Definition
Plain language
Getting better answers out of an AI by letting it check its work against the real world and fix what it got wrong.
As stated in the literature
A test-time compute axis where a proposer's output is measured by an external instrument and revised in a loop; unlike reasoning-longer or best-of-N, it imports information not present in the frozen weights or prompt.
Why it matters: It lets a model pull in real-world feedback it could never get from just thinking harder, so answers improve on tasks where guessing alone falls short.
For example, an AI writing code can run its program, see the error message, and rewrite the faulty part until it actually works.
Heard on the show
“They call it interaction scaling.”Episode 217 — Why an AI Called Fourteen Broken Figures Perfect, And What It Reveals About Test-Time Compute