Not just Amazon, too. It feels like all of big tech (and some smaller firms) have simultaneously gone insane. Imagine if your CEO woke up one day and told the company: "We need to encourage travel spending. Please book as many business trips as you can, and spend as much money as possible. Fly first class to our satellite offices! Take limos instead of Ubers! Eat at fine restaurants! Make sure you are constantly traveling. In fact, we are going to make Travel Spending part of your annual performance review: If you don't spend enough on business travel, you'll get a low rating!"
I know some that was told to try and use AI more on the job so they created some agent to just burn tokens and ended up using about 10x what the next highest employee used. Buddy expected to get shit but instead got an accolade and was asked to give a short talk to the other employees about how they could match their success.
In my first job ever, I used to get my work done on time and leave. There were a few people who’d stay in the office until late and show up on weekends. Same output, but they got the promotions and my bonus got prorated.
If you've never seen this level of perverse incentive, you have been lucky. The creation of and subsequent exploitation of them aren't new. For pre computer examples: https://freakonomics.com/podcast/the-cobra-effect-2/
The worse example I know is the time the Belgians forced the Congolese to harvest more rubber by cutting their hands if they haven't reached the correct quota, ensuing a cross-tribe hands trading economy
At my company we were told AI spend was part of perf review and that the "singularity" had happened. Now 20% of our infrastructure spend is tokens. The average number of pull requests per dev per week increased with all this spend. From 4.2 to 5.1. And that includes a huge chunk of PRs that are just agents changing a line or two in a config. It's all magical thinking
My dad worked at a company that had their own travel agency (early 90s when you needed a travel agent for reasons that no longer apply), and he was often booked on the more expensive flight because the travel agency made more money. More than once he could have got first class for less on a different flight but company policy didn't allow him to fly first class.
Most big companies still have travel agencies/companies manage their corporate travel. I can’t remember who we used when I was at Amazon, but I made a similar complaint to my manager once given I could fly cheaper in a higher class on a different airline (also one I had heaps of points with so I would have preferred it because I’d be able to upgrade further and/or use the lounge).
Turns out the price I saw in the booking portal isn’t actually what Amazon paid. It’s kinda more like a rack rate listing. But then there’s all kinds of discounting/cash back that happens on the backend based on the amount of travel booked each month.
I used to know someone whose parent worked at travel agency (also 90s) and their whole immediate family could book trips wherever, but only economy class.
And the fact that it is an industry-wide meme at this point makes bright red flashing lights and klaxons go off on my mind that a catastrophic reckoning can't be too far. There's not enough money in the world to keep this up for too long.
I worked for an international (mothership in the UK, later acquired by the US) company, which had... sort of a similar policy.
So, the (mothership) company acquired a lot of satellite companies, all in banking business. All over the world. Then they figured their CEO was corrupt, got in problems with the law, got kicked out. While they were waiting for the new "real" CEO to step in, they let some "interim" CEO to take his place.
New new (interim) CEO didn't seem to have a clue about the business she was supposed to run, nor did she care. She knew her time was running out, and she figured she'd spend it traveling the world and partaking in fine dining in every corner of the world the company's tentacle could reach. But, to make it seem more plausible, she, sort of, created a policy of "experience exchange", which sent random troupes of select individuals from different branches of the company to "exchange experience" with another similarly randomly assembled troupe. Of course, the company picked the bill when it comes to lodging and dining.
Our inconsequential branch in Israel saw a pilgrimage of high-ranking banking managers from all over the world, but, mostly the wealthier parts of it. Some didn't even bother to show up in the office though, and proceeded straight to the banquet hall of the most expensive hotel on the Tel Aviv beach.
To be fair though, the interim CEO got the boot even before her time was supposed to end, but it was serendipitously close to the acquisition by the US company, and so she was let go as part of a "restructuring" and "optimization"... but it was a crazy year!
I kind of get what they're thinking in trying to make sure all engineers use AI. For myself, and for the engineers working with me, I saw everyone go through an initial aversion and resistance to AI, and then an instant productivity boost when we started using them. So there's definitely a good reason to get everybody to start using AI. You don't want a good engineer resisting AI indefinitely if you know it will make them more productive.
Incentivizing people who are already using AI to use as many tokens as possible does seem a little crazy, though.
It's worth reflecting on why it's so hard to convince hold outs to discover how AI might help them. The fundamental issue is that there really aren't many convincing demonstrations that hold outs can relate to and there remains basically no evidence of real value gained.
Users attest to higher productivity and point to material but intermediate factors like token use, generated lines of code, pr counts, etc, but there doesn't seem to be a convincing revolution in the quantity or quality of mature software being delivered.
Combine that puzzling impressions of outcomes with a sense, for many, that they don't feel like they have a personal problem that warrants a new tool, and you end up with a pretty earnest and defensible indifference.
To get hold out engineers using AI, the industry needs to be focused on demonstrating relatable workflow improvements and demonstrating practical improvements to finished work product. Instead, policies like token use incentives just rely on luring them into pulling the slot machine handle with the expectation that once they do, they'll join the cadre of other converts who justify their transition with subjective improvements and intermediate metrics.
There is a limit somewhere, but I keep finding more and more ways to use AI.
Not just coding, but things like "here is my teams mandate, go through all my company's slack channels, linear tasks, notion pages, and recent merges in got, summarize any work other teams are doing that intersect with my team's work."
That'll burn a lot of tokens.
Set that up to run once or twice a week and give a report.
Sure, findings ways to burn tokens is not hard. Even finding ways to burn tokens on things (like your example) which are actually useful is not hard. But what is the ROI on that from the company perspective. I mean, you could have also hired an intern to do the job of collating this report every week. But if you went to your boss and asked to hire someone to do something, they would, reasonably, ask what the value of that thing is and whether it justifies more headcount. But we're in this bizarro world where the bosses are basically saying "go hire more people, even if you don't have specific high-value things for them to do. Just create make-work jobs for them!" It's wild.
This would be hilarious if a bunch of companies did not already do exactly this with exec travel. And academics do this all the time when travel has to be funded from grants.
One reason it works out like that for travel funding is that it’s often the ‘use it or lose it’ kind of funding. If you do not use all of the funds allotted, you can’t ask for more and could realistically get less.
> It feels like all of big tech (and some smaller firms) have simultaneously gone insane.
Some companies might just have been scammed by the marketing that told them that AI would make all their employees 10,000x more productive and save them billions and when that didn't happen the assumption was that it's because employees weren't using the new all powerful AI as often as they should be.
Other companies, especially those working on their own AI products, might want employees to use AI as much as possible because they hope it will provide them with the training data they'll need to eventually replace most or all of those employees with the AI. Punishing workers who refuse to train their AI replacement might make sense to them because even if it's costly right now they expect the savings down the road to be much much greater.
It's more like "We really value face-to-face interaction, so we're going to track that with your total travel spend. We don't want to get in the way, so there's no budget."
Even as a very happy NVDA shareholder I agree with you. It's comical that managers are being so naïve as to think that you can crap out a dashboard of "tokens consumed per week" and get any useful signal at all from it, beyond learning who's not using AI.
Incompetent use of a coding agent, or just general shenanigans, can burn tokens all day but it's not going to get tickets done.
Just looking at the work output - how many story points, tickets, how many new bugs are opened, etc. has not become any less relevant a metric for productivity with AI. If you're a skilled and proper user of AI those numbers would be changing in the right direction, compared to before you had it.
> It's comical that managers are being so naïve as to think that you can crap out a dashboard of "tokens consumed per week" and get any useful signal at all from it, beyond learning who's not using AI.
If some guy decides to spend a bunch of money bringing AI tools into the company things might get very uncomfortable for him if they're seeing zero return on that investment. He's sure not going to get recognition and a massive bonus for it. If on the other hand, he can put some numbers in a spreadsheet or powerpoint showing that employees are using AI all the time and profits are up again this quarter, maybe he can take some credit for that or at least keep his boss or the company's shareholders from questioning the wisdom of dumping so much cash into those AI products.
All those numbers are equally gameable and terrible metrics for productivity. With any of those, as with AI spending, you've got to look at actual results qualitatively. There's no shortcut.
When I was at Amazon last year, the bragging (from the AI poo-bah in my section of Amazon, note) about AI included "look at the total line count of commits from the heaviest AI users!"
So if AI screws something up and re-writes it and then screws it up again, needing another re-write, that counted as more positive than if it was done correctly, and simply, the first time.
I don't know where you're working but LLM enhanced development has skyrocketed our rate of feature development. As an example, a project roadmapped to take 7 months was delivered in only 4.5 because of CC/Codex.
I'm confused how anyone could believe it isn't an enhancer, unless they have refused to use any of the technologies.
If we suddenly went from rail travel to jets that's exactly what would happen. We'd go from 0 to all the business flights that happen today. Everyone would be under enormous pressure to not be a laggard.
