Hahahah good luck with that!
For many of us, what is happening now was super obvious. Telling a new formed crack addict (who you wanted to become addicted) to be more thoughtful about their consumption of crack... yeah not gonna work is it.
and governments will keep running massive data centers with classified frontier models for intelligence and propaganda purposes
No company with good engineering leadership should act like this is remotely a good idea.
For example, at more than one company I've worked for, if you wrote shitty code but got it into "testing" faster than anybody else, you are considered a superior programmer. And then, if you fixed the hundreds of bugs found in your code seen as an extraordinary programmer going above and beyond the call of duty.
Management is always measuring the wrong thing.
I assume the execs perspective is something like: if the top 20% of worker produce 80% of the code with LLMs and the company still works then we can get rid of the bottom 80% of devs and save money
But it's just one signal out of many, and more isn't somehow inherently better beyond a certain point.
but looking at the number of people who had taken leave, it suggests otherwise.
You get what you measure.
Everyone is oddly confident despite all of the conflicting explanations.
"I am currently averaging about 10k LOC per day (35% of the lines are tests) so wow, 15k/day is #goals"
Garry Tan - YC CEO and LLM psychosis victim (but he doesn't realise that yet)
There was a cost to it though. Codes greatly reduced redundancy, and caused large miscommunications from very small errors. As Glieck explains it, this was the opposite of the African drumming practice of adding redundancy to strengthen the relationship between the rhythm and the language that the drums mimic.
What you were describing would be token-minimizing, not maxxing.
AI quality outputs are fine for backoffice work now, but they are awful to read and reason about. Hallucinated features are also difficult to work with.
When will Uber (or your favourite company) be 'done'? They've been writing software for 16 years.
They match drivers to passengers. More software isn't going to increase the chance that I seek them out instead of taking a bus or train.
Will their software be finished in 20 years? 80?
Airports: different regulations, different rules for pickup/dropoff. Also scammers who pretend to be in a car, walk with their phones around pick-up ares in airport and do bait-and-switch (saw that in Istanbul SAW and in Dubai Al Maktoum)
Every different country Uber operates in is a moving legal and regulatory target
TL;DR: Managing a taxi service (that's what Uber is in my mind, not whatever "ride share" means) that spans cities and states, never mind countries, is extremely complicated. To their credit, Uber manages to make it look simple to the end user, prompting such comments as "meh it's just a few screens how hard could it be", which is triumph of product engineering as far as I am concerned.
Related: this blog from Uber talks about the problem of serving market-specific configuration data at scale: https://www.uber.com/us/en/blog/how-we-unified-configuration...
I took a ride from SEATAC to my hotel in downtown Seattle and besides the ride itself, there were 5 other items on the bill, 4 of which are specific to the place I used Uber.
Then I had the return trip from my hotel to SEATAC, on this one I got EIGHT items on the bill, on top of the ride fare. Some specific to Seattle itself, some specific to the road that the Uber took (a tunnel fee - which is different based on the direction you take it in), etc.
So the real question is what is NOT different between two locations. Less than 15% of the bill.
I also took Uber in India, where you have to share a one-time password with the driver for example, which I've never seen in any other country.
In some other countries the Uber app exists but Uber drivers are actually taxis, so you're actually ordering a taxi via the app.
Essentially every single airport in the world is custom UI and custom walking path guides and pickup instructions, and rules for where pickups/dropoffs/etc can occur can change multiple times in a day, much to everyone's enjoyment. They're almost all private property, and are so valuable that whatever they want is what they get.
And food. Most/~all? major brands get custom integrations.
Hundreds (iirc) of identity verification providers, most or all custom, and constantly weighed against cost and accuracy because it ain't cheap and it ain't good but it is far better than none (both legally and ethically).
No idea how many payment sources they accept, but it's definitely a lot more than anyone who hasn't lived on 5+ continents thinks.
And remember that this is all international. So scale is huge and law changes are constant and frequently conflicting. Darn near every useful feature is illegal somewhere, at some time, for both good and bad reasons.
---
This is not at all to say I think Uber is efficient, clearly it is not. Not by an enormous margin. But there is a legitimate need for truly absurd complexity, because the world is not consistent. You see similar things happen anywhere [thing] tightly interacts with humans.
I was prompted for this on the US west coast this week. If that feature was ever India-only, I don't think it is now.
They actually had 5,000 engineers in the tokenmaxxing blog post. That's a lot of engineers for the rest of Uber's business activities.
I suppose it becomes easier once the browsers, Android and iOS have been frozen for a little longer than 16 years. Nevermind the changing regulatory field and new products (when was Uber Eats launched?).
In that 16-year period, Covid-19 emerged, as did viable self-driving and partnership with Waymo. A networked, people-facing app can't ever be "done", unless you have perfect prescience. Internal tech-stacks are a living thing: keeping a service that on the outside appears to be unchanging is a lot of work! Scaling is a lot of work! Scaling services and maintenance feed off each other.
