Strength history · 10

Ten years at the Go AI frontier from AlphaGo to transformer KataGo

Reconstruct the 2016–2026 frontier from papers, released weights, and direct matches. Compare width, depth, parameter count, training, search, and why 5,185, 14,554, and relative +300 Elo do not share a scale.

1. Define SOTA first: model, search, and Elo form one experiment

   A Go engine is not a neural network in isolation. A reproducible comparison fixes weights, engine version, rules and komi, colors, visits or wall time per move, hardware backend, threads, and batching. Elo is only a coordinate fitted to wins inside that match pool: if A scores p against B, the gap is about 400·log10(p/(1−p)). Scores of 60%, 75%, and 90% correspond to about +70, +191, and +382 Elo. Absolute values also depend on anchors and participants, so only within-pool gaps or direct matches transfer cleanly; subtracting ratings from unrelated papers does not.

2. 2015–2016: AlphaGo joins convolutional intuition to value and tree search

   The original AlphaGo used separate policy and value networks, each built around roughly 13 convolutional layers and 192 filters. A supervised policy first reached about 57% expert-move prediction, was strengthened by self-play, and was joined by a fast rollout policy. MCTS combined policy priors, value evaluation, and rollouts. The Nature paper's distributed AlphaGo Fan used 1,202 CPUs and 176 GPUs and rated 3,144 in its BayesElo pool; the later AlphaGo Lee was a different system and appears at 3,739 in the 2017 paper's reconstructed pool. Neither was one small model playing alone: strength emerged from several networks, search, and distributed inference.

3. 2017: AlphaGo Master and Zero compress two networks into one residual tower

   AlphaGo Zero replaced separate networks and rollouts with one residual network producing both policy and value, using only eight positions of stones plus the side to move. Its named 20/40-block towers are an input convolution followed by 19/39 residual blocks, each with two 3×3 layers at 256 channels: roughly 23M/46M parameters. Search used 1,600 simulations per move. The three-day 20-block system beat AlphaGo Lee 100–0. The forty-day 40-block system reached 5,185 Elo in the paper's common five-second pool versus 4,858 for AlphaGo Master, a +327 gap, and won their direct match 89–11. The real leap combined a dual-head network, pure self-play, and cleaner PUCT—not depth alone.

4. 2018–2019: after the closed peak, the frontier becomes reproducible

   AlphaZero generalized one algorithm to Go, chess, and shogi, but its public Go result was 60–40 against the three-day 20-block AlphaGo Zero, not the forty-day peak. Open engineering moved quickly: PhoenixGo won the 2018 World AI Go Tournament with a 20-block network. ELF OpenGo trained a roughly 23M-parameter 20×256 model on 2,000 V100s, reached superhuman play in nine days, completed about sixteen days, and beat top professionals 20–0. KataGo's 2019 work still began from random self-play, but ownership and score targets, auxiliary policy, global pooling, forced playouts, and policy-target pruning let fewer than 30 V100s surpass ELF in nineteen days—about 50 times less training compute.

5. 2020–2023: KataGo makes each evaluation worth more instead of only scaling size

   By 2020, g170 b20c256 matched or exceeded Leela Zero's final 40-block nets at moderate-to-high search while adding Japanese rules, handicap strength, and better score estimates. Distributed training then scaled through b40c256 and b60c320; b60 was stronger per evaluation but slower and heavier. In 2022–2023, the nested bottleneck residual block shrank channels by half with 1×1 projections, ran two inner residual pairs of 3×3 convolutions, then expanded back. Together with fixed-variance initialization, one batch normalization, short-term value uncertainty, and a soft policy target, b18c384nbt ran near b40c256 speed while being only slightly weaker per evaluation than the roughly twice-as-heavy b60c320. Block count had stopped being a proxy for effective depth or strength.

6. 2024–2025: the open frontier expands from best play to human modeling and robustness

   The 2024 Human SL model b18c384nbt-humanv0 learned from human records and can condition on rank_20k through rank_9d, preaz_20k through preaz_9d, or proyear_1800 through proyear_2023. It is normally a second network providing humanPolicy, not a stronger replacement for the main SOTA model; multiple visits can solve tactics that the target human rank would miss. The main self-play line meanwhile expanded to the 73M-parameter b28c512nbt. A broader SOTA definition must also include robustness: adversarial cyclic positions achieved very high win rates against superhuman settings, showing that ordinary Elo, out-of-distribution behavior, and rules correctness are three distinct capabilities.

7. 2026: transformer KataGo replaces layer-by-layer relay with global attention

   KataGo v1.17 brought transformers to the major backends; v1.18 significantly optimized CUDA and added ROCm and ONNX. A single 3×3 convolution maps 22 input planes into the trunk, then nested bottlenecks repeatedly apply attention+MLP. Learnable two-dimensional RoPE frequencies inject position per head. Three sizes were released: the small b10c384h6nbttflrs is already stronger per evaluation than the strongest b18; the medium b10c512h8nbt3tflrs has 10 blocks, three attention layers per block, eight heads per layer—240 heads and 29M parameters; the large b11c768h12nbt3tflrs has 396 heads and 71M parameters. With the 73M b28c512 convnet anchored at zero, the medium transformer is about +50 Elo at equal evaluations and +120 at equal time; the large is about +460/+300. A 234M b40c768 reaches about +400 per evaluation but only +70 per time because each evaluation is slower. The training site's current strongest confidently rated b40 is 14,554.2±20.9 in its equal-visits internal pool, not a portable rating. The large transformer's +300 equal-time gap implies roughly an 85% expected score against b28: that is the operational meaning of stronger with fewer parameters.