I washed a former intelligence agency person get interviewed on a youtube talk show and (tangential to the policy subject being discussed) they they said that's basically how it was after 9/11. We couldn't onboard people fast enough to figure out how to spend the money so while we were doing that we flew first class half way around the world to waterboard people with bottled water. The people authorizing it didn't care. They were spending X to fight terrorism. The public was never gonna see the nitty gritty breakdown.
That's basically how it seems to be with AI. Just replace "spent X fighting terrorism" with "spent X implementing AI workflows" or "invested X in AI" or whatever. Nobody actually knows or cares just how far the dollars are going.
I think this version is getting very close to The Emperor's New Clothes Subscription in terms of how transparently the leadership are displaying their delusions.
because it's come to CFO's as "free debt" aka fiat printing. They need to spend thisfree fiat to keep buble going. I'm sure some inv. banking team internally assured too. $Trillion instuitions have access to free printer now, you and I don't. This is different world since unlimited printer started in 2020. All debt math is fake now because they can create fiat money out of nothing,
literally.
I've definitely been in situations where managers tell me to "spend X amount before the end of the year." They don't want higher ups to think they can cut our budget.
> Imagine if your CEO woke up one day and told the company: "We need to encourage travel spending. Please book as many business trips as you can, and spend as much money as possible.
I had a manager like this once. He didn't last very long, but it was without a doubt the most fun six months of my career.
It's the state of modern capitalism. Money must flow from one entity to another even if nothing of tangible value is produced. The flows of money prove the growth of both businesses.
You mean like using lines of code as a metric to rank engineers [1]?
Managers love metrics. Bad managers particularly love metrics. Tokens used was almost the obvious bad metric that was going to be used.
I would argue that tokens used has actually exposed a useful metric: any manager who focused on this, demanded this or ranked based on this should be fired, for being a bad manager.
LoC can occasionally give you signal. For instance, imagine you are joining a new team or company so you don't know how much oversight your predecessor did. If you ask an engineer how they spend most of their time and they say "Mostly just writing code" and you look at GitHub and it says they've made 3 minor commits in the past quarter, that person is lying and your predecessor was incompetent (quite possibly both of them have been MIA from their responsibilities for months).
No, I'm not talking about the engineer who can point to significant contributions outside of code: writing technical specs, leading architecture discussions, etc. I'm talking about the ones who just say they're just coding, but are actually not working at all.
TL;DR LoC and commit count etc can be used only to flag for review likely cases of quiet quitting.
IMO, the investors behind AI play the Uber game: they subsidise the AI costs and inject it into all facets of society they can get their hands on. They can tell the execs to increase AI usage at any cost. Their bet is that we'll become AI addicts with athrophied brains before they run out of money.
Also, don't forget that their datacenters will burn our electricity and boil our rivers at rates much cheaper than what we are billed in our homes. So while you're happy generating mountains of AI slop, somewhere there is a datacenter boiling a river.
I'd compare this to a new patented formula of water that's nobody asked for, and the patent owners are trying to replace all water supply with their crap before we wake up.
No need to invoke a hypothetical water example, just look to how Nestlé pushed baby formula in developing countries¹:
>For example, IBFAN claims that Nestlé distributes free formula samples to hospitals and maternity wards; after leaving the hospital, the formula is no longer free, but because the supplementation has interfered with lactation, the family must continue to buy the formula.
Like six months ago we got a presentation from an AWS guy on the AI tooling available and how it fit with our particular use cases.
At one point seemingly out of nowhere he pointed out on his screen share "Look at how many tokens I've used this month. I run so much Opus." It was a number that was offensively large.
I remember thinking "That's a really odd flex, this crap is so expensive the fact that you use so much should be a red flag"
He demonstrated a number of Claude Code use cases he had to manage and tweak AWS infrastructure that made me, the old greybeard sysadmin older than the internet think "You've used AI to do something that was a single command."
So this story makes sense. They were being encouraged to just blast away at it six plus months ago.
I notice a lot of Cursor's suggestions are just stuff a linter should auto-fix.
But if you hit "tab" it'll claim that as an AI-edited line, LOL.
(A lot of the rest of it is stuff I could already have been doing just as fast if I'd ever bothered to learn to use multiple cursors, learned vim navigation, or set up some macros—I never did because my getting-code-on-the-screen speed without those has never been slow enough to hold anything up, in practice)
Cursor absolutely tries to maximize what they claim is "AI-edited" and it's nonsense a lot of the time. If it writes a function and then I got in and edit that function, it claims my edits _and_ any net-new lines I add above or below the function.
I still don't know how to reconcile these reports with what other people say about GenAI-agentic assisted engineering being the only way of working nowadays, especially in startups.
Probably there is no dichotomy going on and it depends on multiple factors, but it seems so weird to see reports that are so different between each other.
It's not required for startups. But if you are building trashy, brittle products and your main metric is speed to market, and have the expectation of high failure chances (e.g. most yc startup batches) - then yes you have to do agentic eng.
If you are making extremely specific, high quality products over a long time window and your founders are deeply experienced in that field of engineering, then no, you don't need agentic engineering and probably want very little llm code in general (outside of some boilerplate, internal toolings, etc).
I think GenAI-agentic assisted engineering is the only way of working nowadays, and it's the only way I personally have worked for months. I still think that an outright majority of presentations on AI tooling I've seen have been in the nonsensical "Look how many tokens I can burn" genre. Had to sit through one guy recently who explained why you need a complex agentic team with 6 different roles in order to ask Claude to investigate a bug, which you most definitely do not.
I think you'll find that a lot of big investment companies are buried to the hilt in a lot of tech companies and also OpenAI and Anthropic. So you can do the math on where the directive is coming from and why it's not particularly careful or measured.
> "You've used AI to do something that was a single command."
As time passes and the layers of abstraction pile up, later generations won't understand the underlying layers of the abstraction. This is a huge weakness in our systems development -- and a huge potential attack surface for adversaries.
> You've used AI to do something that was a single command
Yes, and that’s a good thing! This is in fact where a lot of AI value lies. You dont need to know that command anymore - knowing the functional contract is now sufficient to perform the requisite work duties. This is huge!
Not even joking that the main benefit I've seen from "AI" for editing code is that it lets me quickly do all the things I could already have been doing just as quickly if I'd ever bothered to learn to use my tools.
Of course I lose about as much time as I save to its fuck-ups, so I'd still have been better off learning to actually use a text editor properly. Though (as I mentioned in a another post) part of why I've never done that in 25ish years of writing code for pay is that my code-writing speed has never been too slow for any of the businesses I've worked in, i.e. other things move slowly enough it never mattered.
Look, I feel for junior admins, I was one 35 years ago and the only reason I'm where I am today was because I had to learn the hard way, repeatedly and often.
I use the shit out of opencode to do things as a force multiplier, not as a way to keep me from knowing what its doing.
The point at which we're optimizing for "we don't need to know that anymore" is the point at which everything blows up, because agentic work is not fully deterministic, models hallucinate even simple things.
Blindly relying on your agent weapon of choice to just do the right thing because you didn't take the time to understand how the lego fits together is an actual problem.
Replace agent with 'direct report' and you've just described middle management. For better or worse, companies have always run on non deterministic tasks doled out by persons who barely understand the work.
I can't tell if this comment is sarcasm or not. If you let AI run commands you don't understand (especially in production) you may end up with some nasty surprises.
With a comment like that, it's no wonder you're dramatically below our minimum guidance for tokens consumed.
If AI breaks production this way, you just tell AI to fix it! And look, now you've consumed tokens twice. Think on that and I'll see you at the end-of-year performance review.
I work at a FAANG (not Amazon), and have heard this a lot, both internally and publicly. Except, never officially from anyone that mattered (leadership). It always starts with a rumor and/or someone (internal) creating a dashboard/metric, and blows up from there. I've even heard leaders proclaim that it's NOT what they're looking at, and that you better NOT be wasting those expensive tokens.
Now, they might be; they've certainly used silly metrics in the past (LoC, commit count, etc.) without ever fully acknowledging it. But I don't believe that it's as simple as more tokens = more better.
Fellow FAANG. We have weekly manager meetings where leadership encourages us to increase token usage. We do push back, and leadership acknowledges that token spend is not a great metric and people are likely to game it... and then go right back to encouraging us to increase token spend in our teams.
We have token tracking dashboards that leadership is looking at. I know because they show us in these manager meetings. Haven't opened them to everyone yet as some kind of leaderboard, so at least that's nice.
Lots of rumors token spend will be involved in perf reviews. Leadership denies it... but then holds more meetings telling us how important it is to increase our token spend and discussing inadequacies from the token spend dashboards.
Interesting. When you say leadership along with manager meetings, are you referring to managers, who might just be exacerbating the rumors I mentioned, or actual company leadership, like Directors, Vps, etc? And are they saying “AI usage”, or explicitly “Token count”?