Each country has their own laws around what uber is and isn’t allowed to do. This needs to be formalized in code. For example you actually call a taxi, though the uber app, and the amount you pay is per mile, not a fixed fare decided ahead of time. To add to this complexity, some cities will have their own laws. What happens if you take an uber from town a to b, where each one has different laws ? A lawyer probably has an answer but the app needs to adhere to that. On top of that laws change all the time.
Optimization, well you can always optimize something. speed, costs, paths etc. In a way this never ends.
I think the part we interact with as consumers is a tiny sliver of the complexity those services have to build and operate.
I think this is partly a problem with companies that have had heavy investment. Uber’s value isn’t based on what they are doing, it is based on the idea that they are going to render ideas like owning your own car or taking public transit obsolete (I mean that’s an exaggeration but less of one than it ought to be).
This always happens when the metric becomes the goal, companies should nurture and foster an environment where AI is used in the most efficient way possible, first asking "do we really need an agent for this" and if so, what kind of agent is needed, what model, reasoning level, etc.
They should also promote projects that aim at saving tokens, increasing cache hits, codifying the information in ways such they use as less context as possible (graphs of knowledge are pretty good for this!)
It's like trying to win a race by setting a gas station on fire.
They are leeches that are just extracting wealth from their respective monopolies.
> They should also promote projects that aim at saving tokens, increasing cache hits, codifying the information in ways such they use as less context as possible (graphs of knowledge are pretty good for this!)
My understanding is that most big "tokenmaxxing" companies do have teams who are working on this in the background.
If you want incredibly fast adoption of AI within a company, the best thing you can do is to signal from the top that tokenmaxxing will be rewarded (or at least not be punished for it).
1. It forces everyone including the lazy ones who normally wouldn't invest their time in learning anything new to actually install codex/claude and learn to use them.
2. It prevents any middle manager from putting up blockers for adoption/experimentation ("this is new, I don't trust this, let's do it the old familiar way", "this might be expensive, we care about efficiency here", etc). Once the C-suite dictates tokenmaxxing is allowed, every middle manager will fall in line instantly.
3. Tokenmaxxing is not choice you have to live with the rest of your life. A year or two from now, once C-suite is satisfied with the rate of AI adoption within their org/company, they can just as easily switch the focus to efficiency. Teams will be asked to justify their token spend and start to optimize.
I would argue that you have an unreasonably optimistic view about corporate culture. There is a substantial amount of adverse selection and political maneuvering going on all the way up to the top. Tokenmaxxing just goes to show this.
That's part of the reason why this website is hosted under the YCombinator name, after all. Hackers are strongly meritocratic, which is not something you will find in a big company.
Good one!
I do not believe that engineers who are tokenmaxxing are truely productive and I have not seen any evidence whatsoever (perhaps the opposite).
I've personally found that with the right flow and codebase knowledge, that's achievable with sustainable levels of effort.
Why nobody talks about those points, which are actually the only interesting points of all this AI craze?
I can output 5 useless/bad features in a day with Claude or I can output 1 useful feature per 2 day period. Which one has better impact on ROI?
In this example, it might seem like it's an easy answer. But, in the real world, it is a lot more nuanced and much more difficult to measure and so not many are bothering to do it and are opting in for the simple solution of following the hype.
A more nuanced view would be something like:
* AI lets you achieve your roadmap somewhat faster, but:
* You incur tech debt that's similar to if you hired a dev temporarily for the features. You don't necessarily have someone on the team that understands the new code.
* Similarly, you aren't upskilling your junior team members. So you aren't getting skill/wage arbitrage as much as before.
* You will complicate the product. P2 features are P2 for a reason, but AI can cause them to be included and complicate the product for lower marginal gain.AI maximalists like to compare the technology to electricity. Imagine if in the early days of electrification, a CEO had rewarded staff for increasing the amount of electricity they consumed rather than finding ways to use it for business impact. Institutionalizing people who showed signs of mental illness was popular in those days, and I suspect that would have been the outcome.
Imagine if engineers were ranked based on their AWS spend. People allocate VMs and fill databases with terabytes of random bits, to get to the top of the AWS leaderboard. If you don't do this, you're ranked at the bottom, and good luck at the next review cycle. Who could have expected that this is not the road to success?
Anyone who can find the actually valuable portions of the space early has a potentially huge competitive advantage. Even if the result of the experiment is the negative that AI is actually mostly not that useful, that is still extremely useful information in a time of great uncertainty regarding outcomes.
The bottom line is that this approach may be expensive, but if you have the money to burn, it's far from the worst strategy if you are trying to position yourself correctly for the future.
If that was the intent, the messaging at many companies failed to communicate that. The message was "increase this metric", not "explore this space".
OTOH maybe we’re in for a future of patenting prompts.
Which absolutely isn’t the case. Even if someone would manage to overtake a market leader on tech merit alone, within 1-2 years, thanks to AI, markets don’t swing on such short notices. The fake urgency is absolutely psychotic.
