The abrupt swing in many non-technology company IT departments from "hey developer, you aren't using enough tokens" to this is just too funny.
And I'm seeing almost no self-awareness from leaders. They are making decisions about things that they just don't understand. And are completely unworried about it. Just blindly following whatever the news cycle is about AI.
The closer people live to the consequences of their decisions the more rational they become. Until leaders(and I use that term loosely) are held accountable, the insanity will continue.
In addition to being true, this observation is profound. When designing any multi-step system that relies on humans making decisions, whether in governance, organizations or economies, placing root causes as close to end effects as possible is almost always better.
I’m sorry you are used to working with out of touch leadership. Not all companies are like that. Even big ones can have smart, empathetic leaders. Although very often money gets in the way of empathy.
I've been enjoying journalist Ed Zitron's recent diatribes about how impossible it is to find a business leader who had a plan for measuring their ROI from adopting AI coding.
What he says he's consistently hearing from them mirrors what I saw at my own employer: they thought they had ROI metrics, but they actually only had usage metrics such as "lines of code committed" or "number of pull requests". The only way those could possibly work as an ROI measure is if your business charges customers by the line of code.
During ZIRP they discovered that the way to lead companies nowadays is to become a maxxer of whatever current fad is, and the more you maxx the better. And then when things change and you're wrong, you'll be a strong leader and, in ZIRPs case fire everyone you over-hired, with AI will be similar.
Why be a normal guy that waits to see what happens and is measured and pragmatic when you can get attention basically through the whole cycle by being the earliest adopter, adopt it to the maxx, then also be the loudest big brain when the tide changes and be praised for "taking hard decisions" when you revert everything you said so far?
A special case of the more general cringe economy we're in. The dumbest, most outrageous ideas win, amplified by social media. Say stupid sh*t loudly, be wrong, profit.
Perhaps, but the change you get (if any) is most likely to be what you push for and reward/punish.
It's irrational to push for tokenmaxxing (literally "please increase our AI spending") and not expect that this is the result you are going to get. You won't get productivity increase, since that is not what you are pushing for - you will get token usage maximization (engineers running inane agentic tasks against your code base to increase usage, using company paid AI for their side projects, etc, etc).
I feel like most successful businesses have such a moat of required capital to compete with them that even tho in theory poor decisions like this is supposed to give opportunities for entreprenuers to hit when the big dogs make a wrong move, it doesn't end up happening.
Don’t play their game and call them leaders. They are management, bosses, executives.
> They are making decisions about things that they just don't understand. And are completely unworried about it.
Clowns, even.
> Just blindly following whatever the news cycle is about AI.
But followers might be most apt.
——
This is such a huge pet peeve of mine. Describing management goofs using their language that makes them sound all-so-brilliant. We constantly watch these people do the dumbest shit and then they go around describing themselves as “thought leaders” and “servant leaders”. When, really, most are just clowns with fragile egos.
And, while I’m rambling, they’ve tried to take away the fact we are workers by calling us individual contributors. Using language to attempt and hide the hierarchy and power dynamic at play. It just…bothers me so much.
> Don’t play their game and call them leaders. They are management, bosses, executives.
You're falling into a common trap here: the ambiguity of the English language.
"Leader" means multiple different things. Yes, it means someone who has leadership qualities—who genuinely inspires those around them to do better, or who boldly marches into the unknown and gets people to follow them.
But it also means "someone in charge of a thing."
Now it's certainly true that many people in charge of things who are also really bad at actually inspiring or getting people to follow them (aside from with threats of destitution) also play on that ambiguity to try to convince people that because they're in charge of things, they must also be Good Leaders, and that's crappy...but yelling at others for using the term casually is very much an "old man yells at cloud" situation.
The worst part is that techies can still work around the insanity if they keep their opinions private. For the serious average Joe the AI mandates must be feeling like hell on earth.
I once worked in a company that had soviet-level efforts to push LLMs into everything, someone eventually made the classic "Natural Language -> SQL Query -> Magic Result in webpage" and got promoted, the tool got mandated for every non-tech employee as part of an AI-boosting effort (people pushing metrics up).
One day I wake up with a product person in despair because the tool couldn't handle what looked like a very simple aggregation, I stopped what I was doing, crafted a 30-line SQL query over HORRIBLE TABLES, a couple CTEs and window functions here and there got him what he wanted. I found out later that single query that took 30 minutes to make saved him from inheriting a 6-month effort to create a microservice dedicated to patching said tool.
Having studied control theory I think it makes perfect sense. When trying to make a system target a new level it's quite natural for there to be overshoot that needs to be reigned in. It's also natural for the correction to go too far and need to be corrected in turn. This is not indicative of stupidity it's completely normal.
It would only be laughable if they waited way too long to reverse course, but I don't think that's the case.
Suppose I'm driving at 20 kph, and I set my cruise control to 40 kph. My car then goes WOT, overshoots my target speed and hits 120 kph, at which point it slams on the brakes[0], dropping my speed to 15 kph. It repeats until it finally settles at my target speed. (Rhetorical question) would that be considered "completely normal"?
Over/undershoots and corrections are of course unavoidable and normal; the absurdity is at the magnitude and rate of change. Furthermore, this is giving it the benefit of the doubt, that measuring AI spend is a good indicator; that's arguably also in dispute. To stretch my car analogy a bit more: it would be like the cruse control system has to hit the target speed, but it only has data from the O2 sensors.
[0] I know that the "classic" cruise control system cannot apply the brakes, but hey no analogy's perfect.
It's not like they accidentally overshot, they were telling people to tokenmax, they didn't even know you could overshoot they thought it was exponential gains all the way. Subtle ideas like balance were not on their minds.
How much that makes it into enterprise pricing is TBD, since none of the hyper scalers are making money yet of selling AI inference.
Almost all businesses are ahead of the gun. For most of their use cases, AI is either not yet good enough on its own, or good enough but too expensive.
No one wants to get left behind, so everyone's trying to get onto it now, even though it's not ready for what most enterprises want to do with it.
It's easy for them to look at a small startup without billions of lines of legacy business logic debt and see them having success and wonder why they can't have just as much - or more - why they're bigger so they should have better and more success, right???
