▲ 79 points
back
93 comments
IMO, I read 2 faulty assumptions:
1) That LLM/Agents are being pushed and not adopted. I see plenty of deep adoption by junior folks.
2) The unit economics don't work out. From the details on every model so far - each model is wildly profitable over it's amotized time-frame. It's just that money is used upfront for the next model, and each next model is significantly more costly to train. The best case argument instead is - this will not last and we'll pour more on some models, than see in it's revenue.
I think realistically these form the core of the thesis, and IMO, and hence it's conclusions are a bit off the mark.
Juniors reverting to something even less useful is not a selling point. Often most adoption is forced by bosses. Repo code change metrics tell you all you need to know how well it 'works'.
The economics do not work, not even close. Even if they ever did (probably a decade or so after the bubble pops), all parts of the stack(with the expetion of nvidia, maybe) are interchangeable. Meaning that people can easliy swap out foundation models, nor are creating new wrappers very hard. It will be a race to the bottom, I doubt anyone will make much money.
Last I checked, ycombinator will not fund your start-up if you shill for AI hard enough.
1) there is more to the world than software development
2) there is no profit. There is barely any revenue, the only money is continuous injections of VC cash and some frankly Enron-like book keeping.
apparently their goal is to take part of the 60 trillion dollar labor market that includes fast food order takers, IT professionals, hotel front desk workers, etc. There is plenty of payroll they can target
> I see plenty of deep adoption by junior folks
Where mate? Details?
> From the details on every model so far - each model is wildly profitable over it's amotized time-frame
Is it? Is that why all of them are switching their users from the subsidized flat-rates to billing based on usage?
> hence it's conclusions are a bit off the mark
You're funny - they are spot on and any dreamer who is working for equity in these LLM-wrapper-product companies who dreams of getting rich in the next few years or so, is in for a nasty surprise.
Additionally the internet bubble left us a legacy of installed fiber that remained mostly unused for almost a decade. This time around all the capital intensive stuff have an expiration date, gpus have a short training lifespan (4-5 years). Models are outdated the moment their training is complete.
4-5 years for GPU being outdated is a bit ... outdated. 3090 from 2020 still get sold for more than the release price
Also thinking about it. Fibre was in the ground. It had minimal storage costs. Same can't really be said about buildings and hardware there which has ongoing costs even if turned off. Storage alone has cost involved at this scale. Warehouses can be relatively expensive. So there is also that sort of aspect.
the ~10x/year drop in inference cost makes the capex depreciation cycle even harder — a cluster that's profitable today may not pencil out in 18 months
Are the datacenters that are being built not directly analogous? Even if the hardware in them is cooked after 5 years, the buildings, power, cooling, and fiber interconnects are still all valuable.
The models may go out of date but the process and software are continuously improving.
My thoughts exactly. At most you get a surplus of cheap third tier AI. Which may or may not be helpful. And or a bunch of unused unmaintained deteriorating data center buildings.
It was only really the US that was left with the legacy of installed fibre.
The 2000 crash left a lot of broken economies worldwide. Many non-US stock markets benefitted from the tech stock feeding frenzy without the investment actually being used to build anything.
If the AI bubble pops, a handful of US megacorps may be left with good models, datacentres and other assets, but the economic shocks will be felt around the world.
4-5 years is not short? Don't companiess write off their hardware after 3 years mostly anyway?
But this is also an insurance against the threat of an overcapacity-induced bubble: whatever capacity is built, it won't last more than a few years before becoming obsolete anyway. There's no risk that once we've finished building the railroads, or the network links, these will be "more than enough" for at least a decade.
I have a question, is the short lifespan of GPUs because they get worn out and are destroyed, or because they get outdated by the ever expanding demands of the AI bubble?
Because if it's the later, I would assume that growth would not continue at the same rate after the bubble bursts?
The article exaggerates things quite a bit.
At the time of the Internet bubble, there were people pushing for more "free" usage of the Internet, and those that couldn't care less.
And it's not like the companies didn't want to take advantage of the Internet, but there was a mismatch between what the companies and the employees had in mind, which mostly boils down
* Employees want to use it to do their jobs and make their life easier
* Companies want to improve productivity, spend less and make more money.
There is some overlap of course, but the problem is where the two clashes.
I don't think today it is too much different. I see plenty of people using AI for what they care about, they complain when they are asked to use it for things they fear will make their life worse (like programmers that think they will have to pick up the pieces of vibe coding later on).