The incentive structure of this type of decision is 'absolutely under no circumstances existentially mess up'. Ostensibly with respect to the organisation, but in actual reality much more so with respect to the individual(s) involved in the decision.
If everyone else is doing something that kind of obviously makes no sense, and you decide to break from the crowd by instead doing what does make sense, then there's a pretty solid chance of gaining a temporary edge while reality resolves the truth. But those gains probably won't matter all that much for the organisation, or indeed your position within it. It's a solid chance of an unimportant gain.
However on the other hand, there's a tail risk that something very unexpected happens and the thing everyone's doing that makes no sense actually turns out to make sense - sometimes even for entirely unpredictable incidental reasons - and then, well, you're in trouble. Not necessarily 'you' the organisation.. they'll likely be able to catch up and it won't matter that much. But for 'you' personally, the decision maker, it's very much not good.
As a bonus, in the much more likely scenario that the thing that makes no sense turns out to indeed make no sense, you're in the same boat as everyone else, there's no relative loss, and most importantly you don't stick out as someone who did something as risky as to go against the prevailing, albeit pretty clearly nonsensical, sentiment.
So basically, game theory tells you pretty quickly to just go with the thing that makes no sense if you're optimising for some (weighted) cross of what's best for the organisation and yourself as the decision maker.
Isn't it more likely that they simply don't in fact care about the "thing they care about", only the metric?
They can plot the metric on a chart and receive praise, so that's what they're interested in.
We aren’t there yet, so far it is just a COO questioning the investment
But it's not. Some FAANGs are doing amazing things with unlimited tokens. Other companies have no clue what to do with tokens, they've just told their engineers to max them.
It really depends on how you're using the tokens. If you're just using them for Codex and Claude Code - yeah, tokenmaxxing is incredibly dumb.
Unlimited tokens is different from “use AI a lot or we will fire you, and we are counting token consumption as usage”. Obviously the latter is stupid and yet it was done in many places.
Surely the anonymous employee feedback polls are totally anonymous too. BigCorp loves you, is family, and would never harm you.
Would love to know what things!
Unsubstantiated claims of 3x-10x productivity at allegedly a FAANG: https://news.ycombinator.com/item?id=48174666
More claims of the same, still unsubstantiated: https://news.ycombinator.com/item?id=48173995
Even more claims of the same, still nothing of substance: https://news.ycombinator.com/item?id=48173975
Outlandish claim of using 300$ in tokens in 1 hour at allegedly a FAANG: https://news.ycombinator.com/item?id=48158438
Ostensibly great things with unlimited tokens at allegedly a FAANG: https://news.ycombinator.com/item?id=48269185
Complaints about HN "not getting" AI: https://news.ycombinator.com/item?id=48245673
An attempt to validate vibe-coding in production: https://news.ycombinator.com/item?id=48243651
Whitewashing Claude degradation: https://news.ycombinator.com/item?id=48211638
More Claude degradation whitewashing: https://news.ycombinator.com/item?id=48245784
Whitewashing Anthropic: https://news.ycombinator.com/item?id=48199951
Promises of coming AI revolution: https://news.ycombinator.com/item?id=48188104
All of this serves to hype up the LLM technology with absolutely outlandish claims, while also propping up Anthropic online.
And now you’re asking them to somehow disprove they aren’t a shill? How would that even work. You seem unnecessarily antagonistic towards that user
I did not have to go through "so much of their history", this is just the last 9 days. There are considerably wilder claims from them earlier, when the account was solely focused on propping up LLM-hype and defending Anthropic.
We are living through a period of time with one of the potentially most disruptive technologies ever being developed. A lot (A LOT) of money is invested into it, a lot of livelihoods are and might be affected by it, and some people stand to gain A LOT from it. So there are significant interests to sway public opinion in favour of LLMs and AI, to hype it up to unreasonable extent, to muddle the waters of a reasonable discourse. Accounts such as solenoid0937 are unleashed on public forums to achieve that, and because of that we have to take everything they write with a huge grain of salt, or even ignore completely, as there is just no true information in their comments.
You yourself got baited by them by considering what they wrote seriously regarding "amazing things with unlimited tokens". Now the idea that "LLMs 1) are used in one of FAANGs massively and 2) are used to produce amazing things" is planted. Will you remember later that they did not actually provide any evidence of that? The account has been doing this trick multiple times over the last few weeks.
For me, as someone who is actually using LLMs in their work, a single balanced comment on the matter would have been more than enough to consider them not being a shill. Unfortunately, instead they have claimed recently that they went completely full-on with agentic coding (https://news.ycombinator.com/item?id=48245721) skipping reviews and pushing to prod directly (https://news.ycombinator.com/item?id=48243651) at a FAANG, no less. And they claimed it in such a manner that this is objectively the only proper way to do the work, and all other approaches are doomed. How is this anything else than peddling unfounded LLM-hype, I do not know.
