Intrinsic discovery, introduced by Induction Labs in August 2026, uses reinforcement learning to push a model toward observations unlike those found so far, then trains a world model from the same base on only those self-discovered command-observation pairs — without human-chosen tasks or task-specific verifiable rewards during exploration. In their terminal setting, the approach reported about 5× more diverse states than a frozen baseline and a world model (Terminal-35B-A3B) that edged GPT-5.6 Sol (xhigh) on AgentWorldBench-Terminal-V2.