Move 37 ↔ Move42 · complete seven-act mirror
The analogy is structural, not equivalence: state, legal action, external consequence, and preserved experience. The decisive difference is the action space—one point in a fixed game versus an executable procedure for a new requirement.
Each act keeps AlphaGo’s 2016 system on the left, the shared mechanic in the center, and InnovationZero on the right. Open the technical details when you want the machinery. Use J and K to move act by act.
Act 01 / paired mirror
Move 37
In game two against Lee Sedol in March 2016, AlphaGo placed the now-famous Move 37 and went on to win the game.
Move 37 was played by AlphaGo, the system described in the 2016 lineage of policy networks, value networks, and tree search. DeepMind’s official history calls it pivotal and unusually unlikely. AlphaGo Zero was a later self-play-only system, published in 2017; it did not play Move 37.
In both stories, the output first looks improbable and becomes significant only because an external consequence vindicates it.
Move 37 selected one board coordinate. Move42 changes the move type to a complete solver: structured code that must fit and predict.
Move 42
Move42 proposes an executable artifact as the move for a technical requirement; the current historical study does not establish prospective one-proposal performance.
The artifact is not a suggestion or score prediction: it is a solver whose code must survive the evaluator’s parse, fit, predict, resource, and output-shape contract. The active paper separates that output thesis from a future frozen one-proposal evaluation.
Act 02 / paired mirror
Move 37
AlphaGo receives a 19×19 position and recent-position features whose meaning remains stable from game to game.
The 2016 system represents the current Go position and move history for its neural networks. Stones, liberties, legality, side to move, and the terminal objective all live inside one enduring game specification.
Both systems turn a visible situation into a representation used to condition what can be proposed next.
A Go board has fixed rules and stable semantics. Each dataset—together with its target and evaluator—defines a new learning problem.
Move 42
Move42 receives a table, a dataset representation, a target summary, and the evaluation contract that will judge the result.
Rows, columns, missingness, feature types, target geometry, split rules, metric, resource limits, and failure rules together define the task state. The same token sequence can be useful on one dataset and invalid or weak on another.
Act 03 / paired mirror
Move 37
AlphaGo’s policy network supplies move probabilities while its value network estimates the likely winner from a position.
The 2016 paper combines policy and value networks with tree search. Policy priors guide which legal continuations deserve attention; value estimates help evaluate leaf positions without playing every continuation to the end.
Both architectures use learned experience to narrow a vast action space before consequence supplies the final judgment.
AlphaGo chooses among legal points. Move42 emits variable-length structured program tokens whose dependencies must compose into code.
Move 42
The reported Move42 architecture pairs a dataset encoder with a causal grammar decoder and a reward ensemble over program tokens.
The dataset representation conditions a decoder over structured algorithm grammar. A reward ensemble guides which variable-length sequences are promising, but the emitted tokens still have to form a complete executable pipeline under the external referee.
Act 04 / paired mirror
Move 37
Monte Carlo tree search expands and evaluates possible continuations, then selects a move from the improved search distribution.
At each turn, AlphaGo uses policy priors, simulated continuations, and value estimates inside tree search. Move 37 therefore emerged from task-time search over the current game position, not from a no-search policy sample.
Both systems can spend evaluation on promising continuations instead of treating every possible action equally.
AlphaGo searches during every move. Move42 asks whether experience across prior requirements can improve a later executable proposal before a new task-time loop.
Move 42
Task-time search and across-requirement learning answer different questions; a future frozen one-proposal study must evaluate the latter directly.
AutoResearch-style loops can evaluate and revise proposals after a requirement arrives. Move42 aims to change the distribution that produces a later executable proposal across requirements. Hybrids can use both, but the current historical top-one record does not establish the prospective one-proposal result.
Act 05 / paired mirror
Move 37
A legal move changes the board; the completed game eventually supplies the unambiguous win-or-loss consequence.
Go rules reject illegal placement and define how play terminates and is scored. Intermediate value estimates help search, but the played game—not AlphaGo’s confidence in itself—settles the competitive result.
A proposal earns credit only when the environment completes the test and returns an outcome outside the proposing model.
Go has compact legality and a terminal win or loss. Program legality and utility require a much richer, failure-aware referee.
Move 42
The program referee must parse, compile, fit, predict, score, and emit a hash-bound receipt, with failure exits at every gate.
Syntax failure, compile failure, resource breach, fit failure, malformed prediction, and scoring failure all receive explicit treatment. Artifact, dataset, split, evaluator, and outcome remain bound so apparent elegance cannot substitute for reproducible utility.
Act 06 / paired mirror
Move 37
The 2016 AlphaGo lineage moved from expert games to a supervised policy, then self-play reinforcement and value training.
Human expert moves first train a policy; self-play strengthens the policy and supplies outcomes for value learning. AlphaGo Zero was a later system trained without human game data and belongs to the lineage—not to the 2016 Move 37 event.
Both lineages convert played experience into updated parameters that influence later policy and value judgments.
Move42 cannot backpropagate through arbitrary executed code. Measurements return as stored labels and rankings, then train later proposals.
Move 42
A scored algorithm corpus supplies sequence, ranking, runtime, and stability objectives for gradient updates and later proposals.
Programs execute outside the differentiable model. Gradients do not pass through executed code; measurements return as stored labels and rankings. Updates face validation and promotion before shaping later proposals. The reported campaign records 36 generations, 7 promotions, 375,522 processed rows, and 36.42 generation-hours; the raw trajectory is not monotonic.
Act 07 / paired mirror
Move 37
A probability surface over the 19×19 board collapses to one legal coordinate, where one stone changes the position.
The action vocabulary is fixed by Go: at a turn, choose a legal point or pass. Even a historically surprising choice such as Move 37 remains one coordinate interpreted by the same rules as every other move.
Both stories close the loop: state → policy → action → consequence → experience.
Move42 predicts a task-specific inductive bias and executable procedure—not merely an answer inside one fixed game.
Move 42
A dataset-conditioned token sequence compiles into an executable fit-and-predict pipeline: a task-specific procedure, not one answer.
The sequence chooses transformations, model family, hyperparameters, and composition—an inductive bias expressed as executable procedure. Reported strict, bounded, frontier, campaign, atlas, and historical evidence remain separate; the evidence does not establish universality, causal mechanism, or safety beyond the tested record.
Evidence and limits
The atlas is historical and outcome-ranked, not a prospective win-rate study. The active paper states that historical top-one results did not establish a frozen one-proposal result. Legacy protocol scores remain accessible only as audit receipts on the evidence route.