Steering the future of AI

Summary

This discussion features Yann LeCun, Meta’s Chief AI Scientist and one of the “godfathers of AI,” in conversation with Nicholas Thompson from The Atlantic about steering the future of artificial intelligence. LeCun clarifies his controversial stance that large language models (LLMs) are a “dead end,” explaining that while LLMs are useful tools, they are insufficient for achieving human-level AI or artificial superintelligence. He argues that LLMs lack four essential capabilities for true intelligence: understanding the physical world, persistent memory, reasoning, and planning. LeCun identifies two fundamental limitations of current LLMs: they use a fixed amount of computation for every response regardless of complexity, and they cannot engage in the kind of prolonged thinking that humans use for difficult problems.

As an alternative, LeCun advocates for JEPA (Joint Embedding Predictive Architecture), which would allow AI systems to understand the world through abstract representations rather than trying to predict every pixel detail in video. He expresses optimism that human-level AI could be achieved within 5-10 years, though likely through architectures beyond current LLMs. Regarding AI safety, LeCun believes future AI systems will be designed to be objective-driven with built-in guardrails, making them safe by construction, comparing the risk level to airplane safety through careful engineering and regulation.

LeCun strongly advocates for open-source AI development, arguing it leads to faster progress, better safety through transparency, and prevents dangerous concentration of AI power in a few companies. He envisions a future with a small number of dominant open-source AI platforms, similar to how Linux dominates computing infrastructure today, potentially requiring international cooperation to train models on distributed data while preserving sovereignty.

Keypoints

**Major Discussion Points:**

– **Limitations of Large Language Models (LLMs):** LeCun argues that while LLMs are useful tools, they are insufficient for achieving human-level AI because they lack essential capabilities like planning, reasoning, persistent memory, and understanding of the physical world. He clarifies that he considers them a “dead end” only in the context of reaching artificial superintelligence, not as useful technology.

– **JEPA Architecture as an Alternative:** LeCun proposes Joint Embedding Predictive Architecture (JEPA) as a superior approach to current LLMs. This architecture would make predictions in abstract representation space rather than at the pixel level, allowing AI systems to understand the physical world more effectively by eliminating unpredictable details.

– **AI Safety and Design Philosophy:** LeCun expresses confidence that AI systems can be made safe by design, comparing the risk level to airplane safety. He advocates for objective-driven systems with built-in guardrails, arguing that the likelihood of AI causing major harm is extremely small when properly engineered.

– **Open Source vs. Closed AI Development:** LeCun strongly advocates for open-source AI development, arguing it leads to faster progress, better safety through more scrutiny, and prevents dangerous concentration of AI power. He discusses how Meta’s open-source Llama models have accelerated global AI development, including in China with DeepSeek.

– **Future AI Landscape and Global Cooperation:** LeCun envisions a future dominated by 1-2 open-source AI platforms (similar to Linux in operating systems) and proposes international partnerships for training foundation models where countries contribute data and computation while maintaining data sovereignty.

**Overall Purpose:**

The discussion aimed to explore LeCun’s perspective on the current state and future direction of AI development, particularly his contrarian views on LLMs, his advocacy for open-source development, and his optimistic stance on AI safety compared to other prominent AI researchers.

**Overall Tone:**

The conversation maintained a consistently intellectual and engaging tone throughout. LeCun was confident and articulate in defending his positions, while Thompson served as an effective moderator by asking probing questions that challenged LeCun’s views. The tone was respectful but direct, with Thompson pushing back on potential contradictions in LeCun’s arguments, particularly around open-source development and geopolitical implications. The audience appeared engaged and supportive, as evidenced by positive reactions to certain points about data sovereignty.

Speakers

– **LJ Rich**: Role/title not specified, appears to be involved in conference organization and introductions

– **Yann LeCun**: Chief AI Scientist at Meta, known as one of the “godfathers of AI” for helping create the architecture that led to current AI systems, conducts research at FAIR (Fundamental AI Research lab at Meta)

– **Nicholas Thompson**: Moderator from The Atlantic

Additional speakers:

None – all speakers in the transcript are included in the provided speakers names list.

# Discussion Report: Yann LeCun on the Future of Artificial Intelligence

## Introduction and Context

This discussion featured Yann LeCun, Meta’s Chief AI Scientist, in conversation with Nicholas Thompson from The Atlantic, moderated by LJ Rich. LeCun, who described his role as “I have the best job in the company because I have nobody reporting to me. I don’t run any organization,” focused on his views regarding current AI limitations and future development paths.

## LeCun’s Position on Large Language Models

### Clarification on the “Dead End” Statement

LeCun clarified his widely reported criticism of large language models: “I didn’t say they [LLMs] were a dead end. I said they were a dead end if you are interested in reaching human-level AI or artificial superintelligence… They’re certainly a useful tool… But if we want to reach the type of intelligence that we observe in humans and animals, we are going to have to invent other methods.”

### Four Missing Capabilities for Human-Level AI

LeCun identified four critical capabilities that current LLMs lack:

1. **Understanding the physical world**: “Intelligence needs to be grounded in some reality,” noting that “a four-year-old child has gotten as much information through vision as the biggest LLMs today that are trained on all the publicly available texts on the internet.”

2. **Persistent memory**: Current LLMs cannot maintain long-term memory across interactions.

3. **Reasoning capabilities**: While LLMs can simulate reasoning, they lack deep reasoning abilities.

4. **Planning abilities**: LLMs cannot engage in strategic planning.

### Architectural Limitations

LeCun highlighted fundamental problems with current LLM architecture:

**Fixed Computational Capacity**: “The way an LLM produces an answer is by running through a fixed number of layers in the neural net and then producing a token… And the amount of computation that goes into producing this word is constant… when we have a hard question to answer, we tend to think about it for a long time, right?”

**Language vs. Reality**: LeCun argued that “language is simple. It’s simple because it’s discrete, and it’s discrete because it’s composed of discrete words and there is only a finite number of them,” contrasting this with the complexity of understanding the physical world.

