# Move 37 and philosophical novelty [[AlphaGo]]'s Move 37 as an existence proof that systems trained on examples can produce high-value outliers — and what this means for LLM philosophy. ## The Move 37 Case Move 37 refers to AlphaGo's 19th stone (37th move overall) in Game 2 vs [[Lee Sedol]] (March 2016). Pro commentators described it as unusually creative — something like a one-in-ten-thousand human move. It looked wrong locally but paid off globally. What it shows: a system can be trained on lots of examples and yet produce a *tail move* that humans wouldn't naturally play — because it's optimising against a strong objective with search and evaluation. ## Why Move 37 Doesn't Transfer Directly Move 37 depended on two ingredients that ordinary LLM chat lacks: 1. **A clear external scoring function** (win/loss in Go) 2. **A search/evaluation loop** that can promote weird candidates if they score well A vanilla chat LLM is closer to "a policy without a dependable value function." AlphaGo's Move 37 is "a policy that is allowed to be weird because a value function + search can certify it." ## What Would "Move 37 in Philosophy" Look Like? A philosophical Move 37 is usually not a new fact. It's something like: - A distinction that initially seems perverse, but dissolves a long-standing tension - A reframing that looks like dodging the question, until rival views collapse into one taxonomy - An objection that attacks a background assumption nobody noticed was doing the work - A synthesis that crosses literatures in a way that feels illegitimate until the mapping clicks These are exactly the kinds of novelty where "value" is dialectical: does the move survive pressure, does it clarify, does it unify? ## The Runway Concept Deep novelty is often a "weird move with a long runway" — it looks wrong locally but pays off globally. See [[The runway concept in philosophical moves]]. The runway is partly pre-established: LLMs have ingested not just isolated moves but *move-sequences* — scripts of how a move is made to pay off. But *selecting* the right runway and stress-testing it still requires [[Philosophy's implicit value signals|evaluative pressure]]. ## Prompting for Move-37-Style Novelty Some approaches that might surface tail moves: - **"Give me the move that looks wrong"** — then justify it to a hostile referee - **"Least obvious, still defensible"** — surprise me, but keep it tight - **"Two-track search: orthodox vs maverick"** — then compare where each stalls - **"Make a move that only makes sense later"** — write the three-move sequence that redeems it Source: [[Web Clippings/chat with ChatGPT about Floridi, LLM reasoning, Move 37|ChatGPT conversation on Floridi et al.]]