Induction Labs

Induction Labs is a San Francisco research lab building foundation models that learn on their own, without human demonstrations or labeled data. This launch introduces intrinsic discovery, a training method that lets a model explore open-ended computer environments and learn from what it observes and predicts, and the team is using it to train world models that simulate software behavior. The company describes itself as building intelligence that is intrinsically motivated to learn about the world. The first concrete result of the method is a world model that outperforms a leading frontier LLM at simulating terminal environments, extending the direction Induction Labs began earlier this year with imagination models, a foundation model architecture that learns from internet video by predicting the future in a learned representation space, with reinforcement learning layered on top to improve competence . That earlier system, Photon-1, a sparse 106B active 5B MoE transformer trained on 18 years of computer demonstration video , showed the approach could produce useful computer-use agents, and intrinsic discovery is the training loop that pushes the same models to keep learning once they are dropped into unfamiliar software. The relevance for founders and operators is practical. Anyone building agents that need to act inside real desktops, shells, or web apps eventually hits the wall of scarce action-labeled data, and a self-directed learner that treats a Linux VM or a terminal as its own curriculum is a different bet on how that wall gets cleared. Induction Labs was founded by Jonathan Li, CEO, and Jeffrey Li, CTO, and went through Y Combinator in the Summer 2025 batch.