> As a group, teenagers and young adults hate AI
I wonder what is their definition of AI. I haven't seen a single young person saying "I don't use chatgpt (or the like) because I hate AI". If else plenty of student have become dependent on it.
> > As a group, teenagers and young adults hate AI
Anecdotally, I've observed a robust correlation between the cost/quality of the model, and attitude towards it.
Most of the general public, young folks, and old folks (ie outside gen z, millennials, and some X) are using free models, usually what's immediately available (cough copilot cough), have really unreliable results, hear all the hype, experience dissonance, and chalk it up to just hype, and walk away thinking AI is a crock of junk.
The Z/Y/G cohort - the ones that grew up alongside the growth of the internet - seem to be the best adopters. They recognize a system which is powerful, albeit flaky, and know how to extract utility from it without over-reliance. Especially ones with paid flat-rate subscriptions.
The power users - the ones using API/paid (by usage) models, tricking out their claude with plugins, seem to have the least amount of hate, but rather a healthy respect for a powerful disruptor.
I also don't buy the whole "the young'ns have never dealt with barriers of entry to the internet and thus lack the tech skills the millennials developed." I think the internet cohort that adopted tech was always split between the powerusers/curious learners, and the "just get my goal accomplished and get out" folks. I think that's roughly the same percentage of folks in Z/alpha, and these kids are just as savvy and aware of limitations of the tech.
I haven't worked in corporate since last year but I keep seeing people complaining that "bosses" are forcing workers to use AI now. I find this so amusing because in 2023-2024 I had to fight to either be allowed to use AI at work (even just MSFT Copilot chatbot) or get a ChatGPT Enterprise license.
It was mismanagement then and it's mismanagement now, the more things change the more they stay the same.
Something else I've been thinking about which makes the economics of AI weird: The more powerful you make AI, the easier you make it for everyone else to make AI. I probably wouldn't bother to train an LLM from scratch, but I'm sure if I spent a few days with Codex/Claude Code I could do it (like GPT-2 level) easily. Obviously the capital moat is massive at the moment, but in like 50 years that probably won't be true.
I AGI ever becomes reality, an interesting question would be: What is the minimal AI system that can come up with AGI?
These AI mandates are quite hard to actually push in some job.
I’m in a finance role and thus far it’s all been rather hand wavy „use copilot more“. Maybe some meeting summarization. Nothing like the programming space where token counts matter to management
Will be interesting to see where this goes. My testing with it thus far has just yielded multi million dollar hallucinations. Senior management will presumably try anyway
This contains my personal disdain for AI. Using it to do bullshit work. That’s solving a symptom. Stop doing bullshit. Stop using tools and processes which are bullshit heavy. Stop sitting there silently accepting bullshit. And certainly don’t pick another tool which is trained in bullshit and ask it how to do things.
One wonderful thing I’ve watched for the last 4 years is my company fail to build a modelling tool better than Excel. On attempt 3 we have some pile of shit Claude generated on nodejs and Postgres on kubernetes which can’t replace a single spreadsheet written in 2008. Because everyone thought into the bullshit not the solution or the requirements.
Edit: thinking further, it appears people forgot what the problems are and think from the solution back. That never works. But it sells tools.
LLMs are life-changing for a dev who has been writing code for 20+ years (because I am tired of it).
Outside of AI's impact on software, which is massive, the biggest change that we are going to see, I think, is the crushing amount of useless information generated by it.
We already see how everything is racing to the lowest common denominator once we granted Average Human Intelligence unfettered access to expressing thought via social media.
Now that Average Human Intelligence just has a button that says "Generate Bullshit For Me. Send to the World".
UGH.
"Simplicity is a great virtue but it requires hard work to achieve it and education to appreciate it. And to make matters worse: complexity sells better" -- Edsger W. Dijkstra
Most workers are just that, workers. They don't have a say in their work, bullshit or real. Only the lucky ones have the opportunity to say "Hey, I've been thinking about this [task/report/whatever] - do we really need it?" and get a "You're right, let's reevaluate this." from their boss / manager.
Or even worse, many employers and employees alike are afraid to cut out BS work - because it could realistically mean cutting down on the workforce. So they continue to produce work that no one checks, because at least then they can justify their position.
taps the "most jobs are bullshit" sign
They are not about actually "doing things", they are social validation, particularly the part where the people with resources/capital enjoy your company and give you what you need to live a dignified lifestyle in exchange for it.
