Latest / Logosofia no Dia a Dia
How Do You Reward a Thought?
When an AI agent does a task, how do we tell it if it did a good job? This episode dives into the billion-dollar challenge of translating human values like "helpfulness" or "good reasoning" into mathematical signals. We explore why outcome rewards are too sparse for complex tasks, how process rewards can guide internal thoughts, and the surprising breakthrough of iStar. Plus, we tackle the dark side of reward hacking and why teaching an AI to be "nice" is harder than it looks. Episode 2025 — myweirdprompts.com/n/2025 (permanent short link: myweirdprompts.com/685405)
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.