The example above might seem like a singular episode, but they have been doing it over and over for the last year and a half, so this is now a pattern. No actual evidence for any of their claims is provided, so the only thing left is AI-hype, and pretty wild at that. So why would a reasonable person, with ostensibly enough money to retire (https://news.ycombinator.com/item?id=48252297), ostensibly working at a FAANG, spend all their days spreading unfounded AI-hype in a degrading manner on HN and defending Anthropic? Given the vested financial interest in the technology, the most plausible answer here is that they are paid to do so.
> Unfortunately, instead they have claimed recently that they went completely full-on with agentic coding skipping reviews and pushing to prod directly at a FAANG, no less.
This is like the most normal thing in my org, lmao. The fact that this aggravates you so much, is proof to me that many people/industries are way behind the curve on agents. Still bullish!
> as someone who is actually using LLMs in their work,
You should consider that your way of using LLMs is not the only correct way, and may in fact be severely limited, unimaginative, and/or close-minded.
Moreover, you refuse to even name the organization you allegedly work at. This is because people who actually work there will immediately call you out on your bullshit, so you have to stay intentionally vague there.
This is just yet another evidence of you being a shill.
And give away my org, for what? So some HN loser can doxx me?
Anyways, there's no point. If you learn to use agents, you'll understand eventually. If you don't, well, that's a you problem.
No, you are intentionally being obtuse, because people have already called you out on your lies. Saying "my team worked on <X>" is not under NDA in 90%+ cases, and no one is asking you for the implementation details for the <X>.
Name your company so others from it can confirm what you have claimed regarding AI and LLMs or deny it. Each FAANG being 10k+ (and in most cases 50-100k+) makes doxxing you unlikely, as others have done so on this forum without any problem.
Giving someone unlimited access to a resources is not the same as directing or incentivizing them to use it for the sake of using it which is what the parent comment criticized.
As for the other FAANGs, Meta and Google have (not good but still) frontier models of their own, so they are very different from a company paying API costs per token.
AI is an accelerator that engineers should know and have access to, but it's not something that should have mandated usage and quotas around. It's also absolutely dangerous for young engineers and the like - it fundamentally denies you of the "learning" aspect. I'm now seeing in interviews young graduates being given AI tasks to complete and they come back with a correct solution and no concept of how it is working.
You learn and reinforce learning by DOING and reading in depth. High level summaries don't teach anything and are the kinds of things only VPs care about. So, unless the intention in the future is for everyone to be a VP using AI to do the work, we need some middle ground here and some real thought around implementation of these tools or there's going to be a generational canyon gap of knowledge between being able to "say" and being able to "do".
When something is abundant, people tend to waste it.
I’m perfectly happy with my base subscriptions. I have Claude Code and Codex monthly subs, plus a yearly Google AI Pro account because it was a logical upgrade from the cloud storage plan I already had. I think it worked out to something like an extra $10/month for the AI features.
I constantly rotate between them during the week, managing tokens carefully, cleaning sessions and contexts as soon as possible, and being intentional about usage.
I honestly don’t understand the appeal of these ultra-expensive max subscriptions.
It reminds me of that flying orb toy I bought for the kids a few years ago. The battery only lasted about 10 minutes, and the kids would go ape shit crazy while it worked. Then it needed a 30-minute recharge, which created a natural cooldown period.
I actually considered that a good feature. I would never want the thing running nonstop.
Still very valuable. They just need to have strategies that match what the tools are capable of - not strategies that involve "rub the magic lamp and increase profits 80%".
If the market is rewarding companies going after the "rub the lamp" strategy, they're going to say they're doing that to juice stock prices.
Maybe the market is finally realizing blindly spending billions on LLMs with almost no strategy is not a good strategy.
Who knows.
You sure about that?
Both labs and tech companies have been desperate to show ROI on LLM use and nobody can seem to
He's saying that like it's some grand epiphany and not the most self-evident, obvious thing I've heard this month. Some of the literal dumbest people on earth are in charge of these major companies.
not only this month, but it is the basic statement of the single most well known 50 year old book in software project management lol. At this point we need to wipe the slate clean and start over, the industry is run by illiterates.
Pretty sure I know what you're saying, but the visual on this one doesn't match the point you're making.
I can see how Uber could burn unbelievable amounts of tokens if they start running internal features that run a bunch of prompts against every completed ride, or every customer profile, for example.
Or maybe this is about employee usage, but they introduced some stupid "you get evaluated on how many tokens you used" thing a couple of months ago when that was trendy and are just beginning to notice how much that cost?
The number of product teams who have shipped expensive-to-operate AI features is wayyyy up there, and for many of the scenarios I've seen, customers simply don't care or are unwilling to pay significant premium for access to it.
At the same time I'm starting to see some direction from people in leadership that I should "use the right model for the job" and things along those lines, which is a very, very different line from what I was hearing 12 months ago.