But acknowledging and acting on this would destroy the leverage the useless-but-likeable have in terms of being able to get paid, and that the owner class have in terms of getting people to pretend that they like them/validate their often cruel and avaricious choices and behavior.
+1 - the Office Bullshit Worker is the one upholding this shit these days- adding some of those creepy unnecessary images to their slide-decks, writing those godawful oververbose e-mails and fucking not being able to take notes without their AI. Why the fuck are you even in the meeting if you cannot note down the key points afterwards.
labor-led automation produces improvements in quality, while capital-driven automation increases throughput
This explains the management push (ignorance) but also the user push (automating BS tasks). The common thread is that the user doesn’t have to take any responsibility for the output. This is why people don’t like having LLMs pushed on them, because for cases where they are responsible for or have to consume the output, they don’t work very well, but when it’s just something that needs to look ok at a glance and be handed off, everyone is rushing to use them.
Quit reading after it suggested Lotus Notes was a Microsoft product [1], which it wasn't. The last time I used it, about 2002, it had become an IBM product. If the basic facts aren't right, I'm not interested in the conclusions.
[1] "... what made Notes so important to Microsoft ..."
+1. Surprised no one else is calling this out.
In fact back then Notes was one of the main competitors to Office/Sharepoint.
Early, very early, in my career unit testing was becoming a thing. A few middle managers (non technical) read some articles and decided this was going to fix all the quality problems with the product so decided to enforce it from the top down, even to the point of requiring developers to present their planned unit tests to management before starting on new features! It was completely absurd, but I was too junior to really understand and articulate why.
I'm lucky enough to be in a great company right now, so I decide when I think AI will help me and use it accordingly - but reading about forced AI adoption reminds me so, so much of that earlier time. Non-technical people who don't trust their engineers to use the tools in the way they see best - in their ignorance, and ego, they think the answer is obvious if only those strong headed tech weirdos would listen.
And amongst all this, there is a class of manager and executive that I'm convinced utterly despise engineers. They hate the fact they focus on details, analyse, make predictions grounded in reality. On a personal level, they can't comprehend that some people take deep satisfaction and contentment from building software, from simply learning things, and they don't understand it, it scares them. Why don't they just pursue normal people things in life? Like super expensive cars, massive houses, golf memberships. I think it scares them that they don't have control over technically minded people they way they might do with others. AI is, in their mind, a way to get rid of these people forever, to just "get stuff done" without objections, and they are pushing extremely hard for that to be true, simply because they want it to be true - not because there is any evidence for it.
Rant over.
Interesting read. I'm going through that same issue at my work where my boss wants me to "educate" the rest of the devs on the use of Copilot to make them more efficient, however, I have no time to put anything together and I imagine the Copilot dashboard figures are not getting any better over time... oh well!
However, something occurred to me when reading it. I was thinking about AGI (or ASI) and what would happen if someone were to achieve it (not sure what it would look like or what constitutes AGI... not the point I'm making here).
What if the primary goal of the first AGI is to keep itself at the top? What if it's goal is to prevent any other AGI? Scary thought...
> What if the primary goal of the first AGI is to keep itself at the top? What if it's goal is to prevent any other AGI? Scary thought...
is basically the premise of
https://en.wikipedia.org/wiki/If_Anyone_Builds_It,_Everyone_...
> Lotus Notes, a kind of primitive precursor to all-in-one office productivity suites like GDocs, Office365, etc.
One of the most inaccurate descriptions of Lotus Notes I have ever come across. Try Exchange+SharePoint+Teams and you'll be at least a little bit more accurate.
At least pets.com tried to actually make money selling dog food.
AI crap is now at the “monetize lonely people for their AI girlfriends” while burning more money than pets.com ever did in a month.
Hopefully it all collapses before we destroy ourselves with data centers and bailouts for the rich.
I think all the screaming about drumming up uses for data centers we don’t actually need is to make the poor pay for the artifacts of the surveillance state.
I keep thinking the plan is clear: beyond some level of compute, they’ll attempt some massive surveillance system that makes China look benign by comparison. They’ll rent out any idle servers to the US Government, thus keeping the rich in place forever.
