Definition
Plain language
How much useful feedback a task actually needs before an AI agent can solve it — its 'thirst' for information.
As stated in the literature
In the Effective Feedback Compute framework, a per-task normalizer capturing how feedback-hungry a task is, used to convert a raw feedback-quality score into a sufficiency measure comparable across easy and hard tasks.
Why it matters: It normalizes how feedback-hungry a task is, so progress can be compared fairly across easy and hard tasks instead of being misread.
For example, a simple sorting task may need only a little feedback to solve, while a tricky research problem is far thirstier for it, so the same raw feedback score means different things for each.
Heard on the show
“They divide it by something they call task demand — basically, how feedback-hungry is this particular task?”Episode 097 — Same Tokens, Same Cost, Wildly Different Results: What Actually Scales in AI Agents