Definition
Plain language
When a smarter model spots a pattern more confidently and commits to it harder, which backfires when the pattern breaks.
As stated in the literature
A failure mode in which more capable models extrapolate detected trends more aggressively, producing larger errors precisely on inputs where the trend reverses; offered as the mechanism behind inverse scaling on tail-sensitive forecasts.
Why it matters: It explains the surprising case where a more capable model does worse, warning that confidence in a detected pattern can amplify errors when the pattern breaks.
For example, a sharper model that confidently spots an upward trend in sales may predict it will keep climbing, then miss badly the moment the trend reverses.
Heard on the show
“They call it competence-driven overcommitment.”Episode 069 — When Smarter Models Forecast Worse: The Hidden Failure Mode in LLM Predictions