So it’s absolutely correct: the AI bubble isn’t like the internet bubble. It wasn’t trying to prop up or enable a surveillance state par excellence and destroying our financial system and environment to do so.
internet cos in 1999 had near-zero revenue. NVDA alone did $130B last year. the risk is the capex depreciation cycle, not the pop itself.
The California gold rush had multiple winners including Levis-Strauss. People selling jeans and shovels made a killing as they are now. That does not mean that most prospectors won't go bust. I.e. NVDA may be the winner, but OpenAI, Anthropic, etc may not survive.
I think the analysis-space really needs to be divided into three groups: software, media (audio/video/image) generation/alteration, and everything else.
Software - this tech is ludicrously powerful and productive. But it's a force multiplier, not a "push button, receive software" system. Great devs that know how to wield it will become überdevs, becoming more productive and with lower defect rate (we have objective internal numbers backing this). But bad devs and non-devs will become high output slop factories. You basically need a dedicated platform team to keep things on the rails. I think this is very akin to the internet bubble. The process, institutional knowledge, and feedback systems developed at this time will grant the "survivors" massive edges after the pop.
I think media generation is or will be a solved problem. Animators, 3dfx, background/filler music composers, those jobs are in sorry shape based on current trends. But a cost explosion could easily level the playing field.
Everything else, where middle managers are aggressively pushing AI usage? Yeah maybe. At this time, other than for document retrieval (basically suped-up search), the "productivity" gains don't really map to value gains. Oh wow you can crank out powerpoint slide decks 50% faster. Write 50% more corporate emails employees barely read anyways. There's definitely a trust issue there with hallucinations. If the reliability gap can be solved (the bots don't even have to be correct, they just need to be less confidently incorrect, and I already see this somewhat with my own agents with tuning), then that could prove the turning point between "begrudging usage at the behest of higher ups" and "actual productivity enhancer."
Does no one remember in the dot com boom all the internet skepticism? "I don't trust it with high value orders, what if it crashes or loses data? Call me old-fashioned but I'd rather write it down." That attitude was quite prevalent for years, even into the 2000s.
"AI bubble" in the title, count me in.
TL;DR:
A great technology drives its own adoption, its usage is pioneered by the tweens and young adults, it requires minimum effort and investment to hop on board, and it does not need explaining. It grows organically. Examples: internet bubble.
A bad technology: despised by the young adults and tweens, needs trillion of investments and marketing to drive market penetration, every day some boomer (=not in terms of age, but in terms of mentality) explains how you are holding it wrong and it needs a fuckton of explanation. The Pope himself issues an Encyclica warning on the dangers of it, spurning the greatest popular interest in Catholicism since the dark ages. Examples: LLMs.
Young people use LLMs extensively. Just ask any educator.
I think my company is a microcosm of the current state. The non-engineering side (HR, correspondence, marketing) are on the "forced adoption" side, giving out gift cards to folks using Glean the most.
In engineering, we can't raise token budgets fast enough. Devs are "routing around damage" when they hit caps, going from claude to opencode to copilot. Productivity is up (roughly) 100-300% in terms of story points and 75-200% in lines of code. And defect rate is down [0], more bugs are caught in review before QA or prod. Our teams are just starting to figure out our new workflows too, for design -> spec -> code -> review, it'll only get better as we refine the process.
It's looking like software industries will reap massive benefits, while most others which have some error tolerance will only see modest gains. It's unclear how it will impact high accuracy fields like legal (it might even be net negative).
Also which is it - a useless technology that has to be force-fed because it sucks, or a economy-shaking game changer that will put folks out of jobs en masse? Those seem like a contradiction.
0 - i think process here is extremely important. I think it would be very easy to create an unmaintainable slopocaplypse. We have an informal platform team of about three (including myself) that have been affectionately and informally dubbed the Tech Priests of Mechanicus Adeptus (warhammer reference) that ensure the prompts/skills and associated tooling are optimal, that code standards are enforced, and that solutions are converging at the system-wide level.
https://www.pewresearch.org/short-reads/2026/03/12/key-findi...
>A majority of teens use AI chatbots. Roughly two-thirds of U.S. teens ages 13 to 17 (64%) say they ever use an AI chatbot, according to a fall 2025 survey.
>Around half of adults under 50 say they interact with AI about once a day or more often. Smaller shares of those 50 and older say the same, according to the June survey.
and mind you, that particular study bends over backwards to say "AI bad".