Tokenmaxxing was just a way to force employees to start leveraging AI in a meaningful way.
For companies that have measured performance based on token spend, they can now dial it back. Employees have learned to leverage AI for things they wouldn’t have prior. Now they know what’s possible and what’s not.
No one is stupid enough to always measure performance based on token spend and have unlimited budget. It was always a temporary thing to transition the employees to a new world.
Management felt like employees weren't leveraging AI fast enough. That's why in 2025, there were many mainstream articles about how CEOs were forcing their employees to use AI or get fired. Tokenmaxxing was just the other extreme. Companies will arrive at an equilibrium.
There's no need to overthink this.
Edit: One reply cited this X post as an example of why management needed to do this. Trying to change a company with hundreds/thousands/tens of thousands of employees is hard. You have to send one simple message at a time. https://x.com/danluu/status/1487228574608211969?lang=en
Its not _just_ that. Orgs aren't remotely sensible at measuring anything that isn't counted in dollars.
employees who are on the ai bandwagon are there for the free management attention.
Management is cooked because the damn market is hard, money is tight and they can't afford to fight the top down love and $$$ thrown at AI.
If you zoom out, all the real money spent on energy to keep AI alive isn't going to be held in nvidia stock for too long. it will burst, but its stupid to time it.
> Management felt like employees weren't leveraging AI fast enough.
If my productivity is in line with their expectations, I don’t understand why management cares what tools I’m using to do it. No employer ever told me to use emacs instead of vi, even though I’m 10x more productive in one vs the other. So why all of a sudden does management need to micromanage my tools?
Did they see productivity gains that they're now calibrating for? Why have these productivity gains not been reflected externally in any measurable way?
It's simpler. Management felt like employees weren't leveraging AI fast enough. They chose to measure "AI leveraging" in the easiest way they could: how many tokens each employee was using. Goodhart's Law ("When a measure becomes a target, it ceases to be a good measure") immediately triggered.
This is obviously wrong. Management has never cared about how engineers do their job. There's never been a push for any other productivity boosting technology: better languages, better editors, automated refactorings, paid code intelligence tools, etc. But suddenly AI comes along and the CEO says "all developers need to write code entirely with LLMs".
They really don’t IMO. Hell most of the companies pushing these tools don’t even agree what LLMs are for or are capable of. Too many people are trying to use it to cut too many corners on their work (making more work for everyone else) or are using it to attempt things they don’t know how to do, which means they are incapable of vetting the results, (vibe coding anyone?) which means more instances of the first case or even getting hurt.
Were there cheaper and humane ways to get more employees to use AI? (yes). Did many people JUST burn tokens (goodbart law)? (Yes). Would people revert back to the mean? (I think yes). Do many professionals hate AI because of this push? (Yes). Was the org net productive?
I wish there was an independent body truly assessing the impact of big tech decisions and running counterfactuals. Instead of accepting nice stories like this as a given.
Interesting take - upvoted you. I'm not convinced it's been the optimal management strategy, but you're succinctly explaining what they have done, not what they should have done, and in that sense you have a good point.
Still leaves huge questions about ROI ($26tln of TAM, anyone???) and doesn't quell the concerns brought forward by AI detractors though.
>I’ve basically never heard a business leader say that they were going to set a bunch of money on fire because it made them feel good.
Really? ~4 years ago our CEO hired a consultant to fly out several times to do team building exercises. We can't afford to do our 3-year server refresh cycle, but the consultant was no problem to pay.
We just recently had branding consultants come in and also spent thousands of dollars (AWS charges) on rebranding all our photos. We operate in a captive market, if you want to operate in our market you are required to subscribe to our service, and if you aren't in our market you can't subscribe. Branding at the end of the day drives 0 sales.
Heck, reminds me of the time a company I was working with hired a new CTO and one of the first things he did was as "server renaming scheme" using obscure (to the US-centric staff) city names from around the world (database servers are Swiss city names, web servers are Denmark, storage is Finland). We went from cattle naming to pet naming, for a CTO that lasted ~6 months.
In my experience company leadership is not quite as thrifty as this article likes to think they are.
The implication that tokenmaxxing was an intentional and thoughtfully considered approach rather than blind hype-following by an overpaid manager class who are too far removed from value to understand the downsides of LLMs is hysterical beyond belief.
Eh, it’s less hysterical than the ever-pervasive belief among junior devs that they are the smartest people in the room and that all managers everywhere are dumb.
Though I understand that gets social validation from other people with no actual experience.
The retconning is absurd. Companies CxOs's exhibit herd-line behaviors all the time: hiring scrum masters, mandatory return to office, and now tokenmaxxing. FOMO is a sufficient explanation of the behavior, any trend gaining momentum that they fear will give their competition an edge, they will reflexively force adoption with no further reasoning needed. There is no cost to boarding hype trains that go nowhere.
