Skip to main content

Becoming an Assistant

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

State

A pretrained predictor becomes an assistant through post-training — instruction tuning on demonstrations of ask-and-answer, then preference tuning, where human judgments steer further descent.

Show

Ask the raw predictor a question and its instinct is to continue the text — perhaps with more questions.

Watch for

Any named system's recipe details are the company's to disclose — the pipeline here is the published shape.