The CEO building AI that actually models how the world works refuses to use the words everyone else is fundraising with.
The Summary
- Alexandre LeBrun, CEO of AMI Labs (backed by Yann LeCun), won't call his company's AI "AGI" or "superintelligence" despite building world models
- LeBrun argues these terms are marketing fog that obscure what AI actually does and doesn't do
- AMI Labs is focused on world models: AI that predicts physical reality rather than just generating text
- The framing fight matters because it determines what gets funded, what gets regulated, and what people expect from AI
The Signal
Alexandre LeBrun sold his company to Facebook in 2015, worked under Yann LeCun at Meta AI, and now runs AMI Labs, a stealth-mode startup building world models. World models are AI systems trained to predict what happens next in physical space, not just text space. Drop a ball, the model knows it falls. Push a glass near the table edge, the model sees the problem. It's the difference between an AI that can write about physics and one that understands it enough to navigate a kitchen.
LeBrun won't use "AGI" or "superintelligence" to describe what they're building. His reasoning cuts through the hype cycle: these terms promise general intelligence and superhuman capability without defining what either means. They're fundraising words, not engineering specs. When you call something AGI, you're claiming it can do anything a human can do. When you call it superintelligence, you're claiming it does everything better. Neither is measurable. Neither ships.
"Marketing fog obscures what AI actually does and doesn't do."
The world model approach sidesteps this entirely. AMI Labs isn't trying to build artificial general intelligence. They're building AI that models physical reality well enough to be useful in the real world. That means:
- Robots that understand object permanence and physics
- Agents that can predict consequences of actions in physical space
- Systems that know when they don't know, because the model breaks down
This matters more than it looks like. The AI field is splitting between companies chasing the AGI marketing narrative and teams building specific, measurable capabilities. OpenAI calls everything a step toward AGI. Anthropic hedges with "helpful, harmless, honest." Google's Gemini positioning shifts every quarter. Meanwhile, LeBrun and LeCun are building models that either predict physical reality accurately or they don't. Pass/fail. No narrative required.
The fundraising implications are obvious. If you call your AI "AGI," you get larger rounds at higher valuations because you're selling the dream of general intelligence. If you call it a world model, you get smaller rounds from people who understand what physical prediction engines unlock. The former attracts growth equity. The latter attracts people who've built robots or worked in manufacturing automation or dealt with real-world AI deployment hell.
"World models either predict physical reality accurately or they don't. Pass/fail."
LeBrun's bet is that world models become the foundation layer for the agent economy. Language models got us ChatGPT and copilots. World models get us agents that do things in physical space: manufacturing, logistics, construction, elder care. The agent that can predict where the wrench is, how the bolt will turn, and what breaks if it's overtightened. That agent is worth more than one that writes poetic emails about the concept of wrenches.
The Implication
Watch where the term "world model" pops up in the next six months. If AMI Labs is right, other labs will quietly shift from AGI messaging to physical prediction capabilities. The companies that matter in 2027 won't be the ones promising superintelligence. They'll be the ones whose agents can navigate a warehouse, assemble furniture, or pack a moving truck without breaking your grandmother's china.
If you're building or funding in this space, the question shifts from "how close to AGI" to "what physical tasks can your model predict reliably." That's a harder question with a clearer answer. Which means it's probably the right one.