From AlphaGo to AGI ft ReflectionAI Founder Ioannis Antonoglou

Ioannis Antonoglou, DeepMind founding engineer who built AlphaGo through MuZero, argues LLMs have had their AlphaGo moment but not yet their AlphaZero moment.

• LLMs hit their AlphaGo moment with ChatGPT’s launch, but the AlphaZero moment—where more compute directly yields more intelligence without human data—has not happened yet.
• AlphaGo’s move 37 (game 2 vs Lee Sedol) was initially flagged internally as a hallucination before being recognized as a genuinely creative unconventional move.
• Lee Sedol’s move 78 in game 4 exposed a real blind spot: AlphaGo misread it as a mistake, didn’t respond correctly, and lost the game.
• MuZero is currently used inside YouTube’s recommendation algorithm and was reported as part of Tesla’s self-driving system at their AI Day.
• AlphaGo ran on 1,000 CPUs and 176 GPUs; the Lee Sedol match version used 48 first-generation TPUs, described as primitive and unstable.
• Antonoglou estimates 50% on SWE-bench in 1–3 years, 90% in 3–5 years; the data wall for text is roughly one year away, with extra modalities buying another year.
• The three unsolved problems he cites: reliability and robustness at AlphaGo-level consistency, test-time planning, and in-context learning that lets agents adapt on the fly to new tools and environments.

2025-01-28 · Watch on YouTube