## Alternative Architecture: Joint Embedding Predictive Architecture (JEPA)

LeCun advocated for Joint Embedding Predictive Architecture (JEPA) as an alternative approach. This architecture would make predictions in abstract representation space rather than attempting to predict every detail in video or text. The key advantage, according to LeCun, is that it eliminates unpredictable details by working with abstract representations, allowing AI systems to understand the physical world more effectively.

## AI Safety Philosophy

### Safety by Design Approach

LeCun expressed confidence in AI safety through careful design: “AI systems will be designed with objectives and guardrails, making them safe by construction similar to how jetliners are engineered for safety.” He argued that “the likelihood of AI causing harm is extremely small, comparable to jetliner accidents.”

Thompson challenged this aviation analogy by pointing out “the difference in regulatory approaches between highly regulated aviation industry and current minimal AI regulation.”

### Unresolved Safety Questions

Despite his optimism, LeCun acknowledged: “How do we design or train objectives and guardrails so that AI systems do not do things that are intrinsically dangerous?” He admitted this was “not a solved problem.”

## Democratic and Cultural Risks

LeCun identified concentration of power as a primary threat: “The biggest danger is that there is a future in which every single one of our interactions with the digital world will be mediated by AI assistance… We need to have a very wide diversity of AI assistance that gives people choice about what type of bias, linguistic capabilities, cultural knowledge and cultural value systems… for the same reason we need social diversity in the press.”

## Open Source Development Advocacy

### Arguments for Open Source

LeCun strongly advocated for open-source AI development, arguing that:

– Open source platforms are historically safer due to more people examining them

– Open research accelerates progress by enabling broader contribution

– It prevents concentration of power in a few companies

He noted that “FAIR has open-sourced around 1,000 research projects over 11 years.”

### Challenges and Responses

Thompson raised concerns about national security implications, specifically asking about “models like Llama being used by DeepSeek.” LeCun responded by highlighting the international nature of AI development: “Lama the first version of Lama was built by a team of 13 people, 12 of whom were in Paris… it’s not an American model, it’s a French model… ideas can pop up from anywhere.”

## Future Predictions and International Cooperation

### Timeline for Human-Level AI

LeCun provided a specific timeline: “The absolute shortest time I can imagine for at least being on a good path to reaching human-level intelligence is no less than five years. It could be 10, it could be more.”

### Market Structure and Cooperation

LeCun predicted a future with “2-3 dominant foundation models similar to operating system distribution, with 1-2 being open source.” He proposed that “Future foundation models should be trained through international partnerships preserving data sovereignty… each country or region will provide computation and its own data but will preserve its own data.”

### AI Assistants and “Artificial People”

When asked about Meta potentially creating “artificial people,” LeCun clarified that the focus would be on AI assistants: diverse AI systems that could serve different cultural and linguistic needs while preventing any single company from controlling all digital interactions.

## Key Areas of Disagreement

The main tensions in the discussion centered on:

1. **Regulatory approaches**: LeCun favored industry self-regulation through design, while Thompson highlighted the need for more robust oversight

2. **Open source vs. national security**: Thompson questioned whether open access undermines Western technological advantages

3. **Evidence for safety claims**: Thompson consistently asked for empirical evidence supporting LeCun’s theoretical arguments about open source benefits

## Conclusion

The discussion revealed LeCun’s comprehensive critique of current AI approaches while proposing specific alternatives through JEPA architecture and open-source development. While acknowledging significant unresolved technical and safety challenges, LeCun expressed optimism about achieving human-level AI within 5-10 years through international cooperation and diverse, democratically-oriented AI development.

The conversation highlighted ongoing debates about the best paths forward for AI development, particularly regarding the balance between innovation, safety, and democratic values in an increasingly global technological landscape.

LJ Rich: covering from seeing him speak at World AI Con. So, Jann and you will also be in really good hands with our fabulous moderator from the Atlantic. It’s Nicholas Thompson. Everybody is ready now, so it’s a pleasure to introduce the next session on how we can steer the future of AI. Thank you.

Nicholas Thompson: All right, Jann, you ready to be information-dense? That was a good introduction. How are you? I’m pretty good. I’m as ready as I’ll ever get. Wonderful. Jann is the chief AI scientist at Meta. He is also known as one of the godfathers of AI because he helped create the architecture that led to all the AI that we have today. And he is incredibly interesting and fun to talk to. And given that last panel, I should verify that man is real, and you can tell because there are no Pikachus behind him. So, let’s get cracking. Jann, first thing I want to ask you is, you have often said that large language models are a dead end. That is quite contrary to what many other people on this stage have been saying all day, many other people in the world, many other people investing billions of dollars. Tell me whether you still believe they’re a dead end and why.

Yann LeCun: Okay, it’s a bit of a fake news due to the soundbite habit. I didn’t say they were a dead end. I said they were a dead end if you are interested in reaching human-level AI or artificial superintelligence, if you want to call it this way, in the sense that they’re not sufficient. They’re certainly a useful tool. There’s a lot of people working on LLMs. I use them. Everybody now in the world uses them for something. There’s a lot of computer technology, including AI technology, that is very useful, yet is not necessarily a good path towards human-level intelligence. So, LLMs are very limited in some ways. That doesn’t make them useless. It doesn’t mean that we shouldn’t be working on them. In fact, we are, and many of my colleagues are working on it. It’s something we should do. But if we want to reach the type of intelligence that we observe in humans and animals, we are going to have to invent other methods to be the basis of the architecture of the systems. Admittedly, LLMs will still have a role to play because they’re extremely good at transforming thoughts into language. So, we’re still going to use them for that. But what about thoughts? How do we elaborate thoughts? How do we understand the physical world? How do we have persistent memory? How do we reason? How do we plan? Those are four essential things for intelligence that LLMs currently are not really capable of doing to the extent that we’d like. Part of the reason I ask you is because you are so optimistic about AI, and in some ways, your optimism seems to stem from a different belief that some of your colleagues have, and your belief that large language models under current architecture will never become all-powerful or incapable of becoming all-powerful. Is that a correct assessment of your views? Yeah, it’s a correct assessment. I’m certainly optimistic in the sense that I have no doubt that at some time in the future, we’ll be able to produce systems that match or surpass human intelligence in pretty much all domains, and that may arrive fairly soon, possibly within the next decade. The absolute shortest time I can imagine for at least being on a good path to reaching human-level intelligence is no less than five years. It could be 10, it could be more. It’s probably more because it’s almost always harder than we think. There are obstacles on the way that we’re not seeing yet, but I’m certainly optimistic that we will reach that goal. It’s just that LLMs, as they are practiced today, do not have the essential characteristics that are necessary, at least that’s my opinion, and that of an increasingly large number of my colleagues,

