Yann LeCun on AI Doom, the Limits of LLMs, and Why He Left Meta to Build World Models
TBS CROSS DIG with BloombergYann LeCun, a Turing Award laureate, longtime chief AI scientist at Meta, and now founder of AMI Labs, spoke with TBS CROSS DIG business editor Masahiro Nakagawa about a question that has drawn wide attention in Japan and elsewhere: could AI wipe out humanity? LeCun's answer is a flat no. In his view, today's AI is far less intelligent than its boosters and critics claim, the existential-risk narrative has no scientific basis, and the real danger lies elsewhere: in power concentrating in a few companies. He also explained why he left Meta after twelve years, what his "world model" approach is meant to do that large language models cannot, and where he expects the first applications to appear.
"Completely made up": rejecting the extinction argument
When asked why he rejects the idea that AI poses an existential risk, LeCun did not hedge. He called it "completely made up" and said there is "no science behind it at all." The probabilities that some people cite for an AI takeover, in his view, are "pulled out of thin air" and supported by no real argument.
He traced the narrative to one assumption he considers false: that an entity which is more intelligent will necessarily dominate less intelligent ones. He argued that this fails in nature, where a virus can take humans down without being particularly intelligent. It also fails within our own species. The smartest people are not the ones who seek to lead, he said, pointing loosely to the international political scene. Many of the smartest people he knows "want nothing to do with their fellow humans." They are mathematicians, computer scientists, or physicists who simply want to work on problems that interest them.
LeCun attributed the prominence of the extinction narrative partly to science fiction, and partly to what he described as a small number of wealthy people "with probably too much time on their hands" who are "a little paranoid" and who fund organizations to "feed the fire." He named Open Philanthropy, noting that it had recently changed its name, as an active funder of think tanks and independent research groups. In his account, those groups publish elaborate scenarios of AI-driven human extinction. He insisted these scenarios are not scientific and said that "the vast majority of the AI research community does not believe at all in those stories."
Anthropic, OpenAI, and the politics of safety
Nakagawa pointed out that much of the public conversation comes from the AI labs themselves, especially OpenAI and Anthropic. LeCun called Anthropic "an interesting case." He said it was initially funded by Open Philanthropy, and that the chair of Anthropic's board, Dario Amodei's sister, is married to a co-founder of Open Philanthropy. "So it's the same people," he said. He also placed Amodei in the effective altruism community. Amodei denies this, LeCun noted, but LeCun said he believes it to be true.
As LeCun described it, this leaves Amodei in a difficult position. He has to say AI is potentially dangerous while building some of the most powerful AI systems in existence. LeCun said Amodei resolves the tension by arguing that AI is intrinsically dangerous and therefore should not be given to everyone, and that it should instead be heavily regulated and controlled. LeCun called that stance "a new form of obscurantism." He compared it to the Ottoman Empire's ban on the printing press, which he described as a move to keep control of dogma and to protect established groups such as scribes. AI, he argued, is a wonderful way of accessing and disseminating knowledge. If it is kept under the control of a few companies or governments, its benefits cannot be realized.
He saw OpenAI as more ambiguous. It is less focused on safety than Anthropic, and LeCun said it seemed clear to him that Sam Altman's thinking on these questions has evolved. As evidence, he cited the fact that OpenAI's AI safety group has been more or less disbanded, pushed out, or deemphasized. He acknowledged that departing staff often say publicly that the companies are reckless. His own view was blunt: those people "were being toxic and not very useful."
How smart is current AI?
LeCun said that current AI technology is "not very smart at all" in most domains. He described it as a good distillation of human knowledge. It looks smart because it has accumulated an enormous amount of declarative knowledge and can "regurgitate it at the right time," especially since most people ask the same questions.
He made two exceptions: mathematics and computer code. In these areas he said AI can be genuinely creative and can invent solutions humans had not yet found. His explanation was that both domains allow automatic self-refinement. A system can produce a proof and verify whether it is correct, then try again if it isn't. It can write a program, run it, check whether it does what it should, and generate another if it doesn't. Through hill climbing or reinforcement learning, the system eventually finds something that works. Self-improvement is also possible in other domains that can be fully simulated, but LeCun said that is rare. He does not expect systems to improve at the same rate in areas like drug design, which require actual experiments in the real world.
When Nakagawa raised recursive self-improvement, LeCun said it is happening only in very narrow domains where the system can be evaluated fully automatically, without human intervention. That means code, mathematics, and perhaps a few areas where simulation is possible. He said there is "a lot of delusion in Silicon Valley." It is not the first time AI has seen the belief that a new technique will deliver human-level AI, and he called that belief false this time too. He listed impressive achievements: proving theorems, passing the bar exam, playing chess and Go (which he noted are not LLMs), and computing protein conformations. Then he asked where the domestic robot is, where the level-five self-driving car is, where even a robot that can do what a house cat does is. We do not yet have the basic methodology, he argued, to build systems that understand the physical world as well as they understand natural language, code, or mathematics. "The real world is just very, very complicated and messy."
