what you are truly after. Because right now, you effectively have a giant, as of yet poorly explored space of potential uses.
Anyone who can find the actually valuable portions of the space early has a potentially huge competitive advantage. Even if the result of the experiment is the negative that AI is actually mostly not that useful, that is still extremely useful information in a time of great uncertainty regarding outcomes.
The bottom line is that this approach may be expensive, but if you have the money to burn, it's far from the worst strategy if you are trying to position yourself correctly for the future.
What’s the huge advantage though? Adopting workflows that give big productivity gains is relatively easy even for big corporations. It’s only an advantage if you can keep it secret.
OTOH maybe we’re in for a future of patenting prompts.
The thing I don't get though, is that most people just don't have that much work they need to do. I can use AI to pretty easily get my work done just via the regular chat interfaces. But because of the tokenmaxxing metrics that leadership tracks, I end up just having the AI deliberate for hours on random things just so that I can boost my token numbers. I think tokenmaxxing for the end goal you described is only realistic when the engineers are truly buried under a backlog of work.
Not being buried under a backlog of work is one aspect, and the other is that the sheer _urgency_ of these efforts makes it look like companies like Uber could be displaced in a year or two by someone who gets lucky with AI use.
Which absolutely isn’t the case. Even if someone would manage to overtake a market leader on tech merit alone, within 1-2 years, thanks to AI, markets don’t swing on such short notices. The fake urgency is absolutely psychotic.
I think unfortunately it's not about what seems obvious, or even what seems more likely, but about what seems retrospectively justifiable regardless of outcome.
The incentive structure of this type of decision is 'absolutely under no circumstances existentially mess up'. Ostensibly with respect to the organisation, but in actual reality much more so with respect to the individual(s) involved in the decision.
If everyone else is doing something that kind of obviously makes no sense, and you decide to break from the crowd by instead doing what does make sense, then there's a pretty solid chance of gaining a temporary edge while reality resolves the truth. But those gains probably won't matter all that much for the organisation, or indeed your position within it. It's a solid chance of an unimportant gain.
However on the other hand, there's a tail risk that something very unexpected happens and the thing everyone's doing that makes no sense actually turns out to make sense - sometimes even for entirely unpredictable incidental reasons - and then, well, you're in trouble. Not necessarily 'you' the organisation.. they'll likely be able to catch up and it won't matter that much. But for 'you' personally, the decision maker, it's very much not good.
As a bonus, in the much more likely scenario that the thing that makes no sense turns out to indeed make no sense, you're in the same boat as everyone else, there's no relative loss, and most importantly you don't stick out as someone who did something as risky as to go against the prevailing, albeit pretty clearly nonsensical, sentiment.
So basically, game theory tells you pretty quickly to just go with the thing that makes no sense if you're optimising for some (weighted) cross of what's best for the organisation and yourself as the decision maker.
Limits are beneficial. They should be treated as a design feature, not just a stopgap.
When something is abundant, people tend to waste it.
I’m perfectly happy with my base subscriptions. I have Claude Code and Codex monthly subs, plus a yearly Google AI Pro account because it was a logical upgrade from the cloud storage plan I already had. I think it worked out to something like an extra $10/month for the AI features.
I constantly rotate between them during the week, managing tokens carefully, cleaning sessions and contexts as soon as possible, and being intentional about usage.
I honestly don’t understand the appeal of these ultra-expensive max subscriptions.
It reminds me of that flying orb toy I bought for the kids a few years ago. The battery only lasted about 10 minutes, and the kids would go ape shit crazy while it worked. Then it needed a 30-minute recharge, which created a natural cooldown period.
I actually considered that a good feature. I would never want the thing running nonstop.
I think companies are reluctantly realizing that AI is not a magic genie in a bottle, and is instead a tool.
Still very valuable. They just need to have strategies that match what the tools are capable of - not strategies that involve "rub the magic lamp and increase profits 80%".
If the market is rewarding companies going after the "rub the lamp" strategy, they're going to say they're doing that to juice stock prices.
Maybe the market is finally realizing blindly spending billions on LLMs with almost no strategy is not a good strategy.
"He said that, based on talks with Uber's senior engineering leaders, he realized higher token usage did not translate into a proportional increase in useful consumer features."
He's saying that like it's some grand epiphany and not the most self-evident, obvious thing I've heard this month. Some of the literal dumbest people on earth are in charge of these major companies.
not only this month, but it is the basic statement of the single most well known 50 year old book in software project management lol. At this point we need to wipe the slate clean and start over, the industry is run by illiterates.
This is also Uber we’re talking about. The company that famously developed a massively engineered ledger to track every event across the entire company, globally consistently, forever, in a single database. This definitely adds enormous value to the bottom line!
The fact that a company with such a ledger has trouble advocating for AI-maxxxing will make watching the "ur holding ur AI wrong bro"-reactions all the more hilarious.
As with many things, users will discover a happy medium. There is scope for a lot of productivity gain here if the C-suite is willing to understand the tech and work with engineers rather than whatever Dario Amodei is selling.
tokenmaxxing is becoming harder to justify could be a change in the labor market => when capital was free the companies optimized aggressively around retention and internal status spending but high rates + slow growth oblige firms to back toward productivity and operating leverage.
I have Opus 4.7 at work at 15x. Burns through tokens like water. It feels like one of these new mega datacenters is just for me. I'd love to know what the bill is, but we're just encouraged to do as much AI as possible.
I'd be interested to know if this is about individual employee AI usage, or use of AI tokens in production features, or both - and assuming both, what the split is.
I can see how Uber could burn unbelievable amounts of tokens if they start running internal features that run a bunch of prompts against every completed ride, or every customer profile, for example.
Or maybe this is about employee usage, but they introduced some stupid "you get evaluated on how many tokens you used" thing a couple of months ago when that was trendy and are just beginning to notice how much that cost?
The number of product teams who have shipped expensive-to-operate AI features is wayyyy up there, and for many of the scenarios I've seen, customers simply don't care or are unwilling to pay significant premium for access to it.
At the same time I'm starting to see some direction from people in leadership that I should "use the right model for the job" and things along those lines, which is a very, very different line from what I was hearing 12 months ago.
My continued prediction is that we are going to see a tweak on the SaaS model where the sweet spot moves to metered usage pricing of really fine-grained API-based access for apps which traditionally have been operated solely via the UI. Long term the trend is going to be "we'll house the data, enrich it, maintain it, provide fine-grained API access over it tailored to model usage, and you bring the model" with some services opting to give you the model interaction layer/harness. IOW I don't think SaaS is dead. Far from it. However, I do think that a lot of people are going to be looking to interact with SaaS apps via their own models with APIs that support those use cases better than a lot of those APIs do today.
Surprisingly, Uber hasn’t had a mass layoff since 2020. The company currently has ~34,000 FTEs, which I personally think is insanely bloated for what amounts to a taxi + food delivery app.
No wonder they need to extract such a massive cut. I really have no hope we will ever get to efficient middle-men who take least they can for good of both sides beyond them.
Replace Tokens with Gas, or water or healthcare or anything - and it's foolish. You shouldn't let the seller dictate what amount you need of something.
Smart engineers are figuring out how to best use their tokens - as tokenmaxing is just as silly as gasmaxing your car.
On token consumption and efficiency... AI-champion guy in my prev company made a metric, like how many tokens are spend per line of generated code, and even put a leaderboard based on that metric, praising guys with the cheapest LOC.
Are you telling me, it did not make them "productive" in ways most of (us non-AI-boosters) "cannot even begin to imagine"? Who could've thought - a lot of average stuff, still ends up producing average result?
The black bill that is coming that nobody is prepared for is that the value of a token varies greatly depending on the human. Companies will quickly find out its much better to give your top 10% engineers a lot more tokens and lay off your average engineers. The 10x engineer will become the 1000x engineer.
I actually do think token maxing is good, but they should have limited it per user. I find it reallly hard to get people to max out the Claude $100 plan, let alone the $200 plan. I understand the enterprise plans are different and more expensive, which is how you get these kinds of issues. But encouraging people to try things with AI is very important, and some amount of token maxing is importsnt.
The business. Employees are hesitant to learn new tools that are very different from what they are used to, so if your business believes that AI is a productivity multiplier, it behooves it to incentivize individual employees to learn to use the tool.
I think the key word is “believes”. There is no proof that AI usage improves productivity. Token maxing is essentially customers paying to try and prove a business’s unsubstantiated claim. The AI companies should be proving their claims themselves not the other way around.
I do think AI has value and is useful but the idea of token maxing is ridiculous.
Sure; I described it that way deliberately. I think you can reasonably disagree with whether or not AI improves efficiency, but regardless, you can agree that if a business believes AI does, it will logically conclude that it should incentivize employees to learn to use AI.
> Employees are hesitant to learn new tools that are very different from what they are used to...
That simply isn't true for technical employees (like software devs). They are so hungry to get stuff done that you have to hold them back from adopting new tools which they think can make them work more effectively. Tech guys will set up entire shadow IT departments just to get around corporate restrictions that are limiting their productivity.
No, if software devs are not using LLMs for programming, that is proof that the tool isn't actually useful for them. It doesn't mean "they need to be forced to use it", because they didn't need to be forced to use any of the tools which came before it.
It's not hard for most people now. 6 months ago when agents first started getting big, I genuinely didn't know enough about AI tools to understand how it was possible to use so many tokens, and I don't think I would have bothered to find time to learn without a kick.
Do you find it hard to max out, or do you find it hard to productively max out?
It's like paying drivers per gallon of fuel consumed and then acting all surprised that you see them revving their engine while waiting at a red light.
>"He said that, based on talks with Uber's senior engineering leaders, he realized higher token usage did not translate into a proportional increase in useful consumer features."
Goodhart's law strikes again at someone with enough power to be both ignorant of it and make others suffer their ignorance. You cannot simply measure productivity by tokens spent just like you can't measure it by hours spent in a chair at a desk.
What if... we stop for a moment, and then, after thinking for a moment, we stop hammering nails with a microscope, and stop using token usage as a metric of productivity?
The crazy thing is their salary does not actually benefit from riding these trends. Unless it's equally/even more clueless board level pressure with ulterior motives (i.e., lifting their other AI investments or the sector as a whole).
Every c suite in the country is panicking about being left behind, from their perspective it’s either token max or fade into obscurity, or at least that’s what they were sold
I don't think that's accurate. I think every C suite in the country is looking to do away with labor's leverage as much as possible. I think this is a cultural thing more than anything else, C suite + investors looking to get rid of those pesky humans required to prop up their lifestyles. AI is the most credible path toward that. Short, medium or long term returns be damned, this is a reconfiguration of society and they want to shed what they consider to be baggage.
Like anything it's a mixed bag. I am certainly working with people who I think truly believe the "max out on AI usage or become irrelevant" line. There are people who will privately let you know they're just working with the current meta the best way they can, but others who are drunk on kool aid.
Trying to operate as a rational, thinking person in a lot of environments right now feels impossible. Rational thought is being treated like AI skepticism.
Please. These are the same people that force their employees to use Microsoft teams because slack is $5 an employee a month. They're not going to sit idly by while employees burn thousands a month in tokens.
It depends on which people you're referring to. The allocation toward AI budget has been so massive that I think a lot of businesses are way behind on trying to assess value for dollar for the AI-related crud they're shelling out for.
Everyone is feeling it out but the vast majority of spend has been subscription based. Some outliers may have used a massive amount of tokens but companies didn't pay for that.
That VC funded gravy train is likely coming to an end. But fortunately there are also reasonably efficient models now so that the tokenmaxxers can still make the (much cheaper) tokens go brrrr.
Those reasonably efficient models assume you use a harness that supports them well, the one size fits all harness of Claude desktop or codex does not support what you want well, and that’s intentional. It’s contradictory that these AI companies will continue to brrrr to the moon and return on AI spend requires discipline…
The next recession (and there's always a next recession) will clean up this AI bubble. The actually useful products and companies will make it the rest goes down.
Yes, but unlike the dot com bubble we’ll be left with half finished (or not even started) abandoned data center projects, instead of reasonably reusable fiber and ISP infra
its a herd mentality, its a lot easier to follow the louder voices than to spend time understanding how it impacts your own particular business. Because google does this way, or apple does this way is a common argument in lot of feature/business decisions
I deeply believe this but have no strong evidence. Revenue has always been a cure all remedy. This will keep model providers alive along with the very wide range of companies that are experiencing growth with them (from chips to backhoes), for a time anyway. If/when that house of cards starts going in the other direction there’s going to be widespread pain. By analogy the nonsense of the dotcoms and that crash had a very direct impact on their suppliers (e.g. telecoms). My only advice is to let the Microsoft’s and Meta’s do the tokenmaxxing, and don’t get suckered into the idea you (startup, individual, etc) should be playing that game.
They get paid for saying whatever VCs want to hear and now that thing is "we have now become an AI-native company". The thing I'm still trying to understand is who is scamming whom
This is actually pretty well-understood if people wanted to do it.. contrary to popular belief it's more about responsible governance than economics and not really a pro/anti capitalism thing
What if the goal of an economic system was to support everyone instead of maximizing the upside for winners? Perhaps that's the sort of change necessary for improvement. Perhaps having billionaires is the failure state.
A goal fails - who sets a goal? The keyword is system.
An economic system needs something like a Nash equilibrium where defectors are sufficiently discouraged (and cooperators are rewarded as you imply). https://en.wikipedia.org/wiki/Nash_equilibrium
There is a complete lack of courage in the leadership of tech companies today, and top-down AI mandates are just another manifestation.
True visionaries think outside the box, but most tech executives are forcing their employees into black boxes, out of fear of not doing exactly what their competitors are doing.
We have lemmings for leaders, and that means that—much like the LLMs that are being shoehorned into everything—there isn’t room for original thinking. Everyone’s strategy looks exactly the same.
If one is a CxO who's looking out for one's job security, herd-like behavior is the safest option, due to the (near universal) structure of "performance"-based executive remuneration.
First is that despite a lot of waste, some innovation will arise from an enterprising employee finding some interesting use case. A lot of the tokenmaxxing is just waste, but out of that waste may arise a small number of genuinely powerful use cases.
Second is that many workers will be entrenched in their ways. If your executive goal is to achieve the above (find innovative ways of using AI), then you need to move everyone to use it. Most will just waste tokens, but someone may find a novel and useful way of using it that benefits the organization. It is difficult to achieve these without forcing people to act since their default is to follow the well-worn grooves.
So mandates like these are a top-down forcing function like a slime mold feeling out different paths to find resources.