Nicholas Thompson: I’m happy to say. And they don’t have the capacity. So large-language models, it is impossible for large-language models to reach human intelligence because they do not have the capacity for planning, they do not have some fundamental capabilities that humans have,

Yann LeCun: and not only that, you think they will never get there. Well, something will get there, and at this point, I think we will not be able to call them LLMs anymore. So here is the two things that I think are essential that LLMs really are not capable of. The first one is the type of inference that they produce. The way an LLM produces an answer is by running through a fixed number of layers in the neural net and then producing a token, which is a word, essentially. And the amount of computation that goes into producing this word is constant. So it’s kind of a problem because when we have a hard question to answer, we tend to think about it for a long time, right? And so if you ask a system, an LLM, to answer by yes or no, the amount of computation you can devote to the answer is fixed. And that doesn’t make any sense because if the answer is simple, of course you can answer it directly, but if it’s complicated, you need to think about it. So the process of thinking about something, generally in humans at least, involves kind of manipulating some mental model of the situation, if you want, and sort of imagining scenarios. And that takes place all in your head. It’s not something you have to write down on paper unless you do mathematics, which is a special case, or perhaps coding. But most of that process takes place in your head using your kind of abstract mental model of the situation you’re thinking of. And you do this for reasoning, for planning, for all kinds of tasks, for everything that we do in our daily lives for which we’re not kind of capable of just taking an action subconsciously without thinking about it, which is what psychologists call the system two type of thinking. So the first characteristic an AI system should have if you wanted to reach human intelligence is this ability to kind of think for a long time and basically produce an answer by search, searching for an answer, not just computing it, but searching for it. This is a very sort of intrinsically different process. That’s the first thing. The second thing is, I’m of the opinion, as many of my colleagues, that intelligence needs to be grounded in some reality. It doesn’t need to be a physical reality. It could be virtual, it could be a situation, but some underlying reality in the sense that the information you get about this underlying reality is much higher bandwidth than what you get through purely text, which is human produced. Okay, so we get a huge amount of information through vision, through touch, through audition. And a four year old child has gotten as much information through vision as the biggest LLMs today that are trained on all the publicly available texts on the internet. And so that tells you, you know, we’re not gonna get to human level AI without having systems that understand the physical world. And here is the problem, with the current architectures that we have, which are basically trained to produce the next tokens, and those tokens have to be discrete, essentially, we cannot use this architecture to get a system to understand the world. We have to use a different architecture. I’ve been advocating for one called JEPA, that means Joint Embedding Predictive Architecture. We explained just a few weeks ago, released an example of this. This is a different architecture. It’s not LLS, it’s completely different. Why don’t you explain the JEPA architecture for the room and explain why that will lead to a more profound form of intelligence than the large language model path that we’re pursuing, so many companies are pursuing. Okay, so really something that may be shocking for a lot of people, language is simple. It’s simple because it’s discrete, and it’s discrete because it’s composed of discrete words and there is only a finite number of them. And so to some extent, it’s relatively simple to kind of predict the sequence of words that will follow, which is what LLMs do. Now the real world, as it turns out, is much more complicated. And of course, the idea that we’re going to train an AI system to understand the physical world by predicting what’s happening in a video is an old idea. I’ve been working on this for almost 20 years now, and my colleagues and I admit I’ve been working on this for the better part of the last 10 years. It turns out it doesn’t work. If you train a system to try to predict all the details of what happens in the video, you don’t get much kind of useful, the kind of representation of the world the system produces are not that useful. And it’s because predicting all the details of what happens in the video is essentially impossible. You know, I’m not seeing the room that with the audience at the moment, I can infer some characteristics of it at the abstract level. I can’t tell how many people are there, but I’m guessing most of the seats are occupied and there is, you know, equal proportion of men and women and, you know, the age distribution. There’s actually several thousand people and they’re all kind of dancing somewhat wildly but mutely, but go on. Okay. Well, that violates my prediction. I guess my distribution of what was happening was wrong. So anyway, there is no way I can predict, you know, the details of what everybody looks like and, you know, what the texture of the carpet is or whatever. Like, you know, there’s a lot of things in sort of video prediction that you just cannot predict. And if you ask a neural net, if you train a neural net to predict all those details, it basically gives up. I mean, it kind of, it can’t really solve the problem. And so it really doesn’t do a good job. Now, of course, if you want to train a system to actually produce cute looking video, you can do that. All right. And we can do that at the moment. But that’s a very far cry from actually understanding the sort of underlying physics if you want or the structure of what goes on. So the idea of JPEG is you have a system that looks at video and you may train it to predict what’s going to happen in a video in time or in space. But instead of making the prediction at the pixel level, you make it in an abstract representation level. So the system is trained to compute a representation, basically a neural net that computes an abstract representation of the world. And it makes prediction in that representation space where a lot of the details that are not predictable have been eliminated, essentially. All right. So you have a different architecture for how we should learn about the world.