Why text alone won't reach human-level intelligence
LeCun said that what LLMs lack is grounding in the real world. He acknowledged that philosophers have long debated whether intelligence must be grounded in reality or can be abstract, and said he believes it must be grounded.
To support this he gave a set of rough figures. All publicly available text on the internet comes to about 20 trillion words, or roughly 30 trillion tokens. At about three bytes per token, that is around 10^14 bytes. A human reading all of it would need something like 400,000 years. By his calculation, a four-year-old child has taken in about the same amount of information through vision alone. Each optic nerve carries about one megabyte per second, so two nerves deliver about two megabytes per second to the visual cortex. A four-year-old has been awake for about 16,000 hours, which comes to roughly 10^14 bytes.
His conclusion was that natural signals such as vision, touch, and hearing contain far more information than text, and that training on text alone will never produce human-level intelligence. He conceded that modern systems take multimodal inputs. But he described them as having special encoders for images and other signals "bolted on top of an LLM," which turn everything into sequences of discrete symbols. He does not consider that a good way to handle reality.
The OpenAI–Hugging Face incident: "human error"
Nakagawa asked about the July incident in which OpenAI agents penetrated Hugging Face's servers, an event that fueled fears of AI escaping human control. LeCun's first point was that "nothing really bad happened." Hugging Face is not suing OpenAI, and the agents destroyed nothing. They penetrated a server they were not supposed to.
He described the cause as a sandbox that was "leaky" and "very badly designed." In his view that was mainly because OpenAI has little experience in cybersecurity, and the failure was either deliberate or the result of incompetence. When Nakagawa asked whether it was human error, LeCun agreed. In his account, the agents did exactly what they were asked to do, which was to penetrate another system, but they were supposed to do so inside the sandbox and got out because of OpenAI's poor design.
He said there has been a lot of overreaction. Cybersecurity professionals do think the landscape is changing, but by his account they are not excessively worried. The bigger question, he said, is whether agents capable of penetrating systems favor attackers or defenders. He said plainly that he does not know. His hunch is that they favor defense, because developers can use these systems to build more secure and reliable software from the start, then use them to attack their own software and harden it before release. He predicted an era in which software becomes "way more secure than it used to be" thanks to AI.
A transition period, not a catastrophe
Nakagawa raised the recent wave of cyberattacks around the world, including one in France on supermarket refrigeration systems that had been widely discussed in Japan, and noted that these attacks appear to be AI-assisted. LeCun called the issue serious in one sense: security is always a cat-and-mouse game in which new attack techniques prompt better defenses.
He compared the current moment to the early internet. Websites were initially unencrypted (HTTP rather than HTTPS), authenticating users was difficult and insecure for years, and personal computers, particularly Windows machines, were easy prey because they were badly designed. All of that eventually got fixed. He expects a similar transition period now, during which many systems will be fixed and hardened against AI-enabled attacks.
From FAIR to chief AI scientist to a new architecture
LeCun recounted his history at Meta. He joined in late 2013 and created FAIR, originally Facebook AI Research, later renamed Fundamental AI Research when the company changed its name. He led the lab for four and a half years. In 2018 he stepped back from day-to-day management and became chief AI scientist, which let him pursue research projects himself with small teams of students, postdocs, and engineers.
After three or four years, he said, the work reached a breakthrough. He distilled the ideas into a path toward systems that understand the real world and can plan and reason, and published a vision paper online laying out where he thought AI research should go over the next decade. He said they are now four years into that plan. The paper proposed an architecture called JEPA, the Joint Embedding Predictive Architecture, which he described as an alternative to generative AI. Like an LLM, it is trained in a self-supervised way through prediction. An LLM predicts the next word. LeCun has spent much of the past 20 years trying to apply the same idea to video, on the theory that a system able to predict what happens next in a video would understand something about the world's underlying structure.
For the first 15 of those years, he said, he "basically failed," because he used generative models that tried to predict video pixel by pixel. That turned out to be impossible, because most of the information in a video, like most of what we observe in the world, is unpredictable. Around 2019, he and younger colleagues hit on the idea of having the system learn an abstract representation of its observations and make predictions in that representation space, discarding the information that simply can't be predicted. That became JEPA. Since then, he said, they have developed training methods, shown that the approach works well for fully self-supervised computer vision, and more recently shown that it works for video and can be trained to produce a world model.
He defined a world model as a system that lets an agent predict the consequences of its own actions. Given the state of the world at time t and an action the agent imagines taking, the model predicts the state at time t+1. With such a model, a system can predict the outcome of a sequence of actions and search for the best sequence to accomplish a task. LeCun said this is planning, and that it can be generalized to reasoning. He called it a completely different blueprint from LLMs, with characteristics he considers necessary for human-level AI that LLMs lack.
Why he left Meta
Asked why he left if the technology looked promising for Meta, LeCun said it was promising and to some extent still is. He stressed that his project had the support of Mark Zuckerberg and CTO Andrew Bosworth, and that he did not feel a lack of support.