In my current company nobody forces you to use more tokens, but you're encouraged to write a 300 lines markdown skill.md which takes 8 minutes and costs 5 bucks to execute. That, instead of writing a 200 lines bash script doing all the same thing, but in a deterministic fashion, completing in under 5 seconds and costing 0 if you're not careful with rounding.
Beyond getting momentum going for a cmpany, Tokenmaxxing is lighting money on fire.
The idea of tokenmaxxing reaches different companies in different waves, so it will be discovered in waves and outgrown in waves in companies and industries in their own cycle.
In the long run, tokenmaxxing is like drunken sailor spending. Scaling is almost always about a large component of efficiency, and lighting money on fire in the street can only last so long.
Your comment implies no ROI on spent tokens. I get a lot more work done tokenmaxxing so the cost is negligible to me but YMMV. Of course there's no point in tokenmaxxing if you don't have enough work available to scale beyond yourself, or you're unable to use AI to do so.
I predict startups will continue to tokenmaxx while 40,000+ person companies will become a little more conservative.
Brute forcing positive outcomes by spending more tokens until a happy path manifests does not solve the underlying comprehension (and liability) problem.
I fear a world where critical software is stood up with
increasingly non-human governed abstraction because it [seems like it] works.
Software engineers as the review terminal in a conveyor of business-led code mass production... coming to a company near you?
You're right, but you'd be lucky if a real human actually reviews any code. At my company, merging a PR still requires 2 humans to press "Approve" but I've been instructed that I don't need to read the PR, I only need to click "Approve". This is what 30 years of SWE experience is being used for now.
It's AI usage mandates now, but rather than focusing on how the current hot topic has ripped through the business world, often without benefit nor repercussions at leadership, I'd prefer to analyze the higher pattern. We've recently experienced such ripples as the metaverse, blockchain/nft/web3, 'the cloud' (and a minor wave of cloud gaming). There was even a teacup buzz of 'apis', oddly disconnected from the semantic web.
Why do such fever dreams occur at all? Are they getting more prevalent? More damaging? Do they jepaordize the global economy? Should they be regulated in some fashion?
I can't prove my case, but I think it's a symptom of media manipulation/consolidation, the 'fiduciary duty' delusion, and that shareholders can hold the puppet strings tighter than they used to. More and more, they place their sillytown bets and expect the plebs to dance to them.
The dominance of finance capital over industrial capital reaching its absurd conclusion. NFT mania was only possible because we don't make anything here, no one has a serious plan to reshore and start making things here again, and we can't indefinitely maintain control of production we've exported to the 3rd world indefinitely. So you might as well play these symbolic games and increase your slice of the pie while the music is still playing.
This seems to happen with most big tech adoption in the first few years. The big data boom in the early 2010's had execs just buying up spark clusters and data lakes before they even had a clear analytical use case or governance.
Folks have been saying “things are different now, the agents are now compounding success instead of error” for at least a year now, but I just don’t see it. I was lucky enough to receive a weeklong $50k per head AI training from the people saying these things, and one of their few helpful concrete recommendations was to constantly clear context all the time, to avoid things going off the rails.
However, I think finding security vulnerabilities is one use case where it doesn’t matter. Tokenmaxxing is absolutely effective for that. We as an industry are in the middle of adopting very expensive, complex continuous fuzzers.
> That’s no longer true. We’ve entered a different regime, where spending more tokens generally results in better results. We call this “compounding correctness” — the more tokens you spend on getting a task correct, the more likely you’ll get a good outcome. We talked about this a bit at the last in person Agentics meetup:
Have we? Is it generally the case that the more tokens you spend, you better results you get? This take is so weird I suspect author somehow financially benefits from tokenmaxxing.
This is more likely the junior camper version of "not everything that counts can be counted, and not everything that can be counted counts."
In the early days of LLMs, we saw the classic hype-driven bi-modality of opinions. Folks were in the "fake news, fad" camp, or they were in the "omg, take over the world" camp.
Those of us closer to the space, with the awareness to know that there was some truth (and a lot of misjudgment) to go around, were in the middle of nowhere. When I co-wrote some driver code with Chat GPT, other engineers (and even one of our directors) told me to keep it quiet. At the same time I had directors and VPs asking me how we could accelerate adoption. For a while, I had access to a cheat code just because I had the audacity to not ask for permission. Folks were sure I would get in trouble for spending thousands per month in LLM operation, but a handful came along for the ride, burning tokens like firewood and learning along the way.
