AMI Labs puts Yann LeCun’s world model agenda into startup form

Yann LeCun has spent years arguing that next-token prediction won’t get AI very far in the physical world. Now there’s a startup built around that view.

AMI Labs, short for Advanced Machine Intelligence, has surfaced with a plan to build foundational world models for domains where bad guesses are expensive: robotics, industrial control, wearables, and healthcare. LeCun is executive chairman. The CEO is Alex LeBrun, former Nabla chief, Wit.ai cofounder, and a onetime LeCun colleague at FAIR. Reports have also linked former Meta Europe VP Laurent Solly to the team.

The company is reportedly circling a valuation around $3.5 billion. That looks aggressive, but the market is already treating this category like a land grab. Fei-Fei Li’s World Labs has reportedly been raising at around $5 billion. Investors are clearly willing to bet that AI beyond chatbots needs a different stack.

That may prove right. It’s early, but the premise here is sturdier than most AI startup branding.

Why LeCun keeps pushing world models

LeCun’s position has been consistent. Language is a thin slice of intelligence. Humans and animals learn by interacting with the world and building internal models of cause, effect, space, time, and consequence. If you want machines that can plan, control systems, or act safely in messy environments, next-token prediction is a weak foundation.

A world model tries to learn how an environment changes over time, especially when actions are involved. In practice, the idea is straightforward:

• take observations from cameras, sensors, audio, telemetry, or clinical systems
• compress them into a latent state
• learn how that state changes when an agent acts
• use that learned dynamics model for prediction, planning, and control

This sits closer to model-based reinforcement learning and classical control than to the standard LLM product stack. It also lines up with LeCun’s long-running work on JEPA, or Joint Embedding Predictive Architecture, where the model predicts future representations instead of generating every pixel or token.

That matters. If you’re training a robot or a medical assistant, you care about state, constraints, and outcomes. A plausible paragraph about those things doesn’t help much.

AMI is aiming at the right problems

The company’s website says it’s building systems around perception, memory, reasoning, and planning. That language is vague enough to fit on any AI homepage. The target sectors make it more concrete.

Industrial process control, robotics, wearables, and healthcare all have the same problem: the system has to operate under uncertainty, over time, with real consequences. "Usually sounds right" is not a serious standard there.

That’s where world models look appealing, at least on paper.

A decent world model stack usually includes:

• a multimodal encoder for video, sensor streams, time series, or text
• an action-conditioned dynamics model, something like f(s_t, a_t) -> s_(t+1)
• memory for partial observability and long horizons
• a planner or controller, often latent-space MPC, CEM, or learned policy layers
• safety constraints, either hard-coded, learned, or attached as costs and invariants

The practical advantage is simple: these systems can evaluate possible futures before acting. In robotics, that means trying action sequences in latent space instead of crashing into the world and hoping recovery works. In healthcare, it could mean reasoning over longitudinal signals and workflow state instead of reacting to isolated text prompts. In industrial settings, it means catching excursions before an operator gets an alarm flood at 2 a.m.

That fits closed-loop systems better. It’s also much harder to build.

The hard part is stable, useful dynamics

Anyone can claim to model the world. The difficult part is learning dynamics that remain useful outside a benchmark.

A lot of work in this area tries to avoid modeling raw sensory detail directly. Predicting future pixels is expensive and often beside the point. You end up optimizing texture fidelity when what you need is semantic state. JEPA-style setups target future embeddings instead, which gives the model room to learn what matters for action.

A stripped-down training loop looks roughly like this:

s_t = encoder(obs_t)
s_pred = dynamics(s_t, act_t)
s_target = stop_grad(encoder(obs_t_plus_1))

loss = consistency(s_pred, s_target) + safety_penalties(s_pred) + physics_regularizers(...)

That’s the tidy version. Real systems get messy quickly.

You need synchronized multimodal data. Timestamp drift between video, IMU, force sensors, or action logs can wreck the whole setup. You need decent memory because the system rarely sees full state. You need planners that don’t fall apart when the learned model is slightly wrong. And if you’re operating in healthcare or industrial control, you need bounded failures, audit trails, and some explanation for why the system took an action.