Nicholas Thompson: And your view is that it will make AI closer to human intelligence, more powerful, down the line maybe we’ll be using JEP architecture, maybe we’ll be using something else. Let me ask you another question. So one of the exercises I did, initially we were going to have you and Jeff Hinton on stage to debate each other about the risks of AI. It didn’t work out. And instead you’re on today, Jeff is on tomorrow. And I was trying to think about what the fundamental differences are between you and Jeff. One, of course, is your view that there are limits to AI’s intelligence, whereas Jeff has a different view. The second, which I think is really important, is your view that it’s highly unlikely that AI would ever have bad intentions. It’s highly unlikely that any AI system, whether a large language model, whether something else, would ever act against the interests of its creator. Explain why you’re confident in that view and why it gives you confidence that we’re not going to head to a bad scenario with AI. Okay. What I’m confident about is that we will design an AI system once we figure out how to potentially

Yann LeCun: reach human level intelligence or something approaching it. And by the way, this is not going to be an event. It’s going to be kind of progressive innovations. And we’ll start with something at the level of a rat or a cat or something like this before we get to humans. So we’re going to have some time to figure out how to make those systems safe. But what I’m convinced of is that we’ll design those systems so that they are objective driven. You give them an objective, they accomplish it. And you give them guardrails. And by construction, they will have to abide by those guardrails and produce actions, sequences of actions or whatever, that are safe. Now we’ll design them for safety. It will be very different in the design from current LLMs that really don’t have any objectives. The objective of an LLM is specified by the prompt, and it’s a very weak specification because it entirely depends on the type of training the LLM has gone through. In the architecture I’m proposing, there is an objective that needs to be satisfied and the system can do nothing else but accomplish this objective subject to the guardrails. So that’s kind of an architecture that would be safe by construction. That leaves the question of how do we design or train those objectives and those guardrails so that the system does not actually do things that are intrinsically dangerous. And it’s not a solved problem, so I don’t want to minimize the complexity of this. But it’s not an infeasible problem either. The likelihood that some AI system will do something deleterious or dangerous to at least some humans is about the same as the likelihood that the next jetliner you fly on will blow up, which is an extremely, extremely small probability. And it’s very small because the people designing it have been working on the reliability of jet engines and airplanes for decades and kind of fine-tuning them. It’s going to be the same thing. AI systems, there’s going to be a design that makes them safe, and then within this, you’ll have to actually find the details of how you make them safe. Wait, Jan, let’s stay on this metaphor for a second because with jetliners, it’s one of the most regulated industries in the world. Before Boeing wants to build a new engine, they have to go through a million specs. There’s regulation to the left, there’s regulation to the right. Before the jetliner goes off, you check everything. The pilot makes you deplane if there’s extra moisture on the windshield. With AI, we’re building things that are so much more powerful, and there is no regulatory structure right now or minimal regulatory structure right now. We’re charging ahead. I’m not saying that’s a bad thing. My job is just the moderator. But isn’t there quite a difference between the way we regulate the thing on the left and the thing on the right? No. If you want to deploy an AI-based system to do a medical diagnosis or a medical image analysis or a driving assistance system for your car, it also has to go through careful testing and basically clinical trial for medical systems and authorization of deployment by government agencies. I mean, I think it’s pretty much the same. Now, there is no such regulation for chatbots because what they can do at the moment is… I mean, the risk is tiny and extremely… The risk of major harm is small. I mean, people have been sort of playing around this for the last two years. There were a lot of discussions as to whether LLMs are the type that we know today. Because of concerns of this type, it turns out LLMs have been around for a long time. They’re increasing in power, but not by huge amounts. And the world has not ended. And in fact, not much bad things have happened because of them.

Nicholas Thompson: Let’s talk about one of the architectural things that you’ve been a very vocal, powerful… Proponent of which is open source. And one of the arguments you’ve made from the very beginning is that open source architecture will lead to More safety if you have more Companies more countries more individuals involved in decisions. You’ll get much safer outcomes Is there any empirical evidence two years in that that is true or that it is false. There’s two. Yeah There’s two factors. The first thing is that historically it is the case that open source

Yann LeCun: Platforms and AI is going to be a platform, right? It’s going to be some sort of common platform that a lot of people are going to use and build upon So it is the case that open source platform software Are safer and more portable More flexible. I mean there’s all kinds of advantages and the reason why for example the entire Software architecture of the Internet and the mobile communication network is open source Is because it’s safer more portable, you know There’s more eyeballs looking at it and you know bugs are fixed faster and everything like that There is a pressure from the market for this to happen So that’s the evidence now for for AI systems is a lot of academic studies of safety or studies in general of safety that are enabled by the fact that a number of AI systems are open source. It’s very difficult to assess the the safety Of a system if you don’t have access If you can’t kind of, you know, run it locally kind of modify it in various ways So yes Now there is a more important factor though, because I think the biggest danger of AI going forward is not At least in the short term not whether you know AI systems are going to you know, do bad things or take over the world The biggest danger is that there is a future in which every single one of our interactions with the digital world Will be mediated by AI assistance Okay, so our entire digital diet will come from AI systems It’s already the case that a lot of our digital layer comes from machine learning systems But, you know, they do ranking and filtering and you know all that stuff right and content moderation But in the future, you know, our entire digital diet will come from from AI system. We need to have a very wide Diversity of AI assistance that gives people choice about what type of bias linguistic Capabilities cultural Knowledge and you know cultural value systems, etc Our political bias those assistants have for the same reason we need social diversity in the press If our AI system come from a handful of companies on the west coast of the US or China First of all, a lot of Countries in the world a lot of governments in the world would just not accept that because they will think it’s a danger to their Culture and democracy and so we need high diversity and that’s only enabled by

LJ Rich: Platforms this I mean, this is an extraordinarily important argument. It’s an extremely interesting argument Let me ask you a question. You may not be able to fully answer I know that you are the chief research scientist at Metta You are not in charge of Metta’s product roadmap But Metta has also talked about creating or Mark Zuckerberg on the Dworkish podcast talked about creating You know fully fake individuals because people have a demand for more friends than they currently have so to populate the social platforms that Metta runs with individuals Do you influence product decisions like that with the philosophy? You just gave like if you’re going to make people make sure there’s a diversity of people from different backgrounds with open source