Some devs in my org have fully embraced AI; some would not even use AI if not for leadership mandates and linking usage to performance reviews (I know, I think this is stupid, too). I can see why mandates could be useful since some folks definitely won't be inclined to use AI.
Absolutely, but most management are not leaders, the moment someone pushes the idea to stack rank based on token usage, it gets approved and some genuine people will be impacted.
Post-ZIRP era proved there are very few strong leaders, before that everyone was behaving like they're most amazing leader because they read some books and raised $10M
> but out of that waste may arise a small number of genuinely powerful use cases.
Imagine you employ me as a hotel manager, and I come to you and say: "sure I spent all our food budget internationally in three months, and sure I have nothing really to show for it, but for those three months, we had a lot of food fights"
Your manager then goes on to explain they not only need more money to cover the food budget, but also they need to quituple the cleaning budget too.
Oh and the service level has dropped, because not all clients liked being in the middle of a food fight.
However "we might have some innovation in the food delivery system of our hotel chain"
> we might have some innovation in the food delivery system of our hotel chain
This is really relative to the size of that innovation, isn't it?
> Imagine you employ me as a hotel manager, and I come to you and say: "sure I spent all our food budget internationally in three months, and sure I have nothing really to show for it, but for those three months, we had a lot of food fights"
This is exactly how startups and VC funding works, isn't it? You have an idea, give you cash to burn to prove the idea and business model. Many teams and ideas fail. But some small number of unicorns produce outsized returns to keep the whole thing going.
It's not how it should work, because food fights are stupid and have no upside.
Even if everyone else is having them.
It's not a fair analogy because AI isn't completely stupid, and there are situations where it does provide a benefit.
But a rational business would ask if the upside is worth the cost, if the pipeline can be restructured to concentrate and amplify the benefits, if some elements are better being done the old way, if there are strategic threats if tokens become much more expensive, and so on.
Instead we're getting a wave of "Cut workers, cut costs, derp" and that's as far as the "thinking" goes.
The worst thing about AI is that it shows how shallow and stupid current C-suites are.
The US used to have real tech visionaries. Now it has tech cargo cultists, all chasing an IPO cash out and hoping the music doesn't stop before they get their bag.
Imagine you employ me as a hotel manager, and I come to you and say: "sure I spent all our food budget internationally in three months, but we invented this new dish and now our restaurant is the hottest in town. Sure 95% of the food was wasted but now we can stop the waste and keep the popular dish."
The intention was to force everyone to experiment with the new ingredient monsanto recently GMO'd. Of course a lot of our employees suck, so food fights were expected, but luckily some of the employees created something great.
Indeed, but that's not a bad thing. If monkeys can produce the next Shakespeare, that will be wildly popular and profitable for the company that did it, justifying the initial waste, just as VC does with companies as a whole.
> Some devs in my org have fully embraced AI; some would not even use AI
So if the people who embrace AI areore successful then the others will follow. Just like every other new tech. Why does AI have to be forced? What's the hurry? Especially when there's no clear example of a company jumping ahead because of their use of it.
It's idiots being driven by FUD. That's the reason.
There are definitely key windows here for innovation driven by competition.
There's also a need to quickly adopt and understand the technology; take the Internet for example. If we were talking about the Internet, forcing teams to build and publish web pages would be one valid way to get teams comfortable with the tech, the workflow, the shift in how to propagate and convey information to an audience.
Without a mandate, many teams won't adopt the Internet as a medium of information exchange because their processes work just fine and have worked for the last 20 years.
I think it's fair to put AI in a similar light. Unless teams adopt it and use it, it's hard for an org to understand how to get value out of this technology and how it affects existing processes and assumptions.
And you think forcing blockbuster's software teams to use the Web would have changed that? You don't think they were using the web for all their corporate communications systems? I very much think they were, and getting blindsided by streaming had probably nothing to do with blockbuster's existing engineering teams not understanding the Internet. Their product teams didn't understand it, but they wouldn't be the ones being "forced to write webpages" either
> There are definitely key windows here for innovation driven by competition.
Those were always there, and will always be there. The type of time frames people are getting anxious about now rarely work in the real world, though, where potential customers don’t just switch products/service provides unless they’re facing catastrophic outcomes if they don’t.
And AI is not making the difference there that people think. I worked on a product that entered the market as a newcomer, wooed plenty of customers, and even though everyone _wanted_ it, only customers _urgently_ looking for a solution got on board quick (within <6 months).
Ironically enough, the product pivoting to Agentic AI hard killed a ton of momentum and interest from customers, despite exciting investors.
Seriously. No mandates at my company. In 2023 and 2024 i had access to Claude, but frankly it wasn't until 2025 that i found the models useful enough, now i use them every day. Nobody forced me. Had they forced me, I'd probably have quit. Once the tools were sufficiently mature and verifiably helpful, people like me all over the company naturally picked up the tools too.
Sure, indiscriminate tokenmaxxing is a gamble that can pay off sometimes. However, I think that the decision to take any gamble should be made by someone who will bear responsibility for the downside as well as the upside. I would prefer to search for new usages in a more strategic way. I agree that experimentation is a great way to learn if done intelligently and with limits. Full “Monte Carlo” makes sense when ops are cheap enough. It seems some orgs don’t think tokens are cheap enough yet.
> I would prefer to search for new usages in a more strategic way
I think this is very, very hard for orgs to do.
Looking back at the Internet, who would have thought that it would eventually create a Netflix, Amazon, Shopify, Spotify, Google Maps, etc. Just wild the things that ended up coming out of pushing bits over a wire with few simple protocols.
In an ideal world, you make strategic bets, but I can also see the case for the opposite this early in the lifecycle of a technology. You just don't know until you try.
Mid/late 2023, it wasn't at all obvious that it would take over coding that fast.
If it were that simple and obvious, Blockbuster would have beat everyone to streaming. Sears would have digitized their catalog and used their vast brick-and-mortar stores as fulfillment centers for same-day shipping.
None of these shifts were obviously the right bet and many organizations lost because they missed the opportunity. Now orgs are on the same horizon and I can see why they don't want to miss this window.
Blockbuster actually did try to beat everyone to streaming. Notably, Blockbuster and Enron [1] entered into a 20-year partnership for online video delivery.
Sears was a different story, in that they were a real estate company with a store front and retail real estate took a nosedive due to ecommerce. But that's a different discussion.
There was an amusing post about judging developers based on token usage where some user on HN here was pushing this idea “ICs don’t like it but this is the best way to evaluate” (something like that).
They have a whole management team and can’t seem to find a way to judge or god forbid encourage developers…
Because the higher up you go in management, the more "strategy" is a Plato's Cave like interpretation of what better/bigger/whatever competitors are doing.
And they know things like “Dude isn’t a high performer metrics wise but his work is solid.” Arguably some of the most difficult things to know from a management perspective.
Yeah courage will get you fired. Whether it be about idiot product decisions, or about how your bosses treat your coworkers. That’s the consequence of letting sociopaths get in charge.
I feel like individually, if you sat down with literally any reasonable person on the planet they would arrive at and/or agree with the tenor here.
I'd be curious to hear from people well versed in group psychology/dynamics and/or just a lot of leadership/people experience: what leads people to this type of thinking once they get in a group setting? It just... seems endemic at this point.
Obviously nobody here is going to know what I do or don't know, but I'm just increasingly curious what I am not understanding about this type of thing. It seems so obvious, yet that makes me ever more suspect that I'm oversimplifying it, or just totally ignorant about the problem in general.
It's because the average organization has lots of people who don't care about their own productivity and won't adopt new tools or processes unless forced to. This is true of most new tech - lots of workers had to be forced into using computers - but AI also has some other bumps to cross like lots of people who tried early models and then wrote them off, not realizing how fast they'd improve. And most orgs have no infrastructure or processes for allocating individuals token budgets, and most employees have no experience of properly deploying budgets.
Roll it all together and saying "just use it dammit" has some obvious advantages:
1. It's clear.
2. It's simple.
3. It eliminates all excuses employees might come up with for not using it.
The people at the top of these companies aren't stupid. They might have miscalculated how many tokens people can actually use, but that's very hard to calculate because usage is opaque and tools/processes change on a nearly weekly basis. They will eventually build out processes, tools, social conventions and performance metrics that take into account efficiency of token usage. But this is hard! Most managers aren't really assessed on the precise productivity of their teams, for instance, because productivity is often poorly defined.
> what leads people to this type of thinking once they get in a group setting
Game theory! The downside of being brave vastly outweighs the upside. For the C-suite, there is no cost to herdlike-behavior, regardless of the outcome. However, there is a very high personal downside to being a maverick, and your board later discovers you made the wrong choice against the grain. The upside of being maverick and right is very limited.
Once a behavior has become mainstream, hopping on the bandwagon is no longer individually attributable to decision-makers, but is seen (and reported) as a macro-economic phenomenon: Nadella, Zuckerberg and Bezos didn't overhire - the American tech industry overhired.
This is a consequence of elements of monopoly power existing in your organization. When you don't have to compete for income you honestly forget how. Then the company becomes a cargo cult of bad ideas driven by managers struggling to differentiate themselves.
we are going through our second AI transformation, the first one didn't work that well because the tools were shit.
Whats happening now and whos driving it is interesting. The CEO has a license for this new tool (think one of the top 4, Qwen Claude, Gemini, openAI) and really likes it. So much so that they (non coder) are making lots of little single page web apps.
The COO is bollocks deep in AI, and is saying that we cannot buy any SaaS products anymore. We must make it ourselves.
The engineering manager has seen this as an opportunity to build out a brand for engineering (its a small department in a medium sized company) by delivering quickly what the large year long efforts cant.
This has formed a slopnexus where PoCs are spun up left right and centre, but there isn't much time or thought going in to making them sustainable.
What started out as a (simple ish) asset management tool, neatly scoped into a deliverable PoC has morphed into a 5 product as one monster.
Its a mess that will either lead to burn out or disaster.
Just...wow. That sounds awful, and I'm sorry for you, but I believe many other companies will or are already following that same path.
And myself being an infrastructure guy that needs to maintain all these PoCs that are now suddenly critical for production, it's the perfect nightmare.
And mind you, that dynamic always existed to a certain degree (laptop on a desk that runs some ugly Python script that does half the work of the BizOps team? Check. GCP account attached to the GSuite running a random instance for finance when the company is 101% on AWS? Check. Spreadsheet with macros that sends emails via Outlook as a mailing list manager? Check.) but at least when you discovered that you could scold them and tell them that we need to migrate this to a proper system because security.
But nowadays with vibe-coded internal apps...it's a challenge.
There is probably some opportunity here for a centralized, internal only LLM proxy which injects AGENT.md and skills and permits switching backend providers.
> what leads people to this type of thinking once they get in a group setting? It just... seems endemic at this point.
Large and fascinating topic I'm researching, very relevant for agentic AI and ML too. One way that groups can fail is that they just don't work to dampen / vote out individual errors properly (see PAC learning, Condorcet). Other kinds of errors only occur in groups, and can occur even when constituents individually aren't actually wrong. Some related stuff is:
The last is probably the most relevant here and made worse by the negative effects of hierarchy. To quote one section:
> The negative effects of informational cascades sometimes become a legal concern and laws have been enacted to neutralize them. Ward Farnsworth, a law professor, analyzed the legal aspects of informational cascades and gave several examples in his book The Legal Analyst: in many military courts, the officers voting to decide a case vote in reverse rank order (the officer of the lowest rank votes first), and he suggested it may be done so the lower-ranked officers would not be tempted by the cascade to vote with the more senior officers, who are believed to have more accurate judgement;
For token-maxxing, our "senior officers" are just executives, and line workers aren't going to vote. Who is the senior officer for those senior officers? It's not shareholders! It's really the executives of even bigger companies, because that is the actually applicable promotion ladder. It's all kind of obvious, but also a genuinely better explanation than "monkey see monkey do". These are just the simpler things, and there's more gnarly dilemmas in https://en.wikipedia.org/wiki/Common_knowledge_(logic)
Wow, amazing answer! I have a lot of reading and then thinking to do, but if you are documenting your research anywhere, I'd greatly appreciate somewhere to follow it.
If there are any tech CEOs out there reading, I can offer my services. I will pointlessly burn unfathomable amounts of tokens, in parallel, 24 hours a day, 7 days a week, all for you. Think big big big numbers of tokens, you know whats cooler than a trillion tokens, a quadrillion tokens.
Lets talk my bonus, I will open the bidding at $1 per token.
> and stop using token usage as a metric of productivity
I participate in some management-focused online communities. It’s crazy how many threads there are from frustrated managers trying to get their teams to stop thinking that their token use will be used as a proxy for their performance.
I think a few dumb companies did this and then it spread across social media, triggering a mass panic from engineers afraid their companies will be doing the same thing.
It’s getting so bad that the conversation is shifting to how to identify and coach the token-maxxers to stop wasting the team’s budget every week.
> managers trying to get their teams to stop thinking that their token use will be used as a proxy for their performance.
Because it is going to happen. Do you think metrics are tracked for fun?
Even if current leaders don't do it, next people might do it, how do you tell new leaders that we don't look at this metric? Metric exists to take action based on it
You missed the second half of my comment: There’s more to a metric than making one number go up. It’s becoming a real problem when people use 10X more tokens to get similar work done because they’re tokenmaxxing.
Nobody wants 90% of their token budget spend going to the 10% of people wasting them for number-go-up purposes.
Inefficient token use is going to become a metric.
Sounds to me like you are advocating the decimation of the technology sector and a global recession that could last the better part of a decade, buddy!
You mean… increasing AI budget has no direct relationship with productivity and therefore revenue? It’s not that simple? But… my TEDx talk and LinkedIn ramblings…
> ... stop using token usage as a metric of productivity?
and tokenmaxxing is even worse due to https://en.wikipedia.org/wiki/Goodhart%27s_law because whatever you measure with tokens, once you start "tokenmaxxing" you have no measure to look at
Tokenmaxxing is so dumb. You should never show your team how exactly you're measuring their performance; people will optimize for the metric, not the actual performance.
Classic Goodhart’s Law: when a measure becomes a target, it ceases to be a good measure.
LLMs are great, I can understand using them in general. I can even understand chasing 100% weekly usage if you're using the gacha-like subscriptions since that's how you get the most value out of what you paid for.
The way these corporations are going about it is completely insane though. They're essentially ordering their employees to set money on fire or be fired themselves. The more money you burn on tokens at insane API rates, the better an employee you are. Absolutely mind boggling.
