Yann LeCun's $3.5B bet on world models

The tech world’s worst-kept secret is officially out in the open.

Turing Award winner and former Meta chief AI scientist Yann LeCun confirmed on Thursday that he has launched a new startup called Advanced Machine Intelligence (AMI Labs). LeCun will take the role of executive chairman rather than CEO.

The top executive job goes to Alex LeBrun, the co-founder and outgoing CEO of French healthcare AI startup Nabla. LeBrun’s move was first announced in a Nabla press release and later confirmed by LeCun on LinkedIn.

Big money chasing big ideas

AMI Labs is aiming sky-high before releasing a single piece of code.

According to the Financial Times, the startup is looking to raise €500 million (around $586 million) at a €3 billion valuation (about $3.5 billion), with some reports pointing to aspirations pushing past $5 billion.

While eye-popping, astronomical valuations have become the norm for elite AI figures. Mira Murati’s Thinking Machines Lab secured a $12 billion valuation for its seed round last year, while Fei-Fei Li’s World Labs debuted with $230 million at a $1 billion valuation back in mid-2024.

LeBrun brings proven operational chops to the table. He cut his teeth on voice systems at Nuance, built and sold NLP startups to Facebook, and later led Facebook’s AI unit before founding Nabla. While stepping back to chairman and chief AI scientist at Nabla, LeBrun has already signed a commercial pact to deploy AMI’s future models inside Nabla’s clinical transcription stack.

Ditching text prediction for causal physics

To no one’s surprise, AMI Labs is focused entirely on world models.

LeCun has spent years arguing that large language models are fundamentally flawed because they rely on probabilistic token matching rather than an intrinsic understanding of the physical world. In his view, hallucinations are a structural defect of LLMs that scaling alone cannot fix.

World models, by contrast, are designed to construct an internal representation of the environment, simulate cause-and-effect dynamics, and run what-if scenarios before acting. AMI joins a competitive field already occupied by Google DeepMind and World Labs, all betting that spatial and causal reasoning represents the true path toward superintelligence.

Why it matters

For enterprise builders and developers, a functional world model could finally unlock reliable agentic workflows and embodied AI, where hallucinations carry real-world penalties. However, there is zero utility for developers right now; AMI is still putting its core team together, and commercial access is nowhere in sight.

Across the broader landscape, this is the most high-profile challenge yet to OpenAI’s text-first paradigm. If LeCun can prove that causal simulations beat brute-force pre-training, venture capital and compute budgets across the industry will aggressively rebalance away from standard LLM scaling.

We think tapping LeBrun as CEO is a shrewd pragmatic move that prevents the company from becoming just another academic ivory tower. But the risks are massive: world models lack a standardized training recipe, data pipelines for causal reasoning are notoriously difficult to scale, and justifying a multi-billion-dollar valuation without an existing product will test even the most patient investors.