Gary, I thought you might be interested in how fusion has in some respects failed in the same way as AI. My father was a professor of plasma physics for 30 years, and his opinion was that, instead of focusing on the tokamak, we still needed much more science to find a better model. In other words, the fusion sector jumped on the first technology that had some kind of return (eg, the ITAR project in France).
The same thing is happening in AI. Instead of trying many different paths, they jumped on one that, while very flawed, yielded some results, and since there was so much money sloshing around and venture capitalists like to hear good stories, the entire industry is wasting trillions on the wrong approach.
And the AI companies think they can use fusion to power their data centers! https://time.com/7328213/nuclear-fusion-energy-ai/. A technology that doesn't exist yet, to power a technology that has no long-term business plan. Great combination!
A difference is that fusion projects have continued to evolve, such that net positive energy (a technical measurement, not a commercial possibility, see https://lift.llnl.gov/) has occurred at the Livermore facility. With advances in superconductors, the Commonwealth Fusion project (https://cfs.energy/) is moving at a pace and with a design that is far different and probably ahead of the ITAR. It is unfortunate that the US government is not pushing the development of fusion at anywhere near the pace of China.
The Livermore facility concentrates on simulating nuclear weapons, and they use lasers focused on a nodule (my technical understanding is general). They have tended, historically, to ocassionally announce breakthroughs that weren't really applicable, just a warning. But thanks for the info. It's certainly a field to keep an eye on, it's just that I am very skeptical. It is the quintessential 'we are 15 (or pick your number) years away' technology
I'm not convinced that the comparison between fusion energy and LLMs is correct. The Tokamac approach being pursued by the international ITER collaboration is by no means unique, and the fundamental impediment- production of sufficiently strong magnetic fields- eluded researchers until the advent of superconducting magnets. I am not trying to minimize the difficulty of the task. The physics is basically understood. But as yet there is no consensus on what constitutes consciousness or6 thought, and moreover, unlike ML, for which error performance is characterizable, LLMs are fundamentally unpredictable.
I couldn’t agree with your closing thought more. I am excited about how Seed IQ’s active inference based AI agents acted as a control layer and created stability on IBM’s NISQ hardware. If Quantum leaps ahead we might not need all the data center compute.
I think the pundits in the West simply haven't got it when it comes to understanding how China has integrated the 1st, 2nd and 3rd Industrial revolutions. Whereas Western capitalism simply jettisoned the 1st and 2nd industrial revolutions largely through financialisation/export of production, driven by short-term profit goals, the 'secret' to China's success has been to possess the entire spectrum of the production process, of which AI is the culmination. After all, it's ALL ABOUT ECONOMICS! What's more, automation isn't leading to unemployment in China (AI is automation by another name), far from it, AI when applied humanely, opens up the possibility of upskilling, of transforming the relationship between labour and production, this is what socialism is all about.. Thus the CCP just passed a law making it illegal for AI to create redundancies. Two different worlds.
These IPO's will be interesting. SpaceX was fortunate and lucky to be first though its value seems absurd and its AI component dubious. Being 2nd is better than being third and if Anthropic can sort out its spat with the USG, which seems to be happening they are positioned to be 2nd.
but
agree this energy and capital intensive "boil the ocean" approach to AI seems low ROI.
In the standard marketing model, there's only room for two dominant brands in a category (the leader and the alternative to the leader, along with a few less-profitable wannabes, epitomized by the old Hertz vs. Avis rental car market, where the second player had to 'try harder'). This is due primarily to the bandwidth of human recall, so it'll be interesting to see how many of the current AI competitors can hang on and achieve profitability as an actual market emerges, and for how long...
LLMs look increasingly like they're following the gene transcription model: once a couple of groups can replicate the initial basic transcription and the enthusiasm for all the new data has worn off, the real fun (and money) is in making transcription easier, faster, more reliable, more flexible and most important, much less expensive. But no one makes any money on the v1.0 of the initial transcription process. Maybe it's a PPP but you need government research dollars to find the 1st path.
The hyperscalers and the frontier labs seemed to think they could build walled gardens like their socmed ad businesses or cloud environments around the initial transcription process, by analogy. They had all the money they needed and now they're trying to monetize the 1st path. The Chinese models trained on those 1st paths are on the 2nd path - they need the 7 trillion token frontier model to get the gist, but it's not much more than a training library at this point. You might need to refresh it once on a while, but it's way to expensive to do it right without a consortium.
Enter: the untenable conflict between (a) open-access models, or the collecting and sharing of research over long periods of time; and (b) currently conceived and implemented models of (ego-self-serving goals of predatory/"zero-sum") capitalism.