My continued prediction is that we are going to see a tweak on the SaaS model where the sweet spot moves to metered usage pricing of really fine-grained API-based access for apps which traditionally have been operated solely via the UI. Long term the trend is going to be "we'll house the data, enrich it, maintain it, provide fine-grained API access over it tailored to model usage, and you bring the model" with some services opting to give you the model interaction layer/harness. IOW I don't think SaaS is dead. Far from it. However, I do think that a lot of people are going to be looking to interact with SaaS apps via their own models with APIs that support those use cases better than a lot of those APIs do today.
isnt this just mcp servers hosted by the saas provider?
For me that's insanity for so many reasons...
Smart engineers are figuring out how to best use their tokens - as tokenmaxing is just as silly as gasmaxing your car.
Wrote about this and the impact of to jobs here: https://x.com/deepwhitman/status/2058324179506831372
It's like paying drivers per gallon of fuel consumed and then acting all surprised that you see them revving their engine while waiting at a red light.
I do think AI has value and is useful but the idea of token maxing is ridiculous.
That simply isn't true for technical employees (like software devs). They are so hungry to get stuff done that you have to hold them back from adopting new tools which they think can make them work more effectively. Tech guys will set up entire shadow IT departments just to get around corporate restrictions that are limiting their productivity.
No, if software devs are not using LLMs for programming, that is proof that the tool isn't actually useful for them. It doesn't mean "they need to be forced to use it", because they didn't need to be forced to use any of the tools which came before it.
Goodhart's law strikes again at someone with enough power to be both ignorant of it and make others suffer their ignorance. You cannot simply measure productivity by tokens spent just like you can't measure it by hours spent in a chair at a desk.
Which is why two identical jobs with the same real life output have drastically different productivity.
A nursing home in Luxembourg has 5 times the productivity of one in Romania despite the services being identical and tech-unrelated.
I know it's sounds stupid, but what if
True visionaries think outside the box, but most tech executives are forcing their employees into black boxes, out of fear of not doing exactly what their competitors are doing.
We have lemmings for leaders, and that means that—much like the LLMs that are being shoehorned into everything—there isn’t room for original thinking. Everyone’s strategy looks exactly the same.
First is that despite a lot of waste, some innovation will arise from an enterprising employee finding some interesting use case. A lot of the tokenmaxxing is just waste, but out of that waste may arise a small number of genuinely powerful use cases.
Second is that many workers will be entrenched in their ways. If your executive goal is to achieve the above (find innovative ways of using AI), then you need to move everyone to use it. Most will just waste tokens, but someone may find a novel and useful way of using it that benefits the organization. It is difficult to achieve these without forcing people to act since their default is to follow the well-worn grooves.
So mandates like these are a top-down forcing function like a slime mold feeling out different paths to find resources.
Some devs in my org have fully embraced AI; some would not even use AI if not for leadership mandates and linking usage to performance reviews (I know, I think this is stupid, too). I can see why mandates could be useful since some folks definitely won't be inclined to use AI.
Imagine you employ me as a hotel manager, and I come to you and say: "sure I spent all our food budget internationally in three months, and sure I have nothing really to show for it, but for those three months, we had a lot of food fights"
Your manager then goes on to explain they not only need more money to cover the food budget, but also they need to quituple the cleaning budget too.
Oh and the service level has dropped, because not all clients liked being in the middle of a food fight.
However "we might have some innovation in the food delivery system of our hotel chain"
> we might have some innovation in the food delivery system of our hotel chain
> Imagine you employ me as a hotel manager, and I come to you and say: "sure I spent all our food budget internationally in three months, and sure I have nothing really to show for it, but for those three months, we had a lot of food fights"
It's not how it should work, because food fights are stupid and have no upside.
Even if everyone else is having them.
It's not a fair analogy because AI isn't completely stupid, and there are situations where it does provide a benefit.
But a rational business would ask if the upside is worth the cost, if the pipeline can be restructured to concentrate and amplify the benefits, if some elements are better being done the old way, if there are strategic threats if tokens become much more expensive, and so on.
Instead we're getting a wave of "Cut workers, cut costs, derp" and that's as far as the "thinking" goes.
The worst thing about AI is that it shows how shallow and stupid current C-suites are.
The US used to have real tech visionaries. Now it has tech cargo cultists, all chasing an IPO cash out and hoping the music doesn't stop before they get their bag.
Thats the problem here. The idea is that we can build more stuff, quickly.
However in uber's case, they just burnt loads of money to push a metric that wasn't really related to productivity.
Absolutely, but most management are not leaders, the moment someone pushes the idea to stack rank based on token usage, it gets approved and some genuine people will be impacted.
Post-ZIRP era proved there are very few strong leaders, before that everyone was behaving like they're most amazing leader because they read some books and raised $10M
> I would prefer to search for new usages in a more strategic way
Looking back at the Internet, who would have thought that it would eventually create a Netflix, Amazon, Shopify, Spotify, Google Maps, etc. Just wild the things that ended up coming out of pushing bits over a wire with few simple protocols.
In an ideal world, you make strategic bets, but I can also see the case for the opposite this early in the lifecycle of a technology. You just don't know until you try.
Mid/late 2023, it wasn't at all obvious that it would take over coding that fast.