Wrong...
But when it gets ~99% cheaper for local inference over the next 4 years, at the same time the price per watt improve 4x -> a lot of those cases will start to pencil out.
Going from Opus 4.5 to 4.7 secretly required 6x more compute to run. 4.8 is apparently 30% more on top. I haven't seen any optimizations lately aside from distillation. Nobody's optimizing, they're just scaling up.
Do you mean the marginal cost by the producer, or the cost on the consumer? I can't see the price of electricity falling much, and the demand curve is apparently exponential if the hype is to be believed.
I don't see how this is even remotely true. Unless there's some super breakthrough into a fundamentally different architecture, there's not really a path to a 50% reduction in price, much less a 99% reduction.
Prices have been very obviously trending up, not down. Even open weights models are becoming more expensive with every release. Computer hardware is ballooning in price.
> The actual cost is going to drop 99% in ~4 years.
We have little visibility into current frontier model costs at mass scale. As a broad historical trend, tech costs tend to fall over longer time periods but your claim far exceeds Moore's Law rates in its heyday - and that heyday is long gone.
In 2021 TSMC announced it was increasing it's price per gate for new nodes for the first time in its history. In the past five years cutting edge nodes have delivered ~8-15% real-world performance gains on average at costs at least 10-20% more than the last node. If you're positing a string of unprecedented efficiency breakthroughs in LLM algorithms - such extraordinary claims require extraordinary evidence.
AI is overhyped. I have yet to see an end user product that in itself isnt a wrapper around LLMs that is impressive created by LLM assistance. I have also yet to see dramatic increases of revenue of companies using LLMs that don't involve selling things in its supply chain. Is it a nice affordance? Sure. 1T capex good? No.
If it was so good I would expect to see 2005-2015 advancements yearly.
Meanwhile China is blowing past the world with real improvements in the real world- solar, EVs, etc. meanwhile people keep making their fancy sans serif websites about todo apps, faster than ever before. Useless.
> I have yet to see an end user product that in itself isnt a wrapper around LLMs that is impressive created by LLM assistance.
I don’t disagree that AI is overhyped. But I think you are probably looking in the wrong place.
I think most software that is written isn’t really a product, at least not a public product. It’s an in-house tool or a one-off project needed to complete some larger task. People everywhere are always writing small programs that make their life or job just a bit easier (and explains why so many corporate projects are little more than an excel spreadsheet).
And there are a lot of people who have made custom software just for themselves with AI. Not a product, just a tool or project that finally made sense to build.
> Meanwhile China is blowing past the world with real improvements in the real world- solar, EVs, etc. meanwhile people keep making their fancy sans serif websites about todo apps, faster than ever before. Useless.
Very little about the American economy even makes sense for keeping the edge on LLMs beyond a few years. All the things I would think would be required: energy, research, construction capacity, labor costs -- it's pretty hard to deny who's on the upswing these days. China cranking out current generation microchips will be the last nail in the coffin.
In the time before LLMs, humans made satellites, Concorde, life-saving medical surgery, James-Webb Space Telescope, communication at the speed of light, the list goes on.
What changes have LLMs (Not AI, not machine learning in general, I'm not going to waste time discussing the definition), LLMs made in the past 4 years that indicate anything close to the above? Solving a whiteboard math problem?
In my opinion, the problem is not even the cost. The problem is that people are using AI for running recurrent stuff instead of writing code to automate it.
For example. Imagine that you are comparing two documents (let's assume diff doesn't exist). You could ask an AI to compare the differences from you or you could use AI to write a tool to do it. For whatever reason, people are starting to go with the former not realizing that now they basically have to pay to compare documents.
I have exposure to AI initiatives at several companies including a few F500's. I have seen teams dump huge logs into frontier models that took hours to get so-so results that we were able to replace with a few lines of python code at 1000 times the speed and 100% accuracy. When asked why they were doing this they literally said "because we don't understand the subject matter so we were depending on the AI". I saw one team file a complaint with a vendor about a frontier backed coding harness and it's inability to consistently format headers because they were using it as a reporting engine. When I recommended they just use the coding tool to write code to generate reports you would have thought I had just cured cancer from their response. I frequently see people complain about the fact that AI is going to take their jobs and then see them gripe about the fact that AI is 'worthless' because it can't do more of their job than it already does. It's easy to see the difference between the people seeing 10x productivity gains from leveraging AI and those who aren't and it's not the AI.
i have trouble understanding these situations, e.g. the AI itself would presumably make the suggestion to write a python script for such a task. It seems to me that there two huge problems right now
* understanding which category of problems an LLM is an appropriate solution for (rather than throwing LLMs at any and all problems)
* matching model capability (and therefore cost) to the problem at hand. You can easily overspend massively by using a model that's too powerful
Someone asked me if I was using models for fantasy sports, and if it was smart enough to help make decisions about drafting.
My answer: no, but it was able to help me find the website and social handles for every beat writer for every team, and generate a simple website where I can do a daily skim of teams/players and draw my own conclusions.
Laziness, pure and simple. The inevitable consequence of “the LLm is the compiler now”. And what do you even expect people to do when they are forced at threat of termination to use AI for everything as much as possible? Not to mention people are being pressured to do insane thing like review hundreds of pull requests per day and deliver like 15 features per week so OBVIOUSLY there isn’t time to build out proper tooling. Just shove everything in a prompt and call it a day. Some people have families to feed, just do what you’re told.
Agreed. I’ve been telling my team to build up internal packages so we can push all that ad hoc reinvention into something more tangible and deterministic. Invest the $$$ in inference into something the agent can reach for next time that’s neutral and consumable by other code to reduce future spend.
Because you look at the work from the perspective of a programmer, not the perspective of a regular person.
Normal people have never gone around automating their work. The most automation they do is dynamic tables on excel sheets.
I obviously know building a tool that can programmatically do something is a better solution, but I think that requires a fundamental shift in how people work. People need to be told by someone "this is how you should be using the AI" but right now they're simple told "use the AI".
