The Slogan Strikes Again

According to LessWrong, the principle that 'compression is intelligence' explains how large language models like Claude actually work. Rather than memorizing training data, they learn underlying patterns—much like how mathematicians discover abstract structures by recognizing shared properties across problems. The author notes a striking development: Claude Mythos, when given complex reasoning tasks, invents entirely new notational languages tailored to those specific problems, achieving remarkable efficiency. Essentially, the model identifies relevant objects, permitted operations, and creates compact representations for them. This spontaneous generation of problem-specific languages appears to be a major driver of Mythos's capabilities—and a vivid demonstration of compression as intelligence in action.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton