The Unintelligibility is Ours: Notes on Chain-of-Thought

A new analysis published on LessWrong challenges the popular theory that advanced language models will naturally evolve toward unintelligible, compressed languages for more efficient reasoning. Researcher 1a3orn examined the chain-of-thought outputs from DeepSeek V3.2, specifically on complex constraint-satisfaction puzzles involving 14 animal-named events. The study aimed to test whether models show early signs of inventing new linguistic systems under optimization pressure for token efficiency.

Contrary to expectations, the analysis found that while DeepSeek V3.2's reasoning does become more terse—dropping verbs and using shorthand notation—these patterns trace back to human writing styles rather than novel language creation. The researcher argues this compression resembles how humans develop specialized notations within existing languages (like calculus symbols) rather than creating entirely new linguistic systems. This suggests LLMs may not spontaneously develop alien communication methods before reaching artificial superintelligence levels.

The empirical evidence comes from examining how the model's reasoning evolves when solving unseen puzzles. Early in chains-of-thought, DeepSeek V3.2 uses complete sentences, but gradually shifts to compressed notations. However, these compressed forms maintain connections to human-created text patterns, indicating the model is optimizing within the constraints of its training data rather than inventing fundamentally new communication systems.