experience replay

Demis Hassabis argues that AGI will require active agents, with continual learning, long-term reasoning, and better memory still among the major gaps. He describes how reinforcement learning and search ideas from AlphaGo remain relevant to current foundation models, and how smaller distilled models can make AI faster, cheaper, and more practical on devices. He sees AI’s scientific impact growing through specialized tools such as AlphaFold, while emphasizing that major breakthroughs need clear objectives, large search spaces, and enough data or simulation. For builders, he recommends pursuing deep technical problems and planning for AGI to arrive during long development cycles.