Specialized Generalists
Build the generalist first, then distill it down — a specialized generalist beats a native specialist almost every time.
Fan's strategic argument for why GEAR pursues a general humanoid foundation model despite specialists being faster to show results. The NLP precedent: before ChatGPT, NLP was a zoo of task-specific pipelines (translation, math, coding); GPT-3/ChatGPT unified them into one generalist, which you then prompt, distill and fine-tune back down to tasks — the 'specialized generalist.' Historically the specialized generalist is far stronger than the original specialist, and easier to maintain (one API). The bet: the same trajectory will play out in robotics.
- Specialists are faster to demo but a dead end for general capability.
- Unify into a generalist, then prompt/distill/fine-tune back to specialists.
- The specialized generalist beats the native specialist and is cheaper to maintain.
- Generalist-first is slower and harder — but it's where the future is.
Made in the 'Specialized generalists' section in response to roboticists who think a general approach can't work and to the recurring Sutton 'Bitter Lesson' theme.