I'm talking about programmers doing this. That's what's sad. These were normal people before,but it feels like they have some kind of AI schizophrenia now. They don't use their brains anymore
Same, even opus favor short term solution and scripts with a billion flags that constabtly require rescanning to understand how to launch it is a constant struggle to get it to build sane default and reusable scripts that run with minimal parameters
Yeah, and what's up with adding dry run to everything? I saw some code that doesn't write anything but still the AI added a dry run which had a completely different codebase
I'm curious if you could give me an example of something that couldn't be down deterministically. We have fuzzy search/matching too ? Regex is a monster when used correctly.
It's this and worse. To use your example, it's like people using AI to write a diff algorithm, incorrectly, then using AI to fix it, because they don't know that diff exists already. Lazyness and starting development with a very low level of understanding. People think lowering the barrier to entry is a good thing, when in reality there are just fundamentals and things you just have to know before you can start using a tool like llms properly.
Isn't that the supposed point of it though? At least how it is marketed/hyped. Don't use your brain, you don't need one, spend all your thinking energy on... dunno, something else, and leave all the "mundane" stuff to AI. Just pay for the tokens, it's going to make you 10x more efficient, the $1000/month is worth it.
100% this. For my own company I mostly build deterministic workflows that may have a simple AI step in the middle using an appropriate Chinese model in a very limited way. I wouldn't want to burn tokens to satisfy some metric.
With this AI is a fallback and not the default. Sounds like large companies have it backwards.
I agree, but even this use case isn't the most wasteful. The interwebs says Agentic consumes 50% of token use, but I'd hazard this number is north of 90% for many shops. My cynical view of Agentic is its sole purpose is to make "number go up".
The cost is a problem, but IMHO more important is delegating so much of your internal knowledge, thinking, and systems to a 3rd party.
We are very close to the point where if Claude and ChatGPT APIs are down, companies cannot function. How is that introduced so quickly into so many critical places without taking that specific fact in consideration? What is the plan for all those companies whose workflows now depend heavily on a remote LLM whenever the services get cut? What if your company account gets banned?
In some ways it is worth than depending on a company for hosting, because even your debugging tools are based on AI. MCP is great to go through datadog, sentry, until your agent or the MCP server are down and you don't know how to look for the issue yourself because you do not actually understand how your systems work.
> We are very close to the point where if Claude and ChatGPT APIs are down, companies cannot function.
Contrast with Gmail/Gsuite/Outlook365/QuickbooksOnline/etc are down, though.
What you cite here isn't a direct attack on AI but on centralized service provision in general. Unfortunately that battle has been lost for decades, now.
None of those are doing the actual development. Here we are talking about a technology people delegate judgement, technical expertise to. It’s way, way deeper of an integration than a standard saas
Those sound like problems for another quarter. The people making the decisions ride the AI hype wave, and if in the worst case the company tanks one day, they take their severance package and leave.
There's an old saying, "in the land of the blind, the one-eyed man is king."
Here we have the opposite: In the land of the one-eyed, the blind are leading.
The blind in this case are all those executives and managers who don't understand much about AI's current potential and limitations, and so far have treated it like a magic button that will solve everything. The one-eyed are rank-and-file employees who maybe sort of know a little more about AI.
Executives and managers are the ones who correctly understood which game was being played. The game we are playing is not one of making good products, it's one of getting money from people who both have more money and are stupider than you. They're succeeding at that. We're also doing it, but we're not getting as much money.
In many cases, the people who have more money and are stupider than you are other executives. Sam Altman is arguably one of the executives who know how the game is played. OpenAI is at the front. Microsoft's executives are an example of the ones who got played.
Would have been nice to see 'soaring costs' with numbers. WSJ could do better here. Hundreds of thousands of dollars a month is nothing compared to how much they take with better financial models.
My org has a monthly team plan, and because I don't use what I would call unsupervised agents, I rarely come close to exceeding limits. I guess I am one of those "chat only users" that so many articles limit my output. A split terminal with Vim on the left snd Claude on the right has been a great combination. Neanderthalic AI.
Personally I'd consider efficient and economical use of AI to be a key skill for a good developer. Using AI for everything at full throttle would appear to be a crutch. I wonder if this will eventually become a hiring criteria for companies without unlimited budgets (most of them).
On the one hand, organizations are without question using LLM's well beyond what is actually necessary, and as reality kicks in they're forced to scale back accordingly. However at the same time, on intervals counted in months, we're seeing breakthroughs both in hardware and software that dramatically reduce the cost of inference.
Between corporate FOMO and the rapidly decreasing costs of actually running LLM's I'm interested to see at which side of the spectrum these two meet
Is Anthropic rushing to an IPO because investors also fear it’s the peak? I guess it can grow even after the IPO but is there a fear that the spending will rationalize and hurt the valuations?
Like a few other commenters, other than reasonably making a software developers work easier I haven’t seen a lot of value creation or revenue boost from AI. I know AI companies make money, but the companies consuming these services, are they meaningfully making a revenue to justify the continued spend?
I’ve seen comments on other threads on this subject the general idea that these article headlines are overstating the pullback from AI.
In other words, the news cycle is looking for an AI story that lands with readers, and that the example
of Uber blowing through its AI budget and Microsoft discontinuing use of Claude internally are not good indicators.
I agree that those aren’t good indicators.
However, at some point we have to remember that CEOs and boards of directors are just regular morons who read the news the same way everyone else does.
At some point, if a lot of corporate leaders associate AI with mediocre results, high costs, and public backlash, they might just start saying “this juice isn’t worth the squeeze.”
This phase that all these companies went through doesn’t seem that bad. Before these places had a big problem where all their employees didn’t understand how to us ai for their work. Now they’ve overspent and tokenmaxxed and haven’t seen much from it. The next phase is to set the goalpost lower and set quotas based on who uses ai more effectively. Eventually the folks that use it well and are productive will bring in roi. Then you can fire all the folks that aren’t using it effectively and replace them with people that know how to use it. We’re already starting to see this.
They are likely also starting to realize that the end result of their anthropic contract is that nobody but anthropic knows how to run their business. Why would anthropic not treat their business like a utility in the future?