Protip: skunkworks type side projects are a great way to do tokenmaxxing when you don’t have enough work coming in, but still need to burn tokens to look productive. And because side projects are only governed by you, you can truly go nuts and let scope creep run wild. Soon enough, you’ll be one of those engineers burning six figures a month on AI and people will be in awe of your abilities, probably even elevating you to key AI evangelist positions within your company. And if you actually create something cool, you’ll be praised for your use of AI, and you can just say you built it all in a day or two instead of slacking off for months on your real work.
AI productivity hasn't been well studied yet, but I'm betting that we'll end up with some variation on Price's Law, I.E. some small subset of workers get most of the benefit, while most just burn tokens with little to show for it.
I also want to call out the false productivity opportunities AI offers. There are whole teams building their own "gas town" and not shipping features.
Not all tokens are created equal. It's easy to use a ton of tokens by having agents work together in parallel. That's basically the equivalent as people spending time in meetings, hardly a productivity win. As with everything in development, results matter, how you get there doesn't (unless you're a bad manager).
I hereby suggest you take the fragmentary excerpts of the infamous erotic stage play The Lusty Argonian Maid shown in The Elder Scrolls series of games and extrapolate them to 100,000 additional full-length acts.
many of these leading AI companies are operating at large losses and subsidizing users with VC money. Profitability will entail having to impose greater limits and raising prices, so this will reduce to some degree the value proposition of AI compared to humans.
As soon as tokens stop stop being subsidized, heavy agentic use will become as least as expensive than paying an (entry level) employee. When this happens many companies will trade off havy tolen usage for (maybe a bit slower, bit less accurate) employees again.
DeepSeek is an open weights model. It's possible the hosted versions are subsidized, but we know what it costs to run locally. And it's expensive, but it's also pretty clearly cheaper than an employee.
Of course, the latest DeepSeek models are not as good as Claude, but they're not super far off either.
They're not far off, getting the same seamless integration as hosted models is a full time job. I think what just happened is that devops is about to explode. What will naturally follow is local hosting of all the things when people realize subscription costs for cloud-whatever are absurd.
Gitlab is going to take off? This is not investment advice.
> What will naturally follow is local hosting of all the things when people realize subscription costs for cloud-whatever are absurd.
Even acknowledging we don't know exactly what costs would look like in a world without VC money, wouldn't hosting models logically be cheaper to do at scale in a data center?
When I compared to the cost of running DeepSeek locally, I meant that we can treat that cost as a price ceiling, not the floor.
Like how server hosting at scale in a datacenter is cheaper than running your own datacenter? Despite ~every company consistently concluding that hosting their own stuff is several multiples cheaper?
No, I think local stuff using also-useful-for-other-things hardware will vastly undercut cloud hosting when the free money pipeline shuts down, and will stay that way for roughly forever. That doesn't mean cloud stuff isn't useful, clearly it is, but adding another company in the middle is rarely the solution for reducing costs.
When you use DeepSeek’s first-party API, you are giving them your token stream. This has some training value, but it also has incredible amounts of, well, business intelligence value. When you tell AWS your secrets or your customer data, you can be fairly confident they won’t abuse that knowledge. When you give this data to, say, OpenAI, they more or less promise not to abuse it if you’re on an appropriate business plan. If you give it to DeepSeek, even incidentally as something your agent reads, I would be quite surprised if DeepSeek doesn’t mine it for whatever purpose they or the government feel is appropriate.
The risk of letting your agent read .env goes far beyond the risk that the agent itself does something you don’t like with the contents.
I have been saying the same for while. Someone always says "but Anthropic is making money on their API" or "But it's inference will get cheaper". But I don't believe it. first all the investments have to payed off at some point and second of all there are other things that cost money. I don't believe that any of them have a positive balance sheet.
I also don't think that blitz scaling will work like with Uber. The engineers are still there. We can work without the LLM tools.
If by "investments will pay off" you mean major profits, that's never going to happen as long as scaling laws hold. All revenue will just go to financing more compute, and either we hit AGI or have the greatest economic collapse in modern history.
The world will look drastically different 5 years from now; for the better or worse, so save every penny (especially if you work in tech).
Maybe this just counts as “light use” since I’m a hobbyist programmer and I only run one coding agent session at a time, but I get about as much done as I did back when I was working while spending a lot of time browsing the Internet, etc.
I’ve spent $10-$20 a day using Claude to write code and closer to $5 a day now that I mostly use Deepseek and GLM, using API pricing (no subscriptions) since I don’t use Claude Code.
This is a rounding error for a company. So I think there’s plenty of room to use AI extensively while being more cost-conscious.
What's funny is that this apparently wasn't something that the Uber COO seemed to think about when their company is arguably one of the most successful ever at the "subsidize to drive down costs until you capture nearly the entire market" strategy.
More straightforward to talk about the hardware directly. Full Kimi K2.6 needs an 8x H200 node to run and serve around 20 heavy users. You can rent an 8x H200 node for around $30/hr.
I'd imagine GPT-5.5 and Claude Opus 4.7 could run just fine on a 16x H200 node and serve at least 10 heavy users without the token output getting choppy.
You're assuming the price won't come down as the tech matures. That seems like a big assumption, considering how quickly open weights models are catching up to frontier models, and how little effort has been invested so far in optimizing inference costs.
It's especially a crazy assumption to make relative to the costs of employing a human. The costs of paying an entry level employee are unlikely to go down at all, and even if those costs do decline, there's a floor they can't drop below (minimum wage at the extreme end), whereas companies are free to optimize agentic costs as close to zero as possible.
So you are assuming that a cost which is extremely susceptible to optimization but which no one has yet seriously attempted to minimize will remain perpetually above a cost which is much less susceptible to optimization, is already subject to enormous efforts to minimize, and has a legally mandated floor. That seems like a bad bet.
A significant caveat is that there is a pricing mismatch that makes it so first party's can subsidize quite heavily.
Agents are expensive in large part because tool calls require round trips. It's because these APIs are stateless and not streaming so you have to resend the whole context each time. This means you have roughly #tool calls x 1/2 context size cached input tokens over any given session. Most API providers overcharge you by a huge amount for cached tokens. A exception being Deepseek. Paying OpenAI $0.05 for 100k cached GPT5.5 tokens during a possibly 2 second round trip agent tool call is like paying $100/hr for what is likely to be ~10 to 20 GB of VRAM residence (holding the KV cache).
Or it got offloaded to NVME and you are paying $0.05 for that much PCIe bandwidth.
I think if local models catch up with current SOTA then that might not happen. Either way, I'm don't think the long-term for OAI, Anthropic etc. really holds up.
Now we are going to get a new profession. Token Engineer! They will be experts on tokenmaxxing! The job growth that the billionaire CEOs promised us from AI is finally here!
I like this too. I have been intentionally -maxxingmaxxing to get the meme out there. It's a good canary to sort out who gets the spicy takes from the pedestrians who probably still copy-paste into the ChatGPT web app like a psychopath.
But on a more serious note, do we know how much Uber spent per technical employee/month? I assume it is far more than even any of those $200 "max ai" plans.
And the other question is how much the public would be willing to spend, in my estimation this is as "cheap" as it will ever get (main-stream at least).
> I assume it is far more than even any of those $200 "max ai" plans.
Am in a random small company, colleague spent 100 EUR a day on Sonnet through AWS Bedrock (needed to use a EU region). Paying for tokens will get you in a deep hole financially compared to any of the subscriptions, unless it's like DeepSeek or one of the other models that are priced a bit better, though that's also a tradeoff in what they can/cannot do and also where the data goes. Ended up trying out the Mistral subscription for the US stuff btw, it was fine.
Exactly. But I did find an article ([1]) and spend doesn't seem that high per engineer ($150 to $250 per eng) - at least on average, I assume the costs were skyrocketing towards the end.
> Adoption climbed from 32 percent of engineers in February to 84 percent classified as agentic coding users by March. By spring, 95 percent of Uber engineers used artificial intelligence tools monthly, and roughly 70 percent of committed code originated from those tools. About 11 percent of live backend updates were written by agents with no human in the loop, according to Uber's own disclosures.
> The numbers behind the spend are what make the story instructive rather than anecdotal. Monthly cost per engineer ranged from $150 to $250 on average, with power users running between $500 and $2,000.
My guess is that the reason to rethink AI-spend was probably the exponential growth in cost over time, and tokenmaxxing payoff not being immediately obvious as mentioned in the article.
Except you won’t because they will threaten to fire you and force you to route all of your AI through data protection proxy to stop exfiltration by filtering and tracking prompts/response tokens.
Probably long term each dev gets their own GPU and runs a model locally I expect. Seems like a more sustainable approach, even if a local model is not absolute SOTA.
GPUs are much more efficient at parallelizing requests for LLMs so it's going to much more efficient to centrally host. Maybe big companies it would make sense to get their own though.
Would you decide its usefulness based on how high the bill is, or how many things you get done while using it?
The former is the issue, and how many companies have been operating. It's like a trucking company ranking driver effectiveness by fuel used instead of by cargo moved.
I remember at Google at around 2007 - 2009, as Google was massively expanding its data centers, there was a lot of unused capacity, especially during off-hours. Any engineer could run as many jobs as they wanted at zero priority, which means the job would be first in line to be killed if a more important task needed the resource.
I did so many interesting experiments with MapReduces that would run overnight.
For a while, I would even build internal services that were basically "free" because I'd just run them all at priority 0.
Over time those services got less and less reliable as overall usage started to increase, so I was forced to either justify the resources or scale back - but that was a good thing.
I feel like something similar would be a good model for AI token use: big tech companies ought to have their own self-hosted LLM data centers to power their own needs, then let employees use off-hours capacity to experiment.
Outside of experimentation, we should be encouraging token efficiency for everyday tasks. Rather than having a certain number of tokens, engineers should be evaluated based on how much they actually get done.
Using a lot of tokens to automate a process that used to require hours of human labor every week? Good use of tokens, should be encouraged.
Using a lot of tokens to debug an easy frontend bug that could have been fixed by hand, and still took you 4 hours to complete? Waste of tokens, should be discouraged.
"Using a lot of tokens to debug an easy frontend bug that could have been fixed by hand, and still took you 4 hours to complete? Waste of tokens, should be discouraged."
Hahahah good luck with that!
For many of us, what is happening now was super obvious. Telling a new formed crack addict (who you wanted to become addicted) to be more thoughtful about their consumption of crack... yeah not gonna work is it.
Most AI front ends seem to be designed for interactive jobs, so they make it hard to define a job that should be done eventually with zero priority. It makes much more sense to do that with spec-driven development (have work done with the human on the loop rather in the loop), but as far as I know that just isn’t well supported by any front end yet (would be happy to be proven wrong, my experience is with Google front ends).
There’s a lot of places where an llm can improve a data pipeline. Like if Claude sonnet was free I’d do a lot of data enrichment.
You’re right of course, but wouldn’t it be more likely that everyone will embrace 10x cheaper Chinese models?
my money is on: eventually frontier model dev and training becomes basic research funded by governments, and LLM operators become essentially private utilities a la ISPs, competing mostly on data center operational costs and occasionally new chip tech to run models cheaper
and governments will keep running massive data centers with classified frontier models for intelligence and propaganda purposes
Don’t think we’ll see similarly logically behaviour from LLM users tbh. A sizeable portion of the user base seems to insist on through opus at every trivial task
Isn’t this like batch processing offered by oai? Your request gets done within 24 hours at half off
https://developers.openai.com/api/docs/guides/batch
If any company announces that they use token consumption as an employee performance signal, for me that's close to a red flag to stay away from that company.
No company with good engineering leadership should act like this is remotely a good idea.
Tokens are the new "lines of code per engineer". Easy to graph, easy to "manage".
...and easier to bill! Back, then noboday had the idea to charge per "lines of code", but today it seems accepted to charge per words processed?
The new TPS reports!
Oh, so that was actually a Token Per Second report! Wild!
I worked at a YC company that was doing this and left last month. I wonder where this all started from, VCs and tech execs are such a monoculture
The where may be the decision makers chasing social media trends. A friend sent me a link to this, this morning, about devs rather than managers, but I suspect it's the same: https://youtu.be/IW3Sbe0Hbgg
Meta does this. Guess what one of the criteria for their recent layoffs was.
Meta tracks token consumption, but has explicitly stated that it is not a primary performance metric. Instead, employees are evaluated on "impact."
Indeed, they also said that previous time off for ill health wasn't a reason either.
but looking at the number of people who had taken leave, it suggests otherwise.
You can believe whatever conspiracy theories you want, of course, but the most straightforward explanation is that when you lay off X,000 or XX,000 people, some number of them will be on leave.
My understanding is that they must layoff some workers on leave, otherwise they would be discriminating
In Europe people on sick leave are protected (under certain conditions).
Megacorps being dishonest and human-hostile under capitalism is hardly a theory.
In London, large number of people on sick leave, or were recently on sick leave were fired. I was there, I saw the people, did the numbers.
If you did the numbers, would you share them with the rest of us instead of just alleging secret facts?
Sure, and I have a bridge to sell you. Or alternately refer you to the inevitability of Goodhart's law.
And my CTO insists that PR count isn't a performance metric. But guess what number gets used the minute people are forced to stack rank (of course they don't call it that, but... that's basically what it is)?
You get what you measure.
Famously honest and on-the-level company Meta, who we can parrot the word of uncritically and unquestioningly.
Do you have any source for this at all? I’ve seen so many different exonerations for Meta’s layoff criteria including claims that engineers using the most AI were laid off because Meta had them build AI tools to replace themselves.
Everyone is oddly confident despite all of the conflicting explanations.
Without any evidence, I would be shocked if performance rating wasn't a factor in the layoffs. But performance rating is not the same thing as AI tool use.
You'd be surprised, I know a few devs in very big tech companies, not faang but you definitely know them, and they all have some kind of token leaderboards, a few told their dev "we don't want you to write a single line of code manually anymore", etc.
I assume the execs perspective is something like: if the top 20% of worker produce 80% of the code with LLMs and the company still works then we can get rid of the bottom 80% of devs and save money
I think there's probably something to token use as some kind of metric. If you aren't using these tools much, you're definitely not going to remain a top contributor. The world is evolving quickly here.
But it's just one signal out of many, and more isn't somehow inherently better beyond a certain point.
Problem is that those 20% depend on the code reviews of the 80% for some form of pushback.
But even if the end goal was to lay off 80% of programmers, shouldn't the 20% to keep be the developers delivering the 80% of the code, regardless of whether they spent the most to do it? Like what if the 20% of workers spending the most tokens were actually the bottom 20% in terms of delivery because they were using the worst prompts and having AI constantly implement 5 different versions of everything, then throwing it all out because their prompt was so bad anyway?
