Yann LeCun reveals he is starting a new startup.
The NYU professor has admitted to the rumors that he is creating a new AI-driven startup focused on “world model” type AI called Advanced Machine Intelligence, and has named longtime collaborator Alex LeBrun as the CEO.
LeCun, who will be executive chairman, is said to be seeking a bet of over $5B as investment in next‑gen AI architectures heats up.
(The people I’ve spoken to wouldn’t put a figure on how much AMI is raising, but the various stories point to raises at the hundreds of millions of euros mark and valuations in the multi‑billion range.)
The Financial Times, which has been cited by industry observers as a publication tracking early valuations, highlights robust demand for frontier AI bets backed by star scientific leadership.
What AMI Aims to Build with World Model Architectures
World models seek to learn an internal model of how the world changes, letting systems predict outcomes, plan many steps ahead, and test potential “what if” scenarios before acting. This contrasts with today’s preeminent large language models, which are good at pattern recognition but tend to confidently produce fabrications when pushed beyond their training distribution.
LeCun has repeatedly made the case that predictive architectures and self‑supervised learning are keys to the robust reasoning of which human intelligence is capable. His lab’s research on joint embedding predictive architectures is an important step toward AI that learns from streams of video, audio, and sensors to form common‑sense priors. If AMI is successful, it will address hallucinations by conditioning outputs closer to the flexibility of the learned dynamics than surface correlations.
Precedents exist. The research at Google DeepMind on MuZero and Dreamer showed that model‑based systems could plan effectively. World Labs from Fei‑Fei Li also walks the world‑model line, with applications in robotics, scientific inquiry, and embodied agents. AMI enters this contest by suggesting that richer, multimodal prediction will enable more dependable autonomy and agentic behavior.
A Big Bet in a Pricey Race for Advanced AI Funding
Constructing a good‑enough world model is expensive. Training at scale routinely involves clusters of tens of thousands of top‑of‑the‑line GPUs, with specialized data production pipelines that run for months on end, resulting in nine‑ or even ten‑figure price tags. That’s the reality that drives startups to shoot for multi‑billion‑dollar valuations out of the gate—they need balance sheets that can withstand multiple training cycles.
Investors have demonstrated an appetite to support category‑defining efforts. Anthropic locked up commitments worth several billion from strategic partners, while xAI announced a record‑breaking $6B round at a nosebleed valuation. OpenAI and other established players are still hoarding capital and compute, which will raise the bar for entry for any would‑be competitor concerned with capability, instead of merely niche specialization.
Leadership and Early Partnerships at AMI Labs
At the operational helm of AMI is Alex LeBrun, an engineer‑founder with deep speech and natural language qualifications. He was at Nuance during the dawn of Siri and sold a startup to Facebook, where he went on to head three AI teams before co‑founding his current healthcare AI company, Nabla. That résumé makes him a leader with proven experience translating research into production systems and regulated markets.
Nabla has promised to leverage AMI’s models as they arrive, an early sign of commercial traction and a testing ground in clinical workflows where trust is crucial. The Paris‑based startup has raised roughly $120 million from investors such as Build Collective, HV Capital, Highland Europe, and Cathay Innovation; the company says ARR has grown significantly under LeBrun’s leadership.
LeCun is the scientific drawing card. The New York University professor and former Meta chief AI scientist was awarded the A.M. Turing Award for seminal contributions to deep learning, in particular pioneering and sustained work on convolutional neural networks. His public criticisms of LLM‑only approaches, and support for model‑based reasoning, established its philosophical tone.
Competitive Landscape and Key Risks Facing AMI
AMI has a lot of heavyweight competition in both research and deployment. DeepMind is still developing planning‑capable systems. World Labs advances multimodal learning with a focus on real‑world understanding. Groups working on autonomous cars like Wayve and major robotics labs are also investing in world‑modeling as a way to enhance perception and control, making this more of a general platform race than a narrow product battle.
The key may well be execution: how fast AMI can convert theory into benchmarks, demos, and developer‑ready APIs. Look for early indicators such as:
- White papers
- Open‑sourced components or evals
- Cloud and silicon partnerships
- Showcase apps where prediction and planning offer compelling safety and cost advantages
For now, the lines are drawn: a high‑quality scientific founder, an operator‑CEO with shipping experience, a somewhat over‑the‑top fundraise, and a thesis that the next wave of AI will do things like understand the world instead of merely autocomplete it. If AMI can make good on that promise at scale, the $5B‑plus valuation stands to look a little short.