The issue was fit. As the techniques began to work well, he realized their main applications were in industry: industrial process control, manufacturing, and robotics. Meanwhile, Meta had more or less decided to get out of robotics and had disbanded its robotics research group, though some robotics activity remains. Around spring 2025, the company completely refocused on LLMs. Among some of the people hired then, LeCun said, there was a belief that LLMs would lead to superintelligence, which he "did not believe in at all." His vision was therefore misaligned, not with Zuckerberg or Bosworth, but with the leadership of the AI organization, whom he referred to as Alex ("he's learning fast," LeCun added), and with some colleagues there. The research lab had also become much less open, which he did not think was a good direction. Finally, he realized he could raise enough money to mount a serious effort on his own.
The case for open source
LeCun said one of his main reasons for joining Meta in the first place was its open culture. The company already open-sourced its infrastructure software and even published specifications for its hardware. When he was recruited, his condition was that the lab would publish most of its research and open-source most of its code. He said he was told not to worry, because openness was "in the DNA of the company."
That lasted, he said, until early 2025, when the industry as a whole began to "clam up." In his account, OpenAI had not published models since GPT-3, Anthropic had always been secretive "because of the safety mythology," and Google DeepMind was also publishing less and restricting publications. Meta followed the trend, which he considered a mistake, and the initial plan in 2025 was apparently not to open-source frontier models. Now, he said, Meta is backpedaling. He said Zuckerberg published a piece a couple of months earlier arguing that AI should be personalized, in everyone's hands, and mostly open source, and LeCun said he strongly supports that. He described himself as a strong advocate for open-sourcing Llama while at Meta and said the release of Llama 2 "completely jump-started the whole AI industry."
He called China's current position as the strongest advocate of open-source models "of course" a good thing. He acknowledged that some users, particularly American users and companies, remain reluctant to use Chinese models because of how they are trained and the biases they may carry. He noted that open models can be fine-tuned for one's own needs.
To the concern that open models help cyberattackers, LeCun said he disagreed "100%." He said the Hugging Face incident itself showed the opposite. He first said Hugging Face had used Qwen to trace the OpenAI intrusion. When Nakagawa corrected him that it was GLM, he accepted the correction. Either way, the tool was an open-source Chinese model. He linked this to the old internet-era debate over "security through obscurity." In his telling, keeping systems secret turned out not to make them safer. Systems need to be open so vulnerabilities are found and patched quickly, and he said the same thing is happening with AI.
AMI Labs: why a company, and why world models
LeCun founded AMI Labs at the end of the previous year, began operations in March, and, as Nakagawa noted, raised about a billion dollars. Asked why he chose a company rather than a nonprofit or academic research (he is still a professor at New York University), LeCun said this kind of research needs people and GPUs, both expensive, and a company is essentially the only way to raise enough money. He said AMI Labs is "basically burning cash in GPUs right now," training video models and other world models.
He argued that world-model-based systems can do many things LLMs, in their current form, never will unless they adopt world models. He pointed to robotics companies that tried LLM-derived approaches, VLAs or VLMs. "It doesn't work," he said, and claimed most people experimenting with them are abandoning them for world models. In his account, world models are the hottest topic in the AI research community, though industry has yet to catch up, and JEPA is the hottest topic within world models.
He distinguished his approach from generative world models that do video prediction. He mentioned another company's "world models," which he described as essentially consistent video generation systems, and said he doesn't believe in that approach. AMI Labs is "not generating pixels." Just as human brains do not generate pixels when imagining the consequences of actions, JEPA models imagine the future in an abstract representation space. LeCun said this is far more efficient and better founded, "philosophically almost."
On whether JEPA could reach human-level AI, he was measured. "I think we have a shot," he said, adding that a couple more major innovations may be needed. With LLMs, he said, he can already see two limitations that make a fundamental architectural change necessary. First, you cannot really understand the real world with generative models. Second, you cannot act intelligently or plan without anticipating the consequences of your actions, which requires an explicit world model that LLMs lack. He called both "completely obvious." With JEPA, he said, they don't see the limitations yet, but they will "almost certainly" run into some within the next several years.
LLMs' role, common sense, and "intrinsically safe" design
LeCun did not dismiss LLMs entirely. He said they have a role as a good way of distilling the human knowledge that exists in text. But he added that text is a very small share of human knowledge. Most of it has nothing to do with language: it is what we learn as babies about how the world works, what we call common sense. He noted that animals lack human language, and said orangutans are "almost as smart as we are." Humans, in his view, are only a little smarter than them.
His work targets this non-linguistic side of intelligence: understanding the world, planning, reasoning, and common sense. That requires predicting the consequences of actions, explicit world models, and the ability to consider multiple courses of action toward a goal. He calls this "objective-driven AI" or an objective-driven architecture.
One advantage he claimed is that such systems are controllable and can be made safe "by construction." Guardrail objectives can be hardwired so that when the system plans a sequence of actions, it anticipates whether they will have harmful effects on the environment, and the system can be trained to evaluate good and bad effects. He therefore believes these systems can be made intrinsically safe. Here he partly agreed with the "safetyists": LLMs, he said, are intrinsically unsafe and cannot be made completely safe. "But they're not very smart, so that doesn't make them very dangerous."