Ah, but "who uses the most tokens" is a number, a number generated by a computer no less. Questions like who delivers lots of high quality work require you to do research and make judgements, which is work.
I worked at Uber from 2022-2025. The engineering culture was pretty abysmal, so it checks out.
As I have snarkily observed at work: if I go $100 over the meal allowance on my business trip, I'll have to have an unpleasant conversation with my manager or finance. If I use $500 in AI tokens unproductively I'll be recognized for being a top AI adopter.
I have seen this type of behavior happen many times in different companies.
For example, at more than one company I've worked for, if you wrote shitty code but got it into "testing" faster than anybody else, you are considered a superior programmer. And then, if you fixed the hundreds of bugs found in your code seen as an extraordinary programmer going above and beyond the call of duty.
Management is always measuring the wrong thing.
The problem is that many companies which had reasonable leadership in the past with the advent of LLM AI started to make rushed (and dubious from my point of view) decisions - using token usage to evaluate an employee performance is just one of them.
There is little new under the big fusion reactor in the sky. I just read a chapter in James Glieck's "The Information" about tokenmaxxing in the telegraphy industry. There used to be a big market for code books to reduce the per-character charges for sending telegrams. Compression was cash in the pocket. The telegraph companies discouraged the practice but were forced to accept it. The telegraph code industry started with the initial commercialization of telegraphy and didn't end until the 1920s.
There was a cost to it though. Codes greatly reduced redundancy, and caused large miscommunications from very small errors. As Glieck explains it, this was the opposite of the African drumming practice of adding redundancy to strengthen the relationship between the rhythm and the language that the drums mimic.
That is interesting but tokenmaxxing is not maximizing token usage _efficiency_. It is maximizing its usage.
Thanks, that's so odd that I assumed it was about efficiency, which is how I treat tokens. It's hard to imagine a 19th century business man ordering his staff to send as many long winded telegrams as they can.
I don’t think the analogy works too well for one specific reason: you can increase your number of used tokens without ever „sending a telegram“! Run a bunch of Claude sessions, ask them to review various docs sites, create random prototypes, they just throw all of that away. Congrats, you’re a token maxxer
Think of it like this: telegraphs are the hot new thing. The more you send, the more modern and relevant your company. No more Pony Express. You can either have employees sending 1-2 a day. Or, 100 per day. Wow so advanced, so modern, invest now.
This will probably lead to some balancing act like ye olden days of big data etc. Companies want AI native engineers who will use AI to do their work, but don't want AI quality outputs and don't want to drop 200k per year per employee.
AI quality outputs are fine for backoffice work now, but they are awful to read and reason about. Hallucinated features are also difficult to work with.
Isn't that the exact opposite of tokenmaxxing - instead, the telegraphy analogy would be if telegraph operators were ranked by how many hours per day they tied up telegraph lines (highest number of tokens burned/highest $ spend wins) instead of by customers served (programmers delivering features).
What you were describing would be token-minimizing, not maxxing.
What you describe is practically the opposite of tokenmaxxing.
I don't like using AI. I don't find it particularly helpful. But my employer insists that we use it and tracks metrics so I make sure to give it pointless busywork daily. That way I show as using it even if it causes more problems than it fixes.
I always used to wonder this about software stacks even prior to LLMs, but it seems more relevant now somehow:
When will Uber (or your favourite company) be 'done'? They've been writing software for 16 years.
They match drivers to passengers. More software isn't going to increase the chance that I seek them out instead of taking a bus or train.
Will their software be finished in 20 years? 80?
Most of the codebase is custom integrations for local markets. You can systematize some of it but most of the complexity comes from there.
Can you provide an example? What is different about running Uber services in Chicago vs. Indianapolis?
For example in Seattle you pay county fees, and then state fees, and then maybe special fees if you were picked up in the airport.
I took a ride from SEATAC to my hotel in downtown Seattle and besides the ride itself, there were 5 other items on the bill, 4 of which are specific to the place I used Uber.
Then I had the return trip from my hotel to SEATAC, on this one I got EIGHT items on the bill, on top of the ride fare. Some specific to Seattle itself, some specific to the road that the Uber took (a tunnel fee - which is different based on the direction you take it in), etc.
So the real question is what is NOT different between two locations. Less than 15% of the bill.
I also took Uber in India, where you have to share a one-time password with the driver for example, which I've never seen in any other country.
In some other countries the Uber app exists but Uber drivers are actually taxis, so you're actually ordering a taxi via the app.
Ah local regulations and fees. Not so much the core service algorithms. That makes sense.
Uber has also been public transit: https://www.theguardian.com/cities/2019/jul/16/the-innisfil-... (like actually public transit, not "lol they reinvented busses again" (though, lol, yes that too))
Essentially every single airport in the world is custom UI and custom walking path guides and pickup instructions, and rules for where pickups/dropoffs/etc can occur can change multiple times in a day, much to everyone's enjoyment. They're almost all private property, and are so valuable that whatever they want is what they get.
And food. Most/~all? major brands get custom integrations.
Hundreds (iirc) of identity verification providers, most or all custom, and constantly weighed against cost and accuracy because it ain't cheap and it ain't good but it is far better than none (both legally and ethically).
No idea how many payment sources they accept, but it's definitely a lot more than anyone who hasn't lived on 5+ continents thinks.
And remember that this is all international. So scale is huge and law changes are constant and frequently conflicting. Darn near every useful feature is illegal somewhere, at some time, for both good and bad reasons.
---
This is not at all to say I think Uber is efficient, clearly it is not. Not by an enormous margin. But there is a legitimate need for truly absurd complexity, because the world is not consistent. You see similar things happen anywhere [thing] tightly interacts with humans.
Use Link next time. Only $3
Twice as long and walking with suitcases when my company's paying? No thanks.
> I also took Uber in India, where you have to share a one-time password with the driver for example, which I've never seen in any other country.
I was prompted for this on the US west coast this week. If that feature was ever India-only, I don't think it is now.
I’ve seen it extremely intermittently over the past 5 years so I have no clue where it’s enabled, but it’s seemingly not consistent
Sri Lanka does the OTP too. But then again India does OTPs for literally everything because your entire identity is tied to your SIM card.
Vegas: ordering a tax "to a hotel" - hotels have different entrances, pickup / dropoff there during crazy times is hard. Uber UI for Vegas is unique / some features are designed to make it easier for driver and passanger to find each other
Airports: different regulations, different rules for pickup/dropoff. Also scammers who pretend to be in a car, walk with their phones around pick-up ares in airport and do bait-and-switch (saw that in Istanbul SAW and in Dubai Al Maktoum)
Additionally each city often has their own fares they impose, restrictions / etc
There's an excellent HN thread that talks about this very question (that comes up on HN every now and then - what _does_ company X do that needs so many engineering resources?): https://news.ycombinator.com/item?id=25375921
TL;DR: Managing a taxi service (that's what Uber is in my mind, not whatever "ride share" means) that spans cities and states, never mind countries, is extremely complicated. To their credit, Uber manages to make it look simple to the end user, prompting such comments as "meh it's just a few screens how hard could it be", which is triumph of product engineering as far as I am concerned.
Related: this blog from Uber talks about the problem of serving market-specific configuration data at scale: https://www.uber.com/us/en/blog/how-we-unified-configuration...
There’s been a bug when ordering an Uber in my (quite small, circa 20k pop) town recently in that they think our annual festival (which is in July) is on now and try to force a stupid pickup point which in my case is about the same distance from my local pub as my house is but in the opposite direction. I guess things like this need some sort of maintenance (which apparently they’re not getting!).
The New Delhi airport has a single pickup location for all Uber rides (which moved recently). Also the Indian government requires separate reporting from Uber depending on whether the rider is using an Indian or foreign credit card to pay. They also have to report to the government anyone who hailed the cab on WiFi rather than LTE (for "security" reasons they can never quite articulate).
Every different country Uber operates in is a moving legal and regulatory target
Sure, but custom integrations seem unlikely to explain the majority of Uber's technical headcount. Let's say they average a dedicated engineer for each of their 1000 largest markets/locations. Let's assume another 200 across the countless smaller markets. Let's assume 50% overhead atop this for things like infra, tools, and management. These all seem like exceedingly generous estimates to me.
They actually had 5,000 engineers in the tokenmaxxing blog post. That's a lot of engineers for the rest of Uber's business activities.
There are always newer technologies and techniques to be implemented. Better algorithms. Larger deployments. Better reliability. There are also almost always bugs to fix. So, so many bugs.
Weren’t they trying to do their own self-driving thing?
I think this is partly a problem with companies that have had heavy investment. Uber’s value isn’t based on what they are doing, it is based on the idea that they are going to render ideas like owning your own car or taking public transit obsolete (I mean that’s an exaggeration but less of one than it ought to be).
AFAIK they gave up on doing self-driving themselves a while ago. I'm sure they are still hoping to be able to get rid of human drivers somehow.
If they didn’t have human drivers, they’d have one less human to exploit per ride.
That's true, I guess possibly they figured out that humans using their own vehicles turn out to be cheaper than contracting or owning self-driving vehicles.
Uber is regularly offering 50¢/mile trips now (which they charge significantly more than that to the rider for). There's no way an autonomous vehicle is going to get operating costs that low anytime soon.
Considering the IRS mileage rate is 72.5 cents per mile, sounds like there's no way any kind of vehicle is going to get operating costs that low!
They shut it down after they killed a pedestrian. (They also got sued by Waymo for illegally acquiring trade secrets, and settled.)
https://en.wikipedia.org/wiki/Death_of_Elaine_Herzberg
I think you’re missing how complex international operations and optimization are.
Each country has their own laws around what uber is and isn’t allowed to do. This needs to be formalized in code. For example you actually call a taxi, though the uber app, and the amount you pay is per mile, not a fixed fare decided ahead of time. To add to this complexity, some cities will have their own laws. What happens if you take an uber from town a to b, where each one has different laws ? A lawyer probably has an answer but the app needs to adhere to that. On top of that laws change all the time.
Optimization, well you can always optimize something. speed, costs, paths etc. In a way this never ends.
I think the part we interact with as consumers is a tiny sliver of the complexity those services have to build and operate.
Uber is at a large enough scale that this analysis doesn't work. You and I do not care even a tiny bit about "Eats for the Way", one of their planned features this year (https://www.uber.com/us/en/newsroom/go-get-2026/) that lets Uber Black passengers specify that their car should arrive with their Starbucks coffee order. But if 0.01% of users order 1 additional ride a month because of this, that's about 200k rides a year, which may well be sufficient to justify the development costs.
shiny new tools but people only want to use them on the same old problems. how can we innovate the development of crud apps even more?! that was what plagued the web dev landscape for some time. Constantly seeking newer lazier means of producing the same old product. I admit it has an allure but if companies are no longer constrained by dev effort / labour then they can only ponder their own reflection as the source of their failures.
There is always a rewriting around the corner
Well there is a lot of ongoing maintenance cost. There is probably still some marginal gains possible on the matching side. There are new products to launch. So while one specific software can mostly be finished, the total software of a company is always changing.
> When will Uber (or your favourite company) be 'done'? They've been writing software for 16 years
I suppose it becomes easier once the browsers, Android and iOS have been frozen for a little longer than 16 years. Nevermind the changing regulatory field and new products (when was Uber Eats launched?).
In that 16-year period, Covid-19 emerged, as did viable self-driving and partnership with Waymo. A networked, people-facing app can't ever be "done", unless you have perfect prescience. Internal tech-stacks are a living thing: keeping a service that on the outside appears to be unchanging is a lot of work! Scaling is a lot of work! Scaling services and maintenance feed off each other.
It’s death for a tech company to be “done,” since that means no more growth. So they will all bloat indefinitely until they implode or get absorbed. It’s simply the fate of all VC-fueled startups.
Tokenmaxxing makes no sense, it is akin to write extremely inefficient SQL / Spark Jobs, full of cartesian joins, ultra skewed datasets, etc, just for the sake of using as much compute / memory / IO as possible.
This always happens when the metric becomes the goal, companies should nurture and foster an environment where AI is used in the most efficient way possible, first asking "do we really need an agent for this" and if so, what kind of agent is needed, what model, reasoning level, etc.
They should also promote projects that aim at saving tokens, increasing cache hits, codifying the information in ways such they use as less context as possible (graphs of knowledge are pretty good for this!)
It's toddler-level logic. "You can achieve positive outcomes by using X. Therefore, we need to use as much X as possible to maximize positive outcomes."
It's like trying to win a race by setting a gas station on fire.
The argument in favor of "tokenmaxxing" has always been that it's creating space for employees to freely explore the broad and novel space of AI-enabled workflows. I've seen a number of use cases where I'm skeptical any value is being produced, but a number of others where some team or another has finally solved a long-standing problem of theirs with an agentic workflow that would have been hard to justify to a cost review committee.
> They should also promote projects that aim at saving tokens, increasing cache hits, codifying the information in ways such they use as less context as possible (graphs of knowledge are pretty good for this!)
My understanding is that most big "tokenmaxxing" companies do have teams who are working on this in the background.
+1 I find the general disdain for C-suite or senior engineering leadership on HN so silly. These people didn't get promoted or hired because of nepotism. A lot of them moved up the engineering ladder and are familiar with how software engineering works and the incentives involved. Yes, some of them are sheep and will blindly copy what is fashionable but so do a large swath of ICs.
If you want incredibly fast adoption of AI within a company, the best thing you can do is to signal from the top that tokenmaxxing will be rewarded (or at least not be punished for it).
1. It forces everyone including the lazy ones who normally wouldn't invest their time in learning anything new to actually install codex/claude and learn to use them.
2. It prevents any middle manager from putting up blockers for adoption/experimentation ("this is new, I don't trust this, let's do it the old familiar way", "this might be expensive, we care about efficiency here", etc). Once the C-suite dictates tokenmaxxing is allowed, every middle manager will fall in line instantly.
3. Tokenmaxxing is not choice you have to live with the rest of your life. A year or two from now, once C-suite is satisfied with the rate of AI adoption within their org/company, they can just as easily switch the focus to efficiency. Teams will be asked to justify their token spend and start to optimize.
>These people didn't get promoted or hired because of nepotism. A lot of them moved up the engineering ladder and are familiar with how software engineering works and the incentives involved.
I would argue that you have an unreasonably optimistic view about corporate culture. There is a substantial amount of adverse selection and political maneuvering going on all the way up to the top. Tokenmaxxing just goes to show this.