Yann LeCun: Does your philosophy go through the rest of the company or is that something you just believe in the products that you’re building? Okay, so so I don’t build products, right? So Because it’s very confusing I’m chief AI scientist. Okay, and I have the best job in the company because I have nobody reporting to me I don’t run any organization Okay, I I conduct research projects on my own With people who want to work for me. They work with me because they want to know because I’m their boss and I try to Inspire people doing research to kind of work on projects and topics that I think are promising and interesting like for example This JEPA idea I was talking about and and things like persistent memory and reasoning and planning of various types So this is really the mission of FAIR, which is the fundamental AI research lab at Metta Okay And this is really long term. So the horizon of projects we work on is You know on average maybe two to three years sometimes a little shorter Sometimes quite a bit longer. It could be ten years Now I Also have you know a bit of input on strategy about about AI not at the product level but at the kind of more Long-term strategy and perhaps how to approach certain types of market So I was certainly a vocal advocate within the company. I was not the only one for open sourcing Lama for example the Lama series and and various other things in fact FAIR in its 11 year history has open source on the order of 1,000 Research projects and kind of various packages. So it’s actually in the DNA of the company really Open sourcing is is really a practice that preceded my arrival at Metta. You know, it’s not it’s not recent so For reasons of you know discussions of safeties and liability and things like this. We have to do this carefully But but certainly the strategy of open source is something that I’m personally extremely happy that Metta has embraced Yeah Now the you know to the idea that you know, should we build artificial people? No, the the idea is that AI systems are going to amplify human intelligence now the best way you can amplify human intelligence is by providing a system that People are used to interact with okay, which basically are other humans, right? So the best way to kind of Help people in their daily lives is to provide them with assistance that basically behave a little bit like humans or not You know frustrating to talk to because they don’t understand what you’re telling them but in that future, you know, we’ll all of us will behave like Like a boss of a manager of a team of potentially super intelligent people who would be working for us and Assisting us, you know in our daily lives That’s making us more Intelligent more rational Perhaps even wiser we can always move. Yeah I I asked that question in part guy I just think it’s such an interesting it’s such an interesting moment where the company that is building AI that will shape so much of society has a lead AI scientist who has These amazing but also quite different from much of the rest of world views about open source It’ll be extremely interesting to watch. I don’t think anybody knows how it plays out let me ask you another question on open source, which is Your models and the models built by llama Were used as part of the architecture in deep seek which in some way is a great validation of what you’ve been saying and It’s a way that AI has spread around the world to many people in America’s national security this was Not what they were looking for

Nicholas Thompson: Explain your views on that Well, so the magic of open research And open source is that you accelerate progress because you get more people involved and they have different ideas and Every idea every good idea kind of gets picked up by everybody else. So

Yann LeCun: So yes deep seek learn from llama. I mean they learn from you know, the whole history of AI before and but then vice versa when when DeepSea came out and and other open source model in China and by the way China is very aggressive about open sourcing their model which I think is interesting and a good thing we learn from them too we we learn you know a few tricks from from DeepSea and Quen and and and the other models and it’s not just you know my colleagues and I in research it’s you know everyone in the in the industry and so the big question is not you know you know who who has the best ideas because if those ideas are shared then they profit everybody the question is who is most capable of basically building on top of those ideas and making them useful for for people and then what matters is the speed at which the information circulates so you you might think that closing up research in the West in the Western world okay like let’s say you know some catastrophic scenario of banning open source or something like this which maybe you know stop geopolitical rivals from getting those ideas but you know they will seep anyway at some point it will just be with it with a delay the problem is we’ll be slowing us down ourselves we’ll we’ll be progressing slower and so basically the geopolitical rivals would be more you know able to kind of keep up because we’re not going to go as fast so it’s a bit like shooting yourself in the foot right I mean what you we have to do is be the the best country region industry whatever you want to call it at picking up on innovations and then pushing them to the market to become the leader and that goes through fast communication and information exchange but wait yeah let me I feel like they’re two kind of competing ideas and the answer you just gave there one is that it sounds like you think you want the West to be leading in AI and if that’s the case and if that is the goal surely it would be better for China not to be able to just go online and look at the structure of llama or perhaps even get weights that were you know leaked online

Nicholas Thompson: shouldn’t there be restrictions on you know a model as powerful as llama and the access that folks around the rest of the world can have okay you’re gonna be careful what you wish for here because first of all I’m not saying what I wish for I’m just asking whether there’s a contradiction I mean I feel like you can say it doesn’t really matter whether the West is ahead or behind because the world will share in this technology it’s a human technology it’s kind of part of the theme of this event in any event but I feel like if you’re gonna make the position if you’re gonna argue the position that the West should have a lead then when you also make this second position which would run counter to what you said that’s all I’m asking okay so first of all even if

Yann LeCun: you manage to completely isolate intellectually China from the rest of the world China has excellent resources in terms of engineers and scientists and they’re not going to be far behind in fact they may come up with ideas that the West does not come up with and so in some respect they might actually be ahead which is why DeepSea was such a shock for a lot of people in kind of the Silicon Valley bubble if you want but likely okay but let me let me give you a bit of information here Lama the first version of Lama was built by a team of 13 people 12 of whom were in Paris okay it’s not an American model it’s a French model now Lama 2 then was okay Lama 2 was built on top of this and that required a big team of engineers and scientists and you know spread over you know Paris New York Menlo Park Seattle and you know various London and various other places okay when you scale up when you pick it up and you try to make it real it becomes a bigger team DeepSea was also you know a small team 10-15 people in China most projects start this way same with PyTorch which you know Meta also open-sourced and actually transferred the thing Gemini produced or Gemini produced by Google a lot of it you know is developed in London so it’s not an American thing it’s not like you know the the the US is alone in this thing ideas can pop up from anywhere and we should have a system by which the best ideas circulate but then the the big question is how fast can you capitalize on those ideas to kind of put things out the door and China is not as efficient as as the US and Silicon Valley in particular are doing this let

Nicholas Thompson: me ask you what one quick question because we’re almost out of time here Yam which relates to this in five years maybe in three years how many foundation LLMs do you think the world will be using well there’ll be a small number of large winners as tends to happen in tech or will there be an infinite or not infinite but extremely large number of open source models will be like it is today there are there some very big closed source models and some open source models what’s it gonna be I think it’s gonna be very much like the situation in operating systems today you’re gonna have like two or three one with one of which maybe one or two will be open source and dominant and then two

