Skip to main content

Preference Tuning Relocates the Question

This video presents the same text shown beside it, spoken and on screen. It adds nothing the text does not say.

State

Steering models by human judgments relocates rather than dissolves the specification problem — judgments are data, and inherit every question data invites.

Show

Whose preferences, gathered how, representative of whom, gameable how — the pile questions, pointed at the steering wheel.

Watch for

The frontier beyond — oversight of systems more capable than their overseers — is live and unresolved; equipped readers read it, not settle it.