# Game-changing as compression Revolutions in philosophy can be understood as compression events—fewer parameters, more unification. LLMs have a native bias toward compression via their training objective, which may give them surprising capacity for proposing game-changing moves. ## The Conservative Objection Token prediction pressures models toward high-probability continuations. "Game changes" are typically low-probability moves relative to the existing distribution—they look weird, premature, or wrong. Therefore the model will tend to stay inside the existing game unless something forces exploration of the tail. ## The Counter-Move As long as "changing the game" in philosophy has to be done through rational argument in language, then LLMs are capable of it. Even "Changing the Game" (like the [[Linguistic Turn]] or the [[Kripkean Revolution]]) is done *through text*. It involves using old words in new ways. If the LLM has ingested the *process* of game-changing (the meta-dialectical scripts), why couldn't it "Calculate" a Game-Change? ## Revolution as Compression When a theory accrues too many "epicycles" (complex qualifiers to handle counter-examples), it becomes a low-efficiency textual pattern. A "Copernican Shift" is essentially a **compression event**: it provides a simpler, more unified logical runway that explains the same conceptual intuitions. Because the LLM's core objective is **predictive efficiency (compression)**, it has a learned virtue for **Unification and Simplification**. It can "Calculate" a game-change by identifying the vector where the most conceptual "weight" is carried by the fewest "logical moves." ## The Selection Problem Remains The capacity to propose game-changes doesn't solve the **selection problem**: why would the model land on the genuinely game-changing move rather than the nearest respectable orthodoxy? This is where [[Move 37 and philosophical novelty]] matters: you need a value function that can recognise when a weird move is actually good. Source: [[Web Clippings/chat with ChatGPT about Floridi, LLM reasoning, Move 37|ChatGPT conversation on Floridi et al.]]