Tokenmaxxing is probably coming from at least a few things:
1. A course-correction for the practiced frugality that kept folks from jumping in and just learning at the ragged edge.
2. A willful and deliberate recognition that the best innovations in the later phases of a disruptive introduction often come from sparks of ideation in concentrations of activity. In other words, we don't know where good is, and we need to find it. (Charitable interpretation from the article)
3. Recognition that, even if they don't know why, leaders and product owners will get punished for not jumping in and, because of bullets 1 and 2, won't get punished for trying and missing. Even if they have no idea what they're doing, they're going to fake it until they make it (or slide into another job).
This last set is where the pain lives. An organization with healthy and increasing AI tool
usage will see elevated token counts, but so too will one using LLMs to rewrite wikipedia articles without the letter "m" to keep token counts high. These are pathological behaviors brought on by conflated metrics.
We had discussions about this in the early LLM days, where my old team was looking to ship new capabilities for older products. There was a lengthy VP-level discussion about getting to "80% usage" of the new system vs the old. Because the new system was a superset of the old, I eventually said "we can do that immediately, but it's a cost goal, where we're just aiming to make our business more expensive to operate, rather than a value goal for our users". We didn't adopt the target, but folks were understandably frustrated that they didn't have a straightforward way to measure and report progress.
Tokenmaxxing is, inevitably, a conflated goal, but it's what we have right now. Take advantage of the moment, learn, build, and keep an eye on levers for efficiency.
The thing that most disconcerts me isn't the runtime pruning, it's the cold loading. Months ago, I added a few skills and MCPs to test them, partly in the frenzy of free shopping, but then I forgot about half of them.
So after I got tired of choosing by hand, and therefore also a bit blindly, I created a small tool that runs locally and analyzes conversations to tell you which skills, MCPs, or other things are always unused.
347 items never used · ~19354 dead tokens/session · ~$25.49/month A lot of ECC that I never used but always loaded.
If anyone's interested, I've put it on GitHub, thousandflowers/skillreaper.
64 comments
[ 2.6 ms ] story [ 68.8 ms ] threadFor companies that have measured performance based on token spend, they can now dial it back. Employees have learned to leverage AI for things they wouldn’t have prior. Now they know what’s possible and what’s not.
No one is stupid enough to always measure performance based on token spend and have unlimited budget. It was always a temporary thing to transition the employees to a new world.
Management felt like employees weren't leveraging AI fast enough. That's why in 2025, there were many mainstream articles about how CEOs were forcing their employees to use AI or get fired. Tokenmaxxing was just the other extreme. Companies will arrive at an equilibrium.
There's no need to overthink this.
Edit: One reply cited this X post as an example of why management needed to do this. Trying to change a company with hundreds/thousands/tens of thousands of employees is hard. You have to send one simple message at a time. https://x.com/danluu/status/1487228574608211969?lang=en
employees who are on the ai bandwagon are there for the free management attention.
Management is cooked because the damn market is hard, money is tight and they can't afford to fight the top down love and $$$ thrown at AI.
If you zoom out, all the real money spent on energy to keep AI alive isn't going to be held in nvidia stock for too long. it will burst, but its stupid to time it.
Accenture was.
If my productivity is in line with their expectations, I don’t understand why management cares what tools I’m using to do it. No employer ever told me to use emacs instead of vi, even though I’m 10x more productive in one vs the other. So why all of a sudden does management need to micromanage my tools?
Surely for this specific example of managerial stupidity it just is, but I mean more generally, it's a beautiful posting.
I aspire to have this much misplaced belief in any humans at all, let alone CEOs.
They really don’t IMO. Hell most of the companies pushing these tools don’t even agree what LLMs are for or are capable of. Too many people are trying to use it to cut too many corners on their work (making more work for everyone else) or are using it to attempt things they don’t know how to do, which means they are incapable of vetting the results, (vibe coding anyone?) which means more instances of the first case or even getting hurt.
I wish there was an independent body truly assessing the impact of big tech decisions and running counterfactuals. Instead of accepting nice stories like this as a given.
Using Claude, I recently tried to do something similar for the Covid hiring spree: https://claude.ai/public/artifacts/21bba86a-ad5d-439c-861d-0...
Still leaves huge questions about ROI ($26tln of TAM, anyone???) and doesn't quell the concerns brought forward by AI detractors though.
Really? ~4 years ago our CEO hired a consultant to fly out several times to do team building exercises. We can't afford to do our 3-year server refresh cycle, but the consultant was no problem to pay.
We just recently had branding consultants come in and also spent thousands of dollars (AWS charges) on rebranding all our photos. We operate in a captive market, if you want to operate in our market you are required to subscribe to our service, and if you aren't in our market you can't subscribe. Branding at the end of the day drives 0 sales.
