Chapter 02 / The search
What does thinking longer buy?
Explore alternatives, including the opponent’s best replies.
Search spends computation on alternatives
A single policy prediction commits immediately. Search instead makes hypothetical moves, asks the model about the resulting positions, and compares the continuations. The weights stay fixed; the extra work happens at inference time.
Your opponent gets a choice
A promising continuation is not enough if the opponent can avoid it. In minimax-style backup, White selects the highest fixed-White value and Black selects the lowest. The root chooses based on the replies actually explored, not an assumption of cooperation.
A toy example, not model output
Move A leads to White scores 0.8 or 0.3 after Black replies. Move B leads to 0.6 or 0.5. White prefers B: Black can hold A to 0.3, but B to only 0.5.
Selective search sees only a subset of replies. Its result is an estimate, not a proof that no better defense exists.
Depth and breadth compete
Depth is how many plies a line reaches. Breadth is how many alternatives it considers. One deep line can miss an immediate refutation; a wide tree can spend its whole budget near the root.
LittlePly’s uncertainty-guided approach opens more branches when the policy is ambiguous and follows fewer when it is concentrated. This is a compute-allocation heuristic, not a guarantee that low-probability moves are harmless.
How to read a depth claim
Ask whether it means the deepest single leaf, completed iterative depth, median depth across decisions, or depth retained from pondering. Also ask about node count, branch coverage, time budget and deadline misses. A maximum alone can be misleading.
The deployed path and the next experiment
The recorded online harness uses a policy-only opening gate for absolute plies 0–5, then conditional search within time and node limits. Helpers are separate. Individual evaluations freeze their own settings.
The new two-GPU asynchronous design aims to deepen and broaden search and reuse opponent-turn work. It is an implementation and testing effort; the historical 99% CPU cohort is not evidence for its strength or live deployment.