I'm in a large-ish peer group for engineering managers. AI token over-use is a growing problem.
The problem explodes at any company that puts up a token use leaderboard or hints that they might do layoffs for engineers that refuse to use AI tools. This triggers a race to use as many tokens as possible to stay ahead.
Anecdotally, the problem is worst among devs who read a lot of social media. Twitter, Threads, Mastodon, LinkedIn, and others are filled with recycled viral stories about companies going AI-native and firing people who don't use enough AI. Anxieties are high right now so nervous developers see this and think they must burn tokens faster than their peers to avoid an inevitable culling.
It's too bad that they go with a safe enterprise option that is so deficient that the outcomes will be bad and lots of people will learn useless lessons that don't translate to state of the art tools and usage patterns
I tried to have copilot create a powerpoint slide with some content and a rough design idea. It created an empty powerpoint slide with the default template, told me to download it, then created a separate bit of text and told me to copy and paste it in (random bullet point nonsense), no design elements. Whereas claude actually created a slide, with color and formatting, and the content was generated and fit it. Copilot feels like Chatgpt 3.0
Yes the PowerPoint generation was one shot only, it would let you make one and then if you wanted to make changes it would tell you to do it yourself lolllll
Not sure if this is still the case, I rarely use PowerPoint.
I feel like it depends on the leader. I've definitely seen leaders value LoC beyond reason and cause worse, bloated codebases by rewarding cowboys with 10k line PRs.
Big companies have thousands of leaders. Many good, many bad.
My friend at Google says they have a "ai-usage" dashboard that tracks everyone's ai token usage as well as aggregated per team, per org, etc. There's a sign on it that says "don't use this for perf reviews!" but I think everyone knows that that's exactly what they're going to use it for.
Lots of people reporting their "I had to use up my tokens, so I burned them on worthless stuff" stories. Incredible thing to do in a climate emergency. Push harder guys, maybe we can hit 3C warming?
This reminds me of the story of how the USSR nearly made whales extinct to meet a quota for whale meat that nobody wanted to eat.
I've been noticing how our economy keeps getting more Soviet as it becomes more top-down. We basically have central planning now with all the pathologies inherent in that system, but unlike the soviets we just have a bunch of guys who happened to get rich or bribe the right people running our GOSPLAN.
Things definitely feel 'Soviet' at my company. AI usage has been mandated by upper management (despite the fact that it doesn't really make sense or solve any problems in my particular job). They literally call it an "AI revolution." If you dare question the wisdom of the company's 'AI-First' policy, it's like you risk being singled out as a "counter-revolutionary."
Yeah, the stories I've heard from Meta are very Soviet-coded. Like, trying to exceed the plan but not too much, because then the new plan would be hopelessly unachievable and you'd be punished for not meeting the insane expectations.
The problem is that the founding fathers believed in constraining the state because it could be abusive, but they should have understood that all power ought to be subject to the people, not just state power.
I wonder what the largest and most powerful private enterprise the FFs knew about was. I suppose they'd probably heard of the Hudson's Bay Company, but I have no idea how they really felt about the potential that many normal people would feel equal amounts of domination from companies with revenue much larger than most countries' GDP.
No worries, we keep drinking from paper straws, because that is what really matters.
The problem with not burning tokens is when you not meet the performance KPIs, get labelled as luddite and off you go, even before the job gets taken over by AI.
I do agree with the sentiment, that and war mongers destroying the planet.
Bullshit work has hit escape velocity, won’t be long now before we have huge warehouses filled with people doing sudoku for their daily food allowance, and that’s just how our entire economy functions.
How are we sliding face first into “snowpiercer but dumber”?
Whale meat is really bad, now only eaten ceremonially in most places. It was an unrationed meat in the Britain during WWII, and people still didn't go for it.
We didn't. The USSR had 100% employment long ago[0], and all the poverty that goes with it.
This isn't like that, as it isn't funded through taxes. This is private companies experimenting with their money, and risking downstream cost increases that may cause people to go elsewhere, as they do when they try anything new.
This is much better than just funding people regardless of productivity through forced taxes.
Right now there are state govts bending over backwards to provide cheap energy for data centers. The difference is being paid by people who live nearby through increased electricity costs. This is a tax with just extra steps
Are you sure this isn't being funded by our taxes? How many data centers are being built in areas where they have been given a huge tax break? How many banks are loaning money for AI infrastructure knowing that they'll be bailed out by taxpayers if they fail?
This is simply not true, especially when you consider the massive amounts of government support so many parts of this "experiment with their own money" is getting. As a Utah resident its extremely evident in how forcefully they're pushing through what will be one of the largest datacenters in the world despite near universal disapproval from the citizens.
> We didn't. The USSR had 100% employment long ago[0], and all the poverty that goes with it.
I don't think USSR poverty rates surpassed those of Tsarist Russia that preceded them. To their credit, I think ideologic competition between capitalist and communist blocks was part of what allowed improvement of life conditions of workers in capitalist countries, after WWII. Fear of revolutions avoided one-percenters taking all productivity gains in the period. They had to share some to keep guillotines away. As soon as things went south in the USSR, from the 70s onwards, and capitalism took over the whole world, lacking any sort of viable extant competition, we reverted back to the old norm, the workers were denied their share of the productivity gains since then, and here are us now. A regime premised on free competition was undermined by lack of competition to itself.
I'd bet that the goal is for people to 'game' it though. By pushing people to use AI more they'll try it, experiment with it, 'waste' time on it ... and from that they'll learn about it. That's the end goal.
They're using tokens for pointless stuff right now in order to figure out use cases where it helps. You can't do that without also learning where it doesn't help.
That is exactly the point. It may be wasteful, but it's the fastest way to explore how AI may actually be useful to your business. Even if 80% of employees are just wasting tokens, you still have 20% who are figuring it out.
It is difficult to believe that you can cobra effect yourself into greatness. I'd rather say the most useful perk for companies doing this is the AI-washing adoption metrics they can report, which will hopefully (for them) increase valuations.
Even if that were true it'd mean that current AI usage is overshooting actual, productive use by 5x. This is a problem when all the AI projections are that the current state is the minimum and future usage will be 10+x.
I'm sorry, but that's insane. I mean, I guess if you have cash to burn I could think of even worse ways to spend it, but seriously, this is dumb. What other tool have businesses spent millions of dollars and person hours on to try and find something useful the tool can do?? Talk about a solution looking for a problem! If it's not clear in the early stages that this tool solves a problem then ditch it and move on! Give that extra cash to your employees and shareholders instead!
1. so is everyone who is subject to a corporate mandate... but...
2. this may be ok. A good way to learn a piece of software or tool or process is to play with it. We learn lots of general knowledge through play and experimentation. Heck we get better at musical instruments by playing on them.
Mandates are kind of dumb in many ways. But they will force the issue of discovering whether anything useful can come from AI other than coding.
Within Amazon, token usage is gamified if you use Kiro and your team isn't billed for it in the same way you are billed for AWS or have to account for your capacity in older systems. I've credibly heard of people gaming this internal ranking before anyone paid attention to it. There are also tons of enthusiasts doing all kinds of internal projects and sharing them.
There's definitely some pressure from managers when they hear about N00% productivity boosts in internal presentations, but where I am at they would figure out if you were making up tasks rather than working pretty quickly and the pressure comes from aggressive deadlines and a shift from the yearly OP1 process to a more agile one.
I've heard similar stories from AWS and other non-AWS FAANG employees. All of the token leaderboards have a "this doesn't count toward your performance review" disclaimer, but there's an implied nudge nudge, wink wink after that statement.
One person I've talked to has someone in their org who is running GasTown and chews through tokens 24/7. They don't contribute very much, but they're comfortably in the #1 spot.
So we've seen sellers of AI hardware invest in AI software companies to create demand for their hardware. Now we are seeing AI (and/or AI adjacent) companies requiring their employees to use AI to create demand for AI. When does this snake finish eating its own tail?
I'd do this if the other punch to follow didn't appear to be 'justify the expenditure'.
Choosing to wait for the PIP instead, if $EMPLOYER goes this way. Tell me the work I'm not doing and how pieces of ~~flair~~, sorry, tokens might help. Or don't, I don't care.
You're right; the justification is made by dancing. Meme moment... weird flex, but okay [for now].
I'm also hesitant to do this because it only means more B2B monopoly money/juiced stats. Or, God forbid, become the resident Token Expert. That praise you mention is exactly what I don't want!
I work at AWS (disclaimer opinions are my own, do not reflect views of my employer) and i think the existence of a leaderboard has led to folks gamifying it. People see peers in a higher tier on the leaderboard and start burning tokens to catch up.
I think the company realizes this and is actively trying to avoid this, since for the new tools there isn't a leaderboard.
I've done similar at my job where management wants us to use all of our tokens before they expire. I usually set it to documentation tasks and other minor tasks just to eat up tokens.
