There is no question LLMs are truly useful in some areas, and the LLM bubble will inevitably burst. Both can be simultaneously true, and we're just running up the big first slope on the hype curve [0].
As we learn more about the capabilities and limits of LLMs, I see no serious arguments scaling up LLMs with increasingly massive data centers and training will actually reach anything like breakthrough to AGI or even anything beyond the magnitude of usefulness already available. Quite the opposite — most experts argue fundamental breakthroughs will be needed in different areas to yield orders-of-magnitude greater utility, nevermind yielding AGI (not that much more refinement won't yield useful results, only that it won't break out).
So one question is timing — When will the crash come?
The next is, how can we collect in an open and preferable independently/distributed/locally-usable way the best usable models to retain access to the tech when the VC-funded data centers shut down?
(Although I think the utility of server farms will not be high after the bubble bursts: even if cheap they will quickly become outdated. In that respect things are different from railway tracks)
In my uninformed opinion, though, companies who spent excessively on bad AI initiatives will begin to introspect as the fiscal year comes to an end. By summer 2026 I think a lot of execs will be getting antsy if they can't defend their investments
Having been through at least two AI hype cycles professionally, this is just another one.
Each cycle filters out people who are not actually interested in AI, they are grifters and sheisters trying to make money.
I have a private list of these starting from 2006 to today.
LLMs =/= AI and if you don’t know this then you should be worried because you are going to get left behind because you don’t actually understand the world of AI.
Those of us that are “forever AI” people are the cockroaches of the tech world and eventually we’ll be all that is left.
Every former “expert systems scientist”, “Bayesian probably engineers” “Computer vision experts” “Big Data Analysts” and “LSTM gurus” are having no trouble implementing LLMs
LLMs are inappropriately hyped. Surrounded in shady practices to make them a reality. I understand why so many people are anti-LLM.
But empty hype? I just can't disagree more.
They are generalized approximation functions that can approximate all manner of modalities, surprisingly quickly.
That's incredibly powerful.
They can be horribly abused, the failure modes unintuitive, using them can open entirely new classes of security vulnerabilities and we don't have proper observability tooling to deeply understand what's going on under the hood.
But empty hype?
Maybe we'll move away from them and adopt something closer to world models or use RL / something more like Sutton's OaK architecture, or replace back prop with something like forward-forward, but it's hard to believe Hal-style AI is going anywhere.
They are just too useful.
We have a rough draft of AI we've only seen in sci-fi. Pandora's box is open and I don't see us closing it.
Really hard to believe articles like this and even more hard to believe this is the hive mind of hacker news today.
Work for a major research lab. So much headroom, so much left on the table with every project, so many obvious directions to go to tackle major problems. These last 3 years have been chaotic sprints. Transfusion, better compressed latent representations, better curation signals, better synthetic data, more flywheel data, insane progress in these last 3 years that somehow just gets continually denigrated by this community.
There is hype and bullshit and stupid money and annoying influencers and hyperbolic executives, but “it’s a bubble” is absurd to me.
It would be colossally stupid for these companies to not pour the money they are pouring into infrastructure buildouts and R&D. They know it’s going to be a ton of waste, nobody in these articles are surprising anyone. These articles are just not very insightful. Only silver lining to reading the comments and these articles is the hope that all of you are investing optimally for your beliefs.
I'm a little skeptical of a full on 2008-style 'burst'. I imagine it'll be closer to a slow deflation as these companies need to turn a profit.
Fundamentally, serving a model via API is profitable (re: Dario, OpenAI), and inference costs come down drastically over time.
The main expense comes twofold:
1. The cost of train a new model is extremely expensive. GPUs / yolo runs / data
2. Newer models tend to churn through more tokens and be more expensive to serve in the beginning before optimizations are made.
(not including payrolls)
OpenAI and Anthropic can become money printers once they downgrade the Free tiers, add ads or other attention monetizing methods, and rely on a usage model once people and businesses become more and more integrated with LLMs, which are undoubtedly useful.
