Garry Tan interviews Demis Hassabis on what AGI still lacks, why memory is the core unsolved problem, how AlphaGo shaped Gemini, and what founders should build before AGI arrives.
Key Takeaways
Memory and continuous learning remain the main unsolved blockers for AGI; Hassabis explicitly frames this as unfinished science, not just engineering.
AlphaGo’s reinforcement learning approach directly shaped Gemini’s architecture and the decision to build it multimodal from the start.
Smaller models are becoming meaningfully more powerful; cheap reasoning changes the economics of what agents can do.
AlphaFold was released free to scientists worldwide and is now a template for AI-as-science-tool; the stated next target is virtual cell simulation.
Hassabis argues founders should ship now in the window before AGI arrives, targeting problems that current models still fail at reliably.
Hacker News Comment Review
Thin discussion so far; the one substantive comment focuses on leadership character rather than technical claims, expressing preference for Hassabis over other prominent AI CEOs based on public record.