Only thing I can say AI was useful for, in a corporate environment, was learning a new coding language on the fly. Gives me a baseline to work off of and fix.
But I can learn without it, too. A nice tool, but not a need.
An ironic analogy sort of, once media started hiding behind a paywall, I just stopped reading them rather than paying. Same with LLMs - usable if cheap/free.
90%+ of corporate people are not programmers. 1 programmers can cause the same token damage with a bunch of concurrent agents as a couple thousand Karens in compliance asking a chatbot questions
It's much easier to deliver incremental AI ROI on the later even if it's hard to measure/quantify. A 1000 tokens might point this compliance person in the right direction on a key problem. Meanwhile 1000 tokens doesn't get you anything useful on coding
The other day we (wrongly) concluded that product market fit has been achieved and now the rivers of hot molten milk chocolate and honey are all that's in the future etc.
Another reason to favor using AI to build automation instead of relying on it in prod: the risk of war and global instability.
If LLMs are genuinely helpful or even decisive in a military engagement, you can expect any host country to commandeer whatever data centers they need, leaving commercial entities to bid up the prices on the leftover capacity.
Another risk is that data centers are a great target for cyber warfare.
It’s ideal if your business can leverage LLMs when they’re online but continue to operate profitably when they’re offline.
Even regular warfare, if the Middle East AWS regions are an indication. The giant and arguably excessive data centers being built are not hardened physically.
I've noticed as well. A lot of pull requests are just agents running constantly, hoping to have produced something of value. Entropy is at an all-time high, though.
Exactly. The bottleneck becomes human review, not code generation. Agents can generate commits faster than humans can verify whether those commits should exist.
There's a paywall, but it's an interesting question how much of the recent explosion of the AI companies revenues is because of the explosion in prices, and how much their customers will accept the increased prices.
Yesterday I updated our dependency on the sqlx crate and put up a PR, and it failed in the CI build in a way I couldn’t reproduce locally.
I asked codex to take a look, and it:
- Grabbed the CI logs on its own to figure out what the CI error was
- Looked at my local setup
- Looked at the changes in sqlx from 0.8 to 0.9
And figured out that sqlx depends on an updated version of the “whoami” crate but doesn’t specify default features, which causes it to fall back on a stub implementation that makes the default user “anonymous”, which was failing to authenticate to the UNIX socket we use in our CI Postgres server. It patched the environment variable for our docker container to explicitly specify a username and the issue was fixed.
It would’ve taken me probably several hours to figure this out on my own. It took codex maybe 5 minutes.
I agree with your point in the broad sense, but the example might be bad. If sqlx is an important crate, and not stable yet, upgrading it without reading the changelog is honestly a flaw in your team process. Using the AI to fix organisational issues is typically one of the reasons I'm very skeptical of AI improving productivity in the long run.
I'm not taking a shot, to be clear, we had a similar issue a few years ago and we made sure this wouldn't happen again, that's absolutely not a shot, nor do I think it's a character flaw to use AI, au contraire, this is a very good use. I'm just worried that because AI is so good at fixing minor issues caused by governance/organisation flaws, we will be stuck using it to fix those and be trapped in mediocrity (that's not an issue for me, mediocrity is where I work best, but I'm a bit sad for the great Devs I've worked with.)
You used it in a way where the result was simple and you could verify its correctness. You used it as a super-search tool, it's good at that. It's a different use case than having it generate a lot of code from scratch.
It sounds like the real problem was that the programmer was not familiar with the tools they were using and decided to dig themselves out of a hole of their own making by turning to AI instead of learning to use their tools better.
They cannot be trusted to produce output that works (let alone works well) because they are just statistical models, without any actual understanding of what they produce. That means that you have to carefully review every single line of code they produce, because you don't know where the hallucinations will be. But by the time you do that, you have saved no time at all (indeed, in my experience you lose time), because typing the code was never the part that took time. It was understanding the problem. So if you use an LLM, you spend a bunch of money for zero gain in productivity, or you sacrifice quality and pray there aren't nasty bugs lurking. I certainly think it's fair to call that state of affairs "it doesn't work".
The abrupt swing in many non-technology company IT departments from "hey developer, you aren't using enough tokens" to this is just too funny.
And I'm seeing almost no self-awareness from leaders. They are making decisions about things that they just don't understand. And are completely unworried about it. Just blindly following whatever the news cycle is about AI.
The closer people live to the consequences of their decisions the more rational they become. Until leaders(and I use that term loosely) are held accountable, the insanity will continue.
Their only accountability is to the stock price. The insanity will continue.
4 replies →
In addition to being true, this observation is profound. When designing any multi-step system that relies on humans making decisions, whether in governance, organizations or economies, placing root causes as close to end effects as possible is almost always better.
I’m sorry you are used to working with out of touch leadership. Not all companies are like that. Even big ones can have smart, empathetic leaders. Although very often money gets in the way of empathy.
4 replies →
I've been enjoying journalist Ed Zitron's recent diatribes about how impossible it is to find a business leader who had a plan for measuring their ROI from adopting AI coding.
What he says he's consistently hearing from them mirrors what I saw at my own employer: they thought they had ROI metrics, but they actually only had usage metrics such as "lines of code committed" or "number of pull requests". The only way those could possibly work as an ROI measure is if your business charges customers by the line of code.
What they really means is they previously had no valid metric to measure productivity of developers before either. AI or not.
3 replies →
During ZIRP they discovered that the way to lead companies nowadays is to become a maxxer of whatever current fad is, and the more you maxx the better. And then when things change and you're wrong, you'll be a strong leader and, in ZIRPs case fire everyone you over-hired, with AI will be similar.
Why be a normal guy that waits to see what happens and is measured and pragmatic when you can get attention basically through the whole cycle by being the earliest adopter, adopt it to the maxx, then also be the loudest big brain when the tide changes and be praised for "taking hard decisions" when you revert everything you said so far?
The fakemaxxing economy.
A special case of the more general cringe economy we're in. The dumbest, most outrageous ideas win, amplified by social media. Say stupid sh*t loudly, be wrong, profit.
Groups resist to change - the bigger the group, the most resistance there is.
