Revisiting AlphaGo Zero: The Real Cost of 100:0 and the Limits of "No Human Knowledge"

In March 2016, Lee Sedol lost 1:4 to AlphaGo. In May 2017, Ke Jie lost 0:3. But both versions of AlphaGo shared one thing: they first learned from humans — supervised learning on amateur and professional game records. Five months later, DeepMind handed over volume 550 of *Nature*: a new program called AlphaGo Zero that read not a single human game, started from random moves playing against itself, and after three days played 100 games against the version that beat Lee Sedol — winning 100:0. The keyword in the paper's title was "without human knowledge," but the truly precise phrase in the abstract was "without human data, guidance or domain knowledge beyond game rules." Those three words — beyond game rules — are where the paper's real complexity lies, and what this re-reading aims to unpack.

How the machine plays

AlphaGo Zero is a hybrid of "intuition plus slow thinking": a neural network provides the intuition; a tree search thinks ahead. The network is a single network with two heads — a policy head looks at the 19×19 board and instantly outputs a win-probability feel for every position; a value head judges the whole board and assesses who is ahead. The search (Monte Carlo tree search) runs 1,600 simulations per move, pushing down branches the network considers promising. This "System 1 proposes candidates, System 2 verifies" structure mirrors a human player's fast and slow thinking in engineering form — except every act of "thinking ahead" by System 2 costs 1,600 simulations of compute.

How it gets stronger

The answer: the student copies the teacher's answers, and the teacher is yesterday's self. For each move, MCTS searches 1,600 plies ahead and produces a more reliable move distribution — the "search's answer." The game ends and the outcome is revealed — the "final win/loss." One head of the network fits the search distribution; the other fits the outcome. The fitted network then plays self-play games tomorrow; the search, with a better prior, becomes sharper; self-play quality rises; so tomorrow's answers are better than today's. It climbs in loops. The elegance: what is learned is the output of search, and search quality depends on the just-trained network — both the data and the teacher are self-produced.

I ran a minimal tic-tac-toe reproduction locally (about 40 lines of core code, rollouts replaced by uniform random play, no neural network): both sides used MCTS with 400 simulations per move over 300 self-play games — first-player win rate 80%, only 20 draws. But perfect tic-tac-toe play is always a draw. That 80% shows 400 simulations are far from a perfect solution; simulation count *is* playing strength. If even tic-tac-toe, with fewer than 6,000 total states, behaves this way, the marginal return of "thinking one move deeper" under Go's ~10^170 branching factor is easy to imagine.

The scoreboard

Three lines suffice: three days, 4.9 million games, 100:0 against AlphaGo Lee (Elo 3739); twenty-one days, reaching AlphaGo Master level (Elo 4858, the Wuzhen version); forty days, Elo 5185, surpassing all previous versions. One easily overlooked entry: hardware. AlphaGo Lee ran distributed on 48 TPUs; AlphaGo Zero's 40-day version ran on a single machine with 4 TPUs. Human knowledge cut to zero, compute cut twelvefold, and strength went up substantially. The paper's sharpest trade is here: human game records are a low-quality prior — discarding them let search find better knowledge more cheaply.

Go theory

The Go-theory section is the paper's most watchable part. Early in training it played nonsense; after tens of thousands of games, human joseki naturally appeared in its self-play — not taught, but converging to the same local optima under the same game structure. Later it adopted new joseki humans don't play, and professionals judged those variations sound. And ladders — the most basic tactic in Go, every beginner's first lesson — it grasped only very late: early in training it even played ladders wrong, and the paper plotted the disappearance curves of these beginner errors (ladders only stabilized correct around 300,000 games). Joseki emerge automatically, ladders arrive late — the ordering itself shows that the sequence of knowledge acquisition is determined by game structure, with no relation to human pedagogy.

What Go knowledge it actually used

Counting carefully, five places. First, legal-move generation rules — which points can be played, how captures work — encoded into the environment. Second, game-end determination and scoring — Chinese rules, 7.5 komi, how wins are counted, hardcoded. Third, board symmetry — data augmentation applied eight symmetric transforms to every position, effectively telling the network the board has no special direction. Fourth, search hyperparameters — 1,600 simulations, the PUCT exploration constant, Dirichlet noise at the root, all human-tuned. Fifth, the self-play protocol — temperature sampling for the first 30 moves to preserve diversity, evaluation every few hundred iterations with 400 games between new and old networks, replacement only if win rate exceeds half.

So the accurate reading of "no human knowledge" is: no human game records or commentary, but the game rules, board geometry, compute budget, and training regime are all human-designed priors. The network learning Go by itself is true — but the physics of the world it learns in was written out of human Go knowledge. *Tabula rasa*: what was wiped was the data, not the foundation.

The aftermath

Along this line, the follow-ups make sense. Two months later, AlphaZero applied the same architecture to chess and shogi; after eight hours of training it beat the three-day AlphaGo Zero 60:40 — remove even Go's rule specialization, and generalization gets stronger. Two years on, MuZero went further: it dropped "rules provided by humans" too, learning its own environment model and planning in worlds whose rules it was never told. The boundary kept pushing outward: data → guidance → domain knowledge → rules → the rules themselves. Each step thins the human-designed prior by one layer.

Back to that 100:0. It is remembered because the machine beat the human again, but what the paper truly established was the mechanism: the learning signal comes not from imitation but from search and network pulling each other uphill in self-play. Incidentally, this "self-produced data, self-produced teacher" structure has a hotter name today — self-improvement. AlphaGo Zero demoted "imitating humans" from a necessary condition to an accelerator, and eight years later the self-improvement loops of large language models walk the same road: wherever verification is cheap enough, machines can climb on their own. Go was simply the first place where verification was cheap enough.

#PaperAnalysis #Xiaokai