AMI Labs is skipping the 'AGI' race to build a 'world model' for robots

Alexandre LeBrun will not enter the AGI race. The CEO of AMI Labs was in Seoul in July 2026 for the ICML machine-learning conference, and the label he wants to avoid is the one most of the industry is chasing. "AGI" and "superintelligence," he says, are vague and trend-driven. "Next time we'll switch to something else."

AMI Labs is building what LeBrun calls a "world model": a model trained to predict the next state of the physical world, not the next word in a sentence. The company frames world models and large language models as "complementary, not replaceable." They sit on different axes. An LLM is a text-prediction machine. A world model is a state-prediction machine. A robot running only on an LLM, LeBrun says, is "not safe right now," because there is "no brain in the hardware" that knows a child is standing in front of it.

The AGI finish line is text-style scaling reaching a single capability ceiling. LeBrun's argument is that it does not. There is a separate physical-world axis that text scaling does not touch, and a model trained only on text is structurally blind to it.

LeBrun's prior company was the AI health startup Nabla, and he argues large language models cover "only 1% of healthcare" because a model that cannot reason about a patient's chart in context will not be a healthcare system. World models, in his framing, cannot be trained in a lab.

AMI Labs closed a $1.03 billion round in March 2026 at a $3.5 billion pre-money valuation. The company has not shipped a model and has not committed to a launch timeline.

LeBrun was in Seoul scouting partners. South Korea's June 2026 industrial plan commits roughly $880 billion to chips, AI data centers, and physical AI, broken down in the underlying coverage as about $518 billion in memory fabs, $52 billion in high-bandwidth memory, and $356 billion in AI data centers, over $900 billion in total. World models, LeBrun argues, cannot be trained in a lab.

AMI is pre-product. If the eventual architecture turns out to be an LLM with a physical-world wrapper, or if LLM vendors close the gap by adding physical context to existing stacks, the world-model-as-separate-stack claim collapses. The bet depends on AMI shipping something genuinely distinct.