I definitely get the impression that many people thought it would eventually create shopping, streaming, and mapping sites.
I think people were less likely to have predicted things like social media or YouTube, though those weren't ideas sprung from a vacuum either.
None of these shifts were obviously the right bet and many organizations lost because they missed the opportunity. Now orgs are on the same horizon and I can see why they don't want to miss this window.
Sears was a different story, in that they were a real estate company with a store front and retail real estate took a nosedive due to ecommerce. But that's a different discussion.
So if the people who embrace AI areore successful then the others will follow. Just like every other new tech. Why does AI have to be forced? What's the hurry? Especially when there's no clear example of a company jumping ahead because of their use of it.
It's idiots being driven by FUD. That's the reason.
> What's the hurry?
There's also a need to quickly adopt and understand the technology; take the Internet for example. If we were talking about the Internet, forcing teams to build and publish web pages would be one valid way to get teams comfortable with the tech, the workflow, the shift in how to propagate and convey information to an audience.
Without a mandate, many teams won't adopt the Internet as a medium of information exchange because their processes work just fine and have worked for the last 20 years.
I think it's fair to put AI in a similar light. Unless teams adopt it and use it, it's hard for an org to understand how to get value out of this technology and how it affects existing processes and assumptions.
Those were always there, and will always be there. The type of time frames people are getting anxious about now rarely work in the real world, though, where potential customers don’t just switch products/service provides unless they’re facing catastrophic outcomes if they don’t.
And AI is not making the difference there that people think. I worked on a product that entered the market as a newcomer, wooed plenty of customers, and even though everyone _wanted_ it, only customers _urgently_ looking for a solution got on board quick (within <6 months).
Ironically enough, the product pivoting to Agentic AI hard killed a ton of momentum and interest from customers, despite exciting investors.
The web took off all by itself because it had a clear value proposition for some use cases.
> The web took off all by itself because it had a clear value proposition for some use cases.
Sears was the OG Amazon. Imagine if Sears had seen it as the new digital catalog.
Blockbuster missed on streaming until it was too late.
Many, many legacy companies did not understand the web and did not understand the impact of the Internet to their business model.
And even more importantly, the companies who went all in early and spent too much money on it too early without good reason went boom. You had to have actual business reason for it to be success.
> And you think forcing blockbuster's software teams to use the Web would have changed that?
A lot of monkeys will also eventually type up Shakespeare?
They have a whole management team and can’t seem to find a way to judge or god forbid encourage developers…
For example, everyone talks about strategy, but when you ask them what's our strategy answer is usually something like:
* let's figure out together
* industry changing is so fast, we should revisit plans every quarter
...
"Welcome to the new contractor who will be the artitect our new infrastructure. What is your dream IT setup?"
"Yeah, we can't afford that. Lets revisit once you wrangle those 2003 Dell R620's running Windows 2008 with no patching."
And that is why after eight months i'm terminating my contract on Friday and swimming back to shore.
If one is a CxO who's looking out for one's job security, herd-like behavior is the safest option, due to the (near universal) structure of "performance"-based executive remuneration.
I mean that's more of an ex-post statement.
Ex-ante they look at things as objects and visualise/simulate what one ought to do independently, as opposed to being a lemming.
https://en.wikipedia.org/wiki/Elinor_Ostrom#%22Design_princi...
What if the goal of an economic system was to support everyone instead of maximizing the upside for winners? Perhaps that's the sort of change necessary for improvement. Perhaps having billionaires is the failure state.
A goal fails - who sets a goal? The keyword is system.
An economic system needs something like a Nash equilibrium where defectors are sufficiently discouraged (and cooperators are rewarded as you imply). https://en.wikipedia.org/wiki/Nash_equilibrium
Lets talk my bonus, I will open the bidding at $1 per token.
I participate in some management-focused online communities. It’s crazy how many threads there are from frustrated managers trying to get their teams to stop thinking that their token use will be used as a proxy for their performance.
I think a few dumb companies did this and then it spread across social media, triggering a mass panic from engineers afraid their companies will be doing the same thing.
It’s getting so bad that the conversation is shifting to how to identify and coach the token-maxxers to stop wasting the team’s budget every week.
Because it is going to happen. Do you think metrics are tracked for fun?
Even if current leaders don't do it, next people might do it, how do you tell new leaders that we don't look at this metric? Metric exists to take action based on it
Nobody wants 90% of their token budget spend going to the 10% of people wasting them for number-go-up purposes.
Inefficient token use is going to become a metric.
I know - slightly off topic - but would you be willing to share this list?
I'd be curious to hear from people well versed in group psychology/dynamics and/or just a lot of leadership/people experience: what leads people to this type of thinking once they get in a group setting? It just... seems endemic at this point.
Obviously nobody here is going to know what I do or don't know, but I'm just increasingly curious what I am not understanding about this type of thing. It seems so obvious, yet that makes me ever more suspect that I'm oversimplifying it, or just totally ignorant about the problem in general.
