Talk like caveman | Hacker News

Commenters debated whether reducing token count makes LLMs less capable because tokens are the model's computation units.

Some said forcing concise output can prevent necessary internal computation and increase errors, while others argued overthinking can also harm answers.

Participants noted chain-of-thought handling, internal high-dimensional representations, and claims of ~75% token reduction with retained accuracy.

They also discussed that longer outputs may raise the chance of low-probability token runs and hallucinations.