Yann LeCun: or three would be proprietary and you know we’ll kind of cater to kind of you know niche markets and this is a situation you you you see today in operating systems almost every single computer in the world today runs Linux okay except for a few a few desktop and a few iPhones and really frankly that’s a small number of computers compared to all the the servers and you know the embedded devices and everything in your car and your phone so it’s gonna be the same situation you’re gonna have a one or two dominant open source platforms AI platforms foundation models on top of which the entire industry would be built and then perhaps it would be space for a few non open source ones that are more proprietary the reason I’m saying this is because the future I’m envisioning is one in which you know LLMs or their descendants will basically constitute a repository of all human knowledge and culture and if you want this to happen you need access to data that currently reside perhaps in digital form perhaps not but in various regions of the world various countries and those countries are not ready to just share the data unless they get something out of it and they want AI sovereignty so the only way out of this that I see is some sort of international partnership in which future foundation models will be trained in a distributed fashion so we’ll have one common big you know model repository of all human knowledge but each country or region will provide computation and its own data but will preserve its own data so the the way the global system will be trained is not by exchanging data but by exchanging parameter vectors basically so we can arrive at a consensus model without every region giving up their data and therefore kind of you know preserving some level of

Nicholas Thompson: sovereignty all right this will have to be open source yeah and I don’t think you heard and you can’t see the room but you got some hoots when you said data sovereignty which tells you a little bit about this conference Jan you said you have one of the best jobs in the world you definitely one of the most interesting jobs in the world I think he’s one of the most interesting minds in AI thank you so much Yann LeCun it’s always a pleasure to talk oh my goodness another amazing conversation thank you very much to Yann LeCun and Nicholas Thompson for a mind-bending discussion

Agreement points

AI systems require careful design and safety considerations

Speakers

– Yann LeCun
– Nicholas Thompson

Arguments

Future AI systems should be objective-driven with built-in guardrails for safety by construction

AI systems will be designed with objectives and guardrails, making them safe by construction similar to how jetliners are engineered for safety

Points out the difference in regulatory approaches between highly regulated aviation industry and current minimal AI regulation

Summary

Both speakers acknowledge that AI systems need careful safety design, though Thompson raises concerns about the current lack of regulation compared to other industries like aviation

Topics

Legal and regulatory | Cybersecurity

Current AI development involves global collaboration and knowledge sharing

Speakers

– Yann LeCun
– Nicholas Thompson

Arguments

Open research accelerates progress by enabling more people to contribute different ideas and learn from each other

Raises concerns about models like Llama being used by DeepSeek and implications for national security

Summary

Both speakers recognize that AI development is inherently global and collaborative, with ideas and models being shared across borders, though they may differ on the implications

Topics

Economic | Infrastructure

Similar viewpoints

Both speakers engage seriously with the timeline and feasibility of achieving human-level AI, with Thompson acknowledging the significant investment and interest in this goal

Speakers

– Yann LeCun
– Nicholas Thompson

Arguments

Human-level AI may be achievable within 5-10 years, starting with animal-level intelligence

Questions whether LeCun still believes LLMs are a dead end given widespread investment and adoption

Topics

Economic | Infrastructure

Both speakers recognize the importance of diversity in AI systems and the potential risks of concentration of power in a few companies

Speakers

– Yann LeCun
– Nicholas Thompson

Arguments

Diversity of AI assistants is crucial to prevent a handful of companies from controlling all digital interactions and cultural biases

Explores how LeCun’s philosophy on diversity and open source influences Meta’s product decisions

Topics

Sociocultural | Human rights

Unexpected consensus

Need for international cooperation in AI development

Speakers

– Yann LeCun
– Nicholas Thompson

Arguments

Future foundation models should be trained through international partnerships preserving data sovereignty

Raises concerns about models like Llama being used by DeepSeek and implications for national security

Explanation

Despite Thompson raising national security concerns, both speakers seem to accept that international collaboration in AI is inevitable and necessary, with LeCun proposing structured partnerships that preserve sovereignty

Topics

Legal and regulatory | Human rights

Recognition of China’s AI capabilities and innovation

Speakers

– Yann LeCun
– Nicholas Thompson

Arguments

China has excellent resources and may develop ideas the West doesn’t, as demonstrated by DeepSeek’s impact

Raises concerns about models like Llama being used by DeepSeek and implications for national security

Explanation

Both speakers acknowledge China’s significant AI capabilities and innovation potential, even when discussing competitive concerns, showing unexpected consensus on China’s technical competence

Topics

Economic | Infrastructure

Overall assessment

Summary

The speakers show substantial agreement on the technical challenges and future directions of AI development, the importance of safety considerations, the need for diversity in AI systems, and the reality of global collaboration in AI research. They also both recognize the significant capabilities of international competitors like China.

Consensus level

High level of consensus on technical and structural issues, with disagreements mainly on policy approaches rather than fundamental assessments. This suggests that while there may be debates about regulation and open source policies, there is broad agreement among experts about the technical trajectory and challenges of AI development.

Different viewpoints

Regulatory approach to AI safety

Speakers

– Yann LeCun
– Nicholas Thompson

Arguments

AI systems will be designed with objectives and guardrails, making them safe by construction similar to how jetliners are engineered for safety

Points out the difference in regulatory approaches between highly regulated aviation industry and current minimal AI regulation

Summary

LeCun argues that AI safety can be achieved through careful design similar to aviation, while Thompson challenges this analogy by highlighting that aviation is heavily regulated whereas AI development currently has minimal regulatory oversight despite building more powerful systems.

Topics

Legal and regulatory | Cybersecurity

Open source vs national security concerns

Speakers

– Yann LeCun
– Nicholas Thompson

Arguments

Restricting open source would slow down Western progress more than it would hinder geopolitical rivals

Questions the contradiction between wanting Western AI leadership while supporting open access to models like Llama

Summary

LeCun believes restricting open source would harm Western progress more than help, while Thompson questions whether supporting open access undermines Western AI leadership, especially given national security concerns about models like Llama being used by competitors like DeepSeek.