There's really no end to dot-language diagrams you can have it make. Call graphs, package dependency maps, let it try to figure out an architecture diagram, whatever.
Giving it busywork that you don't have the time or wherewithal to check carefully sounds like a disaster. Rather than introduce content that will be partially wrong and cause confusion if it's ever read, I'd consume the credits and send the output to /dev/null.
Just title it "draft". Odds are nobody will look at it anyway.
Add a pre-commit hook to re-create the diagrams on every commit (in case anything changed, of course), that way you can really burn tokens and look good to management.
At least that nominally creates some value at the end of the day. Documentation is the thing everyone wants but no one has time/desire to create. My most recent token heavy task was having an agent write unit tests for coverage on a little graphAPI tool I'd written a bit ago to satisfy SonarQube.
People don't want to read LLM-generated docs though. It'll lack the context to justify why things were designed the way they were, and there's always a risk of hallucination so you still have to verify the documentation's claims, since the person who published it likely did not scrutinize it.
There's two major types of documentation "why is this like this" documentation and then there's "here's the features of this library/tool" documentation. LLM stuff is fine for the latter as long as you screen it for hallucinations. Your right the former they can't really do because they don't have access to the reasoning but I've often found even the latter to be lacking in many teams.
...Yes. I didn't say fire and forget but it can handle a lot of rote recitation of library flags and functions perfectly well. The kind of stuff that's autogenerated with javadocs, inputs, outputs and effects that are all in the code are available to the LLMs. Like all things with LLMs generate and review but I've seen some good outputs with minimal errors that saved days of work no one was going to be given the time to do.
This is what happens when you can code faster than you can think. It’s kind of similar to a Facebook hiring 100s of engineers before it even knows what to do with them.
I dont know... this works out until someone approaches you and says: well we see you are using LOTS of tokens so you must be incredibly productive. Please show your results.
The type of leadership judging employee performance by token burn are usually doing it because they don’t have a clue how to judge performance so they’re just taking what they read on linkedin or their local cto roundtable about more AI = more better and turning it into a metric on the thing that makes a simple number.
I have colleagues at prime video who consult AI the way medieval clerks once consulted omens, generating entire chains of speculative labor after ritual examinations of any of their given codebases. no real or new initiatives / innovations are being pushed forward, and thats rumored to be happening in other departments as well.
Hasn't Anthropic being experiencing issues due to extremely high usage? Being their investor, you would think Amazon wouldn't do Anthropic dirty by weakening their ability to handle user traffic
People need to start yelling, throwing things and publicly mocking execs that do this. What is wrong with you all? I do this (except the throwing) and I get nothing but respect. If you've been a good little soldier for years, done nothing but deliver and then you raise your ire people will listen.
If you can't change your company, change your company!
Similar situation here! In fact our team has a no-LLM policy that I'm quite happy with. We did experiment with it, to the point that one of our seniors atrophied so badly we had to let him go, and we're still paying down some of the slop residue...
Being an investor in Anthropic, Amazon must have a preferred billing rate, but others do not. No wonder their revenue shot up so much, so fast, because of BS goals like those.
This is foolish. High token use is associated with worse output. If you fill your models context you are going to be using a lot more context but the labs literally put out charts of how the models degrade at high context use.
This is analogous to measuring productivity by LoC output.
Vicious cycle right here. Making up tasks to burn tokens -> Hey people love to use AI -> More data centers built -> You now have to make up more tasks to burn more tokens.
I think it's mixed. I have seen people with really good use cases and the opposite. It feels like the AWS/GCP situation all over again. Step 1: "this is amazing tech we need to leverage it immediately, use it as much as you can" Step 2: "oh shit this is getting expensive and I'm not sure of the ROI". We are approaching step 2
What's the root cause of these ridiculous decisions being taken at tech corporations? Constantly, they fall into fads like these that everyone with a brain knows make no sense but still many companies decide to follow them. For example: RTO -> what's the point of this shit? we never knew for sure but higher ups at most tech companies suddenly decided that RTO was the way to go forward despite all the downsides. Another example: DEI policies, some of them were very non-sensical.
I believe there has to be some downward pressure on these executives to take these decisions but I would like to know where it's coming from exactly and what's the logic behind them. Is it some big institution like Blackrock which has leverage on many of these companies? That's always been my bet but I never knew for sure.
Crappy managers don’t know (or actively avoid) how to measure business value from individuals. So they need you to be in the office so they can physically see if you are putting in the effort.
Tokens is just yet another proxy for business value.
The problem they face is if everybody is judge by business value in dollars, crappy managers are the first to go
This is what I do. I tell AI to go through every file in my project, identify up to 10 bugs per file, and then write the markdown with the name of the file plus "bugfix". This takes about 2 hours. Then I delete all the files with the suffix "bugfix" and then do it again.
You should probably create an agent to make agents whose jobs are to figure out how to maximize the token usage (and one whose job is to calculate the minimum token usage, so it doesn't look like a boondoggle).
I don’t even understand the point of making up tasks. Surely there’s some moonshot frustration project in your workday you could have an agent plugging away at, even if it’s unsuccessful.
Not just Amazon, too. It feels like all of big tech (and some smaller firms) have simultaneously gone insane. Imagine if your CEO woke up one day and told the company: "We need to encourage travel spending. Please book as many business trips as you can, and spend as much money as possible. Fly first class to our satellite offices! Take limos instead of Ubers! Eat at fine restaurants! Make sure you are constantly traveling. In fact, we are going to make Travel Spending part of your annual performance review: If you don't spend enough on business travel, you'll get a low rating!"
We are living in a totally bonkers time.
I know some that was told to try and use AI more on the job so they created some agent to just burn tokens and ended up using about 10x what the next highest employee used. Buddy expected to get shit but instead got an accolade and was asked to give a short talk to the other employees about how they could match their success.
In my first job ever, I used to get my work done on time and leave. There were a few people who’d stay in the office until late and show up on weekends. Same output, but they got the promotions and my bonus got prorated.
This is the same thing.
I believe it
i call BS on this story
If you've never seen this level of perverse incentive, you have been lucky. The creation of and subsequent exploitation of them aren't new. For pre computer examples: https://freakonomics.com/podcast/the-cobra-effect-2/
The worse example I know is the time the Belgians forced the Congolese to harvest more rubber by cutting their hands if they haven't reached the correct quota, ensuing a cross-tribe hands trading economy
https://en.wikipedia.org/wiki/Goodhart%27s_law
>This article is about statistics and government policy. For Nazi analogies in internet discussions, see Godwin's law.
I call unintended consequences on this KPI culture
They polished the turd more than stating, but the bones are real.
I have seen similar at my company so it is highly plausible.
I don’t.
Things that rhyme with this have indeed been happening at the biggest names.
I call AI on this comment
why?
Imitating your own utter lack of explanation or evidence?
At my company we were told AI spend was part of perf review and that the "singularity" had happened. Now 20% of our infrastructure spend is tokens. The average number of pull requests per dev per week increased with all this spend. From 4.2 to 5.1. And that includes a huge chunk of PRs that are just agents changing a line or two in a config. It's all magical thinking
It's a particularly extreme version of the disease that infects all tech executive decision-making:
1. Objective, quantitative metrics are the only valid basis for decision-making or comparison.
2. Except, hmm, I don't know how to write a SQL query for the outcome I'm actually interested in.
3. We'll use a convenient-to-measure proxy without any loss of enthusiasm for #1.
My dad worked at a company that had their own travel agency (early 90s when you needed a travel agent for reasons that no longer apply), and he was often booked on the more expensive flight because the travel agency made more money. More than once he could have got first class for less on a different flight but company policy didn't allow him to fly first class.
We have always been living in bonkers time.
Most big companies still have travel agencies/companies manage their corporate travel. I can’t remember who we used when I was at Amazon, but I made a similar complaint to my manager once given I could fly cheaper in a higher class on a different airline (also one I had heaps of points with so I would have preferred it because I’d be able to upgrade further and/or use the lounge).
Turns out the price I saw in the booking portal isn’t actually what Amazon paid. It’s kinda more like a rack rate listing. But then there’s all kinds of discounting/cash back that happens on the backend based on the amount of travel booked each month.
I used to know someone whose parent worked at travel agency (also 90s) and their whole immediate family could book trips wherever, but only economy class.
Exactly this.
And the fact that it is an industry-wide meme at this point makes bright red flashing lights and klaxons go off on my mind that a catastrophic reckoning can't be too far. There's not enough money in the world to keep this up for too long.
You'd be surprised...
I worked for an international (mothership in the UK, later acquired by the US) company, which had... sort of a similar policy.
So, the (mothership) company acquired a lot of satellite companies, all in banking business. All over the world. Then they figured their CEO was corrupt, got in problems with the law, got kicked out. While they were waiting for the new "real" CEO to step in, they let some "interim" CEO to take his place.