That's part of the reason why this website is hosted under the YCombinator name, after all. Hackers are strongly meritocratic, which is not something you will find in a big company.
> These people didn't get promoted or hired because of nepotism. A lot of them moved up the engineering ladder and are familiar with how software engineering works and the incentives involved.
Good one!
Tokenmaxxing exists because executives think employees are resistant to change. Thats it, a way to incentivize/force every employee to experiment with a new technology. Obviously once they think everyone is utilizing AI the tokkenmaxxing stuff will end.
Yes. Executives think, correctly, that employees are resistant to change.
But they incorrectly think "employee adoption of LLMs" is a first-order goal
"Correctly" doing a lot of carrying. The last 10 years of tech leadership has proven how out of touch and straight up terrible they are.
They are leeches that are just extracting wealth from their respective monopolies.
I am certain that the max sustainable boost from AI use -- with code review and otherwise all-in -- is approximately 20% with the appropriately skilled senior engineering talent, and the token budget for any engineer should not exceed that.
I do not believe that engineers who are tokenmaxxing are truely productive and I have not seen any evidence whatsoever (perhaps the opposite).
I've personally found that with the right flow and codebase knowledge, that's achievable with sustainable levels of effort.
They are burning money to pay for AI-assisted development. Ok. But what is the ROI of it all? Was it worth the supposed increase on efficiency?
Why nobody talks about those points, which are actually the only interesting points of all this AI craze?
I think it's because not many know how to measure it properly.
I can output 5 useless/bad features in a day with Claude or I can output 1 useful feature per 2 day period. Which one has better impact on ROI?
In this example, it might seem like it's an easy answer. But, in the real world, it is a lot more nuanced and much more difficult to measure and so not many are bothering to do it and are opting in for the simple solution of following the hype.
AI for engineering productivity seems to be widely misunderstood to be a magic button that produces the same result, but faster and more cheaply. And based on that reasoning, you should want to force employees to tokenmax, because, why wouldn't you want to get more results but faster and cheaper?
A more nuanced view would be something like:
* AI lets you achieve your roadmap somewhat faster, but:
I find it shocking that anyone ever thought tokenmaxxing was a good idea.
AI maximalists like to compare the technology to electricity. Imagine if in the early days of electrification, a CEO had rewarded staff for increasing the amount of electricity they consumed rather than finding ways to use it for business impact. Institutionalizing people who showed signs of mental illness was popular in those days, and I suspect that would have been the outcome.
The problem is that it is a good idea at the individual level. Poor management reads it as a signal of productivity.
Regularly experimenting with AI tools as they improve and relying on them where they provide an advantage is a good idea at both individual and institutional levels. Maximizing usage for its own sake is not.
It's amazing that it took months to figure this out. "Well we thought that if engineers are told to maximize costs through AI use, to consume as much as possible of a resource that costs us money, then obviously good things will happen. Imagine my surprise when it didn't turn out that way."
Imagine if engineers were ranked based on their AWS spend. People allocate VMs and fill databases with terabytes of random bits, to get to the top of the AWS leaderboard. If you don't do this, you're ranked at the bottom, and good luck at the next review cycle. Who could have expected that this is not the road to success?
You say "amazing that it took months to figure this out" as if the answer to the question is obvious.
But it's not. Some FAANGs are doing amazing things with unlimited tokens. Other companies have no clue what to do with tokens, they've just told their engineers to max them.
It really depends on how you're using the tokens. If you're just using them for Codex and Claude Code - yeah, tokenmaxxing is incredibly dumb.
> Some FAANGs are doing amazing things with unlimited tokens. Others have no clue what to do with tokens.
Unlimited tokens is different from “use AI a lot or we will fire you, and we are counting token consumption as usage”. Obviously the latter is stupid and yet it was done in many places.
I'm not convinced it actually was done in many places, although I understand why in a bad job market people don't trust that it isn't happening in secret. Every time I've heard of a token leaderboard or such it's come with a denial that the company is using it as an employee performance metric.
> it's come with a denial that the company is using it as an employee performance metric.
Surely the anonymous employee feedback polls are totally anonymous too. BigCorp loves you, is family, and would never harm you.
> But it's not. Some FAANGs are doing amazing things with unlimited tokens
Giving someone unlimited access to a resources is not the same as directing or incentivizing them to use it for the sake of using it which is what the parent comment criticized.
As for the other FAANGs, Meta and Google have (not good but still) frontier models of their own, so they are very different from a company paying API costs per token.
Show me some fang that have made nice outwards facing products through a fully embraced AI workflow?
AI is an accelerator that engineers should know and have access to, but it's not something that should have mandated usage and quotas around. It's also absolutely dangerous for young engineers and the like - it fundamentally denies you of the "learning" aspect. I'm now seeing in interviews young graduates being given AI tasks to complete and they come back with a correct solution and no concept of how it is working.
You learn and reinforce learning by DOING and reading in depth. High level summaries don't teach anything and are the kinds of things only VPs care about. So, unless the intention in the future is for everyone to be a VP using AI to do the work, we need some middle ground here and some real thought around implementation of these tools or there's going to be a generational canyon gap of knowledge between being able to "say" and being able to "do".
> Some FAANGs are doing amazing things with unlimited tokens.
Would love to know what things!
OP (solenoid0937) is an unfounded AI-hype peddler and an Anthropic shill (check their comment history), do not expect them to provide an actual example of their wild claims.
I checked their history and they seem alright ¯\_(ツ)_/¯
Here, just a few examples over the last 9 days, in no particular order of exaggeration magnitude:
Unsubstantiated claims of 3x-10x productivity at allegedly a FAANG: https://news.ycombinator.com/item?id=48174666
More claims of the same, still unsubstantiated: https://news.ycombinator.com/item?id=48173995
Even more claims of the same, still nothing of substance: https://news.ycombinator.com/item?id=48173975
Outlandish claim of using 300$ in tokens in 1 hour at allegedly a FAANG: https://news.ycombinator.com/item?id=48158438
Ostensibly great things with unlimited tokens at allegedly a FAANG: https://news.ycombinator.com/item?id=48269185
Complaints about HN "not getting" AI: https://news.ycombinator.com/item?id=48245673
An attempt to validate vibe-coding in production: https://news.ycombinator.com/item?id=48243651
Whitewashing Claude degradation: https://news.ycombinator.com/item?id=48211638
More Claude degradation whitewashing: https://news.ycombinator.com/item?id=48245784
Whitewashing Anthropic: https://news.ycombinator.com/item?id=48199951
Promises of coming AI revolution: https://news.ycombinator.com/item?id=48188104
All of this serves to hype up the LLM technology with absolutely outlandish claims, while also propping up Anthropic online.
This is so cool, I have my own fan on HN that cherry picks my comment history to call me a shill!
You could have easily disproved my claim by linking your comments with a more balanced or nuanced opinions on the matter, except you cannot, because there is only even more outlandish and wild stuff you say about AI and LLMs.
It‘s perfectly ok to share opinions that aren’t nuanced or balanced. You seem to have something strongly against that specific user, the fact that you felt the need to go through so much of their history, post a massive list of their „suspicious“ comments, and mention in multiple places how they are a shill is pretty concerning imho and doesn’t make you look good at all. Their activity looks fine, they seem to be enthusiastic and optimistic about the technology, but that’s pretty much it.
And now you’re asking them to somehow disprove they aren’t a shill? How would that even work. You seem unnecessarily antagonistic towards that user
It is not just me who finds solenoid0937 account suspicious: https://news.ycombinator.com/item?id=48173975
I did not have to go through "so much of their history", this is just the last 9 days. There are considerably wilder claims from them earlier, when the account was solely focused on propping up LLM-hype and defending Anthropic.
We are living through a period of time with one of the potentially most disruptive technologies ever being developed. A lot (A LOT) of money is invested into it, a lot of livelihoods are and might be affected by it, and some people stand to gain A LOT from it. So there are significant interests to sway public opinion in favour of LLMs and AI, to hype it up to unreasonable extent, to muddle the waters of a reasonable discourse. Accounts such as solenoid0937 are unleashed on public forums to achieve that, and because of that we have to take everything they write with a huge grain of salt, or even ignore completely, as there is just no true information in their comments.
You yourself got baited by them by considering what they wrote seriously regarding "amazing things with unlimited tokens". Now the idea that "LLMs 1) are used in one of FAANGs massively and 2) are used to produce amazing things" is planted. Will you remember later that they did not actually provide any evidence of that? The account has been doing this trick multiple times over the last few weeks.
For me, as someone who is actually using LLMs in their work, a single balanced comment on the matter would have been more than enough to consider them not being a shill. Unfortunately, instead they have claimed recently that they went completely full-on with agentic coding (https://news.ycombinator.com/item?id=48245721) skipping reviews and pushing to prod directly (https://news.ycombinator.com/item?id=48243651) at a FAANG, no less. And they claimed it in such a manner that this is objectively the only proper way to do the work, and all other approaches are doomed. How is this anything else than peddling unfounded LLM-hype, I do not know.
The example above might seem like a singular episode, but they have been doing it over and over for the last year and a half, so this is now a pattern. No actual evidence for any of their claims is provided, so the only thing left is AI-hype, and pretty wild at that. So why would a reasonable person, with ostensibly enough money to retire (https://news.ycombinator.com/item?id=48252297), ostensibly working at a FAANG, spend all their days spreading unfounded AI-hype in a degrading manner on HN and defending Anthropic? Given the vested financial interest in the technology, the most plausible answer here is that they are paid to do so.
Every time I've tried to engage you in debate you just call me a shill so there's no point. Love ya too though!
I have explicitly stated more than once, beforehand, what would have given you a benefit of doubt from my perspective. Even after reading all that you opted for answer evasion, rather than providing any substance, as you do with all the questions addressed to you. I do understand why you had to do it here though, because the claims regarding AI and LLM you have made before are even more outlandish than what has been posted over the last 10 days and similarly without a single shred of proof.
He may not be a shill, he may just be terribly and enthusiastically wrong. Lots of HN posters are.
In other words, people who are productive get more done when you scale up what they're already doing, and people who aren't productive will not magically become productive when you scale up what they're already doing. That's incredibly obvious, because we've seen how this plays out repeatedly in so many different ways (lines of code, commits, tickets closed, etc.), and it has nothing to do with tokens or even programming, but just how trying to manage people works.
Where can I see those amazing things done by FAANGs?
Join one!
Anything more concrete?
You got baited by an LLM-hype peddler and Anthropic shill, do not expect anything of substance. See https://news.ycombinator.com/item?id=48272695
The inability of leaders to understand Goodhart’s Law is always a sight to behold. They see a number go up and pat themselves on the back for how well their employees are making it go up without ever wondering if the thing they care about is happening.
This is one explanation, sure.
Isn't it more likely that they simply don't in fact care about the "thing they care about", only the metric?
They can plot the metric on a chart and receive praise, so that's what they're interested in.
That’s an even worse characterization of them isn’t it? They don’t even care about the end result just the metric. That would take them from clueless to malicious.
I wouldn't describe it as malice. If your job is to make the line go up, you make the line go up, and are rewarded for doing so, then you have done your job.
The point of this was always to explore what is possible with AI as quickly as possible. Obviously, there is going to be a lot of waste, but the 5-10% of employees who are truly thinking about it and discovering novel applications are what you are truly after. Because right now, you effectively have a giant, as of yet poorly explored space of potential uses.
Anyone who can find the actually valuable portions of the space early has a potentially huge competitive advantage. Even if the result of the experiment is the negative that AI is actually mostly not that useful, that is still extremely useful information in a time of great uncertainty regarding outcomes.
The bottom line is that this approach may be expensive, but if you have the money to burn, it's far from the worst strategy if you are trying to position yourself correctly for the future.
What’s the huge advantage though? Adopting workflows that give big productivity gains is relatively easy even for big corporations. It’s only an advantage if you can keep it secret.
OTOH maybe we’re in for a future of patenting prompts.
The thing I don't get though, is that most people just don't have that much work they need to do. I can use AI to pretty easily get my work done just via the regular chat interfaces. But because of the tokenmaxxing metrics that leadership tracks, I end up just having the AI deliberate for hours on random things just so that I can boost my token numbers. I think tokenmaxxing for the end goal you described is only realistic when the engineers are truly buried under a backlog of work.
Not being buried under a backlog of work is one aspect, and the other is that the sheer _urgency_ of these efforts makes it look like companies like Uber could be displaced in a year or two by someone who gets lucky with AI use.
Which absolutely isn’t the case. Even if someone would manage to overtake a market leader on tech merit alone, within 1-2 years, thanks to AI, markets don’t swing on such short notices. The fake urgency is absolutely psychotic.
Ha the real stars of Uber aren't the programmers. It is the lawyers.
> The point of this was always to explore what is possible with AI as quickly as possible.
If that was the intent, the messaging at many companies failed to communicate that. The message was "increase this metric", not "explore this space".
Someday maybe Goodhart's Law will be intuitive to people making decisions like this, but not any time soon I guess
> It's amazing that it took months to figure this out
We aren’t there yet, so far it is just a COO questioning the investment
I think unfortunately it's not about what seems obvious, or even what seems more likely, but about what seems retrospectively justifiable regardless of outcome.
The incentive structure of this type of decision is 'absolutely under no circumstances existentially mess up'. Ostensibly with respect to the organisation, but in actual reality much more so with respect to the individual(s) involved in the decision.
If everyone else is doing something that kind of obviously makes no sense, and you decide to break from the crowd by instead doing what does make sense, then there's a pretty solid chance of gaining a temporary edge while reality resolves the truth. But those gains probably won't matter all that much for the organisation, or indeed your position within it. It's a solid chance of an unimportant gain.
However on the other hand, there's a tail risk that something very unexpected happens and the thing everyone's doing that makes no sense actually turns out to make sense - sometimes even for entirely unpredictable incidental reasons - and then, well, you're in trouble. Not necessarily 'you' the organisation.. they'll likely be able to catch up and it won't matter that much. But for 'you' personally, the decision maker, it's very much not good.
As a bonus, in the much more likely scenario that the thing that makes no sense turns out to indeed make no sense, you're in the same boat as everyone else, there's no relative loss, and most importantly you don't stick out as someone who did something as risky as to go against the prevailing, albeit pretty clearly nonsensical, sentiment.
So basically, game theory tells you pretty quickly to just go with the thing that makes no sense if you're optimising for some (weighted) cross of what's best for the organisation and yourself as the decision maker.