Topics

Economic | Legal and regulatory

Evidence for open source safety benefits

Speakers

– Yann LeCun
– Nicholas Thompson

Arguments

Open source platforms are historically safer, more portable, and flexible due to more people examining and improving them

Challenges whether open source truly leads to better safety outcomes and asks for empirical evidence

Summary

LeCun argues that open source is historically safer based on general software principles, while Thompson specifically asks for empirical evidence that this applies to AI systems after two years of implementation.

Topics

Legal and regulatory | Infrastructure

Unexpected differences

Fundamental disagreement on current AI regulation adequacy

Speakers

– Yann LeCun
– Nicholas Thompson

Arguments

Current LLMs pose minimal risk as evidenced by two years of widespread use without major incidents

Points out the difference in regulatory approaches between highly regulated aviation industry and current minimal AI regulation

Explanation

This disagreement is unexpected because both speakers generally support AI development, yet they have fundamentally different views on whether current regulatory approaches are adequate. LeCun sees current evidence as proof that minimal regulation is sufficient, while Thompson argues that more powerful systems require more robust regulatory frameworks.

Topics

Legal and regulatory | Cybersecurity

Overall assessment

Summary

The main disagreements center around regulatory approaches to AI safety, the balance between open source benefits and national security concerns, and the adequacy of current oversight mechanisms. While both speakers support AI development, they differ significantly on risk assessment and appropriate safeguards.

Disagreement level

Moderate to significant disagreement with important implications for AI governance. The disagreements reflect broader tensions in the AI community between innovation acceleration and risk management, between global collaboration and national competitiveness, and between industry self-regulation and government oversight. These disagreements could influence policy decisions about AI regulation, international cooperation, and open source development practices.

Partial agreements

Similar viewpoints

Both speakers engage seriously with the timeline and feasibility of achieving human-level AI, with Thompson acknowledging the significant investment and interest in this goal

Speakers

– Yann LeCun
– Nicholas Thompson

Arguments

Human-level AI may be achievable within 5-10 years, starting with animal-level intelligence

Questions whether LeCun still believes LLMs are a dead end given widespread investment and adoption

Topics

Economic | Infrastructure

Both speakers recognize the importance of diversity in AI systems and the potential risks of concentration of power in a few companies

Speakers

– Yann LeCun
– Nicholas Thompson

Arguments

Diversity of AI assistants is crucial to prevent a handful of companies from controlling all digital interactions and cultural biases

Explores how LeCun’s philosophy on diversity and open source influences Meta’s product decisions

Topics

Sociocultural | Human rights

Key takeaways

Large Language Models (LLMs) are useful tools but insufficient for achieving human-level AI due to fundamental limitations in planning, reasoning, persistent memory, and world understanding

Alternative architectures like JEPA (Joint Embedding Predictive Architecture) are needed that can make predictions in abstract representation space rather than at pixel level

AI safety can be achieved through systems designed with built-in objectives and guardrails, making them safe by construction similar to aviation engineering

Open source AI development accelerates progress, increases safety through broader scrutiny, and prevents concentration of power in a few companies

Future AI landscape will likely have 2-3 dominant foundation models, with 1-2 being open source, similar to current operating system distribution

International cooperation and data sovereignty will be crucial for training future foundation models that represent global human knowledge

Human-level AI may be achievable within 5-10 years, progressing through animal-level intelligence first

Restricting open source AI development would slow Western progress more than it would hinder geopolitical rivals

Resolutions and action items

N

o

n

e

i

d

e

n

t

i

f

i

e

d

Unresolved issues

How to design and train objectives and guardrails for AI systems to prevent intrinsically dangerous behavior – acknowledged as complex and unsolved

Whether open source AI development truly leads to better safety outcomes – empirical evidence still being gathered

The apparent contradiction between wanting Western AI leadership while supporting open access to powerful models

How to balance national security concerns with the benefits of open source AI development

The technical details of implementing international partnerships for distributed AI training while preserving data sovereignty

How Meta’s AI research philosophy translates into actual product decisions and implementations

Suggested compromises

International partnerships for training foundation models where countries provide computation and data but preserve data sovereignty through parameter vector exchange rather than raw data sharing

A future AI ecosystem with both open source dominant platforms and proprietary niche models, similar to current operating system distribution

Careful but not restrictive approach to open sourcing AI models, balancing safety considerations with the benefits of open development

I didn’t say they [LLMs] were a dead end. I said they were a dead end if you are interested in reaching human-level AI or artificial superintelligence… They’re certainly a useful tool… But if we want to reach the type of intelligence that we observe in humans and animals, we are going to have to invent other methods.

Speaker

Yann LeCun

Reason

This comment is deeply insightful because it reframes the entire AI debate by distinguishing between utility and sufficiency. LeCun clarifies that his criticism isn’t about LLMs being useless, but about their fundamental limitations in achieving human-level intelligence. This nuanced position challenges both AI optimists who see LLMs as the path to AGI and pessimists who might dismiss them entirely.

Impact

This comment set the foundational tone for the entire discussion, moving it away from sensationalized ‘dead end’ interpretations toward a more sophisticated analysis of AI architectures. It allowed Thompson to probe deeper into LeCun’s specific technical objections and alternative proposals, structuring the conversation around the core question of what constitutes a path to human-level AI.

The way an LLM produces an answer is by running through a fixed number of layers in the neural net and then producing a token… And the amount of computation that goes into producing this word is constant… when we have a hard question to answer, we tend to think about it for a long time, right?

Speaker

Yann LeCun

Reason

This is a profound technical insight that makes abstract AI limitations concrete and relatable. By connecting the fixed computational architecture of LLMs to human cognitive processes, LeCun reveals a fundamental mismatch between current AI and human intelligence. The observation that humans allocate variable thinking time to problems of different complexity exposes a core architectural limitation.

Impact

This comment shifted the discussion from general criticisms to specific technical limitations, providing a concrete foundation for understanding why LLMs might be insufficient for human-level AI. It led directly into LeCun’s explanation of ‘system two thinking’ and the need for AI systems that can ‘search’ for answers rather than just compute them, deepening the technical sophistication of the conversation.