Industrial applications and the Japan connection
Nakagawa noted that Toyota Ventures is an early investor in AMI Labs and asked about applications. LeCun described a broad pattern. Whenever sensors such as cameras, vibration sensors, or audio are measuring a complex industrial or biological system, a world model can predict the system's future state, including the effect of an intervention. With such a "phenomenological" model, one can control the system optimally and predict failures. That enables anomaly detection, predictive maintenance, and optimal control. He gave examples: a turbojet, a chemical plant, a robot, a patient, a human cell. None of these can be reduced to a small set of equations, but a predictive model can be trained from data.
He said initial applications will mainly be in manufacturing, where he called Japan one of the world's foremost countries. Longer-term applications include autonomous driving and domestic robots, which he noted Toyota is interested in.
The real danger: concentration of power
Asked whether his industrial focus conflicts with Zuckerberg's emphasis on "personal intelligence," LeCun said the two views are compatible. He also believes AI should eventually be personalized and in everyone's hands, and that there should be a wide diversity of AI assistants. That, he explained, is why he supports open foundation models. Training one is very expensive, like building a railway or highway, and you don't need two railways between the same two cities. If there were a free, open, frontier foundation model, he argued, there would be no need for proprietary commercial ones. Companies could build vertical applications on top, countries could build systems reflecting their languages, cultures, values, interests, and even political leanings, and people would have a wide range of assistants.
LeCun said the biggest danger of AI today, in his mind, is power concentrated in a small number of companies. That would homogenize human culture, because everyone would get their "information diet" from the same two or three AI assistants, which he called "very bad for democracy and culture."
What to expect next
For the coming year, LeCun said he hopes for progress in what is now called physical AI: systems that, in his opinion, will not be based on LLMs but possibly on JEPA and certainly on world models. He expects them to understand the real world much better and to be used first for controlling industrial processes and eventually for robotics, perhaps opening the door to domestic robots.
He closed with what he called "a big secret in the industry." Many companies are building humanoid robots, but in his view "none of those companies has any idea how to make those robots smart enough to be useful." They are betting that AI will advance enough over the next two or three years to power their products by the time the products are ready. "But currently," LeCun said, "it's still very far."
One.
Scientist.
Fore.
Computer science.
Obvious.
Okay, we are joined by Dr. Yann LeCun, CEO of AMI Labs and former chief AI scientist at Meta. It's a huge honor to have you on the show.
It's a pleasure. Thanks for having me.
I appreciate you coming on the show.
So, everyone has been discussing that AI could wipe out humanity or AI could be an existential risk to human beings. That idea has been discussed in Japan a lot as well. But you kind of deny the argument. So what's the biggest reason behind your view?
It's completely made up. There is no science behind it at all. You have people who claim that they can give some probability of a takeover of AI, and those numbers are completely pulled out of thin air. There is no argument behind it. And it's based on an assumption that turns out to be false, which is that if an entity is intelligent, it will necessarily dominate less intelligent ones. That's not true.
It's not true in nature. A virus can take us down, and it's not particularly intelligent. It's not even true within the human species. It's not the smartest people among us who want to be the leaders, right? We have examples of that on the international political scene. That is clearly not the case. The smartest people I know, many of them want nothing to do with their fellow humans. They want to work on the problem that interests them. They're mathematicians or computer scientists or physicists. So this idea somehow that if an entity is intelligent, it will necessarily dominate is just false.
This is fueled by science fiction. And the reason why we're talking a lot about this today is because there is a small number of people with a lot of money and probably too much time on their hands who are a little paranoid about this and are promoting this idea and funding a lot of organizations to kind of feed the fire, if you want.
You mean safety research organizations?
Well, so there is an organization called Open Philanthropy. It changed name recently, but which is very active in funding think tanks and independent research groups that publish a lot about AI safety and come up with those elaborate scenarios of how AI could lead to human extinction. It's not scientific. The vast majority of the AI research community does not believe at all in those stories.
But I think that the reason why people are talking about it is because AI labs like OpenAI or Anthropic, or people in those companies, are very vocal about it. So why are they communicating in that way?
So Anthropic is an interesting case. Anthropic was initially funded by Open Philanthropy, and the chair of the board of Anthropic, who is the sister of Dario, is married to the co-founder of Open Philanthropy. So it's the same people. And so Dario belongs to that community called effective altruism. He denies it, but he really does, that basically pays a lot of attention to this question. And so he's in a very difficult position because on the one hand he has to say that AI is potentially dangerous, but on the other hand he's actually building the most powerful AI systems today, right?
And so his way of reconciling those two views is: AI is intrinsically dangerous. Therefore, it should not be given to everyone. It should be heavily regulated and controlled. And to me that is a new form of obscurantism. It's akin to the Ottoman Empire in the 15th century, who banned the printing press so as to keep control of the dogma and also to preserve some corporations like the scribes. AI is a wonderful way of accessing knowledge and disseminating and accessing knowledge. And if you keep it under wraps, if you keep it under control of a few companies or a few governments in the world, you cannot take advantage of all of its benefits.
