Prism: When an LLM predicts the next token, which training does it relying on? | Dark Hacker News