I don’t think people know what the definition of this bubble is yet. I can provide one:
- AI-first app companies that actually go public on the stock exchange
- Massive influx of investment from retail as the basket of “AI” is just too much to pass up
- This basket is no longer a collection of top tier hardware and software titans, but led by resellers and wrappers like Palantir, like something like Cursor, like Windsurf, and finally rounded out with crud-apps turned publicly traded companies. Figma going public is a very bad indicator of what’s to come. Perplexity going public would be one of my biggest Red Flag moments.
- The basket I’m describing is the package that includes all these “toxic” assets.
- Some really dumb big players will lose here too because they will acquire some of these resellers and wrappers at prices they’ll never recoup (Newscorp buying MySpace).
- And finally, those who know, know, and they will bail first unscathed. Say it ain’t so, the story of our lives.
That will be the vehicle retail will pile into. We’re a little bit aways from that as companies are still building out their AI offerings. We’ll need a flurry of companies like that to go public soon after OpenAI does, sparking the beginning of one of the worst bubbles ever. You won’t be able to make sense of it because the bull market will make it impossible to not FOMO in.
That’s the systemic risk to this entire industry and the broader economy in about a few years.
Remember, humans can’t have nice things. If the secondary companies didn’t rush to the stock market as their prime imperative, we wouldn’t have to worry about it because all sensible investment will be in the large caps. The pursuit of gaudy returns will fail humans again, as always.
Stay safe and right-sized, all. The actual tech is not over-hyped.
So, one difference between this and the dot-com bubble is that it is much, much harder to go public now, and much, much easier to raise funds as a private company. This has lead to loss-making private companies with valuations that would not have been remotely plausible a couple of decades ago. Arguably a more likely end to all of this is that the VCs turn off the tap, which will kill most companies in the space within a year or so, with fairly limited contagion to the broader markets; public companies who've gone heavily into it may be badly burned, but that would be about it.
Retail may never really get to participate at all, beyond trading Nvidia and similar.
> OpenAI began this hype cycle [...] and its death (or, as mentioned, some other kind of collapse, such as acquisition) is the sign that we’re done here, in the same way that FTX signaled the end of the cryptocurrency boom..
The collapse of FTX sent bitcoin from ~$20k to ~$17k. It's now $110k. I imagine the AI boom will 'collapse' in the same sort of way.
A lot of the economics depends on whether you think human level intelligence is coming or not. Zitron kind of assumes not in which case his economic doomerism makes sense. But if it does come you could effectively double gdp which is a lot of financial upside.
15 comments
[ 4.2 ms ] story [ 35.8 ms ] threadAs we learn more about the capabilities and limits of LLMs, I see no serious arguments scaling up LLMs with increasingly massive data centers and training will actually reach anything like breakthrough to AGI or even anything beyond the magnitude of usefulness already available. Quite the opposite — most experts argue fundamental breakthroughs will be needed in different areas to yield orders-of-magnitude greater utility, nevermind yielding AGI (not that much more refinement won't yield useful results, only that it won't break out).
So one question is timing — When will the crash come?
The next is, how can we collect in an open and preferable independently/distributed/locally-usable way the best usable models to retain access to the tech when the VC-funded data centers shut down?
[0] https://en.wikipedia.org/wiki/Gartner_hype_cycle
(Although I think the utility of server farms will not be high after the bubble bursts: even if cheap they will quickly become outdated. In that respect things are different from railway tracks)
In my uninformed opinion, though, companies who spent excessively on bad AI initiatives will begin to introspect as the fiscal year comes to an end. By summer 2026 I think a lot of execs will be getting antsy if they can't defend their investments
Each cycle filters out people who are not actually interested in AI, they are grifters and sheisters trying to make money.
I have a private list of these starting from 2006 to today.
LLMs =/= AI and if you don’t know this then you should be worried because you are going to get left behind because you don’t actually understand the world of AI.
Those of us that are “forever AI” people are the cockroaches of the tech world and eventually we’ll be all that is left.
