Jul 20264 min read
How we won an AI-agent competition
Winning by OverfittingMy brother and I used a language model and the competition’s own evaluator to generate training data and win the NeurIPS 2025 Embodied Agent Interface Challenge.
AI, robotics, and what I’ve learned from building things.
My brother and I used a language model and the competition’s own evaluator to generate training data and win the NeurIPS 2025 Embodied Agent Interface Challenge.
An essay about how advances in AI and cheaper robot hardware could bring capable robots into everyday life. Also adapted into an animated film.
The story of building a tool for novelists through UCSD’s StartR accelerator, and why it never made it to market.