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.
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
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).
"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.
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
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.
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.
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.
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
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?
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.
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.
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.
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
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.
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.
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.
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.
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.
> 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.
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.
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).
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.
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.
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.
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.
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.
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.
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.
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.
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.
"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.
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.
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.
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?
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.
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.
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.
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.
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.
> 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.
> 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".
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.
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.
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.
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.
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.
>"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?
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.
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.
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…
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.
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.
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
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.
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.
> 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)
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.
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
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!
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).
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
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.
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.
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.
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.
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.
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.
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.
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.
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.
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).
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.
“Please don't post comments saying that HN is turning into Reddit. It's a semi-noob illusion, as old as the hills.” --hn guidelines (there are links to examples in the original)
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.
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.
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.
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.
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.
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
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
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.
"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.
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
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.
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.
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.
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
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?
1 reply →
Problem is that those 20% depend on the code reviews of the 80% for some form of pushback.
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.
Tokens are the new "lines of code per engineer". Easy to graph, easy to "manage".
The new TPS reports!
1 reply →
...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 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.
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."
10 replies →
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.
1 reply →
I worked at Uber from 2022-2025. The engineering culture was pretty abysmal, so it checks out.
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
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.
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.
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.
2 replies →
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.
What you describe is practically the opposite of tokenmaxxing.
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?
13 replies →
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.
> 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.
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.
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).
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
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.
4 replies →
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.
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.
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.
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.
There is always a rewriting around the corner
[dead]
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.
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.
2 replies →
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.
2 replies →
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.
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.
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.
"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.
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.
>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.
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?
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.
> 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".
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.
2 replies →
[dead]
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.
Someday maybe Goodhart's Law will be intuitive to people making decisions like this, but not any time soon I guess
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.
2 replies →
> 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
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.
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.
> 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.
3 replies →
> Some FAANGs are doing amazing things with unlimited tokens.
Would love to know what things!
10 replies →
> 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.
Where can I see those amazing things done by FAANGs?
3 replies →
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".
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?
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
Sounds like you want to be in the next round of layoffs?
But the executives need the fanciest models to evaluate how well they can replace the expensive labor costs.
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.
^ Philip K. Dick's unreleased book title
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.
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.
2 replies →
2^30 tokens costs something like 2^10 dollars, order of magnitude, if that helps ballpark.
[dead]
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.
I’m genuinely curious why they don’t cap at $100/month Claude Max per employee. That would be sufficient for 80% of 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...
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?
At what point might it be cheaper to, say, hire a human?
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 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.
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.
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.
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.
3 replies →
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
2 replies →
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
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.
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.
22 replies →
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…
5 replies →
> 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.
Is keeping your company private the easiest way to get around this?
It should not be overlooked that a lot of this fervor is from investors/board members putting pressure on these companies.
Lacking not just courage, but also character. Wasting company money on buzzwords and dubious outcomes is lack of character.
"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.
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.
6 replies →
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
1 reply →
> 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.
Not very Billion Dollar Valuation of you.
Yes, but, I sleep very soundly at night.
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.
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.
3 replies →
> 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)
1 reply →
Won’t be canned for going with the herd. I think it’s that simple, even if the herd is running off a cliff.
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.
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
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!
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
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).
11 replies →
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
2 replies →
Come on, don’t be crazy
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.
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.
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.
1 reply →
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.
6 replies →
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.
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.
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.
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.
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.
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.
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.
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).
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
[flagged]
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.
1 reply →
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.
[dead]
[dead]
[flagged]
Slightly ot, but I really dislike this reddit WSBization of HN.
Adds nothing insightful to these discussions.
“Please don't post comments saying that HN is turning into Reddit. It's a semi-noob illusion, as old as the hills.” --hn guidelines (there are links to examples in the original)
It's unfortunately the WSBification of the entire society.
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.
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.
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
1 reply →
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.
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.
[dead]