From games to biology and beyond: 10 years of AlphaGo’s impact

Source: DeepMind Date: 2026-03-09 URL: https://deepmind.google/blog/10-years-of-alphago/

Summary

Demis Hassabis’s retrospective on AlphaGo’s 2016 victory marks a decade of the reinforcement learning + search paradigm that seeded the entire DeepMind research tree: AlphaZero, AlphaFold 2 (3M+ researchers using the protein database), AlphaProof (silver-medal IMO performance), and AlphaEvolve (novel matrix multiplication algorithms). The post frames these as explicit stepping stones toward AGI, not isolated research artifacts.

Implications

The RL-plus-search thread. DeepMind’s narrative insists that deep RL combined with search — not pure scaling — is the durable path to AGI. That’s a thesis-level claim that runs counter to the “scale is all you need” framing dominant at OpenAI. Ten years of results (Go → protein folding → math olympiad → algorithm discovery) give it empirical weight that’s hard to dismiss.

Positioning ahead of Gemini 3 cycle. Publishing a 10-year retrospective at this point in the Gemini 3 rollout is not accidental. It ties the research credibility of AlphaFold/AlphaProof to Google’s product brand — reminding the market that DeepMind’s science track record predates and exceeds the LLM era.

Science-as-moat. AlphaFold 2’s 3M-researcher adoption figure is the strongest lock-in metric DeepMind has. No competitor has an equivalent. The post signals Google intends to keep compounding that lead (bioacoustics, genomics, weather, fusion) rather than ceding science to general-purpose LLMs.

Watch:

• Whether AlphaProof capabilities get folded into Gemini 3’s Deep Think tier
• Competitor science bets — OpenAI’s o-series math push, Anthropic’s interpretability-to-science plays
• AlphaEvolve publication and whether the matrix multiplication result generalizes to other combinatorial domains