Definition
Plain language
When teaching a model a new skill accidentally wipes out one it had already learned.
As stated in the literature
A failure mode in sequential learning where new updates overwrite parameters encoding previously acquired capabilities; in harness-evolution work, the seesaw/no-regression constraint exists specifically to guard against it.
Why it matters: Without guarding against it, every attempt to add a new skill risks silently destroying old ones, making cumulative learning unreliable.
For example, fine-tuning a model to solve math problems might leave it suddenly worse at writing the poems it used to handle easily.
Heard on the show
“The second is catastrophic forgetting — in learning a new skill, you overwrite an old one you'd already mastered.”Episode 147 — Agents Fail at the Body, Not the Brain: A Self-Rewriting Scaffold That Lifts a 9B Model 44 Points