Yann LeCun is leaving Meta to bet his career on world models. His thesis: LLMs are a dead end, and the real path to intelligence is through systems that learn by watching and interacting with the world, not just reading about it.
I think he’s wrong. And I think his track record on LLM predictions backs me up.
LeCun’s Prediction Problem
A couple of years ago, LeCun said that even a hypothetical GPT-5000, trained only on text, could never figure out that if you put a book on a table and then push the table, the book moves with it. His reasoning was that there’s no text on the internet that explains that.
That prediction has already been falsified. Modern models handle that kind of tabletop physics scenario just fine. Since then he’s doubled down with claims like LLMs are a dead end, we’re nowhere near cat-level intelligence, and nobody in their right mind will be using LLMs like the ones we have today within 3-5 years.
Meanwhile, the actual frontier of AI capability and economic value is almost entirely LLM-centered. Meta itself is pouring billions into Llama. OpenAI, Anthropic, Google, all betting heavily on language models. The market has spoken pretty clearly on this one.
What World Models Actually Are
People use the term world model in confusingly different ways, and this confusion muddies the entire debate.
In robotics, a world model is basically a learned game engine. DeepMind’s Genie models are a good example: given a current frame and an action, they predict the next frame and let you play around in a simulated world. That’s useful for training robots and making games. It’s not a full brain. It’s an environment simulator.
LeCun’s JEPA work is another flavor: a predictive model that learns abstract latent representations of how the world changes over time. The idea is that instead of predicting raw pixels, you predict in a compressed latent space, which should be more efficient and generalizable.
But in the looser sense that alignment researchers use, world model just means an internal model of how the world works that you can use to make predictions. On that definition, frontier LLMs already have one. They compress most of the internet into internal representations that capture facts about physics, software, history, social norms, and more.
Different types of world models excel at different tasks. LLMs already handle abstract reasoning remarkably well.
Why LLMs Already Contain World Models
The structure of the world shapes the structure of our text. If gravity worked differently, the stories people tell, the engineering manuals they write, and the metaphors they use would all look different. By learning to compress and predict that whole distribution of text, a model can infer a lot of underlying regularities that nobody ever spells out explicitly.
The same way you learned most of English grammar long before anyone gave you a formal grammar book, these models learn a lot of physics without needing a sentence that literally says if you put a book on a table and push the table, the book moves with it.
On the claim that LLMs can’t learn anything that isn’t written in text: I think that’s just wrong in principle. The implicit structure is there. The causal relationships are encoded in how we describe events. The physics leaks through in every story about dropping something or pushing something or building something.
By the time a kid is five, they’ve processed vastly more sensory data than any LLM has seen in text. That’s true. But it doesn’t follow that text-trained models can’t learn physics. It just means they learn it differently, through the traces that physical reality leaves in language.
The Future Is Hybrid, Not Replacement
Video-style world models will mostly serve as training media and tools. The systems that actually act in the world and make decisions will be LLM-like for a long time.
You can already see this in robotics. Figure’s humanoid robots use a large language model as the brain for reasoning, planning, and dialog, with other networks handling raw motor control. I expect that pattern to continue: a big multimodal LLM at the center, acting as the decision-making core, calling out to specialized world-model modules when it needs detailed physical simulation.
This isn’t LLMs versus world models. It’s LLMs plus world models. The LLM handles the high-level reasoning, planning, and communication. The world model handles the detailed physics simulation when you need it. They complement each other.
LeCun’s framing of this as an either-or choice is part of why I think he’s missing the mark. The architecture that wins won’t be a pure world model that replaces LLMs. It’ll be a hybrid system where LLMs remain central.
The Superintelligence Question
Whether superintelligence is achievable depends entirely on how you define intelligence.
If you define true intelligence as soul-level, image-of-God rationality, the kind that involves knowing and loving God, being morally accountable, entering covenant, and being redeemed, then AI will never be intelligent or superintelligent in that sense. Machines will stay tools, not persons. As a Christian, I don’t buy the idea that humans are just biological computers. We are body and soul, made in the image of God. No amount of GPU will turn a neural net into an image-bearer.
But if you define intelligence more functionally as the ability to achieve goals effectively across a wide range of tasks, given your resources, then AI is already clearly intelligent. In many narrow areas it’s already superhuman. A calculator is superhuman at arithmetic without having a soul. There’s nothing in the Christian worldview that rules out AI vastly outperforming humans in problem-solving capacity.
On that functional definition, AGI is just a system that can match or beat a competent human at basically any cognitive job you could hire a human to do. Superintelligence is the same thing but far past us: something that beats our best scientists, engineers, and strategists at almost everything that matters. There’s no good argument that this is impossible in principle.
Why I’m Not Worried About AI Doom
This also shapes how I think about AI doom conversations. Christ is not sitting in heaven wringing His hands about GPUs. Psalm 110 says He rules in the midst of His enemies. First Corinthians 15 says He must reign till He has put all enemies under His feet. The Great Commission is to disciple the nations, not to hang on until technology destroys us.
I do not believe in a future where some superintelligent model permanently destroys human meaning or slips out of God’s control. History belongs to Christ, and AI is just one of the tools He lets us develop along the way.
Same on the other extreme. God calls work good. From Genesis 1, the mandate is to be fruitful, fill the earth, and exercise dominion. Ecclesiastes spends a lot of time on the idea that honest labor under God is a huge part of the meaning of life. I don’t buy the secular dream that AI will usher in a world where no one works and robots do everything. That’s not how Scripture talks about human purpose. New tools have always shifted what work looks like, but they don’t erase the basic pattern that people are called to build, cultivate, and rule under God. AI will change a lot of jobs, but it won’t abolish vocation.
My Bet
If you’re betting on a single AI brain for the next decade, it’s not going to be a pure world model. It’s going to be a big, multimodal LLM at the core, one that can reason, plan, and communicate, calling out to specialized world model modules when it needs detailed physical simulation or perception.
LeCun is a brilliant scientist. His past work on convolutional neural networks was foundational. But his predictions about where LLMs would hit walls have been consistently wrong. I respect his contributions, but I don’t treat him as a reliable prophet about where things are going.
The future is hybrid: LLMs plus world models, not one replacing the other. And none of it will produce a soul.