As a leader, pushing for rapid change cannot really be nuanced lest the push dissipates into the organization's entropy.
Perhaps, but the change you get (if any) is most likely to be what you push for and reward/punish.
It's irrational to push for tokenmaxxing (literally "please increase our AI spending") and not expect that this is the result you are going to get. You won't get productivity increase, since that is not what you are pushing for - you will get token usage maximization (engineers running inane agentic tasks against your code base to increase usage, using company paid AI for their side projects, etc, etc).
8 replies →
I feel like most successful businesses have such a moat of required capital to compete with them that even tho in theory poor decisions like this is supposed to give opportunities for entreprenuers to hit when the big dogs make a wrong move, it doesn't end up happening.
> leaders
Don’t play their game and call them leaders. They are management, bosses, executives.
> They are making decisions about things that they just don't understand. And are completely unworried about it.
Clowns, even.
> Just blindly following whatever the news cycle is about AI.
But followers might be most apt.
——
This is such a huge pet peeve of mine. Describing management goofs using their language that makes them sound all-so-brilliant. We constantly watch these people do the dumbest shit and then they go around describing themselves as “thought leaders” and “servant leaders”. When, really, most are just clowns with fragile egos.
And, while I’m rambling, they’ve tried to take away the fact we are workers by calling us individual contributors. Using language to attempt and hide the hierarchy and power dynamic at play. It just…bothers me so much.
I don't hear them refer to themselves as "job creators" much these days.
And many of them still claim they are "risk takers", but have effectively insulated themselves from risk by socializing losses.
> Don’t play their game and call them leaders. They are management, bosses, executives.
You're falling into a common trap here: the ambiguity of the English language.
"Leader" means multiple different things. Yes, it means someone who has leadership qualities—who genuinely inspires those around them to do better, or who boldly marches into the unknown and gets people to follow them.
But it also means "someone in charge of a thing."
Now it's certainly true that many people in charge of things who are also really bad at actually inspiring or getting people to follow them (aside from with threats of destitution) also play on that ambiguity to try to convince people that because they're in charge of things, they must also be Good Leaders, and that's crappy...but yelling at others for using the term casually is very much an "old man yells at cloud" situation.
The worst part is that techies can still work around the insanity if they keep their opinions private. For the serious average Joe the AI mandates must be feeling like hell on earth.
I once worked in a company that had soviet-level efforts to push LLMs into everything, someone eventually made the classic "Natural Language -> SQL Query -> Magic Result in webpage" and got promoted, the tool got mandated for every non-tech employee as part of an AI-boosting effort (people pushing metrics up).
One day I wake up with a product person in despair because the tool couldn't handle what looked like a very simple aggregation, I stopped what I was doing, crafted a 30-line SQL query over HORRIBLE TABLES, a couple CTEs and window functions here and there got him what he wanted. I found out later that single query that took 30 minutes to make saved him from inheriting a 6-month effort to create a microservice dedicated to patching said tool.
I've never seen self-awareness from leaders. They always lead on vibes.
Understanding this was one of the most important things in my career.
That's nothing new though. It's just very obvious this time.
Having studied control theory I think it makes perfect sense. When trying to make a system target a new level it's quite natural for there to be overshoot that needs to be reigned in. It's also natural for the correction to go too far and need to be corrected in turn. This is not indicative of stupidity it's completely normal.
It would only be laughable if they waited way too long to reverse course, but I don't think that's the case.
Suppose I'm driving at 20 kph, and I set my cruise control to 40 kph. My car then goes WOT, overshoots my target speed and hits 120 kph, at which point it slams on the brakes[0], dropping my speed to 15 kph. It repeats until it finally settles at my target speed. (Rhetorical question) would that be considered "completely normal"?
Over/undershoots and corrections are of course unavoidable and normal; the absurdity is at the magnitude and rate of change. Furthermore, this is giving it the benefit of the doubt, that measuring AI spend is a good indicator; that's arguably also in dispute. To stretch my car analogy a bit more: it would be like the cruse control system has to hit the target speed, but it only has data from the O2 sensors.
[0] I know that the "classic" cruise control system cannot apply the brakes, but hey no analogy's perfect.
It's not like they accidentally overshot, they were telling people to tokenmax, they didn't even know you could overshoot they thought it was exponential gains all the way. Subtle ideas like balance were not on their minds.
1 reply →
The actual cost is going to drop 99% in ~4 years.
How much that makes it into enterprise pricing is TBD, since none of the hyper scalers are making money yet of selling AI inference.
Almost all businesses are ahead of the gun. For most of their use cases, AI is either not yet good enough on its own, or good enough but too expensive.
No one wants to get left behind, so everyone's trying to get onto it now, even though it's not ready for what most enterprises want to do with it.
It's easy for them to look at a small startup without billions of lines of legacy business logic debt and see them having success and wonder why they can't have just as much - or more - why they're bigger so they should have better and more success, right???
Wrong...
But when it gets ~99% cheaper for local inference over the next 4 years, at the same time the price per watt improve 4x -> a lot of those cases will start to pencil out.
Going from Opus 4.5 to 4.7 secretly required 6x more compute to run. 4.8 is apparently 30% more on top. I haven't seen any optimizations lately aside from distillation. Nobody's optimizing, they're just scaling up.
10 replies →
> The actual cost is going to drop 99%
Do you mean the marginal cost by the producer, or the cost on the consumer? I can't see the price of electricity falling much, and the demand curve is apparently exponential if the hype is to be believed.
4 replies →
I don't see how this is even remotely true. Unless there's some super breakthrough into a fundamentally different architecture, there's not really a path to a 50% reduction in price, much less a 99% reduction.
2 replies →
What makes you think prices will drop? Everyone I’ve spoken to believes they will only skyrocket. Genuinely curious
6 replies →
Prices have been very obviously trending up, not down. Even open weights models are becoming more expensive with every release. Computer hardware is ballooning in price.
10 replies →
> The actual cost is going to drop 99% in ~4 years.
And fusion power is just 2 decades into the future!
1 reply →
> The actual cost is going to drop 99% in ~4 years.
