About Yann Lecun
Yann LeCun, the Turing Award winner and former chief AI scientist at Meta, has been publicly advocating for an alternative approach to artificial intelligence that moves beyond large language models (LLMs). In talks and interviews from 2025 and 2026, LeCun described LLMs as useful for tasks like code generation and information access but argued they are not a path to human-level intelligence, stating that they lack the ability to predict the consequences of their actions and cannot handle the "messy" real world. He has promoted his Joint Embedding Predictive Architecture (JEPA) and "world models" as a more promising direction, emphasizing that AI systems should learn abstract representations rather than generating pixel-level predictions. LeCun has also been critical of vision-language-action (VLA) models used in robotics, calling them "doomed" and asserting they do not work well without vast amounts of training data.
LeCun left Meta in early 2026 and became executive chairman of a new company, Advanced Machine Intelligence (AMI) Labs, which focuses on "physical AI" for robotics and industrial control. He also serves as chief scientific advisor to the Tapestry project, an open-source AI initiative under the AI Alliance that aims to collaboratively train foundation models without pooling private data. LeCun has argued that a diverse ecosystem of AI assistants is necessary to protect cultural and linguistic diversity, and that current models produced by a handful of companies pose risks to information diversity. He has described his mission as "protecting democracy" by ensuring people have access to a wide variety of information sources.
Source: AI-verified profile updated from Yann Lecun's recent appearances.
Browse all interviews →
✨ AI-enhanced transcript with speaker attribution
A
Audience Member0:07
My question would be related to consciousness. If you believe that consciousness is a requirement, consciousness as humans perceive it is a requirement for building machines that can actually achieve human level intelligence and cognitive abilities, and if yes, if you think consciousness can be computationally reduced to computation basically.
Y
Yann Lecun0:37
Okay, so I have a funny opinion about consciousness. I don't think it's that important. I think it's an illusion due to the way our brains and our minds are organized. Basically, my hypothesis, which is not a theory, it's very much a hypothesis, is the fact that I was talking about the world model in my talk. So the world model in the human brain is the prefrontal cortex, it's the entire front part of the brain. And the thing is, we cannot have a universal world model for every situation we encounter. But we have basically a single world model engine, a single piece of hardware that's prefrontal cortex. And so probably what's happening is that there is some way of configuring the prefrontal cortex to handle the situation at hand. And because we have a single engine for our world model, we can only accomplish a single conscious task at any one time. Right. So we can accomplish a lot of subconscious tasks simultaneously. We can drive a car and listen to the radio, whatever, because that doesn't require our prefrontal cortex to be particularly active, doesn't require a world model. Because those tasks have become kind of subconscious, automatic, where the psychological system went. Now, as soon as we're trying to accomplish a task that we're not used to, or requires some reasoning or planning, then we need to use a world model, and there we can only solve one such problem at any one time. We can't do multiple tasks simultaneously when we do this. So that suggests that we have only one conscious engine, if you want, like a world model that we can configure for the task at hand. So now, if we have this configurable world model, we need some sort of module to configure it. We need a module that kind of sits on top of everything else in our brain and configures our brain for the task at hand. And well, that gives us the illusion of consciousness. But it's a side effect of the fact that we only have one world model engine. If we had a brain that was 100 times bigger, perhaps we would have 100 different world models and we wouldn't need a configurator to configure the brain for that. We wouldn't have consciousness, but we would be smarter. So people associate consciousness with high intelligence. I just think that's false. I think consciousness is actually a consequence of the limitation of our brain.
A
Audience Member3:24
Actually, my question is also regarding the world model, because to me this idea is fascinating, finding a way to abstractly represent our environment, our action space. So do you believe that it is possible to come up with such a theoretical formulation without having a bit more insight into the lower level processes of our brain, so that then we can use it to perform inference and so on, to include it in the pipeline for the objective driven AI that you proposed?
Y
Yann Lecun3:49
Okay, so one characteristic of objective driven AI versus the current most AI systems that are used today, with a few exceptions, is the fact that the output is not computed by running through a fixed number of layers of a neural net. An LLM, any LLM, it's a Transformer with 40 layers or 90 layers, whatever it is, and you give it a prompt and it produces one word by just running through those layers. That's it. So the amount of computation that's devoted to producing a token or word is constant. And so what that means is that those systems cannot devote more resources to solving more complex problems than to simple problems. Means they can't really reason. And that process of running through a fixed number of layers is not Turing complete. You can't represent every single function this way, at least not without a ridiculous amount of resources. So now compare this with objective driven AI. The way the output is computed is by optimization. There is a cost function which may be complicated to compute because you have to go through a world model and objectives, but it's done by an optimization process that finds a set of output variables that minimize a set of objectives. That is Turing complete. You can reduce every computation to an optimization problem. And of course, when the problem is more complicated, the optimization will be more difficult and the system will take longer. So there you have the kind of property you observe in humans and animals, where you have to think about something longer for things that are more complicated. But it's more powerful in the end. So those are fundamental differences between feed forward inference and inference by optimization. Now this concept is not new at all in AI. Things like Bayes networks, factor graphs, they do inference by minimization of a negative likelihood, but it doesn't matter. And that's intrinsically more powerful than just running through a feedforward net. But then, is this what goes on in the brain? Not clear. Do we have to copy the brain? The answer is no, for the same reason airplanes don't copy birds. They use the same underlying principle, but they use it in different ways. So it's the same idea. You want to get inspiration from neuroscience, but not reproduce it, because you don't know which details are important if you don't know the underlying principle. If you look at birds, you might think that flapping wings is important, or that feathers are important. They're not. What's important is how you generate lift by pushing air through the air. So it's the same question for AI. What's the underlying principle that is at the basis of intelligence?
