Distillation in practice often means stealing chain-of-thought, says Dean Ball

deanwball · x · 2026-10-18

Dean Ball, replying to @eliebakouch, argues that a big part of distillation in practice involves stealing full chain-of-thought (CoT) from models — not just training on their normal outputs — which is one reason the industry tends to treat such behavior as an "attack."

The original point from the reply: distillation effectiveness (information gained per sample) correlates with how much reasoning detail is disclosed to users, and it is quite predictable — one could train models with varying disclosure levels, measure samples needed, and calibrate how much reasoning to show.

Original post →

Distillation in practice often means stealing chain-of-thought, says Dean Ball

More from Models