We have little visibility into current frontier model costs at mass scale. As a broad historical trend, tech costs tend to fall over longer time periods but your claim far exceeds Moore's Law rates in its heyday - and that heyday is long gone.
In 2021 TSMC announced it was increasing it's price per gate for new nodes for the first time in its history. In the past five years cutting edge nodes have delivered ~8-15% real-world performance gains on average at costs at least 10-20% more than the last node. If you're positing a string of unprecedented efficiency breakthroughs in LLM algorithms - such extraordinary claims require extraordinary evidence.
AI is overhyped. I have yet to see an end user product that in itself isnt a wrapper around LLMs that is impressive created by LLM assistance. I have also yet to see dramatic increases of revenue of companies using LLMs that don't involve selling things in its supply chain. Is it a nice affordance? Sure. 1T capex good? No.
If it was so good I would expect to see 2005-2015 advancements yearly.
Meanwhile China is blowing past the world with real improvements in the real world- solar, EVs, etc. meanwhile people keep making their fancy sans serif websites about todo apps, faster than ever before. Useless.
> I have yet to see an end user product that in itself isnt a wrapper around LLMs that is impressive created by LLM assistance.
I don’t disagree that AI is overhyped. But I think you are probably looking in the wrong place.
I think most software that is written isn’t really a product, at least not a public product. It’s an in-house tool or a one-off project needed to complete some larger task. People everywhere are always writing small programs that make their life or job just a bit easier (and explains why so many corporate projects are little more than an excel spreadsheet).
And there are a lot of people who have made custom software just for themselves with AI. Not a product, just a tool or project that finally made sense to build.
But where's the revenue from those? It has to add up to a couple trillion dollars to break even on the capital spending.
13 replies →
> Meanwhile China is blowing past the world with real improvements in the real world- solar, EVs, etc. meanwhile people keep making their fancy sans serif websites about todo apps, faster than ever before. Useless.
Very little about the American economy even makes sense for keeping the edge on LLMs beyond a few years. All the things I would think would be required: energy, research, construction capacity, labor costs -- it's pretty hard to deny who's on the upswing these days. China cranking out current generation microchips will be the last nail in the coffin.
Productivity gains seem like it’s at best a wash when you factor in the massive tech debt cleanup and additional time needed to spec and review.
Misuse of AI tools because of continuing a fundamentally broken software development process.
AI is both overhyped but is also revolutionary at the same time.
I would agree that a lot of companies talking a big talk about using LLMs are failing to actually apply it in a sensible way to their business.
In the time before LLMs, humans made satellites, Concorde, life-saving medical surgery, James-Webb Space Telescope, communication at the speed of light, the list goes on.
What changes have LLMs (Not AI, not machine learning in general, I'm not going to waste time discussing the definition), LLMs made in the past 4 years that indicate anything close to the above? Solving a whiteboard math problem?
2 replies →
Oh, war is transforming hard.
In my opinion, the problem is not even the cost. The problem is that people are using AI for running recurrent stuff instead of writing code to automate it.
For example. Imagine that you are comparing two documents (let's assume diff doesn't exist). You could ask an AI to compare the differences from you or you could use AI to write a tool to do it. For whatever reason, people are starting to go with the former not realizing that now they basically have to pay to compare documents.
I have exposure to AI initiatives at several companies including a few F500's. I have seen teams dump huge logs into frontier models that took hours to get so-so results that we were able to replace with a few lines of python code at 1000 times the speed and 100% accuracy. When asked why they were doing this they literally said "because we don't understand the subject matter so we were depending on the AI". I saw one team file a complaint with a vendor about a frontier backed coding harness and it's inability to consistently format headers because they were using it as a reporting engine. When I recommended they just use the coding tool to write code to generate reports you would have thought I had just cured cancer from their response. I frequently see people complain about the fact that AI is going to take their jobs and then see them gripe about the fact that AI is 'worthless' because it can't do more of their job than it already does. It's easy to see the difference between the people seeing 10x productivity gains from leveraging AI and those who aren't and it's not the AI.
i have trouble understanding these situations, e.g. the AI itself would presumably make the suggestion to write a python script for such a task. It seems to me that there two huge problems right now * understanding which category of problems an LLM is an appropriate solution for (rather than throwing LLMs at any and all problems) * matching model capability (and therefore cost) to the problem at hand. You can easily overspend massively by using a model that's too powerful
Someone asked me if I was using models for fantasy sports, and if it was smart enough to help make decisions about drafting.
My answer: no, but it was able to help me find the website and social handles for every beat writer for every team, and generate a simple website where I can do a daily skim of teams/players and draw my own conclusions.
LLMs are a tool, not a panacea.
I've heard this framed as "AI raises the floor by 2x or less but raises the ceiling by 10x or more"
Laziness, pure and simple. The inevitable consequence of “the LLm is the compiler now”. And what do you even expect people to do when they are forced at threat of termination to use AI for everything as much as possible? Not to mention people are being pressured to do insane thing like review hundreds of pull requests per day and deliver like 15 features per week so OBVIOUSLY there isn’t time to build out proper tooling. Just shove everything in a prompt and call it a day. Some people have families to feed, just do what you’re told.
Agreed. I’ve been telling my team to build up internal packages so we can push all that ad hoc reinvention into something more tangible and deterministic. Invest the $$$ in inference into something the agent can reach for next time that’s neutral and consumable by other code to reduce future spend.
Yes. Build compact CLI-driven tools, write a skill for it (you can use your agent to do most of this work for you).
It just requires being willing to think instead of mashing prompts into a keyboard.
Because you look at the work from the perspective of a programmer, not the perspective of a regular person.
Normal people have never gone around automating their work. The most automation they do is dynamic tables on excel sheets.
I obviously know building a tool that can programmatically do something is a better solution, but I think that requires a fundamental shift in how people work. People need to be told by someone "this is how you should be using the AI" but right now they're simple told "use the AI".
