Google DeepMind's Demis Hassabis: How AI could become dangerous (with Reid Hoffman) | Summit '26

The gist

1. Hassabis says AlphaFold 2, finished in 2020, is used by 4 million researchers, and DeepMind wants to extend it to toxicity and other drug properties.
2. He explains AlphaZero went from random play to beating every chess entity in under 12 hours by playing itself millions of times.
3. Hassabis warns that closed-loop recursive self-improvement, with no human reviewing code, is the riskiest path to AGI and could produce an alien intelligence.
4. He calls for hardened sandboxes, better monitoring, third-party auditors, and labs red-teaming each other, warning an AI "oil spill" incident is otherwise likely.
5. Hassabis argues intelligence and consciousness can be separated, so labs should build intelligent tools now and delay any conscious AI until society debates it.