It's not the end of competition, just the end of thinking that "tribal capitalism" is the only applicable and acceptable model for world-influencing content and events.
At some point and for many reasons, the model, to be viable and long-standing, has to have flex built into it, at the service of the principle of sublation--self-consciously drawing from the old what is good about it, and transforming the whole into a new way of doing things that better suits the ongoing context, with the goal of "everyone benefits."
The competition plays itself out, like sumo wrestling or the ideals of the World Cup and other international cooperative ventures--a respect for one's "enemy," a regard for intelligence and excellence and for oneself in terms of civil, rather than old-way tribal (kill or be killed) principles. People shake hands after someone wins.
If that seems pie-in-the-sky, then that's your problem. It's been going on in different facets of highly variable cultures for centuries.
Predictable and, indeed, predicted. If you believe that you need to spend a trillion dollars to create a capability, then there's a decent chance (say, more than 10%) that with the expenditure of say $100million, you can cut the cost of that capability by a factor of 2 every year for 10 years. (This can be and has been done, and no, I'm not talking about Moore's Law). So it's worth spending that $100million to save yourself $970 billion over 5 years, or $999 billion over 10 years with 10% probability: the expected profit is $96.9 billion over 5 years or $99.89 billion over 10. Why would anyone not do that? And it seems that they have.
If the issue would be only that China catches up it would not be as scary as it is if you realize that China can an will use it as a way to undermine the west.
The west used China as cheap labor for a while and no one likes to be someone's else bitch.
So, if I'd be China I'd push over the western investment in a very bad deal by creating and sustaining this false race for AI. What's extraordinary is that all the smarty pants in GenAI here seem not to understand this danger and engage in it as if they are in control.
At this point this is what Carl Sagan predicted a while ago. The question is l will we be able to defuse this investment bomb/bubble before takes us all down?
JEPA is only a partial solution. Yes, it strives to construct a non-trivial internal world model, which is required for proper understanding of (and then reasoning about) the physical universe, and which LLMs fundamentally lack. But it's still neural-net-based, which means that it will almost certainly be impossible to either align or validate to the degree required for an existentially-critical system such as human-level AGI (or above, e.g. ASI).
A culture of greed is all you need to bleed, bleed, bleed. I suppose the Trump regime will now turn into a Red Cross society with transfusion after transfusion, all at taxpayers expense.
The valley is very, very good at the pitch. That's kind of been its sorter since the late 90s.
Not so great at the make money side of things. Worse yet, the tech is too expensive so you can't do VC subsidized dumping to eat someone else's lunch through scale like taxi, delivery, or lodging.
AI is a nice fantasy / plagiarism machine with some limited use cases that are good (copypasta opensource code and pray, templating whatever docs, proofreading, natural language search / OS, and docreview. Not worth trillions in investment costs.
Gary — I think this is the right diagnosis wearing the wrong frame. "Brute force, unreliable, easily replicated" reads like an indictment of the paradigm, but it's really a description of what happens to *any* technology once the people who built it stop being the only ones who can build it. That's not unique to LLMs — it's every fast-follower story since Japanese steel. The "moat" was never the architecture. It was the six-month head start, and head starts close.
Where I'd push further than the price-war framing: the "trillion-dollar IPO" math didn't fail because the paradigm is bad. It failed because it was underwritten on the assumption that the gap would stay proprietary long enough to extract monopoly rents from it — the same assumption every currency peg and labor-cost advantage in the postwar export economies ran on, until the moment someone else learned to do the thing and the story flipped overnight from "comparative advantage" to "unfair competition." I wrote about that pattern at length here, if it's useful: [The Forgotten Theater](https://freeparadox.substack.com/p/the-forgotten-theater-dollar-hegemony)
Wright's point about non-zero-sum games is the real exit, I agree — but it only works if the U.S. stops treating "we lost the price war" as proof the technology was always worthless, instead of proof the monopoly pricing was always temporary.
a match made in charlatan heaven?
Gary, I thought you might be interested in how fusion has in some respects failed in the same way as AI. My father was a professor of plasma physics for 30 years, and his opinion was that, instead of focusing on the tokamak, we still needed much more science to find a better model. In other words, the fusion sector jumped on the first technology that had some kind of return (eg, the ITAR project in France).
The same thing is happening in AI. Instead of trying many different paths, they jumped on one that, while very flawed, yielded some results, and since there was so much money sloshing around and venture capitalists like to hear good stories, the entire industry is wasting trillions on the wrong approach.