I'm talking about programmers doing this. That's what's sad. These were normal people before,but it feels like they have some kind of AI schizophrenia now. They don't use their brains anymore
Same, even opus favor short term solution and scripts with a billion flags that constabtly require rescanning to understand how to launch it is a constant struggle to get it to build sane default and reusable scripts that run with minimal parameters
Yeah, and what's up with adding dry run to everything? I saw some code that doesn't write anything but still the AI added a dry run which had a completely different codebase
2 replies →
AI can do things around semantic analysis that a deterministic diff tool cannot.
I understand and agree with your point though.
I'm curious if you could give me an example of something that couldn't be down deterministically. We have fuzzy search/matching too ? Regex is a monster when used correctly.
6 replies →
It's this and worse. To use your example, it's like people using AI to write a diff algorithm, incorrectly, then using AI to fix it, because they don't know that diff exists already. Lazyness and starting development with a very low level of understanding. People think lowering the barrier to entry is a good thing, when in reality there are just fundamentals and things you just have to know before you can start using a tool like llms properly.
Isn't that the supposed point of it though? At least how it is marketed/hyped. Don't use your brain, you don't need one, spend all your thinking energy on... dunno, something else, and leave all the "mundane" stuff to AI. Just pay for the tokens, it's going to make you 10x more efficient, the $1000/month is worth it.
100% this. For my own company I mostly build deterministic workflows that may have a simple AI step in the middle using an appropriate Chinese model in a very limited way. I wouldn't want to burn tokens to satisfy some metric.
With this AI is a fallback and not the default. Sounds like large companies have it backwards.
I agree, but even this use case isn't the most wasteful. The interwebs says Agentic consumes 50% of token use, but I'd hazard this number is north of 90% for many shops. My cynical view of Agentic is its sole purpose is to make "number go up".
Look at me! I'm the smartest guy. I've wasted 10M tokens! No one has wasted more!
Same with writing boilerplate code. It’s been a solved problem yet here we are.
Recurrent expensive inefficient processes? Sounds familiar !
This is why we have business analysts and software developers.
To help identify inefficiencies and to build technical solutions.
Oh no! People are doing what they've been told to do!
[dead]
it's all about cost at the end of the day. if you're allowed and encouraged to tokenmaxx, then of course this'll happen.
The cost is a problem, but IMHO more important is delegating so much of your internal knowledge, thinking, and systems to a 3rd party.
We are very close to the point where if Claude and ChatGPT APIs are down, companies cannot function. How is that introduced so quickly into so many critical places without taking that specific fact in consideration? What is the plan for all those companies whose workflows now depend heavily on a remote LLM whenever the services get cut? What if your company account gets banned?
In some ways it is worth than depending on a company for hosting, because even your debugging tools are based on AI. MCP is great to go through datadog, sentry, until your agent or the MCP server are down and you don't know how to look for the issue yourself because you do not actually understand how your systems work.
> We are very close to the point where if Claude and ChatGPT APIs are down, companies cannot function.
Contrast with Gmail/Gsuite/Outlook365/QuickbooksOnline/etc are down, though.
What you cite here isn't a direct attack on AI but on centralized service provision in general. Unfortunately that battle has been lost for decades, now.
None of those are doing the actual development. Here we are talking about a technology people delegate judgement, technical expertise to. It’s way, way deeper of an integration than a standard saas
Those sound like problems for another quarter. The people making the decisions ride the AI hype wave, and if in the worst case the company tanks one day, they take their severance package and leave.
[dead]
There's an old saying, "in the land of the blind, the one-eyed man is king."
Here we have the opposite: In the land of the one-eyed, the blind are leading.
The blind in this case are all those executives and managers who don't understand much about AI's current potential and limitations, and so far have treated it like a magic button that will solve everything. The one-eyed are rank-and-file employees who maybe sort of know a little more about AI.
Executives and managers are the ones who correctly understood which game was being played. The game we are playing is not one of making good products, it's one of getting money from people who both have more money and are stupider than you. They're succeeding at that. We're also doing it, but we're not getting as much money.
In many cases, the people who have more money and are stupider than you are other executives. Sam Altman is arguably one of the executives who know how the game is played. OpenAI is at the front. Microsoft's executives are an example of the ones who got played.
Would have been nice to see 'soaring costs' with numbers. WSJ could do better here. Hundreds of thousands of dollars a month is nothing compared to how much they take with better financial models.
My org has a monthly team plan, and because I don't use what I would call unsupervised agents, I rarely come close to exceeding limits. I guess I am one of those "chat only users" that so many articles limit my output. A split terminal with Vim on the left snd Claude on the right has been a great combination. Neanderthalic AI.
Personally I'd consider efficient and economical use of AI to be a key skill for a good developer. Using AI for everything at full throttle would appear to be a crutch. I wonder if this will eventually become a hiring criteria for companies without unlimited budgets (most of them).
On the one hand, organizations are without question using LLM's well beyond what is actually necessary, and as reality kicks in they're forced to scale back accordingly. However at the same time, on intervals counted in months, we're seeing breakthroughs both in hardware and software that dramatically reduce the cost of inference.
Between corporate FOMO and the rapidly decreasing costs of actually running LLM's I'm interested to see at which side of the spectrum these two meet
Is Anthropic rushing to an IPO because investors also fear it’s the peak? I guess it can grow even after the IPO but is there a fear that the spending will rationalize and hurt the valuations?
Like a few other commenters, other than reasonably making a software developers work easier I haven’t seen a lot of value creation or revenue boost from AI. I know AI companies make money, but the companies consuming these services, are they meaningfully making a revenue to justify the continued spend?
These articles are weird because rationing consumption based on price is one of the most fundamental concepts in economics.
I’ve seen comments on other threads on this subject the general idea that these article headlines are overstating the pullback from AI.
In other words, the news cycle is looking for an AI story that lands with readers, and that the example of Uber blowing through its AI budget and Microsoft discontinuing use of Claude internally are not good indicators.
I agree that those aren’t good indicators.
However, at some point we have to remember that CEOs and boards of directors are just regular morons who read the news the same way everyone else does.
At some point, if a lot of corporate leaders associate AI with mediocre results, high costs, and public backlash, they might just start saying “this juice isn’t worth the squeeze.”