Limits are beneficial. They should be treated as a design feature, not just a stopgap.
When something is abundant, people tend to waste it.
I’m perfectly happy with my base subscriptions. I have Claude Code and Codex monthly subs, plus a yearly Google AI Pro account because it was a logical upgrade from the cloud storage plan I already had. I think it worked out to something like an extra $10/month for the AI features.
I constantly rotate between them during the week, managing tokens carefully, cleaning sessions and contexts as soon as possible, and being intentional about usage.
I honestly don’t understand the appeal of these ultra-expensive max subscriptions.
It reminds me of that flying orb toy I bought for the kids a few years ago. The battery only lasted about 10 minutes, and the kids would go ape shit crazy while it worked. Then it needed a 30-minute recharge, which created a natural cooldown period.
I actually considered that a good feature. I would never want the thing running nonstop.
Maybe don't use the most expensive models on the planet? Maybe use AI like a tool and not this black box that grants wishes?
Sounds like you want to be in the next round of layoffs?
I think companies are reluctantly realizing that AI is not a magic genie in a bottle, and is instead a tool.
Still very valuable. They just need to have strategies that match what the tools are capable of - not strategies that involve "rub the magic lamp and increase profits 80%".
If the market is rewarding companies going after the "rub the lamp" strategy, they're going to say they're doing that to juice stock prices.
Maybe the market is finally realizing blindly spending billions on LLMs with almost no strategy is not a good strategy.
Who knows.
> Still very valuable
You sure about that?
Both labs and tech companies have been desperate to show ROI on LLM use and nobody can seem to
But the executives need the fanciest models to evaluate how well they can replace the expensive labor costs.
"He said that, based on talks with Uber's senior engineering leaders, he realized higher token usage did not translate into a proportional increase in useful consumer features."
He's saying that like it's some grand epiphany and not the most self-evident, obvious thing I've heard this month. Some of the literal dumbest people on earth are in charge of these major companies.
>obvious thing I've heard this month
not only this month, but it is the basic statement of the single most well known 50 year old book in software project management lol. At this point we need to wipe the slate clean and start over, the industry is run by illiterates.
This is also Uber we’re talking about. The company that famously developed a massively engineered ledger to track every event across the entire company, globally consistently, forever, in a single database. This definitely adds enormous value to the bottom line!
The fact that a company with such a ledger has trouble advocating for AI-maxxxing will make watching the "ur holding ur AI wrong bro"-reactions all the more hilarious.
As with many things, users will discover a happy medium. There is scope for a lot of productivity gain here if the C-suite is willing to understand the tech and work with engineers rather than whatever Dario Amodei is selling.
Waiting for tokenedging next.
Is this when you type the prompt into the text window, but don't hit enter? Make the GPU see the message "x is typing"? Lol.
As long as there's an RPC connection established and a partially sent request, I think it would count.
^ Philip K. Dick's unreleased book title
tokenmaxxing is becoming harder to justify could be a change in the labor market => when capital was free the companies optimized aggressively around retention and internal status spending but high rates + slow growth oblige firms to back toward productivity and operating leverage.
I have Opus 4.7 at work at 15x. Burns through tokens like water. It feels like one of these new mega datacenters is just for me. I'd love to know what the bill is, but we're just encouraged to do as much AI as possible.
> Burns through tokens like water.
Pretty sure I know what you're saying, but the visual on this one doesn't match the point you're making.
lol yeah I'm not a poet.
Just append a reactive metal. "Like water through sodium"
Or make water relativistic, xkcd What-If style https://www.youtube.com/watch?v=pfbzrrcQZjs&t=155
2^30 tokens costs something like 2^10 dollars, order of magnitude, if that helps ballpark.
I'd be interested to know if this is about individual employee AI usage, or use of AI tokens in production features, or both - and assuming both, what the split is.
I can see how Uber could burn unbelievable amounts of tokens if they start running internal features that run a bunch of prompts against every completed ride, or every customer profile, for example.
Or maybe this is about employee usage, but they introduced some stupid "you get evaluated on how many tokens you used" thing a couple of months ago when that was trendy and are just beginning to notice how much that cost?
IMO, it's undoubtedly both.
The number of product teams who have shipped expensive-to-operate AI features is wayyyy up there, and for many of the scenarios I've seen, customers simply don't care or are unwilling to pay significant premium for access to it.
At the same time I'm starting to see some direction from people in leadership that I should "use the right model for the job" and things along those lines, which is a very, very different line from what I was hearing 12 months ago.
My continued prediction is that we are going to see a tweak on the SaaS model where the sweet spot moves to metered usage pricing of really fine-grained API-based access for apps which traditionally have been operated solely via the UI. Long term the trend is going to be "we'll house the data, enrich it, maintain it, provide fine-grained API access over it tailored to model usage, and you bring the model" with some services opting to give you the model interaction layer/harness. IOW I don't think SaaS is dead. Far from it. However, I do think that a lot of people are going to be looking to interact with SaaS apps via their own models with APIs that support those use cases better than a lot of those APIs do today.
> we'll house the data, enrich it, maintain it, provide fine-grained API access over it tailored to model usage, and you bring the model
isnt this just mcp servers hosted by the saas provider?
Clearly they need more layoffs, and for that matter why keep anyone around? After all, AI will be writing 100% of code in 2026.
By 2025 we will have AGI and software developers don't be necessary. Also next year we will have self driving.
Surprisingly, Uber hasn’t had a mass layoff since 2020. The company currently has ~34,000 FTEs, which I personally think is insanely bloated for what amounts to a taxi + food delivery app.
No wonder they need to extract such a massive cut. I really have no hope we will ever get to efficient middle-men who take least they can for good of both sides beyond them.
Replace Tokens with Gas, or water or healthcare or anything - and it's foolish. You shouldn't let the seller dictate what amount you need of something.
Smart engineers are figuring out how to best use their tokens - as tokenmaxing is just as silly as gasmaxing your car.
On token consumption and efficiency... AI-champion guy in my prev company made a metric, like how many tokens are spend per line of generated code, and even put a leaderboard based on that metric, praising guys with the cheapest LOC.
For me that's insanity for so many reasons...
I’m genuinely curious why they don’t cap at $100/month Claude Max per employee. That would be sufficient for 80% of them.
Are you telling me, it did not make them "productive" in ways most of (us non-AI-boosters) "cannot even begin to imagine"? Who could've thought - a lot of average stuff, still ends up producing average result?
Oof leader of bubble are starting to take a step back?
The black bill that is coming that nobody is prepared for is that the value of a token varies greatly depending on the human. Companies will quickly find out its much better to give your top 10% engineers a lot more tokens and lay off your average engineers. The 10x engineer will become the 1000x engineer.
Wrote about this and the impact of to jobs here: https://x.com/deepwhitman/status/2058324179506831372
Lol, no, no one’s becoming a 1000x engineer.
Feels like they are debating internally whether to cut people or AI spending. Very healthy debate. Let's hope they spare people.
At what point might it be cheaper to, say, hire a human?
At what point is there a difference between a burn rate and tokenmaxxing? Isn't it the same as during the dotcom bubble?
I actually do think token maxing is good, but they should have limited it per user. I find it reallly hard to get people to max out the Claude $100 plan, let alone the $200 plan. I understand the enterprise plans are different and more expensive, which is how you get these kinds of issues. But encouraging people to try things with AI is very important, and some amount of token maxing is importsnt.
Who’s it important for?
The business. Employees are hesitant to learn new tools that are very different from what they are used to, so if your business believes that AI is a productivity multiplier, it behooves it to incentivize individual employees to learn to use the tool.
I think the key word is “believes”. There is no proof that AI usage improves productivity. Token maxing is essentially customers paying to try and prove a business’s unsubstantiated claim. The AI companies should be proving their claims themselves not the other way around.
I do think AI has value and is useful but the idea of token maxing is ridiculous.
Sure; I described it that way deliberately. I think you can reasonably disagree with whether or not AI improves efficiency, but regardless, you can agree that if a business believes AI does, it will logically conclude that it should incentivize employees to learn to use AI.
> Employees are hesitant to learn new tools that are very different from what they are used to...
That simply isn't true for technical employees (like software devs). They are so hungry to get stuff done that you have to hold them back from adopting new tools which they think can make them work more effectively. Tech guys will set up entire shadow IT departments just to get around corporate restrictions that are limiting their productivity.
No, if software devs are not using LLMs for programming, that is proof that the tool isn't actually useful for them. It doesn't mean "they need to be forced to use it", because they didn't need to be forced to use any of the tools which came before it.
Man, it sure isn’t hard for me to max it out.
It's not hard for most people now. 6 months ago when agents first started getting big, I genuinely didn't know enough about AI tools to understand how it was possible to use so many tokens, and I don't think I would have bothered to find time to learn without a kick.
Do you find it hard to max out, or do you find it hard to productively max out?
It's like paying drivers per gallon of fuel consumed and then acting all surprised that you see them revving their engine while waiting at a red light.
Levie’s Law of AI Psychosis
Why do keep doing this? It's the same as measuring by LoC, we know it's not gonna work. Also, see Goodhart's Law[1]
- https://en.wikipedia.org/wiki/Goodhart%27s_law
hah came here to say exactly this
>"He said that, based on talks with Uber's senior engineering leaders, he realized higher token usage did not translate into a proportional increase in useful consumer features."
Goodhart's law strikes again at someone with enough power to be both ignorant of it and make others suffer their ignorance. You cannot simply measure productivity by tokens spent just like you can't measure it by hours spent in a chair at a desk.
You can measure productivity by hours spent at a desk?
You can measure attendance by hours spent at a desk
Well if you're a devshop just billing hours of mostly low impact work then hours are very much equal to productivity.
Next time you're going to work for an hour, ping me, and I bet I can surprise you with how much less productive I am than you
Productivity is measured by economists in $/hour.
Which is why two identical jobs with the same real life output have drastically different productivity.
A nursing home in Luxembourg has 5 times the productivity of one in Romania despite the services being identical and tech-unrelated.
What if... we stop for a moment, and then, after thinking for a moment, we stop hammering nails with a microscope, and stop using token usage as a metric of productivity?
I know it's sounds stupid, but what if
Not very Billion Dollar Valuation of you.
Yes, but, I sleep very soundly at night.
The people who have ascended to leadership positions are deeply divorced from reality.
"It is difficult to get a man to understand something, when his salary depends on his not understanding it." -Upton Sinclair
The crazy thing is their salary does not actually benefit from riding these trends. Unless it's equally/even more clueless board level pressure with ulterior motives (i.e., lifting their other AI investments or the sector as a whole).
Every c suite in the country is panicking about being left behind, from their perspective it’s either token max or fade into obscurity, or at least that’s what they were sold
I don't think that's accurate. I think every C suite in the country is looking to do away with labor's leverage as much as possible. I think this is a cultural thing more than anything else, C suite + investors looking to get rid of those pesky humans required to prop up their lifestyles. AI is the most credible path toward that. Short, medium or long term returns be damned, this is a reconfiguration of society and they want to shed what they consider to be baggage.
Like anything it's a mixed bag. I am certainly working with people who I think truly believe the "max out on AI usage or become irrelevant" line. There are people who will privately let you know they're just working with the current meta the best way they can, but others who are drunk on kool aid.
Trying to operate as a rational, thinking person in a lot of environments right now feels impossible. Rational thought is being treated like AI skepticism.
Please. These are the same people that force their employees to use Microsoft teams because slack is $5 an employee a month. They're not going to sit idly by while employees burn thousands a month in tokens.
It depends on which people you're referring to. The allocation toward AI budget has been so massive that I think a lot of businesses are way behind on trying to assess value for dollar for the AI-related crud they're shelling out for.
Everyone is feeling it out but the vast majority of spend has been subscription based. Some outliers may have used a massive amount of tokens but companies didn't pay for that.
That VC funded gravy train is likely coming to an end. But fortunately there are also reasonably efficient models now so that the tokenmaxxers can still make the (much cheaper) tokens go brrrr.
Those reasonably efficient models assume you use a harness that supports them well, the one size fits all harness of Claude desktop or codex does not support what you want well, and that’s intentional. It’s contradictory that these AI companies will continue to brrrr to the moon and return on AI spend requires discipline…
The next recession (and there's always a next recession) will clean up this AI bubble. The actually useful products and companies will make it the rest goes down.
Yes, but unlike the dot com bubble we’ll be left with half finished (or not even started) abandoned data center projects, instead of reasonably reusable fiber and ISP infra
its a herd mentality, its a lot easier to follow the louder voices than to spend time understanding how it impacts your own particular business. Because google does this way, or apple does this way is a common argument in lot of feature/business decisions
I deeply believe this but have no strong evidence. Revenue has always been a cure all remedy. This will keep model providers alive along with the very wide range of companies that are experiencing growth with them (from chips to backhoes), for a time anyway. If/when that house of cards starts going in the other direction there’s going to be widespread pain. By analogy the nonsense of the dotcoms and that crash had a very direct impact on their suppliers (e.g. telecoms). My only advice is to let the Microsoft’s and Meta’s do the tokenmaxxing, and don’t get suckered into the idea you (startup, individual, etc) should be playing that game.
They get paid for saying whatever VCs want to hear and now that thing is "we have now become an AI-native company". The thing I'm still trying to understand is who is scamming whom
Uber is publicly traded. They're not beholden to VCs any more.
I don't believe the comments I replied to were specific to Uber.
Come on, don’t be crazy
You're now in the last frame of the comic, getting thrown out the window.
Maybe it's time we adopt/design an economic system that isn't so easily co-opted by counterproductive prisoner's dilemmas.
What would such an economic system look like?
This is actually pretty well-understood if people wanted to do it.. contrary to popular belief it's more about responsible governance than economics and not really a pro/anti capitalism thing
https://en.wikipedia.org/wiki/Elinor_Ostrom#%22Design_princi...
Go ahead and start from the small and grow it bigger: you could become a billionaire if you succeed.
Hahahaha....
What if the goal of an economic system was to support everyone instead of maximizing the upside for winners? Perhaps that's the sort of change necessary for improvement. Perhaps having billionaires is the failure state.
> goal of an economic system
A goal fails - who sets a goal? The keyword is system.
An economic system needs something like a Nash equilibrium where defectors are sufficiently discouraged (and cooperators are rewarded as you imply). https://en.wikipedia.org/wiki/Nash_equilibrium
Everyone says this but no one seems to have an answer as to what it'd be.
There is a complete lack of courage in the leadership of tech companies today, and top-down AI mandates are just another manifestation.
