jim fan

Robotics will follow the exact three-step LLM playbook: (1) world model pre-training on egocentric video (Ego-Exo, Veo3) where physics emerges from next-pixel prediction, (2) action fine-tuning (Dreamer) collapsing simulation onto real robot actions, (3) massive RL in neural simulators (Dream Dojo) breaking the environment bottleneck. Physical Turing test in 2-3 years, lights-out factories via physical API, recursive self-improvement by 2040.

Models like Veo3 and Dream Dojo learn gravity, buoyancy, lighting, and visual planning purely from pixel prediction at scale, then function as real-time neural simulators taking continuous actions and outputting next frames/sensor states — no physics equations or graphics engines required.

Massive compute enables synthetic environment generation (real-to-sim-to-real, Dream Dojo) which produces infinite training data, creating a self-reinforcing loop where 'compute = environment = data' — structurally advantaging Nvidia's full-stack GPU/simulation/software platform.

A standardized physical API will abstract robot fleets into configurable software endpoints, enabling frontier AI agents to orchestrate lights-out factories and automated wet labs — turning atoms into programmable primitives.