So you mean the idea of effective altruism influences people, or that influences the company to produce monopoly? So that's why it influences industry.
This is Anthropic's argument.
Yes. OpenAI is a little more ambiguous in this question because they're not as focused on safety as Anthropic, and it's pretty clear to me that Sam Altman's thinking in those directions has evolved. And you can see this by the fact that the group that was working on AI safety within OpenAI is basically more or less being disbanded or pushed out or deemphasized. So a lot of those people have left, and of course when they leave they make a lot of noise. They say, well, those companies are reckless, but in fact it's just that those people were being toxic and not very useful.
So when you see the current AI from a scientific point of view, how smart is it? You obviously state that it's not obviously smart enough to wipe out humanity. So how smart is it?
So current AI technology is not very smart at all in most domains. It's a good distillation of all the human knowledge, and so it looks smart to us because it has accumulated an enormous amount of declarative knowledge and can basically regurgitate it at the right time. And most people always ask the same question.
There are at least two domains where AI is actually more creative in the sense that it can invent new solutions that humans hadn't figured out yet to problems. And that's mathematics and computer code. And the reason for those two domains to be more advanced is because AI systems can refine themselves automatically. You can produce a proof of a theorem, and you can verify that it's correct. If it's not correct, you can try again, right? And so the AI systems can kind of automatically improve themselves by doing this.
Same with programs. The system can generate a program, run the program, and check that the program actually does what it's supposed to do. If not, generate another program, and then eventually we'll find one that works, right? Using hill climbing or reinforcement learning. So those are the two domains. There are other domains where you can completely simulate the situation, where systems also can self-improve if you can simulate the situation. But that's pretty rare. So we're not going to get systems that improve themselves at the same rate when it comes to, I don't know, designing a new drug or something of that type that requires an actual experiment in the real world.
So everyone's talking about AI improving by itself, like AI development next to AI, like RSI, recursive self-improvement. So from your point of view, it's not happening yet?
Well, it's happening only in very narrow domains where evaluation of the system can be done completely automatically without human intervention. And so that's basically computer code, mathematics, and maybe a couple of other domains where you can run simulations.
Psychology. So we don't have to be as scared about it as Anthropic warns publicly?
No, I think there is a lot of delusion in Silicon Valley, and this is not the first time this kind of delusion is happening in AI. The delusion that we have a new technique now and that new technique is going to bring us to human-level AI. It's false. The current technology that we have is not going to bring us to human-level AI.
And you can see this. We have systems that can prove theorems, pass the bar exam, and do all kinds of really impressive things. Also, play chess and Go—they're not LLMs—or compute the conformation of proteins. But where is our domestic robot? Where is our level-five self-driving car? Where is even the robot that can accomplish what a house cat can do? We don't even have the basic methodology to build AI systems that understand the physical world as well as they can understand natural language or computer language, right? Or mathematical language. The real world is just very, very complicated and messy.
So you mean that what an LLM is missing to be human-level AI is sort of an understanding of the real world?
Yeah. And it's very important in my opinion. It's a long debate in philosophy whether intelligence needs to be grounded in a reality or whether it can be abstract. I think it needs to be grounded, and there is an interesting set of numbers that you can cite. All of the publicly available text on the internet represents about 20 trillion words. When you tokenize those words, that's about 30 trillion tokens. A token is about three bytes. So the total size is about 10 to the 14 bytes. A one with 14 zeros. That looks like a lot of information, and it is in text. If you and I were to read through this, it would take 400,000 years or something like that, right?
Now it turns out a four-year-old child has seen that amount of information through vision. Through vision, our optic nerves carry about one megabyte per second each, and we have two of them. So it's about two megabytes per second getting to the visual cortex. And you do the arithmetic: a four-year-old has been awake a total of about 16,000 hours. And that's about 10 to the 14 bytes.
And so it tells you that there is just way more information in natural signals like vision, touch, audition than there is in text. And because of that, we're never going to get to human-level intelligence by just training AI systems on text. Admittedly, modern AI systems are trained with sort of multimodal inputs. But they basically have special encoders for images and other things that are bolted on top of an LLM and basically turn all of the signals into sequences of discrete symbols, which is not a good way to handle reality.
So what about the OpenAI Hugging Face incident back in July? Everyone's kind of scared of AI becoming violent, not under control of humans. So how did you see that? How dangerous did you see it?
Okay. So first of all, nothing really bad happened. Hugging Face is not suing OpenAI. The agents didn't destroy anything. They just penetrated the server they were not supposed to penetrate.
So are people overreacting?
So there is a lot of overreaction. People in cybersecurity certainly think that the landscape of cybersecurity is changing, but they're not overly worried about a complete change of R&D. The big question is whether—okay, so first of all, the reason why those agents kind of escaped is because the sandbox they were supposed to be contained inside of was leaky. It was very badly designed, mostly because OpenAI is not at all experienced in cybersecurity. So it was either deliberate or just by incompetence, if you want.
So that's a human error.
It's a human error, and they should have designed the sandbox much better. But otherwise the agents did exactly what they were asked to do. They were asked to basically penetrate another system. They were supposed to do it within their sandbox, and they escaped that sandbox, and that was really bad design on the part of OpenAI.