True visionaries think outside the box, but most tech executives are forcing their employees into black boxes, out of fear of not doing exactly what their competitors are doing.
We have lemmings for leaders, and that means that—much like the LLMs that are being shoehorned into everything—there isn’t room for original thinking. Everyone’s strategy looks exactly the same.
> Everyone’s strategy looks exactly the same.
If one is a CxO who's looking out for one's job security, herd-like behavior is the safest option, due to the (near universal) structure of "performance"-based executive remuneration.
Lacking not just courage, but also character. Wasting company money on buzzwords and dubious outcomes is lack of character.
Is keeping your company private the easiest way to get around this?
I'm going to offer a contrarian view here:
First is that despite a lot of waste, some innovation will arise from an enterprising employee finding some interesting use case. A lot of the tokenmaxxing is just waste, but out of that waste may arise a small number of genuinely powerful use cases.
Second is that many workers will be entrenched in their ways. If your executive goal is to achieve the above (find innovative ways of using AI), then you need to move everyone to use it. Most will just waste tokens, but someone may find a novel and useful way of using it that benefits the organization. It is difficult to achieve these without forcing people to act since their default is to follow the well-worn grooves.
So mandates like these are a top-down forcing function like a slime mold feeling out different paths to find resources.
Some devs in my org have fully embraced AI; some would not even use AI if not for leadership mandates and linking usage to performance reviews (I know, I think this is stupid, too). I can see why mandates could be useful since some folks definitely won't be inclined to use AI.
> some innovation will arise
Absolutely, but most management are not leaders, the moment someone pushes the idea to stack rank based on token usage, it gets approved and some genuine people will be impacted.
Post-ZIRP era proved there are very few strong leaders, before that everyone was behaving like they're most amazing leader because they read some books and raised $10M
> but out of that waste may arise a small number of genuinely powerful use cases.
Imagine you employ me as a hotel manager, and I come to you and say: "sure I spent all our food budget internationally in three months, and sure I have nothing really to show for it, but for those three months, we had a lot of food fights"
Your manager then goes on to explain they not only need more money to cover the food budget, but also they need to quituple the cleaning budget too.
Oh and the service level has dropped, because not all clients liked being in the middle of a food fight.
However "we might have some innovation in the food delivery system of our hotel chain"
This is really relative to the size of that innovation, isn't it?
This is exactly how startups and VC funding works, isn't it? You have an idea, give you cash to burn to prove the idea and business model. Many teams and ideas fail. But some small number of unicorns produce outsized returns to keep the whole thing going.
It's how it does work, often.
It's not how it should work, because food fights are stupid and have no upside.
Even if everyone else is having them.
It's not a fair analogy because AI isn't completely stupid, and there are situations where it does provide a benefit.
But a rational business would ask if the upside is worth the cost, if the pipeline can be restructured to concentrate and amplify the benefits, if some elements are better being done the old way, if there are strategic threats if tokens become much more expensive, and so on.
Instead we're getting a wave of "Cut workers, cut costs, derp" and that's as far as the "thinking" goes.
The worst thing about AI is that it shows how shallow and stupid current C-suites are.
The US used to have real tech visionaries. Now it has tech cargo cultists, all chasing an IPO cash out and hoping the music doesn't stop before they get their bag.
Imagine you employ me as a hotel manager, and I come to you and say: "sure I spent all our food budget internationally in three months, but we invented this new dish and now our restaurant is the hottest in town. Sure 95% of the food was wasted but now we can stop the waste and keep the popular dish."
Ok, but was that your intention in the first place? or was it to have food fights.
Thats the problem here. The idea is that we can build more stuff, quickly.
However in uber's case, they just burnt loads of money to push a metric that wasn't really related to productivity.
The intention was to force everyone to experiment with the new ingredient monsanto recently GMO'd. Of course a lot of our employees suck, so food fights were expected, but luckily some of the employees created something great.
> A lot of the tokenmaxxing is just waste, but out of that waste may arise a small number of genuinely powerful use cases
A lot of monkeys will also eventually type up Shakespeare?
Indeed, but that's not a bad thing. If monkeys can produce the next Shakespeare, that will be wildly popular and profitable for the company that did it, justifying the initial waste, just as VC does with companies as a whole.
> Some devs in my org have fully embraced AI; some would not even use AI
So if the people who embrace AI areore successful then the others will follow. Just like every other new tech. Why does AI have to be forced? What's the hurry? Especially when there's no clear example of a company jumping ahead because of their use of it.
It's idiots being driven by FUD. That's the reason.
There are definitely key windows here for innovation driven by competition.
There's also a need to quickly adopt and understand the technology; take the Internet for example. If we were talking about the Internet, forcing teams to build and publish web pages would be one valid way to get teams comfortable with the tech, the workflow, the shift in how to propagate and convey information to an audience.
Without a mandate, many teams won't adopt the Internet as a medium of information exchange because their processes work just fine and have worked for the last 20 years.
I think it's fair to put AI in a similar light. Unless teams adopt it and use it, it's hard for an org to understand how to get value out of this technology and how it affects existing processes and assumptions.
I was programming desktop applications when the web came along. I don't remember anyone ever saying they had been mandated to program for it.
The web took off all by itself because it had a clear value proposition for some use cases.
Many enterprises became legacy because of the web, many enterprises failed because they didn't understand the impact of the tech.
Sears was the OG Amazon. Imagine if Sears had seen it as the new digital catalog.
Blockbuster missed on streaming until it was too late.
Many, many legacy companies did not understand the web and did not understand the impact of the Internet to their business model.
And you think forcing blockbuster's software teams to use the Web would have changed that? You don't think they were using the web for all their corporate communications systems? I very much think they were, and getting blindsided by streaming had probably nothing to do with blockbuster's existing engineering teams not understanding the Internet. Their product teams didn't understand it, but they wouldn't be the ones being "forced to write webpages" either
Yes; non-zero chance that had they been more aggressive in pushing the web, someone would have landed on the right answer.
> There are definitely key windows here for innovation driven by competition.
Those were always there, and will always be there. The type of time frames people are getting anxious about now rarely work in the real world, though, where potential customers don’t just switch products/service provides unless they’re facing catastrophic outcomes if they don’t.
And AI is not making the difference there that people think. I worked on a product that entered the market as a newcomer, wooed plenty of customers, and even though everyone _wanted_ it, only customers _urgently_ looking for a solution got on board quick (within <6 months).
Ironically enough, the product pivoting to Agentic AI hard killed a ton of momentum and interest from customers, despite exciting investors.
Seriously. No mandates at my company. In 2023 and 2024 i had access to Claude, but frankly it wasn't until 2025 that i found the models useful enough, now i use them every day. Nobody forced me. Had they forced me, I'd probably have quit. Once the tools were sufficiently mature and verifiably helpful, people like me all over the company naturally picked up the tools too.
Sure, indiscriminate tokenmaxxing is a gamble that can pay off sometimes. However, I think that the decision to take any gamble should be made by someone who will bear responsibility for the downside as well as the upside. I would prefer to search for new usages in a more strategic way. I agree that experimentation is a great way to learn if done intelligently and with limits. Full “Monte Carlo” makes sense when ops are cheap enough. It seems some orgs don’t think tokens are cheap enough yet.
I think this is very, very hard for orgs to do.
Looking back at the Internet, who would have thought that it would eventually create a Netflix, Amazon, Shopify, Spotify, Google Maps, etc. Just wild the things that ended up coming out of pushing bits over a wire with few simple protocols.
In an ideal world, you make strategic bets, but I can also see the case for the opposite this early in the lifecycle of a technology. You just don't know until you try.
Mid/late 2023, it wasn't at all obvious that it would take over coding that fast.
People talked about streaming years before Netflix. Online maps apps date back to the 1990s. E-commerce as well.
I definitely get the impression that many people thought it would eventually create shopping, streaming, and mapping sites.
I think people were less likely to have predicted things like social media or YouTube, though those weren't ideas sprung from a vacuum either.
If it were that simple and obvious, Blockbuster would have beat everyone to streaming. Sears would have digitized their catalog and used their vast brick-and-mortar stores as fulfillment centers for same-day shipping.
None of these shifts were obviously the right bet and many organizations lost because they missed the opportunity. Now orgs are on the same horizon and I can see why they don't want to miss this window.
Blockbuster actually did try to beat everyone to streaming. Notably, Blockbuster and Enron [1] entered into a 20-year partnership for online video delivery.
Sears was a different story, in that they were a real estate company with a store front and retail real estate took a nosedive due to ecommerce. But that's a different discussion.
[1] https://en.wikipedia.org/wiki/Enron_scandal
There was an amusing post about judging developers based on token usage where some user on HN here was pushing this idea “ICs don’t like it but this is the best way to evaluate” (something like that).
They have a whole management team and can’t seem to find a way to judge or god forbid encourage developers…
Problem is in management, management usually comes up non-sense metric when they themselves lack of good metric.
For example, everyone talks about strategy, but when you ask them what's our strategy answer is usually something like:
* let's figure out together
* industry changing is so fast, we should revisit plans every quarter
...
Because the higher up you go in management, the more "strategy" is a Plato's Cave like interpretation of what better/bigger/whatever competitors are doing.
Ha. Exactly at my current contract job.
"Welcome to the new contractor who will be the artitect our new infrastructure. What is your dream IT setup?"
"Yeah, we can't afford that. Lets revisit once you wrangle those 2003 Dell R620's running Windows 2008 with no patching."
And that is why after eight months i'm terminating my contract on Friday and swimming back to shore.
I don't get why it's so hard for management to see the good devs. All the devs know who are the good devs.
And they know things like “Dude isn’t a high performer metrics wise but his work is solid.” Arguably some of the most difficult things to know from a management perspective.
It should not be overlooked that a lot of this fervor is from investors/board members putting pressure on these companies.
"True visionaries think outside the box,"
I mean that's more of an ex-post statement.
Ex-ante they look at things as objects and visualise/simulate what one ought to do independently, as opposed to being a lemming.
Yeah courage will get you fired. Whether it be about idiot product decisions, or about how your bosses treat your coworkers. That’s the consequence of letting sociopaths get in charge.
I feel like individually, if you sat down with literally any reasonable person on the planet they would arrive at and/or agree with the tenor here.
I'd be curious to hear from people well versed in group psychology/dynamics and/or just a lot of leadership/people experience: what leads people to this type of thinking once they get in a group setting? It just... seems endemic at this point.
Obviously nobody here is going to know what I do or don't know, but I'm just increasingly curious what I am not understanding about this type of thing. It seems so obvious, yet that makes me ever more suspect that I'm oversimplifying it, or just totally ignorant about the problem in general.
Won’t be canned for going with the herd. I think it’s that simple, even if the herd is running off a cliff.
It's because the average organization has lots of people who don't care about their own productivity and won't adopt new tools or processes unless forced to. This is true of most new tech - lots of workers had to be forced into using computers - but AI also has some other bumps to cross like lots of people who tried early models and then wrote them off, not realizing how fast they'd improve. And most orgs have no infrastructure or processes for allocating individuals token budgets, and most employees have no experience of properly deploying budgets.
Roll it all together and saying "just use it dammit" has some obvious advantages:
1. It's clear.
2. It's simple.
3. It eliminates all excuses employees might come up with for not using it.
The people at the top of these companies aren't stupid. They might have miscalculated how many tokens people can actually use, but that's very hard to calculate because usage is opaque and tools/processes change on a nearly weekly basis. They will eventually build out processes, tools, social conventions and performance metrics that take into account efficiency of token usage. But this is hard! Most managers aren't really assessed on the precise productivity of their teams, for instance, because productivity is often poorly defined.
> what leads people to this type of thinking once they get in a group setting
Game theory! The downside of being brave vastly outweighs the upside. For the C-suite, there is no cost to herdlike-behavior, regardless of the outcome. However, there is a very high personal downside to being a maverick, and your board later discovers you made the wrong choice against the grain. The upside of being maverick and right is very limited.
Once a behavior has become mainstream, hopping on the bandwagon is no longer individually attributable to decision-makers, but is seen (and reported) as a macro-economic phenomenon: Nadella, Zuckerberg and Bezos didn't overhire - the American tech industry overhired.
This is a consequence of elements of monopoly power existing in your organization. When you don't have to compete for income you honestly forget how. Then the company becomes a cargo cult of bad ideas driven by managers struggling to differentiate themselves.
we are going through our second AI transformation, the first one didn't work that well because the tools were shit.
Whats happening now and whos driving it is interesting. The CEO has a license for this new tool (think one of the top 4, Qwen Claude, Gemini, openAI) and really likes it. So much so that they (non coder) are making lots of little single page web apps.
The COO is bollocks deep in AI, and is saying that we cannot buy any SaaS products anymore. We must make it ourselves.
The engineering manager has seen this as an opportunity to build out a brand for engineering (its a small department in a medium sized company) by delivering quickly what the large year long efforts cant.
This has formed a slopnexus where PoCs are spun up left right and centre, but there isn't much time or thought going in to making them sustainable.
What started out as a (simple ish) asset management tool, neatly scoped into a deliverable PoC has morphed into a 5 product as one monster.
Its a mess that will either lead to burn out or disaster.
Just...wow. That sounds awful, and I'm sorry for you, but I believe many other companies will or are already following that same path.
And myself being an infrastructure guy that needs to maintain all these PoCs that are now suddenly critical for production, it's the perfect nightmare.
And mind you, that dynamic always existed to a certain degree (laptop on a desk that runs some ugly Python script that does half the work of the BizOps team? Check. GCP account attached to the GSuite running a random instance for finance when the company is 101% on AWS? Check. Spreadsheet with macros that sends emails via Outlook as a mailing list manager? Check.) but at least when you discovered that you could scold them and tell them that we need to migrate this to a proper system because security. But nowadays with vibe-coded internal apps...it's a challenge.
It would be fine if my boss actually followed the docs, or just pointed his agent at the docs.
Whats more annoying is that he changes AI provider lots, so we can never inject rules/skills to make sure he uses the right pathway from the start.
There is probably some opportunity here for a centralized, internal only LLM proxy which injects AGENT.md and skills and permits switching backend providers.
> what leads people to this type of thinking once they get in a group setting? It just... seems endemic at this point.