New new (interim) CEO didn't seem to have a clue about the business she was supposed to run, nor did she care. She knew her time was running out, and she figured she'd spend it traveling the world and partaking in fine dining in every corner of the world the company's tentacle could reach. But, to make it seem more plausible, she, sort of, created a policy of "experience exchange", which sent random troupes of select individuals from different branches of the company to "exchange experience" with another similarly randomly assembled troupe. Of course, the company picked the bill when it comes to lodging and dining.
Our inconsequential branch in Israel saw a pilgrimage of high-ranking banking managers from all over the world, but, mostly the wealthier parts of it. Some didn't even bother to show up in the office though, and proceeded straight to the banquet hall of the most expensive hotel on the Tel Aviv beach.
To be fair though, the interim CEO got the boot even before her time was supposed to end, but it was serendipitously close to the acquisition by the US company, and so she was let go as part of a "restructuring" and "optimization"... but it was a crazy year!
I kind of get what they're thinking in trying to make sure all engineers use AI. For myself, and for the engineers working with me, I saw everyone go through an initial aversion and resistance to AI, and then an instant productivity boost when we started using them. So there's definitely a good reason to get everybody to start using AI. You don't want a good engineer resisting AI indefinitely if you know it will make them more productive.
Incentivizing people who are already using AI to use as many tokens as possible does seem a little crazy, though.
It's worth reflecting on why it's so hard to convince hold outs to discover how AI might help them. The fundamental issue is that there really aren't many convincing demonstrations that hold outs can relate to and there remains basically no evidence of real value gained.
Users attest to higher productivity and point to material but intermediate factors like token use, generated lines of code, pr counts, etc, but there doesn't seem to be a convincing revolution in the quantity or quality of mature software being delivered.
Combine that puzzling impressions of outcomes with a sense, for many, that they don't feel like they have a personal problem that warrants a new tool, and you end up with a pretty earnest and defensible indifference.
To get hold out engineers using AI, the industry needs to be focused on demonstrating relatable workflow improvements and demonstrating practical improvements to finished work product. Instead, policies like token use incentives just rely on luring them into pulling the slot machine handle with the expectation that once they do, they'll join the cadre of other converts who justify their transition with subjective improvements and intermediate metrics.
There is a limit somewhere, but I keep finding more and more ways to use AI.
Not just coding, but things like "here is my teams mandate, go through all my company's slack channels, linear tasks, notion pages, and recent merges in got, summarize any work other teams are doing that intersect with my team's work."
That'll burn a lot of tokens.
Set that up to run once or twice a week and give a report.
Sure, findings ways to burn tokens is not hard. Even finding ways to burn tokens on things (like your example) which are actually useful is not hard. But what is the ROI on that from the company perspective. I mean, you could have also hired an intern to do the job of collating this report every week. But if you went to your boss and asked to hire someone to do something, they would, reasonably, ask what the value of that thing is and whether it justifies more headcount. But we're in this bizarro world where the bosses are basically saying "go hire more people, even if you don't have specific high-value things for them to do. Just create make-work jobs for them!" It's wild.
This would be hilarious if a bunch of companies did not already do exactly this with exec travel. And academics do this all the time when travel has to be funded from grants.
One reason it works out like that for travel funding is that it’s often the ‘use it or lose it’ kind of funding. If you do not use all of the funds allotted, you can’t ask for more and could realistically get less.
> It feels like all of big tech (and some smaller firms) have simultaneously gone insane.
Some companies might just have been scammed by the marketing that told them that AI would make all their employees 10,000x more productive and save them billions and when that didn't happen the assumption was that it's because employees weren't using the new all powerful AI as often as they should be.
Other companies, especially those working on their own AI products, might want employees to use AI as much as possible because they hope it will provide them with the training data they'll need to eventually replace most or all of those employees with the AI. Punishing workers who refuse to train their AI replacement might make sense to them because even if it's costly right now they expect the savings down the road to be much much greater.
It's like if class-based society materialized within the IT. And the manager class collectively pushes the narrative of AI replacing ICs.
Note that it has beaten capitalism, making rational choices to increase earnings has lost to this AI dream.
I think a lot of these execs have equity in Anthropic... and the dumb ones that don't are just "keeping up with the Joneses" so to speak.
It's more like "We really value face-to-face interaction, so we're going to track that with your total travel spend. We don't want to get in the way, so there's no budget."
Even as a very happy NVDA shareholder I agree with you. It's comical that managers are being so naïve as to think that you can crap out a dashboard of "tokens consumed per week" and get any useful signal at all from it, beyond learning who's not using AI.
Incompetent use of a coding agent, or just general shenanigans, can burn tokens all day but it's not going to get tickets done.
Just looking at the work output - how many story points, tickets, how many new bugs are opened, etc. has not become any less relevant a metric for productivity with AI. If you're a skilled and proper user of AI those numbers would be changing in the right direction, compared to before you had it.
> It's comical that managers are being so naïve as to think that you can crap out a dashboard of "tokens consumed per week" and get any useful signal at all from it, beyond learning who's not using AI.
If some guy decides to spend a bunch of money bringing AI tools into the company things might get very uncomfortable for him if they're seeing zero return on that investment. He's sure not going to get recognition and a massive bonus for it. If on the other hand, he can put some numbers in a spreadsheet or powerpoint showing that employees are using AI all the time and profits are up again this quarter, maybe he can take some credit for that or at least keep his boss or the company's shareholders from questioning the wisdom of dumping so much cash into those AI products.
All those numbers are equally gameable and terrible metrics for productivity. With any of those, as with AI spending, you've got to look at actual results qualitatively. There's no shortcut.
Bragging about token usage is like bragging about LoC written.
When I was at Amazon last year, the bragging (from the AI poo-bah in my section of Amazon, note) about AI included "look at the total line count of commits from the heaviest AI users!"
So if AI screws something up and re-writes it and then screws it up again, needing another re-write, that counted as more positive than if it was done correctly, and simply, the first time.
It’s honestly 10x worse than LOC. At least in the human era LOC had correlation to shipping features.
It’s more like bragging about compiler cycles spent.
I don't know where you're working but LLM enhanced development has skyrocketed our rate of feature development. As an example, a project roadmapped to take 7 months was delivered in only 4.5 because of CC/Codex.
I'm confused how anyone could believe it isn't an enhancer, unless they have refused to use any of the technologies.
Obligatory:
Negative 2000 Lines of Code
https://news.ycombinator.com/item?id=44381252
Versus my sibling comment to yours, I actually sent that to some internal folks after the bit about AI+total lines committed was said.
was there any kind of response or reaction to that? it’s something i would have done and probably wouldn’t have gone well. xD
If we suddenly went from rail travel to jets that's exactly what would happen. We'd go from 0 to all the business flights that happen today. Everyone would be under enormous pressure to not be a laggard.
I washed a former intelligence agency person get interviewed on a youtube talk show and (tangential to the policy subject being discussed) they they said that's basically how it was after 9/11. We couldn't onboard people fast enough to figure out how to spend the money so while we were doing that we flew first class half way around the world to waterboard people with bottled water. The people authorizing it didn't care. They were spending X to fight terrorism. The public was never gonna see the nitty gritty breakdown.
That's basically how it seems to be with AI. Just replace "spent X fighting terrorism" with "spent X implementing AI workflows" or "invested X in AI" or whatever. Nobody actually knows or cares just how far the dollars are going.
I think this version is getting very close to The Emperor's New Clothes Subscription in terms of how transparently the leadership are displaying their delusions.
because it's come to CFO's as "free debt" aka fiat printing. They need to spend thisfree fiat to keep buble going. I'm sure some inv. banking team internally assured too. $Trillion instuitions have access to free printer now, you and I don't. This is different world since unlimited printer started in 2020. All debt math is fake now because they can create fiat money out of nothing, literally.
I've definitely been in situations where managers tell me to "spend X amount before the end of the year." They don't want higher ups to think they can cut our budget.
> Imagine if your CEO woke up one day and told the company: "We need to encourage travel spending. Please book as many business trips as you can, and spend as much money as possible.
I had a manager like this once. He didn't last very long, but it was without a doubt the most fun six months of my career.
It’s preposterous, companies are blindly funding slop and the product is fool’s gold.
It's the state of modern capitalism. Money must flow from one entity to another even if nothing of tangible value is produced. The flows of money prove the growth of both businesses.
You mean like using lines of code as a metric to rank engineers [1]?
Managers love metrics. Bad managers particularly love metrics. Tokens used was almost the obvious bad metric that was going to be used.
I would argue that tokens used has actually exposed a useful metric: any manager who focused on this, demanded this or ranked based on this should be fired, for being a bad manager.
[1]: https://evan-soohoo.medium.com/did-elon-musk-really-fire-peo...
