AlphaGo's Ten-Year Legacy: The Main Road Toward AGI
*A ten-year retrospective infographic originally published on zhichai.net, referencing "10 years of AlphaGo: The turning point for AI" by Thore Graepel & Pushmeet Kohli.*
The Copernican Moment: Move 37
Seoul, 2016. While Lee Sedol's "divine move" in Game Four defended human dignity, it was Move 37 in Game Two that marked the true turning point.
The move landed in a position human intuition would never consider, yet it proved the machine had developed an "intuition" that transcends human cognitive paradigms — marking the birth of non-human intelligence.
> "We no longer need to worry about whether machines can play Go; the new question is how to use such tools to answer fundamental scientific problems."
The Machine's "Intuition" and "Big Picture"
AlphaGo was not brute-force computation; it simulated human decision-making through three pillars:
| Component | Role | |---|---| | Policy Network | Intuitive move filtering — "fast thinking" | | Value Network | Position evaluation — "big picture" | | Tree Search | Forward simulation — "slow thinking" |
AlphaZero's Ultimate Leap
Freeing itself from dependence on human data, AlphaZero learned purely through self-play:
1. Rediscovery — it independently derived centuries of human opening theory (joseki). 2. Transcendence and abandonment — it found flaws in human "optimal solutions" and explored new strategies. 3. A new realm — it discovered globally optimal moves in regions of human cognitive blindness.
Gamifying Scientific Discovery
Complex scientific problems can be reframed as "games" with clear rules, then solved with the AlphaGo paradigm to find optimal paths through the solution space:
- AlphaFold — predicted protein 3D structures, solving a 50-year biology problem.
- AlphaTensor — discovered matrix multiplication algorithms better than any known human algorithm.
- AlphaDev — optimized fundamental algorithm libraries, improving global computing efficiency.
- LLMs handle conjecture — generating novel hypotheses
- Agents handle refutation — verifying and filtering through the environment
The Fork in the Road: Shortcut vs. Main Road
| LLMs — the "Shortcut" | The AlphaGo Paradigm — the "Main Road" | |---|---| | Imitates human knowledge | Generates new knowledge | | Compresses and recombines existing human data | Autonomously explores optimal solutions in the solution space | | Struggles to produce disruptive original knowledge | Has a strict verification mechanism (win/loss) | | Hard to verify in open domains (hallucination problem) | Can transcend the limits of human prior knowledge |
The Path of Convergence: Conjecture and Refutation
The future lies in fusing the two roads:
*Source reference: 10 years of AlphaGo: The turning point for AI — Thore Graepel & Pushmeet Kohli.*