Large and fascinating topic I'm researching, very relevant for agentic AI and ML too. One way that groups can fail is that they just don't work to dampen / vote out individual errors properly (see PAC learning, Condorcet). Other kinds of errors only occur in groups, and can occur even when constituents individually aren't actually wrong. Some related stuff is:
https://en.wikipedia.org/wiki/Condorcet's_jury_theorem https://en.wikipedia.org/wiki/Group_polarization https://en.wikipedia.org/wiki/Availability_cascade https://en.wikipedia.org/wiki/Information_cascade
The last is probably the most relevant here and made worse by the negative effects of hierarchy. To quote one section:
> The negative effects of informational cascades sometimes become a legal concern and laws have been enacted to neutralize them. Ward Farnsworth, a law professor, analyzed the legal aspects of informational cascades and gave several examples in his book The Legal Analyst: in many military courts, the officers voting to decide a case vote in reverse rank order (the officer of the lowest rank votes first), and he suggested it may be done so the lower-ranked officers would not be tempted by the cascade to vote with the more senior officers, who are believed to have more accurate judgement;
For token-maxxing, our "senior officers" are just executives, and line workers aren't going to vote. Who is the senior officer for those senior officers? It's not shareholders! It's really the executives of even bigger companies, because that is the actually applicable promotion ladder. It's all kind of obvious, but also a genuinely better explanation than "monkey see monkey do". These are just the simpler things, and there's more gnarly dilemmas in https://en.wikipedia.org/wiki/Common_knowledge_(logic)
Wow, amazing answer! I have a lot of reading and then thinking to do, but if you are documenting your research anywhere, I'd greatly appreciate somewhere to follow it.
Thank you so much! This is why I love HN.
If there are any tech CEOs out there reading, I can offer my services. I will pointlessly burn unfathomable amounts of tokens, in parallel, 24 hours a day, 7 days a week, all for you. Think big big big numbers of tokens, you know whats cooler than a trillion tokens, a quadrillion tokens.
Lets talk my bonus, I will open the bidding at $1 per token.
That was a fun thought experiment while I waited for my ralph wiggum to finish running. Now thinking is over and back to the vibe
> and stop using token usage as a metric of productivity
I participate in some management-focused online communities. It’s crazy how many threads there are from frustrated managers trying to get their teams to stop thinking that their token use will be used as a proxy for their performance.
I think a few dumb companies did this and then it spread across social media, triggering a mass panic from engineers afraid their companies will be doing the same thing.
It’s getting so bad that the conversation is shifting to how to identify and coach the token-maxxers to stop wasting the team’s budget every week.
> managers trying to get their teams to stop thinking that their token use will be used as a proxy for their performance.
Because it is going to happen. Do you think metrics are tracked for fun?
Even if current leaders don't do it, next people might do it, how do you tell new leaders that we don't look at this metric? Metric exists to take action based on it
You missed the second half of my comment: There’s more to a metric than making one number go up. It’s becoming a real problem when people use 10X more tokens to get similar work done because they’re tokenmaxxing.
Nobody wants 90% of their token budget spend going to the 10% of people wasting them for number-go-up purposes.
Inefficient token use is going to become a metric.
> I participate in some management-focused online communities.
I know - slightly off topic - but would you be willing to share this list?
I can personally verify that Cisco does it and they're not exactly at the top of the food chain. It's probably more common than you think.
Sounds to me like you are advocating the decimation of the technology sector and a global recession that could last the better part of a decade, buddy!
You mean… increasing AI budget has no direct relationship with productivity and therefore revenue? It’s not that simple? But… my TEDx talk and LinkedIn ramblings…
> ... stop using token usage as a metric of productivity?
and tokenmaxxing is even worse due to https://en.wikipedia.org/wiki/Goodhart%27s_law because whatever you measure with tokens, once you start "tokenmaxxing" you have no measure to look at
Tokenmaxxing is so dumb. You should never show your team how exactly you're measuring their performance; people will optimize for the metric, not the actual performance.
Classic Goodhart’s Law: when a measure becomes a target, it ceases to be a good measure.
LLMs are great, I can understand using them in general. I can even understand chasing 100% weekly usage if you're using the gacha-like subscriptions since that's how you get the most value out of what you paid for.
The way these corporations are going about it is completely insane though. They're essentially ordering their employees to set money on fire or be fired themselves. The more money you burn on tokens at insane API rates, the better an employee you are. Absolutely mind boggling.
Not the first time supposed leaders ran into Goodhart's law.
Protip: skunkworks type side projects are a great way to do tokenmaxxing when you don’t have enough work coming in, but still need to burn tokens to look productive. And because side projects are only governed by you, you can truly go nuts and let scope creep run wild. Soon enough, you’ll be one of those engineers burning six figures a month on AI and people will be in awe of your abilities, probably even elevating you to key AI evangelist positions within your company. And if you actually create something cool, you’ll be praised for your use of AI, and you can just say you built it all in a day or two instead of slacking off for months on your real work.
AI productivity hasn't been well studied yet, but I'm betting that we'll end up with some variation on Price's Law, I.E. some small subset of workers get most of the benefit, while most just burn tokens with little to show for it.
I also want to call out the false productivity opportunities AI offers. There are whole teams building their own "gas town" and not shipping features.
Not all tokens are created equal. It's easy to use a ton of tokens by having agents work together in parallel. That's basically the equivalent as people spending time in meetings, hardly a productivity win. As with everything in development, results matter, how you get there doesn't (unless you're a bad manager).
I just realized my company is months behind this curve. About to blow my token allocation. Before I do, anyone have requests? Sincerely.
I hereby suggest you take the fragmentary excerpts of the infamous erotic stage play The Lusty Argonian Maid shown in The Elder Scrolls series of games and extrapolate them to 100,000 additional full-length acts.
tangent: anyone have businessinsider subscription. i feel like they've really stepped up their game last few years.
many of these leading AI companies are operating at large losses and subsidizing users with VC money. Profitability will entail having to impose greater limits and raising prices, so this will reduce to some degree the value proposition of AI compared to humans.
The industry has to tokenmax to juice the revenue numbers. Its a big club
As soon as tokens stop stop being subsidized, heavy agentic use will become as least as expensive than paying an (entry level) employee. When this happens many companies will trade off havy tolen usage for (maybe a bit slower, bit less accurate) employees again.
This is what I’m betting on.
The financials don’t make sense now. Based on the expenditure the finances won’t ever make sense.
DeepSeek is an open weights model. It's possible the hosted versions are subsidized, but we know what it costs to run locally. And it's expensive, but it's also pretty clearly cheaper than an employee.
Of course, the latest DeepSeek models are not as good as Claude, but they're not super far off either.
They're not far off, getting the same seamless integration as hosted models is a full time job. I think what just happened is that devops is about to explode. What will naturally follow is local hosting of all the things when people realize subscription costs for cloud-whatever are absurd.
Gitlab is going to take off? This is not investment advice.
> What will naturally follow is local hosting of all the things when people realize subscription costs for cloud-whatever are absurd.
Even acknowledging we don't know exactly what costs would look like in a world without VC money, wouldn't hosting models logically be cheaper to do at scale in a data center?
When I compared to the cost of running DeepSeek locally, I meant that we can treat that cost as a price ceiling, not the floor.
Like how server hosting at scale in a datacenter is cheaper than running your own datacenter? Despite ~every company consistently concluding that hosting their own stuff is several multiples cheaper?
No, I think local stuff using also-useful-for-other-things hardware will vastly undercut cloud hosting when the free money pipeline shuts down, and will stay that way for roughly forever. That doesn't mean cloud stuff isn't useful, clearly it is, but adding another company in the middle is rarely the solution for reducing costs.
Vscode uses undocumented api calls. Right now we are living through the PokémonGo api fiasco. That ended with paid keys and a mostly dead economy.
Sorry, can you fill me in? I don't see the connection.
Which part? The undocumented, shifting api or the intentional friction?
When you use DeepSeek’s first-party API, you are giving them your token stream. This has some training value, but it also has incredible amounts of, well, business intelligence value. When you tell AWS your secrets or your customer data, you can be fairly confident they won’t abuse that knowledge. When you give this data to, say, OpenAI, they more or less promise not to abuse it if you’re on an appropriate business plan. If you give it to DeepSeek, even incidentally as something your agent reads, I would be quite surprised if DeepSeek doesn’t mine it for whatever purpose they or the government feel is appropriate.
The risk of letting your agent read .env goes far beyond the risk that the agent itself does something you don’t like with the contents.
But this shouldn't be a risk if you host the model locally.
I have been saying the same for while. Someone always says "but Anthropic is making money on their API" or "But it's inference will get cheaper". But I don't believe it. first all the investments have to payed off at some point and second of all there are other things that cost money. I don't believe that any of them have a positive balance sheet.
I also don't think that blitz scaling will work like with Uber. The engineers are still there. We can work without the LLM tools.
If by "investments will pay off" you mean major profits, that's never going to happen as long as scaling laws hold. All revenue will just go to financing more compute, and either we hit AGI or have the greatest economic collapse in modern history.
The world will look drastically different 5 years from now; for the better or worse, so save every penny (especially if you work in tech).
Maybe this just counts as “light use” since I’m a hobbyist programmer and I only run one coding agent session at a time, but I get about as much done as I did back when I was working while spending a lot of time browsing the Internet, etc.
I’ve spent $10-$20 a day using Claude to write code and closer to $5 a day now that I mostly use Deepseek and GLM, using API pricing (no subscriptions) since I don’t use Claude Code.
This is a rounding error for a company. So I think there’s plenty of room to use AI extensively while being more cost-conscious.
What's funny is that this apparently wasn't something that the Uber COO seemed to think about when their company is arguably one of the most successful ever at the "subsidize to drive down costs until you capture nearly the entire market" strategy.
More straightforward to talk about the hardware directly. Full Kimi K2.6 needs an 8x H200 node to run and serve around 20 heavy users. You can rent an 8x H200 node for around $30/hr.
I'd imagine GPT-5.5 and Claude Opus 4.7 could run just fine on a 16x H200 node and serve at least 10 heavy users without the token output getting choppy.
You're assuming the price won't come down as the tech matures. That seems like a big assumption, considering how quickly open weights models are catching up to frontier models, and how little effort has been invested so far in optimizing inference costs.
It's especially a crazy assumption to make relative to the costs of employing a human. The costs of paying an entry level employee are unlikely to go down at all, and even if those costs do decline, there's a floor they can't drop below (minimum wage at the extreme end), whereas companies are free to optimize agentic costs as close to zero as possible.
So you are assuming that a cost which is extremely susceptible to optimization but which no one has yet seriously attempted to minimize will remain perpetually above a cost which is much less susceptible to optimization, is already subject to enormous efforts to minimize, and has a legally mandated floor. That seems like a bad bet.
A significant caveat is that there is a pricing mismatch that makes it so first party's can subsidize quite heavily.
Agents are expensive in large part because tool calls require round trips. It's because these APIs are stateless and not streaming so you have to resend the whole context each time. This means you have roughly #tool calls x 1/2 context size cached input tokens over any given session. Most API providers overcharge you by a huge amount for cached tokens. A exception being Deepseek. Paying OpenAI $0.05 for 100k cached GPT5.5 tokens during a possibly 2 second round trip agent tool call is like paying $100/hr for what is likely to be ~10 to 20 GB of VRAM residence (holding the KV cache).
Or it got offloaded to NVME and you are paying $0.05 for that much PCIe bandwidth.
I think if local models catch up with current SOTA then that might not happen. Either way, I'm don't think the long-term for OAI, Anthropic etc. really holds up.
Now we are going to get a new profession. Token Engineer! They will be experts on tokenmaxxing! The job growth that the billionaire CEOs promised us from AI is finally here!
Well there are already offerings like githits (https://news.ycombinator.com/item?id=46105112) that sort of promise optimize bang-per-buck of inference
wtv
It’s funny that “maxxing” entered the common vocabulary.
If you're not tokenmaxxing, you're getting tokenmogged on the AI leaderboard, and your next review ain't gonna be pretty.
A good 80% by volume of the modern vernacular is 4chan language that got sanded down.
Sanding down is how we got goyslop turned into slop.
Slop is a word in its own right which got the goy prefix later in life.
I like this too. I have been intentionally -maxxingmaxxing to get the meme out there. It's a good canary to sort out who gets the spicy takes from the pedestrians who probably still copy-paste into the ChatGPT web app like a psychopath.
what the fuck is this timeline I am stuck living in
I find it useful that if they cut the use altogether I will pay for it out of pocket.
Maybe that's the plan :)
But on a more serious note, do we know how much Uber spent per technical employee/month? I assume it is far more than even any of those $200 "max ai" plans.
And the other question is how much the public would be willing to spend, in my estimation this is as "cheap" as it will ever get (main-stream at least).
> I assume it is far more than even any of those $200 "max ai" plans.
Am in a random small company, colleague spent 100 EUR a day on Sonnet through AWS Bedrock (needed to use a EU region). Paying for tokens will get you in a deep hole financially compared to any of the subscriptions, unless it's like DeepSeek or one of the other models that are priced a bit better, though that's also a tradeoff in what they can/cannot do and also where the data goes. Ended up trying out the Mistral subscription for the US stuff btw, it was fine.
bigCo’s don’t get to do the $200 Max plans, they have unlimited plans but get charged like API
Exactly. But I did find an article ([1]) and spend doesn't seem that high per engineer ($150 to $250 per eng) - at least on average, I assume the costs were skyrocketing towards the end.
> Adoption climbed from 32 percent of engineers in February to 84 percent classified as agentic coding users by March. By spring, 95 percent of Uber engineers used artificial intelligence tools monthly, and roughly 70 percent of committed code originated from those tools. About 11 percent of live backend updates were written by agents with no human in the loop, according to Uber's own disclosures.
> The numbers behind the spend are what make the story instructive rather than anecdotal. Monthly cost per engineer ranged from $150 to $250 on average, with power users running between $500 and $2,000.
My guess is that the reason to rethink AI-spend was probably the exponential growth in cost over time, and tokenmaxxing payoff not being immediately obvious as mentioned in the article.
[1] https://www.forbes.com/sites/janakirammsv/2026/05/17/uber-bu...
Except you won’t because they will threaten to fire you and force you to route all of your AI through data protection proxy to stop exfiltration by filtering and tracking prompts/response tokens.
Probably long term each dev gets their own GPU and runs a model locally I expect. Seems like a more sustainable approach, even if a local model is not absolute SOTA.
GPUs are much more efficient at parallelizing requests for LLMs so it's going to much more efficient to centrally host. Maybe big companies it would make sense to get their own though.
Would you decide its usefulness based on how high the bill is, or how many things you get done while using it?
The former is the issue, and how many companies have been operating. It's like a trucking company ranking driver effectiveness by fuel used instead of by cargo moved.
The former. I’m able to get more Tickets done with it than without.