In many many many cases it's not the manager choosing to do that. Its our brilliant job creator class demanding that he does
LoC can occasionally give you signal. For instance, imagine you are joining a new team or company so you don't know how much oversight your predecessor did. If you ask an engineer how they spend most of their time and they say "Mostly just writing code" and you look at GitHub and it says they've made 3 minor commits in the past quarter, that person is lying and your predecessor was incompetent (quite possibly both of them have been MIA from their responsibilities for months).
No, I'm not talking about the engineer who can point to significant contributions outside of code: writing technical specs, leading architecture discussions, etc. I'm talking about the ones who just say they're just coding, but are actually not working at all.
TL;DR LoC and commit count etc can be used only to flag for review likely cases of quiet quitting.
I wonder where in business school they teach you to "measure inputs and try to maximize them", because that's basically what's happening.
IMO, the investors behind AI play the Uber game: they subsidise the AI costs and inject it into all facets of society they can get their hands on. They can tell the execs to increase AI usage at any cost. Their bet is that we'll become AI addicts with athrophied brains before they run out of money.
Also, don't forget that their datacenters will burn our electricity and boil our rivers at rates much cheaper than what we are billed in our homes. So while you're happy generating mountains of AI slop, somewhere there is a datacenter boiling a river.
I'd compare this to a new patented formula of water that's nobody asked for, and the patent owners are trying to replace all water supply with their crap before we wake up.
No need to invoke a hypothetical water example, just look to how Nestlé pushed baby formula in developing countries¹:
>For example, IBFAN claims that Nestlé distributes free formula samples to hospitals and maternity wards; after leaving the hospital, the formula is no longer free, but because the supplementation has interfered with lactation, the family must continue to buy the formula.
1: https://en.wikipedia.org/wiki/1977_Nestl%C3%A9_boycott
But Brawndo's got what plants crave. It's got electrolytes.
[delayed]
Like six months ago we got a presentation from an AWS guy on the AI tooling available and how it fit with our particular use cases.
At one point seemingly out of nowhere he pointed out on his screen share "Look at how many tokens I've used this month. I run so much Opus." It was a number that was offensively large.
I remember thinking "That's a really odd flex, this crap is so expensive the fact that you use so much should be a red flag"
He demonstrated a number of Claude Code use cases he had to manage and tweak AWS infrastructure that made me, the old greybeard sysadmin older than the internet think "You've used AI to do something that was a single command."
So this story makes sense. They were being encouraged to just blast away at it six plus months ago.
I notice a lot of Cursor's suggestions are just stuff a linter should auto-fix.
But if you hit "tab" it'll claim that as an AI-edited line, LOL.
(A lot of the rest of it is stuff I could already have been doing just as fast if I'd ever bothered to learn to use multiple cursors, learned vim navigation, or set up some macros—I never did because my getting-code-on-the-screen speed without those has never been slow enough to hold anything up, in practice)
Cursor absolutely tries to maximize what they claim is "AI-edited" and it's nonsense a lot of the time. If it writes a function and then I got in and edit that function, it claims my edits _and_ any net-new lines I add above or below the function.
I still don't know how to reconcile these reports with what other people say about GenAI-agentic assisted engineering being the only way of working nowadays, especially in startups.
Probably there is no dichotomy going on and it depends on multiple factors, but it seems so weird to see reports that are so different between each other.
It's not required for startups. But if you are building trashy, brittle products and your main metric is speed to market, and have the expectation of high failure chances (e.g. most yc startup batches) - then yes you have to do agentic eng.
If you are making extremely specific, high quality products over a long time window and your founders are deeply experienced in that field of engineering, then no, you don't need agentic engineering and probably want very little llm code in general (outside of some boilerplate, internal toolings, etc).
I think GenAI-agentic assisted engineering is the only way of working nowadays, and it's the only way I personally have worked for months. I still think that an outright majority of presentations on AI tooling I've seen have been in the nonsensical "Look how many tokens I can burn" genre. Had to sit through one guy recently who explained why you need a complex agentic team with 6 different roles in order to ask Claude to investigate a bug, which you most definitely do not.
Wage workers are evaluated on behaviors, founders are evaluated on growth and revenue. Of course usage patterns and outcomes will be different
I think you'll find that a lot of big investment companies are buried to the hilt in a lot of tech companies and also OpenAI and Anthropic. So you can do the math on where the directive is coming from and why it's not particularly careful or measured.
> "You've used AI to do something that was a single command."
As time passes and the layers of abstraction pile up, later generations won't understand the underlying layers of the abstraction. This is a huge weakness in our systems development -- and a huge potential attack surface for adversaries.
> You've used AI to do something that was a single command
Yes, and that’s a good thing! This is in fact where a lot of AI value lies. You dont need to know that command anymore - knowing the functional contract is now sufficient to perform the requisite work duties. This is huge!
Not even joking that the main benefit I've seen from "AI" for editing code is that it lets me quickly do all the things I could already have been doing just as quickly if I'd ever bothered to learn to use my tools.
Of course I lose about as much time as I save to its fuck-ups, so I'd still have been better off learning to actually use a text editor properly. Though (as I mentioned in a another post) part of why I've never done that in 25ish years of writing code for pay is that my code-writing speed has never been too slow for any of the businesses I've worked in, i.e. other things move slowly enough it never mattered.
Once I learn a command that is both repeatable and useful, I prefer to either keep it in my mind or in my aliases. Thank you.
You can still do this! And AI will teach you that command far far faster than synthesizing it yourself.
That's what Skills are for
:^)
Is it? If the LLMs change broke something do you know enough to fix it?
The same question can be applied to work without AI, so this isn’t a meaningful criticism
Look, I feel for junior admins, I was one 35 years ago and the only reason I'm where I am today was because I had to learn the hard way, repeatedly and often.
I use the shit out of opencode to do things as a force multiplier, not as a way to keep me from knowing what its doing.
The point at which we're optimizing for "we don't need to know that anymore" is the point at which everything blows up, because agentic work is not fully deterministic, models hallucinate even simple things.
Blindly relying on your agent weapon of choice to just do the right thing because you didn't take the time to understand how the lego fits together is an actual problem.
Replace agent with 'direct report' and you've just described middle management. For better or worse, companies have always run on non deterministic tasks doled out by persons who barely understand the work.
> You dont need to know that command anymore
I find it hard to read "You can do things without knowing things" as a positive improvement in work, society, life, anywhere
You are the worse kind of gatekeeper, then! A true reactionary who believes they are righteous for impeding others!
It's also several hundred times more expensive.
I can't tell if this comment is sarcasm or not. If you let AI run commands you don't understand (especially in production) you may end up with some nasty surprises.
With a comment like that, it's no wonder you're dramatically below our minimum guidance for tokens consumed.
If AI breaks production this way, you just tell AI to fix it! And look, now you've consumed tokens twice. Think on that and I'll see you at the end-of-year performance review.
Play stupid games, win stupid prizes.
I work at a FAANG (not Amazon), and have heard this a lot, both internally and publicly. Except, never officially from anyone that mattered (leadership). It always starts with a rumor and/or someone (internal) creating a dashboard/metric, and blows up from there. I've even heard leaders proclaim that it's NOT what they're looking at, and that you better NOT be wasting those expensive tokens.
Now, they might be; they've certainly used silly metrics in the past (LoC, commit count, etc.) without ever fully acknowledging it. But I don't believe that it's as simple as more tokens = more better.
Fellow FAANG. We have weekly manager meetings where leadership encourages us to increase token usage. We do push back, and leadership acknowledges that token spend is not a great metric and people are likely to game it... and then go right back to encouraging us to increase token spend in our teams.
We have token tracking dashboards that leadership is looking at. I know because they show us in these manager meetings. Haven't opened them to everyone yet as some kind of leaderboard, so at least that's nice.
Lots of rumors token spend will be involved in perf reviews. Leadership denies it... but then holds more meetings telling us how important it is to increase our token spend and discussing inadequacies from the token spend dashboards.
Interesting. When you say leadership along with manager meetings, are you referring to managers, who might just be exacerbating the rumors I mentioned, or actual company leadership, like Directors, Vps, etc? And are they saying “AI usage”, or explicitly “Token count”?
I'm in a large-ish peer group for engineering managers. AI token over-use is a growing problem.
The problem explodes at any company that puts up a token use leaderboard or hints that they might do layoffs for engineers that refuse to use AI tools. This triggers a race to use as many tokens as possible to stay ahead.
Anecdotally, the problem is worst among devs who read a lot of social media. Twitter, Threads, Mastodon, LinkedIn, and others are filled with recycled viral stories about companies going AI-native and firing people who don't use enough AI. Anxieties are high right now so nervous developers see this and think they must burn tokens faster than their peers to avoid an inevitable culling.
In our place it is really a thing and comes from leadership. They feel like they spent a lot on copilot and they want to see people using it.
It's too bad that they go with a safe enterprise option that is so deficient that the outcomes will be bad and lots of people will learn useless lessons that don't translate to state of the art tools and usage patterns
Yes agreed. It's actually the office version of copilot which isn't much to write home about.