Roll it all together and saying "just use it dammit" has some obvious advantages:
1. It's clear.
2. It's simple.
3. It eliminates all excuses employees might come up with for not using it.
The people at the top of these companies aren't stupid. They might have miscalculated how many tokens people can actually use, but that's very hard to calculate because usage is opaque and tools/processes change on a nearly weekly basis. They will eventually build out processes, tools, social conventions and performance metrics that take into account efficiency of token usage. But this is hard! Most managers aren't really assessed on the precise productivity of their teams, for instance, because productivity is often poorly defined.
Game theory! The downside of being brave vastly outweighs the upside. For the C-suite, there is no cost to herdlike-behavior, regardless of the outcome. However, there is a very high personal downside to being a maverick, and your board later discovers you made the wrong choice against the grain. The upside of being maverick and right is very limited.
Once a behavior has become mainstream, hopping on the bandwagon is no longer individually attributable to decision-makers, but is seen (and reported) as a macro-economic phenomenon: Nadella, Zuckerberg and Bezos didn't overhire - the American tech industry overhired.
Whats happening now and whos driving it is interesting. The CEO has a license for this new tool (think one of the top 4, Qwen Claude, Gemini, openAI) and really likes it. So much so that they (non coder) are making lots of little single page web apps.
The COO is bollocks deep in AI, and is saying that we cannot buy any SaaS products anymore. We must make it ourselves.
The engineering manager has seen this as an opportunity to build out a brand for engineering (its a small department in a medium sized company) by delivering quickly what the large year long efforts cant.
This has formed a slopnexus where PoCs are spun up left right and centre, but there isn't much time or thought going in to making them sustainable.
What started out as a (simple ish) asset management tool, neatly scoped into a deliverable PoC has morphed into a 5 product as one monster.
Its a mess that will either lead to burn out or disaster.
And myself being an infrastructure guy that needs to maintain all these PoCs that are now suddenly critical for production, it's the perfect nightmare.
And mind you, that dynamic always existed to a certain degree (laptop on a desk that runs some ugly Python script that does half the work of the BizOps team? Check. GCP account attached to the GSuite running a random instance for finance when the company is 101% on AWS? Check. Spreadsheet with macros that sends emails via Outlook as a mailing list manager? Check.) but at least when you discovered that you could scold them and tell them that we need to migrate this to a proper system because security. But nowadays with vibe-coded internal apps...it's a challenge.
Whats more annoying is that he changes AI provider lots, so we can never inject rules/skills to make sure he uses the right pathway from the start.
Large and fascinating topic I'm researching, very relevant for agentic AI and ML too. One way that groups can fail is that they just don't work to dampen / vote out individual errors properly (see PAC learning, Condorcet). Other kinds of errors only occur in groups, and can occur even when constituents individually aren't actually wrong. Some related stuff is:
https://en.wikipedia.org/wiki/Condorcet's_jury_theorem https://en.wikipedia.org/wiki/Group_polarization https://en.wikipedia.org/wiki/Availability_cascade https://en.wikipedia.org/wiki/Information_cascade
The last is probably the most relevant here and made worse by the negative effects of hierarchy. To quote one section:
> The negative effects of informational cascades sometimes become a legal concern and laws have been enacted to neutralize them. Ward Farnsworth, a law professor, analyzed the legal aspects of informational cascades and gave several examples in his book The Legal Analyst: in many military courts, the officers voting to decide a case vote in reverse rank order (the officer of the lowest rank votes first), and he suggested it may be done so the lower-ranked officers would not be tempted by the cascade to vote with the more senior officers, who are believed to have more accurate judgement;
For token-maxxing, our "senior officers" are just executives, and line workers aren't going to vote. Who is the senior officer for those senior officers? It's not shareholders! It's really the executives of even bigger companies, because that is the actually applicable promotion ladder. It's all kind of obvious, but also a genuinely better explanation than "monkey see monkey do". These are just the simpler things, and there's more gnarly dilemmas in https://en.wikipedia.org/wiki/Common_knowledge_(logic)
Thank you so much! This is why I love HN.
and tokenmaxxing is even worse due to https://en.wikipedia.org/wiki/Goodhart%27s_law because whatever you measure with tokens, once you start "tokenmaxxing" you have no measure to look at
"It is difficult to get a man to understand something, when his salary depends on his not understanding it." -Upton Sinclair
That VC funded gravy train is likely coming to an end. But fortunately there are also reasonably efficient models now so that the tokenmaxxers can still make the (much cheaper) tokens go brrrr.
Trying to operate as a rational, thinking person in a lot of environments right now feels impossible. Rational thought is being treated like AI skepticism.
Classic Goodhart’s Law: when a measure becomes a target, it ceases to be a good measure.
The way these corporations are going about it is completely insane though. They're essentially ordering their employees to set money on fire or be fired themselves. The more money you burn on tokens at insane API rates, the better an employee you are. Absolutely mind boggling.
I also want to call out the false productivity opportunities AI offers. There are whole teams building their own "gas town" and not shipping features.