ASI also shares the claims made about fusion which are repeated ad nauseam : “We know how to build it. It’s only X years off. We just need more money”
And now the “promise” of fusion and ASI are joined at the hip, of course. “ASI will solve fusion, which will feed ASI to infinity and beyond”
And the AI companies think they can use fusion to power their data centers! https://time.com/7328213/nuclear-fusion-energy-ai/. A technology that doesn't exist yet, to power a technology that has no long-term business plan. Great combination!
A difference is that fusion projects have continued to evolve, such that net positive energy (a technical measurement, not a commercial possibility, see https://lift.llnl.gov/) has occurred at the Livermore facility. With advances in superconductors, the Commonwealth Fusion project (https://cfs.energy/) is moving at a pace and with a design that is far different and probably ahead of the ITAR. It is unfortunate that the US government is not pushing the development of fusion at anywhere near the pace of China.
The Livermore facility concentrates on simulating nuclear weapons, and they use lasers focused on a nodule (my technical understanding is general). They have tended, historically, to ocassionally announce breakthroughs that weren't really applicable, just a warning. But thanks for the info. It's certainly a field to keep an eye on, it's just that I am very skeptical. It is the quintessential 'we are 15 (or pick your number) years away' technology
I'm not convinced that the comparison between fusion energy and LLMs is correct. The Tokamac approach being pursued by the international ITER collaboration is by no means unique, and the fundamental impediment- production of sufficiently strong magnetic fields- eluded researchers until the advent of superconducting magnets. I am not trying to minimize the difficulty of the task. The physics is basically understood. But as yet there is no consensus on what constitutes consciousness or6 thought, and moreover, unlike ML, for which error performance is characterizable, LLMs are fundamentally unpredictable.
I couldn’t agree with your closing thought more. I am excited about how Seed IQ’s active inference based AI agents acted as a control layer and created stability on IBM’s NISQ hardware. If Quantum leaps ahead we might not need all the data center compute.
Yep. One gigantic uber costly machine that can do everything certainly does not sound like the most efficient way to do anything.
“Triv-AI-al Pursuit”
Triv-AI-al pursuit
Is all the rage, it seems
It’s played by men in suits
And tech bros wearing jeans
It’s played by CEOs
And played by public too
But AI-mperor’s got no clothes
And privates are in view
Wish I could like this 100 times!
I think the pundits in the West simply haven't got it when it comes to understanding how China has integrated the 1st, 2nd and 3rd Industrial revolutions. Whereas Western capitalism simply jettisoned the 1st and 2nd industrial revolutions largely through financialisation/export of production, driven by short-term profit goals, the 'secret' to China's success has been to possess the entire spectrum of the production process, of which AI is the culmination. After all, it's ALL ABOUT ECONOMICS! What's more, automation isn't leading to unemployment in China (AI is automation by another name), far from it, AI when applied humanely, opens up the possibility of upskilling, of transforming the relationship between labour and production, this is what socialism is all about.. Thus the CCP just passed a law making it illegal for AI to create redundancies. Two different worlds.
“AI when applied humanely, opens up the possibility of upskilling, of transforming the relationship between labour and production”
- beautiful thought and an admirable goal.
These IPO's will be interesting. SpaceX was fortunate and lucky to be first though its value seems absurd and its AI component dubious. Being 2nd is better than being third and if Anthropic can sort out its spat with the USG, which seems to be happening they are positioned to be 2nd.
but
agree this energy and capital intensive "boil the ocean" approach to AI seems low ROI.
In the standard marketing model, there's only room for two dominant brands in a category (the leader and the alternative to the leader, along with a few less-profitable wannabes, epitomized by the old Hertz vs. Avis rental car market, where the second player had to 'try harder'). This is due primarily to the bandwidth of human recall, so it'll be interesting to see how many of the current AI competitors can hang on and achieve profitability as an actual market emerges, and for how long...
Agreed. Also network effects. The losers never reach the critical mass to get virtuous circles going
LLMs look increasingly like they're following the gene transcription model: once a couple of groups can replicate the initial basic transcription and the enthusiasm for all the new data has worn off, the real fun (and money) is in making transcription easier, faster, more reliable, more flexible and most important, much less expensive. But no one makes any money on the v1.0 of the initial transcription process. Maybe it's a PPP but you need government research dollars to find the 1st path.
The hyperscalers and the frontier labs seemed to think they could build walled gardens like their socmed ad businesses or cloud environments around the initial transcription process, by analogy. They had all the money they needed and now they're trying to monetize the 1st path. The Chinese models trained on those 1st paths are on the 2nd path - they need the 7 trillion token frontier model to get the gist, but it's not much more than a training library at this point. You might need to refresh it once on a while, but it's way to expensive to do it right without a consortium.