Apparently the github one is more useful for its target audience.
I tried to have copilot create a powerpoint slide with some content and a rough design idea. It created an empty powerpoint slide with the default template, told me to download it, then created a separate bit of text and told me to copy and paste it in (random bullet point nonsense), no design elements. Whereas claude actually created a slide, with color and formatting, and the content was generated and fit it. Copilot feels like Chatgpt 3.0
Yes the PowerPoint generation was one shot only, it would let you make one and then if you wanted to make changes it would tell you to do it yourself lolllll
Not sure if this is still the case, I rarely use PowerPoint.
I feel like it depends on the leader. I've definitely seen leaders value LoC beyond reason and cause worse, bloated codebases by rewarding cowboys with 10k line PRs.
Big companies have thousands of leaders. Many good, many bad.
Enterprise consulting here, it is getting ridiculous, with forced trainings, workshops and hacktons to motivate use of AI in daily activities.
Stuff that could be easily done as shell scripts gets asked how could we make an agent out of it.
My friend at Google says they have a "ai-usage" dashboard that tracks everyone's ai token usage as well as aggregated per team, per org, etc. There's a sign on it that says "don't use this for perf reviews!" but I think everyone knows that that's exactly what they're going to use it for.
Lots of people reporting their "I had to use up my tokens, so I burned them on worthless stuff" stories. Incredible thing to do in a climate emergency. Push harder guys, maybe we can hit 3C warming?
This reminds me of the story of how the USSR nearly made whales extinct to meet a quota for whale meat that nobody wanted to eat.
I've been noticing how our economy keeps getting more Soviet as it becomes more top-down. We basically have central planning now with all the pathologies inherent in that system, but unlike the soviets we just have a bunch of guys who happened to get rich or bribe the right people running our GOSPLAN.
Things definitely feel 'Soviet' at my company. AI usage has been mandated by upper management (despite the fact that it doesn't really make sense or solve any problems in my particular job). They literally call it an "AI revolution." If you dare question the wisdom of the company's 'AI-First' policy, it's like you risk being singled out as a "counter-revolutionary."
Yeah, the stories I've heard from Meta are very Soviet-coded. Like, trying to exceed the plan but not too much, because then the new plan would be hopelessly unachievable and you'd be punished for not meeting the insane expectations.
The problem is that the founding fathers believed in constraining the state because it could be abusive, but they should have understood that all power ought to be subject to the people, not just state power.
I wonder what the largest and most powerful private enterprise the FFs knew about was. I suppose they'd probably heard of the Hudson's Bay Company, but I have no idea how they really felt about the potential that many normal people would feel equal amounts of domination from companies with revenue much larger than most countries' GDP.
No worries, we keep drinking from paper straws, because that is what really matters.
The problem with not burning tokens is when you not meet the performance KPIs, get labelled as luddite and off you go, even before the job gets taken over by AI.
I do agree with the sentiment, that and war mongers destroying the planet.
This is why we're clear-cutting forests to build new data centers? Not even for "real" productivity gains, but just for the sake of using the tokens.
Bullshit work has hit escape velocity, won’t be long now before we have huge warehouses filled with people doing sudoku for their daily food allowance, and that’s just how our entire economy functions.
How are we sliding face first into “snowpiercer but dumber”?
Gotta scale and then IPO those startups, so the VCs can cash out profitably.
> USSR nearly made whales extinct
USSR barely accounted for 15% of the world caught amount (with Japan as the leader).
> that nobody wanted to eat
unsubstantiated.
Whale meat is really bad, now only eaten ceremonially in most places. It was an unrationed meat in the Britain during WWII, and people still didn't go for it.
Yeah but what can we do. I don't want to be punished by work either.
Luckily I work in app management and I know they can only see the last date used so if I just put in one query per day I'm good.
But I'm so sick and tired of this AI hype :(
It's a shame AI now has a universal basic jobs[1] program, but humans still not. Companies are paying AI to dig holes, so other AI can fill them.
[1] https://locusmag.com/feature/cory-doctorow-full-employment/
We didn't. The USSR had 100% employment long ago[0], and all the poverty that goes with it.
This isn't like that, as it isn't funded through taxes. This is private companies experimenting with their money, and risking downstream cost increases that may cause people to go elsewhere, as they do when they try anything new.
This is much better than just funding people regardless of productivity through forced taxes.
[0] https://nintil.com/the-soviet-union-achieving-full-employmen...
Right now there are state govts bending over backwards to provide cheap energy for data centers. The difference is being paid by people who live nearby through increased electricity costs. This is a tax with just extra steps
Are you sure this isn't being funded by our taxes? How many data centers are being built in areas where they have been given a huge tax break? How many banks are loaning money for AI infrastructure knowing that they'll be bailed out by taxpayers if they fail?
> as it isn't funded through taxes
This is simply not true, especially when you consider the massive amounts of government support so many parts of this "experiment with their own money" is getting. As a Utah resident its extremely evident in how forcefully they're pushing through what will be one of the largest datacenters in the world despite near universal disapproval from the citizens.
> We didn't. The USSR had 100% employment long ago[0], and all the poverty that goes with it.
I don't think USSR poverty rates surpassed those of Tsarist Russia that preceded them. To their credit, I think ideologic competition between capitalist and communist blocks was part of what allowed improvement of life conditions of workers in capitalist countries, after WWII. Fear of revolutions avoided one-percenters taking all productivity gains in the period. They had to share some to keep guillotines away. As soon as things went south in the USSR, from the 70s onwards, and capitalism took over the whole world, lacking any sort of viable extant competition, we reverted back to the old norm, the workers were denied their share of the productivity gains since then, and here are us now. A regime premised on free competition was undermined by lack of competition to itself.
I'd bet that the goal is for people to 'game' it though. By pushing people to use AI more they'll try it, experiment with it, 'waste' time on it ... and from that they'll learn about it. That's the end goal.
They're using tokens for pointless stuff right now in order to figure out use cases where it helps. You can't do that without also learning where it doesn't help.
My company is doing the same thing.
That is exactly the point. It may be wasteful, but it's the fastest way to explore how AI may actually be useful to your business. Even if 80% of employees are just wasting tokens, you still have 20% who are figuring it out.
It is difficult to believe that you can cobra effect yourself into greatness. I'd rather say the most useful perk for companies doing this is the AI-washing adoption metrics they can report, which will hopefully (for them) increase valuations.
Even if that were true it'd mean that current AI usage is overshooting actual, productive use by 5x. This is a problem when all the AI projections are that the current state is the minimum and future usage will be 10+x.
I'm sorry, but that's insane. I mean, I guess if you have cash to burn I could think of even worse ways to spend it, but seriously, this is dumb. What other tool have businesses spent millions of dollars and person hours on to try and find something useful the tool can do?? Talk about a solution looking for a problem! If it's not clear in the early stages that this tool solves a problem then ditch it and move on! Give that extra cash to your employees and shareholders instead!
Goodharts Law - When a measure becomes a target, it ceases to be a good measure.
https://lawsofsoftwareengineering.com/laws/goodharts-law/
But does Goodhearts law apply to things that were never a good measure to start with :)
1. so is everyone who is subject to a corporate mandate... but...
2. this may be ok. A good way to learn a piece of software or tool or process is to play with it. We learn lots of general knowledge through play and experimentation. Heck we get better at musical instruments by playing on them.
Mandates are kind of dumb in many ways. But they will force the issue of discovering whether anything useful can come from AI other than coding.
Within Amazon, token usage is gamified if you use Kiro and your team isn't billed for it in the same way you are billed for AWS or have to account for your capacity in older systems. I've credibly heard of people gaming this internal ranking before anyone paid attention to it. There are also tons of enthusiasts doing all kinds of internal projects and sharing them.
There's definitely some pressure from managers when they hear about N00% productivity boosts in internal presentations, but where I am at they would figure out if you were making up tasks rather than working pretty quickly and the pressure comes from aggressive deadlines and a shift from the yearly OP1 process to a more agile one.
I've heard similar stories from AWS and other non-AWS FAANG employees. All of the token leaderboards have a "this doesn't count toward your performance review" disclaimer, but there's an implied nudge nudge, wink wink after that statement.
One person I've talked to has someone in their org who is running GasTown and chews through tokens 24/7. They don't contribute very much, but they're comfortably in the #1 spot.
I have heard from multiple people at smaller to medium sized orgs where token usage and AI adoption are a central part of performance reviews.
So we've seen sellers of AI hardware invest in AI software companies to create demand for their hardware. Now we are seeing AI (and/or AI adjacent) companies requiring their employees to use AI to create demand for AI. When does this snake finish eating its own tail?
When are they going to admit that they over invested in AI and somehow have to justify that spend with usage down our throat?
Just like return-to-office is there to maintain real estate value.
I'd do this if the other punch to follow didn't appear to be 'justify the expenditure'.