Intelligence needs to be grounded in some reality… a four year old child has gotten as much information through vision as the biggest LLMs today that are trained on all the publicly available texts on the internet.

Speaker

Yann LeCun

Reason

This comparison is strikingly thought-provoking because it challenges our assumptions about the scale and nature of learning. The idea that a four-year-old’s visual experience equals the entire internet’s text content reframes our understanding of information density and learning efficiency. It suggests that text-based training, no matter how massive, may be fundamentally insufficient.

Impact

This insight pivoted the conversation toward embodied cognition and multimodal learning, setting up LeCun’s introduction of the JEPA architecture. It provided compelling justification for why alternative approaches are needed and helped the audience understand why purely text-based systems might hit fundamental limits.

The biggest danger is that there is a future in which every single one of our interactions with the digital world will be mediated by AI assistance… We need to have a very wide diversity of AI assistance that gives people choice about what type of bias, linguistic capabilities, cultural knowledge and cultural value systems… for the same reason we need social diversity in the press.

Speaker

Yann LeCun

Reason

This comment is exceptionally insightful because it reframes AI safety from technical risks to democratic and cultural risks. LeCun identifies a more immediate and arguably more important danger than AI takeover: the concentration of AI systems in few hands could create unprecedented control over human information consumption and cultural expression. The analogy to press diversity makes this abstract concern tangible.

Impact

This comment fundamentally shifted the safety discussion away from science fiction scenarios toward concrete democratic concerns. It provided the intellectual foundation for LeCun’s open source advocacy and led to a more nuanced exploration of geopolitical implications. The conversation moved from technical safety to societal governance of AI systems.

Lama the first version of Lama was built by a team of 13 people, 12 of whom were in Paris… it’s not an American model, it’s a French model… ideas can pop up from anywhere and we should have a system by which the best ideas circulate.

Speaker

Yann LeCun

Reason

This revelation is thought-provoking because it challenges nationalist narratives about AI development while providing concrete evidence for international collaboration. By revealing the French origins of Llama, LeCun demonstrates that AI innovation is inherently global, making arguments for technological isolation seem both impractical and counterproductive.

Impact

This comment effectively countered Thompson’s probing about national security concerns and technological competition. It shifted the discussion from zero-sum geopolitical thinking toward a more collaborative vision of AI development, supporting LeCun’s arguments about open source benefits and international cooperation.

The future I’m envisioning is one in which LLMs or their descendants will basically constitute a repository of all human knowledge and culture… each country or region will provide computation and its own data but will preserve its own data… we can arrive at a consensus model without every region giving up their data.

Speaker

Yann LeCun

Reason

This vision is remarkably thought-provoking because it proposes a technical solution to the fundamental tension between global AI development and data sovereignty. The idea of distributed training with shared parameters but protected data offers a potential path forward for international AI cooperation while respecting national concerns about data control.

Impact

This comment provided a compelling conclusion to the discussion by offering a concrete vision for how AI development could proceed globally while addressing sovereignty concerns. It synthesized many of the conversation’s themes – open source, international cooperation, and democratic access to AI – into a practical framework for future development.

Overall assessment

These key comments shaped the discussion by systematically building a comprehensive alternative vision for AI development. LeCun’s insights moved the conversation through several levels: from technical architecture (why LLMs are limited) to safety philosophy (democratic rather than existential risks) to geopolitical strategy (open source and international cooperation). Each comment built upon previous ones, creating a coherent argument that challenged mainstream AI discourse on multiple fronts. The discussion evolved from defensive clarifications about LLM limitations to a proactive vision for global, democratic AI development. Thompson’s skilled questioning helped draw out these insights, but LeCun’s comments consistently elevated the conversation beyond typical AI hype or fear-mongering toward substantive technical and philosophical analysis. The overall effect was to present a mature, nuanced perspective that acknowledged both AI’s potential and its current limitations while proposing concrete paths forward.

How do we design or train objectives and guardrails so that AI systems do not do things that are intrinsically dangerous?

Speaker

Yann LeCun

Explanation

LeCun acknowledged this as ‘not a solved problem’ and emphasized its complexity while discussing AI safety architecture, indicating it requires further research and development.

How do we elaborate thoughts in AI systems?

Speaker

Yann LeCun

Explanation

This was mentioned as one of four essential capabilities for intelligence that current LLMs lack, requiring new methods and architectures to achieve human-level AI.

How do we enable AI systems to understand the physical world?

Speaker

Yann LeCun

Explanation

LeCun identified this as a fundamental limitation of current LLMs and a necessary component for reaching human-level intelligence, requiring different architectural approaches.

How do we implement persistent memory in AI systems?

Speaker

Yann LeCun

Explanation

This was listed as one of the four essential capabilities missing from current LLMs that would be necessary for human-level intelligence.

How do we enable AI systems to reason effectively?

Speaker

Yann LeCun

Explanation

Reasoning was identified as a critical capability that LLMs currently cannot perform to the desired extent for achieving human-level intelligence.

How do we enable AI systems to plan effectively?

Speaker

Yann LeCun

Explanation

Planning was mentioned as one of the four essential missing capabilities in current LLMs that would need to be developed for human-level AI.

What obstacles on the path to human-level AI are we not seeing yet?

Speaker

Yann LeCun

Explanation

LeCun mentioned that reaching human-level AI is ‘almost always harder than we think’ and there are likely unforeseen obstacles that need to be identified and addressed.

How can we implement international partnerships for distributed training of foundation models while preserving data sovereignty?

Speaker

Yann LeCun

Explanation

LeCun proposed a complex system where countries could contribute data and computation without sharing raw data, but the technical and political implementation details require further development.

What is the empirical evidence for open source leading to safer AI outcomes?

Speaker

Nicholas Thompson

Explanation

Thompson specifically asked whether there was empirical evidence supporting LeCun’s claims about open source safety benefits, indicating a need for more concrete data and research on this topic.

How will the JEPA architecture specifically lead to more profound intelligence than LLMs?

Speaker

Nicholas Thompson

Explanation

While LeCun explained the concept, the practical implementation and validation of JEPA’s superiority over current architectures requires further research and development.