Sutton on Experiential Learning and Alignment
机器之心 · wechat · 2026-07-20
During WAIC, Richard Sutton shared his views on the future of AI in a group interview, focusing on 'experiential learning' and a 'complete mind.'
He believes the industry has been overly reliant on human data in recent years, and the next crucial phase is for agents to learn from experience. His new venture, OakLab, aims to build a 'trillion-parameter model' capable of real-time learning and planning with a power consumption of around 20 watts, using short-term commercial revenue to fund long-term research. Sutton also discussed the Robot Kindergarten project in Beijing: enabling robots to learn through trial and error in more rugged environments, much like children, rather than just performing one-off demos.
On the alignment issue, he explicitly questioned whether a unified answer for 'human values' even exists, arguing that fully aligning AI to a specific set of values could be dangerous. He even compared this capability to forced alignment among humans, suggesting it would undermine global diversity and peace.
More from AGI Musings
- FactoryAI’s Enoreyes says model distillation is basically unstoppable — LangChain · 2026-07-21
- Andrew Blumberg says formalization without interpretability is not science — AlexKontorovich · 2026-07-21
- Ken Ono says AI is forcing mathematicians to rethink how discovery works — soumitrashukla9 · 2026-07-21
- Open-source labs could distill a state-of-the-art model to 32GB or 80GB VRAM, the post argues — bookwormengr · 2026-07-21
- Two US companies are now using superintelligence to speed up the next generation of models — yacineMTB · 2026-07-21
- MIT Sloan says information, national security and finance are most exposed to AI — Exp_Mark · 2026-07-21