A lot of startup decks get hazy right there. In consumer chat products, "safety" often means filtering output. In industrial control or clinical workflows, it means constraints the model cannot casually violate.

The Nabla link matters

One of the more interesting details is the exclusive partnership with Nabla, announced in December, which gives Nabla privileged access to AMI’s models for agentic healthcare. LeBrun moved into a chief AI scientist and chairman role at Nabla while taking over AMI.

That suggests AMI already has an applied wedge, not just a research story.

Healthcare makes sense as an early proving ground if the company is serious about memory, multimodal reasoning, and safety-constrained planning. Clinical work is full of longitudinal context, incomplete observations, risk trade-offs, and workflow state. LLMs are useful there for summarization and conversational UI, but they’re shaky as the main reasoning substrate. A system that can track patient state over time, integrate signals beyond transcripts, and operate under policy constraints would be valuable.

It’s also one of the hardest places to prove reliability. If AMI shows real progress there, people will notice.

This probably ends up as a hybrid stack

World-model startups often get framed as anti-LLM, partly because LeCun has been one of the loudest critics of LLM maximalism. That misses the obvious architecture pattern.

For most products, the likely stack is hybrid:

• LLMs for interface, natural language interaction, tool calling, and procedural decomposition
• world models for state estimation, prediction, planning, and control under uncertainty

That division makes sense. LLMs are good at operating over language, code, and loosely structured tools. They are far less convincing when the job is latent state tracking over long horizons in a partially observed physical system.

If AMI works, the company probably won’t "beat" LLMs. It will supply a missing layer for problems where LLMs have always felt bolted on.

Why Paris matters

AMI will be headquartered in Paris, with offices in Montreal, New York, and Singapore. France’s political class has already leaned into the announcement, which is predictable and still worth noting.

Paris is building a real AI cluster. FAIR talent, Mistral, H, national backing, European capital, and LeCun’s name give it some density. If you care where foundational AI research may turn into applied infrastructure over the next few years, Paris belongs in that conversation.

That doesn’t solve the hard parts. Compute costs, data access, evaluation, and distribution still decide whether this category works. World-model companies can burn money very fast.

What developers should watch

The interesting questions around AMI are engineering questions.

Data pipelines

World models are data-hungry in a specific way. The issue isn’t only volume. It’s temporal integrity.

Teams building in this direction should care about:

• precise time alignment, often with PTP or equivalent clock discipline
• storage formats that preserve multimodal sequence data cleanly, such as Parquet or Zarr
• versioned datasets and replayable event streams
• logs of both observations and actions, because actions are part of the state transition problem

Without that, the model learns a fuzzy fiction.

Simulation and sim-to-real

Expect heavy use of tools like Isaac Sim, MuJoCo, PyBullet, and domain-specific simulators. But simulation only goes so far. If the learned dynamics overfit to a clean sim world, deployment gets ugly. Domain randomization helps. Sensor noise modeling helps. Learned latent adaptation helps. None of it fully closes the gap.

Safety and standards

If AMI wants industrial and healthcare revenue, standards will matter as much as model quality. In process control, names like IEC 61508 and ISA/IEC 62443 show up quickly. In healthcare, auditability, data governance, and regulated workflow integration are unavoidable. Any serious product here needs traceability around decisions and constraints.

That work is slow and annoying. It’s also where a lot of real defensibility lives.

Promising category, thin evidence so far

The valuations tell you investors think world models could shape the next phase of AI infrastructure. Maybe. The argument is coherent. Physical systems need models of state and consequence, not polished text generation.

But this field has a habit of looking better in papers than in production. Long-horizon prediction drifts. Learned planners exploit model errors. Real environments are partial, noisy, and adversarial in all the ordinary ways. A hospital workflow changes. A factory line gets a new sensor. A robot’s gripper wears down. Your elegant latent dynamics model is now reasoning about a world that no longer exists.

AMI matters for two reasons. It gives LeCun’s long-running critique of LLM-first AI a commercial vehicle with serious money and experienced operators. And it puts pressure on the market to separate language fluency from actual environmental intelligence.

For engineers, that’s the part worth watching. If your system has to perceive, remember, predict, and act in the real world, chatbot architecture was never going to carry the whole load. AMI is betting a few billion dollars that the industry is ready to accept that.