Enter: the untenable conflict between (a) open-access models, or the collecting and sharing of research over long periods of time; and (b) currently conceived and implemented models of (ego-self-serving goals of predatory/"zero-sum") capitalism.
It's not the end of competition, just the end of thinking that "tribal capitalism" is the only applicable and acceptable model for world-influencing content and events.
At some point and for many reasons, the model, to be viable and long-standing, has to have flex built into it, at the service of the principle of sublation--self-consciously drawing from the old what is good about it, and transforming the whole into a new way of doing things that better suits the ongoing context, with the goal of "everyone benefits."
The competition plays itself out, like sumo wrestling or the ideals of the World Cup and other international cooperative ventures--a respect for one's "enemy," a regard for intelligence and excellence and for oneself in terms of civil, rather than old-way tribal (kill or be killed) principles. People shake hands after someone wins.
If that seems pie-in-the-sky, then that's your problem. It's been going on in different facets of highly variable cultures for centuries.
Predictable and, indeed, predicted. If you believe that you need to spend a trillion dollars to create a capability, then there's a decent chance (say, more than 10%) that with the expenditure of say $100million, you can cut the cost of that capability by a factor of 2 every year for 10 years. (This can be and has been done, and no, I'm not talking about Moore's Law). So it's worth spending that $100million to save yourself $970 billion over 5 years, or $999 billion over 10 years with 10% probability: the expected profit is $96.9 billion over 5 years or $99.89 billion over 10. Why would anyone not do that? And it seems that they have.
If the issue would be only that China catches up it would not be as scary as it is if you realize that China can an will use it as a way to undermine the west.
The west used China as cheap labor for a while and no one likes to be someone's else bitch.
So, if I'd be China I'd push over the western investment in a very bad deal by creating and sustaining this false race for AI. What's extraordinary is that all the smarty pants in GenAI here seem not to understand this danger and engage in it as if they are in control.
At this point this is what Carl Sagan predicted a while ago. The question is l will we be able to defuse this investment bomb/bubble before takes us all down?
https://youtu.be/f9A_qFM9IAk?si=SoTJZ-sbgx0dBlew
Hi Gary, Yann LeCun raised a billion to work away from LLMs into JEPA. Could you please explain if this addresses your concerns?
JEPA is only a partial solution. Yes, it strives to construct a non-trivial internal world model, which is required for proper understanding of (and then reasoning about) the physical universe, and which LLMs fundamentally lack. But it's still neural-net-based, which means that it will almost certainly be impossible to either align or validate to the degree required for an existentially-critical system such as human-level AGI (or above, e.g. ASI).
Even the AI optimists are now calling for doom...
If you thought the global financial crisis was bad…
https://www.economist.com/by-invitation/2026/06/28/if-you-thought-the-global-financial-crisis-was-bad
From The Economist
A culture of greed is all you need to bleed, bleed, bleed. I suppose the Trump regime will now turn into a Red Cross society with transfusion after transfusion, all at taxpayers expense.
Cool demo, mid-product.
The valley is very, very good at the pitch. That's kind of been its sorter since the late 90s.
Not so great at the make money side of things. Worse yet, the tech is too expensive so you can't do VC subsidized dumping to eat someone else's lunch through scale like taxi, delivery, or lodging.
AI is a nice fantasy / plagiarism machine with some limited use cases that are good (copypasta opensource code and pray, templating whatever docs, proofreading, natural language search / OS, and docreview. Not worth trillions in investment costs.
Gary — I think this is the right diagnosis wearing the wrong frame. "Brute force, unreliable, easily replicated" reads like an indictment of the paradigm, but it's really a description of what happens to *any* technology once the people who built it stop being the only ones who can build it. That's not unique to LLMs — it's every fast-follower story since Japanese steel. The "moat" was never the architecture. It was the six-month head start, and head starts close.
Where I'd push further than the price-war framing: the "trillion-dollar IPO" math didn't fail because the paradigm is bad. It failed because it was underwritten on the assumption that the gap would stay proprietary long enough to extract monopoly rents from it — the same assumption every currency peg and labor-cost advantage in the postwar export economies ran on, until the moment someone else learned to do the thing and the story flipped overnight from "comparative advantage" to "unfair competition." I wrote about that pattern at length here, if it's useful: [The Forgotten Theater](https://freeparadox.substack.com/p/the-forgotten-theater-dollar-hegemony)
Wright's point about non-zero-sum games is the real exit, I agree — but it only works if the U.S. stops treating "we lost the price war" as proof the technology was always worthless, instead of proof the monopoly pricing was always temporary.