Heck, reminds me of the time a company I was working with hired a new CTO and one of the first things he did was as "server renaming scheme" using obscure (to the US-centric staff) city names from around the world (database servers are Swiss city names, web servers are Denmark, storage is Finland). We went from cattle naming to pet naming, for a CTO that lasted ~6 months.
In my experience company leadership is not quite as thrifty as this article likes to think they are.
Though I understand that gets social validation from other people with no actual experience.
Ugh. Tell me you're early in career without telling me. Sophomoric take.
In my current company nobody forces you to use more tokens, but you're encouraged to write a 300 lines markdown skill.md which takes 8 minutes and costs 5 bucks to execute. That, instead of writing a 200 lines bash script doing all the same thing, but in a deterministic fashion, completing in under 5 seconds and costing 0 if you're not careful with rounding.
The idea of tokenmaxxing reaches different companies in different waves, so it will be discovered in waves and outgrown in waves in companies and industries in their own cycle.
In the long run, tokenmaxxing is like drunken sailor spending. Scaling is almost always about a large component of efficiency, and lighting money on fire in the street can only last so long.
I predict startups will continue to tokenmaxx while 40,000+ person companies will become a little more conservative.
I fear a world where critical software is stood up with increasingly non-human governed abstraction because it [seems like it] works.
Software engineers as the review terminal in a conveyor of business-led code mass production... coming to a company near you?
Why do such fever dreams occur at all? Are they getting more prevalent? More damaging? Do they jepaordize the global economy? Should they be regulated in some fashion?
I can't prove my case, but I think it's a symptom of media manipulation/consolidation, the 'fiduciary duty' delusion, and that shareholders can hold the puppet strings tighter than they used to. More and more, they place their sillytown bets and expect the plebs to dance to them.
If you were tokenmaxxing you would understand.
Pet peeve of mine is nonsensical usage of the x is dead, long live x.
However, I think finding security vulnerabilities is one use case where it doesn’t matter. Tokenmaxxing is absolutely effective for that. We as an industry are in the middle of adopting very expensive, complex continuous fuzzers.
In Hawaii I assume?
Like ... pivoting to the "metaverse" and changing the company name to show he's serious.
Have we? Is it generally the case that the more tokens you spend, you better results you get? This take is so weird I suspect author somehow financially benefits from tokenmaxxing.
In the early days of LLMs, we saw the classic hype-driven bi-modality of opinions. Folks were in the "fake news, fad" camp, or they were in the "omg, take over the world" camp.
Those of us closer to the space, with the awareness to know that there was some truth (and a lot of misjudgment) to go around, were in the middle of nowhere. When I co-wrote some driver code with Chat GPT, other engineers (and even one of our directors) told me to keep it quiet. At the same time I had directors and VPs asking me how we could accelerate adoption. For a while, I had access to a cheat code just because I had the audacity to not ask for permission. Folks were sure I would get in trouble for spending thousands per month in LLM operation, but a handful came along for the ride, burning tokens like firewood and learning along the way.
Tokenmaxxing is probably coming from at least a few things:
1. A course-correction for the practiced frugality that kept folks from jumping in and just learning at the ragged edge.
2. A willful and deliberate recognition that the best innovations in the later phases of a disruptive introduction often come from sparks of ideation in concentrations of activity. In other words, we don't know where good is, and we need to find it. (Charitable interpretation from the article)
3. Recognition that, even if they don't know why, leaders and product owners will get punished for not jumping in and, because of bullets 1 and 2, won't get punished for trying and missing. Even if they have no idea what they're doing, they're going to fake it until they make it (or slide into another job).
This last set is where the pain lives. An organization with healthy and increasing AI tool usage will see elevated token counts, but so too will one using LLMs to rewrite wikipedia articles without the letter "m" to keep token counts high. These are pathological behaviors brought on by conflated metrics.
We had discussions about this in the early LLM days, where my old team was looking to ship new capabilities for older products. There was a lengthy VP-level discussion about getting to "80% usage" of the new system vs the old. Because the new system was a superset of the old, I eventually said "we can do that immediately, but it's a cost goal, where we're just aiming to make our business more expensive to operate, rather than a value goal for our users". We didn't adopt the target, but folks were understandably frustrated that they didn't have a straightforward way to measure and report progress.
Tokenmaxxing is, inevitably, a conflated goal, but it's what we have right now. Take advantage of the moment, learn, build, and keep an eye on levers for efficiency.
Citation needed
So after I got tired of choosing by hand, and therefore also a bit blindly, I created a small tool that runs locally and analyzes conversations to tell you which skills, MCPs, or other things are always unused.
347 items never used · ~19354 dead tokens/session · ~$25.49/month A lot of ECC that I never used but always loaded.
If anyone's interested, I've put it on GitHub, thousandflowers/skillreaper.