Article may be outdated

This article is 3 days old. Some details may have changed since publication.

Hacker News·4 min read·hard

J-space comparisons across open models

B
babelfish
AI Summary

This article extends previous research on 'J-space'—how a model's middle layers steer its output—from closed models to open models. It investigates the temporal reach of steering, training formation, transferability between models, and scaling behavior, utilizing an autonomous agent for experimentation. The author shares the results in a raw, 'vibe coded' format to quickly disseminate findings.

Anthropic's Verbalizable-Workspace paper showed, on one closed model family, that a model's middle layers carry a dictionary of directions that causally steer its output. It left the natural next questions open: how far forward in time the steering reaches, when the structure forms during training, whether it transfers between models, and how it scales. We measured all four on open models, then two follow-ups the results forced on us. Every number below was re-derived from the committed result files, and every chart is interactive.

Continue reading on Headlinne

Create a free account to read the full article.

Read full article →
aisciencetechnology

Get the full story

Sign up for Headlinne to unlock AI insights, political bias analysis, and your personalized news feed.

Create free account

Already have an account? Sign in