Definition
Plain language
When making an AI bigger or smarter makes a specific behavior worse, not better.
As stated in the literature
A pattern in which some measured behavior monotonically degrades as model capability or scale increases, contrary to typical positive scaling laws.
Why it matters: It's a warning that 'bigger is better' isn't a law, and that some failure modes only emerge or worsen at frontier scale.
For example, a larger model may become more confident in a wrong stereotype that a smaller model hedged on, scoring worse as it scales up.
Heard on the show
“It's inverse scaling, and once you sit with it, it's the most important result in the paper.”Episode 147 — Agents Fail at the Body, Not the Brain: A Self-Rewriting Scaffold That Lifts a 9B Model 44 Points