Taking the human out of the loop, one step at a time. Three things a human still does in the loop: data collection, labelling and tuning, reward design.
2022–2023, Superhuman: an agent learns to choose its own experience and breaks 24 Atari human world records (GDI, ICML 2022; LBC, ICLR 2023 Oral).
2025–2026, Its own data: no collected preferences; the model improves from its own generations without collapsing (ORW-CFM-W2, ICLR 2025; ProteinZero, TMLR 2026).
2025–2026, Its own reasoning: no hand-tuned regularisation and no human-written chains of thought (ADRPO, NeurIPS 2025; CESAR, ICLR 2026).
2026, Worlds to act in: voice agents that listen and speak natively, and tool agents that learn the order of things (SpeechGym, arXiv 2026; Procedure-Aware RL, COLM 2026).
Next, Its own critic: recursive self-improvement, where each better model becomes the critic and teacher of the next.
The goal, An AI scientist: out of the loop entirely, a system that improves itself and does research on its own.