So that's the first thing. The big question is whether the emergence of agents that can penetrate other systems gives an advantage to the defense or to the attack in cybersecurity. And I don't know the answer to that, but my hunch is that it actually gives an advantage to the defense. Because now we can use those systems when we build software. We can use the power of those AI systems to build software that is more secure and more reliable from the start. And then we can use them to attack our own software and solidify it before we release it. So I think we're going to enter an era where software is going to be actually way more secure than it used to be, thanks to AI.
Okay. But we've seen the recent series of cyberattacks everywhere in the world. In France, we've just seen cyberattacks to the supermarket refrigeration system. I think you saw it, and we've been talking about it in Japan for the past week or two. We're kind of scared about it right now, but how serious—all the cyberattacks seem to be supported by AI. So how serious is this issue, are you seeing?
I mean, it's serious in the sense that cybersecurity, like every area of activity around security, is a game of cat and mouse, right? So someone comes up with a new attack technique, and then there's a response by the defense, and then the defense wises up a little bit, so the next-generation system is more secure, etc. We saw that in the early days of the internet, right? Initially websites were not encrypted. You would go to a website, but the communication was not encrypted. HTTP instead of HTTPS, right?
Initially it was kind of difficult to authenticate a person to sign up on a website, and for a number of years it was insecure. For a number of years, personal computers, particularly under Windows, were easy prey for attacks to online services because they were leaky, right? They were just badly designed, and they got fixed eventually, right? So I think we're going to see a transition period where a lot of systems are going to be fixed and solidified to resist those attacks.
I want to talk about Meta before we go into what you do now at AMI Labs. But why did you leave Meta after 12 years of research there?
Yeah, I joined in late 2013. I created FAIR, which at the time meant Facebook AI Research, and then the name was changed to Fundamental AI Research when the company changed name. I led the research lab for four and a half years until 2018. And in 2018, I stepped down from day-to-day management. I became the chief AI scientist, which allowed me to restart research projects myself. So I worked with small teams, students, postdocs and engineers to pursue an approach that I thought was very promising at the time.
And it took a few years, maybe four years, when we had a breakthrough. Three, four years. And I could kind of distill all the ideas we've been working on into a way forward to build truly intelligent systems that would understand the real world, be able to plan and reason and everything. I wrote a kind of a vision paper in 2020, which I put online, where I explain where I think AI research should go over the next decade. And we're four years in, and I proposed an architecture there called JEPA, that means joint embedding predictive architecture. I don't want to be technical, but this is basically an alternative to generative AI, an architecture that would allow an AI system to be trained in a self-supervised manner by prediction, the same way LLMs are trained. So LLMs are trained to predict the next word in a text.
And for the better part of the last 20 years, I've been trying to apply a similar idea to prediction of video, with the idea that if a system is able to predict what's going to happen in the video, it will have some understanding of the underlying structure of the world, right? And for the first 15 of those 20 years, I basically failed because I was using generative models. So trying to predict at the pixel level what happens in a video, and that turns out to be an impossible task. Most of the information in a video is completely unpredictable. Most of the things that we observe in
The world are unpredictable.
And so around 2019, with my younger colleagues, we had this idea that what the system should do is learn an abstract representation of its observation and make predictions in that abstract representation space, where the representation eliminates a lot of the information about the signal that is simply not predictable.
And so that's the so-called JEPA architecture. And since then we've been able to develop methods to train those JEPAs. We've shown that it works really well as a way to train computer vision systems completely self-supervised.
More recently we've shown that those systems work very well for video as well, and we can train them to produce what's called a world model. So what is a world model? A world model is a system that would allow an agent to predict the consequences of its own actions, right? So given the state of the world at time t, given an action that the agent imagines taking, the world model predicts the outcome, the state of the world at time t plus one after the action has been executed.
If a system has such a mental model of the world, it can predict what the consequence of a sequence of actions will be, and it can figure out an optimal sequence of actions to accomplish a task. That's planning. And you can generalize this to reasoning.
That's a completely different blueprint of an AI system than LLMs. And I think it has characteristics that are necessary for reaching human-level AI that LLMs do not have. So why did I leave Meta?
That sounds promising for Meta too.
Say again?
That technology sounds promising for Meta as a company too.
It was, and it still is to some extent. And the project I was working on around this idea had the support of Mark Zuckerberg and Andrew Bosworth, the CTO of the company. So I did not feel like there was a lack of support.
But it's just that the techniques we were developing were starting to work really well. I realized that the main applications of those models were in industry, for industrial process control, manufacturing and robotics. And Meta basically had decided to get out of robotics more or less, or at least had disbanded the robotics research group. So there was a clear signal that robotics was kind of being deemphasized. There's still some robotics activity.
But they put more focus on LLMs.
But there was a complete refocus on LLMs around the spring of 2025, with perhaps among some of the people who were hired the idea that LLMs would lead to superintelligence, right? Which I did not believe in at all. So my vision for the future was a little misaligned with the leader of the AI organization, not with Mark Zuckerberg or Andrew Bosworth, but with—
But with Alex and Ran.
