Paris, 23 March 2026 — Yann LeCun, the Turing-award winning scientist who spent a decade as Meta’s chief AI strategist, has quietly closed one of the largest seed rounds in European tech history. His new company, Advanced Machine Intelligence (AMI), has secured just over one billion dollars to build AI that learns like a child: by predicting how objects, forces, and agents behave in the physical world.
The funding, led by a sovereign-wealth fund from the United Arab Emirates and a consortium of European life-insurance giants, is a direct challenge to the “bigger-is-better” philosophy that has dominated Silicon Valley since the arrival of ChatGPT. Instead of training ever-larger language models on internet text, AMI will focus on what LeCun calls “world models” — systems that continuously update an internal simulation of reality and use it to plan, reason, and adapt.
Why LeCun Thinks LLMs Have Hit a Wall
Large language models are brilliant at next-token prediction, but they remain “aliens from the internet,” LeCun told reporters at AMI’s airy headquarters in the 13th arrondissement. “They can write Shakespearean sonnets about gravity, yet they don’t know that a rock released from your hand will fall to the ground.”
Empirical studies back up his skepticism. In one widely cited experiment, GPT-5 still fails about 38 % of physics questions that an average 11-year-old answers intuitively: will a stack of uneven books topple left or right? How many degrees will a bicycle lean if the rider shifts her weight? These shortcomings matter when AI is asked to control robots, drive cars, or manage power grids.
LeCun’s critique is not new; he has delivered variants of it in keynote slides since 2019. What is new is the money. With a billion-dollar war chest, AMI can afford the compute, talent, and data-collection campaigns that world-model research has always lacked.
Inside AMI’s Blueprint for Common-Sense AI
AMI’s 120-page technical prospectus, shared selectively with investors and reviewed by this blog, outlines a four-stage roadmap:
- Perceptual Abstraction: Convert high-dimensional sensor streams (video, lidar, haptics) into compact symbolic objects and relations.
- Causal Dynamics: Learn differential equations that govern how those objects interact, using self-supervised prediction similar to video compression.
- Counterfactual Simulation: Run fast “what-if” rollouts to evaluate action sequences before executing them in the real world.
- Grounded Language Interface: Attach a small, specialized LLM that translates between natural-language requests and the structured world model, avoiding the hallucinations that plague text-only systems.
The company will open-source stage-one models under a permissive license, betting that community contributions will improve the physics engine faster than any single lab could manage. Commercial licenses for stages three and four will fund further research.
Meet the Startups Already Proving the Concept
AMI is not operating in a vacuum. A handful of research groups have already demonstrated that hybrid architectures—blending neural nets with explicit reasoning—can outperform pure LLMs on tasks that require common sense.
- QuData’s DemonScript: The Ukrainian-American start-up has built a multi-valued logic language that encodes spatial, temporal, and causal constraints. In benchmarks, DemonScript answers “micro-story” questions (e.g., “Will the egg break if the table is flipped?”) with 91 % accuracy, compared with 63 % for GPT-5.
- DeepMind’s Adaptive Agent: Using a similar world-model approach, the London lab trained a quadruped robot to navigate unfamiliar staircases after only 30 minutes of real-world interaction, a task that took a reinforcement-learning baseline six hours.
- Toyota Research’s Physics-GAN: By pairing a generative adversarial network with a physics simulator, Toyota cut the data needed to train an industrial robot arm by 70 %.
AMI’s twist is scale. The company has reserved 40 megawatts of compute in a former nuclear-facility cooling plant outside Lyon and has hired more than 60 PhDs in computational physics, neuroscience, and symbolic reasoning since January.
What a Billion Dollars Actually Buys in 2026
Training frontier models has become astronomically expensive. Industry estimates put the cost of a single GPT-5-scale training run at $250–300 million. LeCun insists AMI will not chase parameter counts. Instead, the budget allocates:
- 48 % to data acquisition: building a 500,000-square-foot studio full of robotic arms, kitchen appliances, toys, and tools whose interactions are filmed by 10,000 high-fps cameras.
- 22 % to compute: a modular cluster of 32,000 NVIDIA B300s, cooled by the plant’s existing water circuit.
- 20 % to talent: luring roboticists from Boston Dynamics, neuroscientists from DeepMind, and symbolic-AI veterans who once worked on Cyc.
- 10 % to safety and governance: continuous red-teaming, formal verification of core physics modules, and an external ethics board chaired by Stuart