This phase that all these companies went through doesn’t seem that bad. Before these places had a big problem where all their employees didn’t understand how to us ai for their work. Now they’ve overspent and tokenmaxxed and haven’t seen much from it. The next phase is to set the goalpost lower and set quotas based on who uses ai more effectively. Eventually the folks that use it well and are productive will bring in roi. Then you can fire all the folks that aren’t using it effectively and replace them with people that know how to use it. We’re already starting to see this.
They are likely also starting to realize that the end result of their anthropic contract is that nobody but anthropic knows how to run their business. Why would anthropic not treat their business like a utility in the future?
Don't have a subscription to wsj.
Only thing I can say AI was useful for, in a corporate environment, was learning a new coding language on the fly. Gives me a baseline to work off of and fix.
But I can learn without it, too. A nice tool, but not a need.
> Don't have a subscription to wsj.
An ironic analogy sort of, once media started hiding behind a paywall, I just stopped reading them rather than paying. Same with LLMs - usable if cheap/free.
Corporate or corporate in programming space?
90%+ of corporate people are not programmers. 1 programmers can cause the same token damage with a bunch of concurrent agents as a couple thousand Karens in compliance asking a chatbot questions
It's much easier to deliver incremental AI ROI on the later even if it's hard to measure/quantify. A 1000 tokens might point this compliance person in the right direction on a key problem. Meanwhile 1000 tokens doesn't get you anything useful on coding
The other day we (wrongly) concluded that product market fit has been achieved and now the rivers of hot molten milk chocolate and honey are all that's in the future etc.
Some related discussions:
https://news.ycombinator.com/item?id=48307098
https://archive.ph/v2dwg
Tokenmaxxing is absurd. Using a fixed cost monthly plan seems sensible. Taking time to review the generated code is a good thing.
Where is the tokenmaxxing chad / chadette that burnt a half a billion dollars in a single month?
Another reason to favor using AI to build automation instead of relying on it in prod: the risk of war and global instability.
If LLMs are genuinely helpful or even decisive in a military engagement, you can expect any host country to commandeer whatever data centers they need, leaving commercial entities to bid up the prices on the leftover capacity.
Another risk is that data centers are a great target for cyber warfare.
It’s ideal if your business can leverage LLMs when they’re online but continue to operate profitably when they’re offline.
Even regular warfare, if the Middle East AWS regions are an indication. The giant and arguably excessive data centers being built are not hardened physically.
As a developer, I don’t think it’s just that costs are going up. I’m also seeing more people lately talk about “vibe slop”.
I've noticed as well. A lot of pull requests are just agents running constantly, hoping to have produced something of value. Entropy is at an all-time high, though.
The bottleneck becomes human review, not code generation. A PR can look plausible and still add more entropy than value.
We have some dude at work who runs their own agent that makes constant commits. We're supposed to review the agent's output.
Exactly. The bottleneck becomes human review, not code generation. Agents can generate commits faster than humans can verify whether those commits should exist.
There's a paywall, but it's an interesting question how much of the recent explosion of the AI companies revenues is because of the explosion in prices, and how much their customers will accept the increased prices.
It will be interesting to see to see Anthropic’s “revenue bubble” pop as this happens. At least it should hopefully free up some capacity.
- Global economy on the verge of depression.
- ChatGPT drops, AI is perfect to be our savior.
- AI glorified as the great messiah.
- Everyone worships stocks even remotely related to AI.
- Execs desperate for relevance boast about tokenmaxxing.
- SHTF
- burst
- last year flagship GPUs and DRAM are sold used for the price of a burger.
- Laidoff people start using local AI as hardware price drops to make actual useful stuff
- New round of bootstrapped tech bros that eventually give birth to the new metaverse/NFT/etc.. hype.
[dead]
[dead]
[dead]
LLM doesn't work, let alone profit.
Yesterday I updated our dependency on the sqlx crate and put up a PR, and it failed in the CI build in a way I couldn’t reproduce locally.
I asked codex to take a look, and it:
- Grabbed the CI logs on its own to figure out what the CI error was
- Looked at my local setup
- Looked at the changes in sqlx from 0.8 to 0.9
And figured out that sqlx depends on an updated version of the “whoami” crate but doesn’t specify default features, which causes it to fall back on a stub implementation that makes the default user “anonymous”, which was failing to authenticate to the UNIX socket we use in our CI Postgres server. It patched the environment variable for our docker container to explicitly specify a username and the issue was fixed.
It would’ve taken me probably several hours to figure this out on my own. It took codex maybe 5 minutes.
Tell me again how LLM’s “don’t work”?
I agree with your point in the broad sense, but the example might be bad. If sqlx is an important crate, and not stable yet, upgrading it without reading the changelog is honestly a flaw in your team process. Using the AI to fix organisational issues is typically one of the reasons I'm very skeptical of AI improving productivity in the long run.
I'm not taking a shot, to be clear, we had a similar issue a few years ago and we made sure this wouldn't happen again, that's absolutely not a shot, nor do I think it's a character flaw to use AI, au contraire, this is a very good use. I'm just worried that because AI is so good at fixing minor issues caused by governance/organisation flaws, we will be stuck using it to fix those and be trapped in mediocrity (that's not an issue for me, mediocrity is where I work best, but I'm a bit sad for the great Devs I've worked with.)
1 reply →
You used it in a way where the result was simple and you could verify its correctness. You used it as a super-search tool, it's good at that. It's a different use case than having it generate a lot of code from scratch.
1 reply →
It sounds like the real problem was that the programmer was not familiar with the tools they were using and decided to dig themselves out of a hole of their own making by turning to AI instead of learning to use their tools better.
3 replies →
elaborate please, how does it not work?
They cannot be trusted to produce output that works (let alone works well) because they are just statistical models, without any actual understanding of what they produce. That means that you have to carefully review every single line of code they produce, because you don't know where the hallucinations will be. But by the time you do that, you have saved no time at all (indeed, in my experience you lose time), because typing the code was never the part that took time. It was understanding the problem. So if you use an LLM, you spend a bunch of money for zero gain in productivity, or you sacrifice quality and pray there aren't nasty bugs lurking. I certainly think it's fair to call that state of affairs "it doesn't work".