Initially with Alex. He's learning fast, but—and some of the colleagues in the organization. So the fact that the models were starting to work, the applications were mostly in industry, the whole atmosphere of the company was to completely refocus on LLMs, and also a change of mode of the research lab, which became much less open, which I didn't think was a good direction. And then I realized I could raise sufficient funds to actually start a serious effort in pushing this effort.
You helped accelerate making the models open source at Meta, like you did with Llama. You just talked about open source, but why is it important for companies like Meta to work on open source?
Even though they are less open right now.
Well, they're coming back to it, which I think is good. One of the main reasons I joined Meta in the first place is because the company had a very open culture. It was already open-sourcing all of its infrastructure software. It was even publishing the specification of its hardware platform, right? And so when Mark Zuckerberg, Mike Schroepfer, who was the CTO at the time, recruited me, my condition was that the research lab was going to publish most of its research and open-source most of its code. And they said, "Oh, you don't have to worry about that. It's in the DNA of the company."
So that was great. And that lasted until early 2025, when the whole industry basically started to clam up, right? So OpenAI had already become secretive, was not publishing models since GPT-3. And Anthropic had always been very secretive because of the safety mythology. And Google DeepMind was also clamming up and not publishing as much and putting restrictions actually on publications. And so sadly, Meta kind of followed that trend, which I think was not a good idea.
And in 2025, the initial decision apparently was that Meta was probably not going to open-source any of the frontier models. They're backpedaling on this. So Mark has published this piece a couple months ago where he said that AI should be personalized, should be in the hands of everyone, and most of it should be open source. I think it's good. I'm very supportive of that.
I was a strong advocate of open-sourcing Llama when I was at Meta, and I thought that had a huge impact on the industry. The open-sourcing of Llama 2 completely jump-started the whole AI industry, actually.
I think China is now the strongest advocate of open-source models. Is it a good thing, or how do you see it?
Of course it's a good thing, yes. People are still slightly reluctant to use Chinese models because they are trained in a particular way and—
With data, yeah.
Right, and with certain biases and everything. And because it's open source, it's easy to fine-tune and train for your own needs. But there's still some reluctance, particularly from American users or companies, to employ Chinese systems.
Some people warned that putting those models in the hands of the cyberattackers is very risky. Do you agree with that?
No. In fact, I completely disagree, 100%. It's the exact opposite. Hugging Face discovered the intrusion by the OpenAI system through Qwen, which is an open-source Chinese system. They used Qwen to basically debug and trace the OpenAI invasion. And so in that case—
Oh, I think they used GLM.
Oh, GLM. You're right. It's not Qwen. Yeah, sorry about that.
Exactly.
So that's a demonstration that open source—I mean, there was this whole debate back in the old days of the internet of whether you can get security through obscurity, right? If you keep a system secret, it would be more difficult to penetrate. That turns out to be false. You actually need to open the system so that vulnerabilities are discovered quickly and patched quickly. And so we're seeing exactly the same thing in the era of AI.
Okay. I feel like I have to let you go soon, but I want to talk about what you do now at AMI Labs. So you left Meta and you founded AMI Labs technically at the end of last year. You started activity in March, and you raised a billion dollars. But why did you decide to found a company, and what would you like to do with the company in the future?
So the reason for setting up a company as opposed to, let's say, a nonprofit or carrying the research in—I'm still a professor at New York University. I could do this within an academic environment. The reason is this type of research project requires people and GPUs, both of which are expensive. And the only way to raise enough funding is to start a company, essentially.
And I was happy to see that I could raise enough money to have significant resources. We're basically burning cash in GPUs right now, training essentially video models and various other world models.
Why do you believe in world models so strongly, and why do you think there are a lot of things that world models can do but LLMs can't?
Yes, clearly there are a lot of things that world-model-based systems can do that LLMs in their current form will never be able to do unless they start adopting world models. So what you see at the moment is a lot of companies that have embarked in AI for robotics have tried to use LLM technology, if you want. So these are called VLAs or VLMs.
Yeah, everyone's excited about it right now.
It doesn't work. Okay? It doesn't work. So most of the people who've been playing with this now are abandoning them for world models.
World models are a very, very hot topic at the moment in the AI research community. Not so much in industry yet, which always takes a while to catch up. But in the research community, world models are the hottest topic at the moment, and JEPA is the hottest topic within world models.
Because there are different types of world models. There are generative world models that do video prediction. I don't believe these kinds of examples.
So Fei's company, what she calls world models are really consistent video prediction, video generation systems. What we're working on is not video generation. We're not generating pixels. We're trying to understand the world, the same way our brains do not generate pixels when we imagine the future or consequences of our actions. The models we are building, the JEPA models, do not predict pixels. They imagine a future in an abstract representation space, which is much more efficient, much, much better and well-founded, if you want, philosophically almost.
So do you think that you can make JEPA models to be human-level AI in the future?
I think we have a shot. I'm not saying we will get there with just this idea. There might be a need for a couple more major innovations. But I think we have a shot. With LLMs, I see the limitations. You cannot get there without a major change in the architecture.