Choosing to wait for the PIP instead, if $EMPLOYER goes this way. Tell me the work I'm not doing and how pieces of ~~flair~~, sorry, tokens might help. Or don't, I don't care.
Anecdotal, but appears to be common among other comments in the thread.
For companies doing this there is no 'justify the expenditure'. Employees are being praised for high expenditure, regardless of actual outcome.
Leadership see the problem as 'people resisting AI'. Embracing AI is seen as the solution, and token usage is seen as the measure of success.
You're right; the justification is made by dancing. Meme moment... weird flex, but okay [for now].
I'm also hesitant to do this because it only means more B2B monopoly money/juiced stats. Or, God forbid, become the resident Token Expert. That praise you mention is exactly what I don't want!
Made this as a joke but maybe it'll get some use
https://token-burner.pages.dev/
I work at AWS (disclaimer opinions are my own, do not reflect views of my employer) and i think the existence of a leaderboard has led to folks gamifying it. People see peers in a higher tier on the leaderboard and start burning tokens to catch up.
I think the company realizes this and is actively trying to avoid this, since for the new tools there isn't a leaderboard.
This is coming to my workplace too. They send us angry reminders if we don't use copilot in ms office every day :( I just type Hello to it.
When your incentive is to tokenmaxx don’t be suprised when people game the system. Measurements something somethjng benchmark something something bad.
I've done similar at my job where management wants us to use all of our tokens before they expire. I usually set it to documentation tasks and other minor tasks just to eat up tokens.
There's really no end to dot-language diagrams you can have it make. Call graphs, package dependency maps, let it try to figure out an architecture diagram, whatever.
Giving it busywork that you don't have the time or wherewithal to check carefully sounds like a disaster. Rather than introduce content that will be partially wrong and cause confusion if it's ever read, I'd consume the credits and send the output to /dev/null.
Just title it "draft". Odds are nobody will look at it anyway.
Add a pre-commit hook to re-create the diagrams on every commit (in case anything changed, of course), that way you can really burn tokens and look good to management.
At least that nominally creates some value at the end of the day. Documentation is the thing everyone wants but no one has time/desire to create. My most recent token heavy task was having an agent write unit tests for coverage on a little graphAPI tool I'd written a bit ago to satisfy SonarQube.
People don't want to read LLM-generated docs though. It'll lack the context to justify why things were designed the way they were, and there's always a risk of hallucination so you still have to verify the documentation's claims, since the person who published it likely did not scrutinize it.
There's two major types of documentation "why is this like this" documentation and then there's "here's the features of this library/tool" documentation. LLM stuff is fine for the latter as long as you screen it for hallucinations. Your right the former they can't really do because they don't have access to the reasoning but I've often found even the latter to be lacking in many teams.
If you have artifacts saved as you develop it can use those when writing docs to capture intent and design decisions.
Inaccurate documentation can be worse than no documentation at all!
...Yes. I didn't say fire and forget but it can handle a lot of rote recitation of library flags and functions perfectly well. The kind of stuff that's autogenerated with javadocs, inputs, outputs and effects that are all in the code are available to the LLMs. Like all things with LLMs generate and review but I've seen some good outputs with minimal errors that saved days of work no one was going to be given the time to do.
This is what happens when you can code faster than you can think. It’s kind of similar to a Facebook hiring 100s of engineers before it even knows what to do with them.
The Casual form of Goodhart's Law... https://unintendedconsequenc.es/new-morality-of-attainment-g...
I dont know... this works out until someone approaches you and says: well we see you are using LOTS of tokens so you must be incredibly productive. Please show your results.
The type of leadership judging employee performance by token burn are usually doing it because they don’t have a clue how to judge performance so they’re just taking what they read on linkedin or their local cto roundtable about more AI = more better and turning it into a metric on the thing that makes a simple number.
I have colleagues at prime video who consult AI the way medieval clerks once consulted omens, generating entire chains of speculative labor after ritual examinations of any of their given codebases. no real or new initiatives / innovations are being pushed forward, and thats rumored to be happening in other departments as well.
Hasn't Anthropic being experiencing issues due to extremely high usage? Being their investor, you would think Amazon wouldn't do Anthropic dirty by weakening their ability to handle user traffic
Amazon runs anthropic models in it's own DCs with Bedrock.
How does this work?
Anthropic sends .gguf and a claude-serve binary?
GGUF is mostly a hobbyist format, but yeah I'd assume it's a big pile of tensors and and executable.
People need to start yelling, throwing things and publicly mocking execs that do this. What is wrong with you all? I do this (except the throwing) and I get nothing but respect. If you've been a good little soldier for years, done nothing but deliver and then you raise your ire people will listen.
If you can't change your company, change your company!
Similar situation here! In fact our team has a no-LLM policy that I'm quite happy with. We did experiment with it, to the point that one of our seniors atrophied so badly we had to let him go, and we're still paying down some of the slop residue...
Let it write unit tests for every single function in the codebase lol
I've chosen the wrong profession.
Being an investor in Anthropic, Amazon must have a preferred billing rate, but others do not. No wonder their revenue shot up so much, so fast, because of BS goals like those.
This is foolish. High token use is associated with worse output. If you fill your models context you are going to be using a lot more context but the labs literally put out charts of how the models degrade at high context use.
This is analogous to measuring productivity by LoC output.
> This is analogous to measuring productivity by LoC output
True, but it looks like productivity to people whose own productivity is measured by how busy their subordinates appear to be.
Good old Goodhart's law. https://xkcd.com/2899/
Dumb bureaucracy with dumb requirements will be met with corresponding response.
Token-driven development
Vicious cycle right here. Making up tasks to burn tokens -> Hey people love to use AI -> More data centers built -> You now have to make up more tasks to burn more tokens.
Waiting for the YC startup in the next batch that provides tokenmaxxing-as-a-service.
Goodhart's Law in effect right there.
Use Vim or you're fired!
this one is legit the right call, though
But how would you compete when people inevitably started mandating superior editors such as emacs[1] or even notepad[2]?
[1] https://www.gnu.org/software/emacs/
[2] https://en.wikipedia.org/wiki/Windows_Notepad
I knew a guy who used Microsoft Word
Corporate tech has accelerated into a preposterous trajectory.
Burn resources at all costs to appear productive and use proxy metrics to measure success.
Fire productive employees to ensure we have resources to fund the proxy metrics.
AI slop fool’s gold is the product.
"When a metric becomes the target, it ceases to be a good measure".
Long live Goodhart!
Love it. This needs to become a new trend, and price per token can't rise soon enough.
Has anyone actually seen true business lift from agents or is this one of those "do stupid things faster" situation?
I think it's mixed. I have seen people with really good use cases and the opposite. It feels like the AWS/GCP situation all over again. Step 1: "this is amazing tech we need to leverage it immediately, use it as much as you can" Step 2: "oh shit this is getting expensive and I'm not sure of the ROI". We are approaching step 2
Narrator: “it wasn’t just Amazon”
What's the root cause of these ridiculous decisions being taken at tech corporations? Constantly, they fall into fads like these that everyone with a brain knows make no sense but still many companies decide to follow them. For example: RTO -> what's the point of this shit? we never knew for sure but higher ups at most tech companies suddenly decided that RTO was the way to go forward despite all the downsides. Another example: DEI policies, some of them were very non-sensical.
I believe there has to be some downward pressure on these executives to take these decisions but I would like to know where it's coming from exactly and what's the logic behind them. Is it some big institution like Blackrock which has leverage on many of these companies? That's always been my bet but I never knew for sure.
Crappy managers don’t know (or actively avoid) how to measure business value from individuals. So they need you to be in the office so they can physically see if you are putting in the effort.
Tokens is just yet another proxy for business value.
The problem they face is if everybody is judge by business value in dollars, crappy managers are the first to go
yesterday's front page: AI is making me dumb. today's front page: employees are making AI dumb. the circle is complete.
This is what I do. I tell AI to go through every file in my project, identify up to 10 bugs per file, and then write the markdown with the name of the file plus "bugfix". This takes about 2 hours. Then I delete all the files with the suffix "bugfix" and then do it again.
You should probably create an agent to make agents whose jobs are to figure out how to maximize the token usage (and one whose job is to calculate the minimum token usage, so it doesn't look like a boondoggle).
I don’t even understand the point of making up tasks. Surely there’s some moonshot frustration project in your workday you could have an agent plugging away at, even if it’s unsuccessful.
When a measure becomes a target....
Especially a measure that's so easily manipulated.
I was just going to invoke Goodhart's Law
https://en.wikipedia.org/wiki/Goodhart's_law
New proposed corporate slogan: "Tokens must roll for victory!"
The original (third reich): "Wheels must roll for victory!"
It will end in the same manner.
There are some secret random seeds that will prevent the end token and just keep generating forever. This will ruin your hardware though.
This seems like AI is the new ponzi scheme.
If GDP is going up, we must be wealthier and more productive, right? Surely? (/s)