Yann LeCun just raised $1.03 billion for AMI Labs to build “world models” instead of large language models. The Turing Prize winner left Meta to pursue AI that learns from reality rather than text patterns. His new company landed a $3.5 billion pre-money valuation betting against the entire generative AI playbook.
LeCun’s co-founder Alexandre LeBrun admits this could take years to commercialize. No product in three months, no revenue in six months, no $10 million ARR in twelve months. That timeline puts AMI Labs completely at odds with the venture capital model that has funded almost every other AI startup.
World models sound like a probable future for AI. Instead of predicting the next word in a sentence, a world model would predict how the world actually works. Said differently, a world model would “understand” causality instead of just correlation. I put understand in quotes because I don’t have a good way to define (or personally understand) how an AI model will understand or even if they ever can understand the way humans do.
I love this journey, but the tech also terrifies me. I keep asking myself what scares me more: a future where AI fundamentally misunderstands how the world works, or a future where it actually understands the world completely. Both scenarios have kept me awake recently.
LeBrun predicts “world models” will become the next buzzword within six months. Every AI company will rebrand itself around world modeling to chase funding. The distinction between real world modeling and marketing world modeling will matter enormously.
LeCun has been arguing against the current LLM approach for years. Now he has $1.03 billion to prove his alternative works. The success or failure of AMI Labs will determine whether the entire AI industry has been heading in the wrong direction.
Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.