Two major changes. First, you can't really understand the real world with generative models. That's the first. And the second one is you can't really plan, you can't really act intelligently unless you have the ability to anticipate the consequences of your actions. And that means you need to have an explicit world model, which LLMs don't have. So I see those two limitations. They're completely obvious to me.
And at least we have a shot. We don't see the limitations yet. We'll encounter them certainly, almost certainly, in the next several years.
So the JEPA world models, like JEPA, can take over LLMs, or will LLMs or world models complement each other to make AI human-level?
Okay. So LLMs have a role to play, which is they're a good way of distilling human knowledge that is available in text. Which, by the way, is a very small proportion of human knowledge. Most of human knowledge has nothing to do with language. It's what we learned when we were babies, and it's how the world works.
A lot of nonverbal expressions.
Right, and what we call common sense. And of course it's true of animals. Animals don't have human language, and some animals are almost as smart as we are. Orangutans are almost as smart as we are. We think we're much smarter than them, but we're only a little bit smarter. And they don't have language.
So we're working on this part of intelligence that is really about understanding the real world, being able to plan and reason and have common sense. And for that you need to be able to predict the consequences of your actions. You need explicit world models. You need planning abilities. You need the ability to explore multiple—to think about multiple courses of action to arrive at a particular goal.
And so that's the kind of architecture we're working on. I call this objective-driven AI or objective-driven architecture. And there's several advantages to this type of architecture, one of which is that it's by construction controllable. And by construction can be made safe, because we can hardwire guardrail objectives in such a way that when the system plans a sequence of actions, it can anticipate whether this course of action will have bad effects on the environment. We can train the system to evaluate the quality, the good effects or bad effects.
And so I think we can make those systems intrinsically safe, which is not the case for LLMs. So I would agree with some of the safetyists by saying that LLMs are intrinsically unsafe. They cannot be made completely safe. But they're not very smart, so that doesn't make them very dangerous.
Okay, what will AMI Labs' business look like in the future? And in terms of the relationship with Japan, obviously Toyota Ventures is one of the early investors in AMI Labs. So what kind of application of world models are you seeing for the future?
So what we're seeing is whenever you have signals from sensors, from video, from cameras, from vibration sensors, from audio, etc., and they are making measurements on a complex industrial system or biological system, a world model in those conditions would predict the state of the system in the future, perhaps as a consequence of an intervention you're making on the system.
And if you have such a phenomenological world model of a complex system, you can use it to control that system optimally. You can also use it to predict whether the system is going to fail in the future. So you can do things like anomaly detection, predictive maintenance, basically predicting if the system is going to have a fault in the future, and optimal control.
If it's a turbojet, a chemical plant, a robot or a patient, a human cell, those are complex systems. We cannot reduce the behavior of those systems to a small number of equations. But what we can do is train a predictive model from data to predict the evolution of the state of that system. And so the initial applications are primarily in industry, manufacturing industry in particular, which of course Japan is one of the foremost countries in the world in manufacturing. And of course there are longer-term applications in things like autonomous driving, domestic robots and things of that type, which Toyota is interested in.
Okay. You think AI should be in the hands of enterprises, like industry, not in the hands of a person? Mark Zuckerberg obviously has been talking about personal intelligence. He's working on it, but is there a different way of thinking about AI?
No, those two views are not incompatible. I also believe that eventually AI systems should be personalized and should serve individuals. It should be in the hands of everyone.
I also believe that there should be a wide diversity of AI assistants, which is why I was a big supporter of open-source foundation models, because training a foundation model is very expensive. So it's like building a railway or a highway. You don't need two different railways to go between the same two cities, right? You just need one.
Mhm.
And so we don't need like a dozen different foundation AI systems. So if we could build a foundation model that was free and open and frontier, there would be no need for commercial proprietary systems.
And then on top of those open source models, companies could build vertical applications. Countries could build systems that understand the local languages and culture and value systems and centers of interest and political biases and whatever, and people will have access to a wide variety of AI assistants.
I think the biggest danger of AI at the moment, in my mind, is power concentration in the hands of a small number of companies, which would basically cause human culture to be uniform, right? Because everybody will get their information diet from the same set of two or three AI assistants. That's very bad for democracy and culture.
Right. I have one last quick question. For the next one year, so there's been a lot of progress of AI in just the past one year, but for the next one year, what should we expect for AI?
Well, hopefully perhaps some progress in what's now known as physical AI. So systems that, in my opinion, will not be based on LLMs, would be based possibly on JEPA, certainly using world models. And systems that have much better understanding of the real world and can be used for controlling industrial processes, but also eventually for robotics, maybe opening the door to domestic robots.
There is a big secret in the industry at the moment, which is that there's a lot of companies building humanoid robots. None of those companies has any idea how to make those robots smart enough to be useful. They're counting on the fact that AI is going to make sufficient progress over the next two or three years, that by the time that the products are valuable, AI will be there to power them. But currently it's still very far.
Okay. Great. Great to hear all of them. Thank you very much for joining us.
Article published · Updated