Every former “expert systems scientist”, “Bayesian probably engineers” “Computer vision experts” “Big Data Analysts” and “LSTM gurus” are having no trouble implementing LLMs
We’ll be fine
LLMs are inappropriately hyped. Surrounded in shady practices to make them a reality. I understand why so many people are anti-LLM.
But empty hype? I just can't disagree more.
They are generalized approximation functions that can approximate all manner of modalities, surprisingly quickly.
That's incredibly powerful.
They can be horribly abused, the failure modes unintuitive, using them can open entirely new classes of security vulnerabilities and we don't have proper observability tooling to deeply understand what's going on under the hood.
But empty hype?
Maybe we'll move away from them and adopt something closer to world models or use RL / something more like Sutton's OaK architecture, or replace back prop with something like forward-forward, but it's hard to believe Hal-style AI is going anywhere.
They are just too useful.
We have a rough draft of AI we've only seen in sci-fi. Pandora's box is open and I don't see us closing it.
Work for a major research lab. So much headroom, so much left on the table with every project, so many obvious directions to go to tackle major problems. These last 3 years have been chaotic sprints. Transfusion, better compressed latent representations, better curation signals, better synthetic data, more flywheel data, insane progress in these last 3 years that somehow just gets continually denigrated by this community.
There is hype and bullshit and stupid money and annoying influencers and hyperbolic executives, but “it’s a bubble” is absurd to me.
It would be colossally stupid for these companies to not pour the money they are pouring into infrastructure buildouts and R&D. They know it’s going to be a ton of waste, nobody in these articles are surprising anyone. These articles are just not very insightful. Only silver lining to reading the comments and these articles is the hope that all of you are investing optimally for your beliefs.
Fundamentally, serving a model via API is profitable (re: Dario, OpenAI), and inference costs come down drastically over time.
The main expense comes twofold: 1. The cost of train a new model is extremely expensive. GPUs / yolo runs / data
2. Newer models tend to churn through more tokens and be more expensive to serve in the beginning before optimizations are made.
(not including payrolls)
OpenAI and Anthropic can become money printers once they downgrade the Free tiers, add ads or other attention monetizing methods, and rely on a usage model once people and businesses become more and more integrated with LLMs, which are undoubtedly useful.
- AI-first app companies that actually go public on the stock exchange
- Massive influx of investment from retail as the basket of “AI” is just too much to pass up
- This basket is no longer a collection of top tier hardware and software titans, but led by resellers and wrappers like Palantir, like something like Cursor, like Windsurf, and finally rounded out with crud-apps turned publicly traded companies. Figma going public is a very bad indicator of what’s to come. Perplexity going public would be one of my biggest Red Flag moments.
- The basket I’m describing is the package that includes all these “toxic” assets.
- Some really dumb big players will lose here too because they will acquire some of these resellers and wrappers at prices they’ll never recoup (Newscorp buying MySpace).
- And finally, those who know, know, and they will bail first unscathed. Say it ain’t so, the story of our lives.
That will be the vehicle retail will pile into. We’re a little bit aways from that as companies are still building out their AI offerings. We’ll need a flurry of companies like that to go public soon after OpenAI does, sparking the beginning of one of the worst bubbles ever. You won’t be able to make sense of it because the bull market will make it impossible to not FOMO in.
That’s the systemic risk to this entire industry and the broader economy in about a few years.
Remember, humans can’t have nice things. If the secondary companies didn’t rush to the stock market as their prime imperative, we wouldn’t have to worry about it because all sensible investment will be in the large caps. The pursuit of gaudy returns will fail humans again, as always.
Stay safe and right-sized, all. The actual tech is not over-hyped.
Retail may never really get to participate at all, beyond trading Nvidia and similar.
The collapse of FTX sent bitcoin from ~$20k to ~$17k. It's now $110k. I imagine the AI boom will 'collapse' in the same sort of way.
A lot of the economics depends on whether you think human level intelligence is coming or not. Zitron kind of assumes not in which case his economic doomerism makes sense. But if it does come you could effectively double gdp which is a lot of financial upside.