Of course, the latest DeepSeek models are not as good as Claude, but they're not super far off either.
The risk of letting your agent read .env goes far beyond the risk that the agent itself does something you don’t like with the contents.
Gitlab is going to take off? This is not investment advice.
Even acknowledging we don't know exactly what costs would look like in a world without VC money, wouldn't hosting models logically be cheaper to do at scale in a data center?
When I compared to the cost of running DeepSeek locally, I meant that we can treat that cost as a price ceiling, not the floor.
No, I think local stuff using also-useful-for-other-things hardware will vastly undercut cloud hosting when the free money pipeline shuts down, and will stay that way for roughly forever. That doesn't mean cloud stuff isn't useful, clearly it is, but adding another company in the middle is rarely the solution for reducing costs.
It's especially a crazy assumption to make relative to the costs of employing a human. The costs of paying an entry level employee are unlikely to go down at all, and even if those costs do decline, there's a floor they can't drop below (minimum wage at the extreme end), whereas companies are free to optimize agentic costs as close to zero as possible.
So you are assuming that a cost which is extremely susceptible to optimization but which no one has yet seriously attempted to minimize will remain perpetually above a cost which is much less susceptible to optimization, is already subject to enormous efforts to minimize, and has a legally mandated floor. That seems like a bad bet.
I’ve spent $10-$20 a day using Claude to write code and closer to $5 a day now that I mostly use Deepseek and GLM, using API pricing (no subscriptions) since I don’t use Claude Code.
This is a rounding error for a company. So I think there’s plenty of room to use AI extensively while being more cost-conscious.
Agents are expensive in large part because tool calls require round trips. It's because these APIs are stateless and not streaming so you have to resend the whole context each time. This means you have roughly #tool calls x 1/2 context size cached input tokens over any given session. Most API providers overcharge you by a huge amount for cached tokens. A exception being Deepseek. Paying OpenAI $0.05 for 100k cached GPT5.5 tokens during a possibly 2 second round trip agent tool call is like paying $100/hr for what is likely to be ~10 to 20 GB of VRAM residence (holding the KV cache).
Or it got offloaded to NVME and you are paying $0.05 for that much PCIe bandwidth.
I'd imagine GPT-5.5 and Claude Opus 4.7 could run just fine on a 16x H200 node and serve at least 10 heavy users without the token output getting choppy.
The financials don’t make sense now. Based on the expenditure the finances won’t ever make sense.
I also don't think that blitz scaling will work like with Uber. The engineers are still there. We can work without the LLM tools.
The world will look drastically different 5 years from now; for the better or worse, so save every penny (especially if you work in tech).
Adds nothing insightful to these discussions.
The former is the issue, and how many companies have been operating. It's like a trucking company ranking driver effectiveness by fuel used instead of by cargo moved.
But on a more serious note, do we know how much Uber spent per technical employee/month? I assume it is far more than even any of those $200 "max ai" plans.
And the other question is how much the public would be willing to spend, in my estimation this is as "cheap" as it will ever get (main-stream at least).
Am in a random small company, colleague spent 100 EUR a day on Sonnet through AWS Bedrock (needed to use a EU region). Paying for tokens will get you in a deep hole financially compared to any of the subscriptions, unless it's like DeepSeek or one of the other models that are priced a bit better, though that's also a tradeoff in what they can/cannot do and also where the data goes. Ended up trying out the Mistral subscription for the US stuff btw, it was fine.
> Adoption climbed from 32 percent of engineers in February to 84 percent classified as agentic coding users by March. By spring, 95 percent of Uber engineers used artificial intelligence tools monthly, and roughly 70 percent of committed code originated from those tools. About 11 percent of live backend updates were written by agents with no human in the loop, according to Uber's own disclosures.
> The numbers behind the spend are what make the story instructive rather than anecdotal. Monthly cost per engineer ranged from $150 to $250 on average, with power users running between $500 and $2,000.
My guess is that the reason to rethink AI-spend was probably the exponential growth in cost over time, and tokenmaxxing payoff not being immediately obvious as mentioned in the article.
[1] https://www.forbes.com/sites/janakirammsv/2026/05/17/uber-bu...
I did so many interesting experiments with MapReduces that would run overnight.
For a while, I would even build internal services that were basically "free" because I'd just run them all at priority 0.
Over time those services got less and less reliable as overall usage started to increase, so I was forced to either justify the resources or scale back - but that was a good thing.
I feel like something similar would be a good model for AI token use: big tech companies ought to have their own self-hosted LLM data centers to power their own needs, then let employees use off-hours capacity to experiment.
Outside of experimentation, we should be encouraging token efficiency for everyday tasks. Rather than having a certain number of tokens, engineers should be evaluated based on how much they actually get done.
Using a lot of tokens to automate a process that used to require hours of human labor every week? Good use of tokens, should be encouraged.
Using a lot of tokens to debug an easy frontend bug that could have been fixed by hand, and still took you 4 hours to complete? Waste of tokens, should be discouraged.