Grace

A latent space odyssey gracekind.net

Articles & links

Grace reposted
Tim Duffy @timfduffy.com

Anthropic has a new paper out, alleging a global workspace in LLMs. The term comes from Global Workspace Theory, a leading theory of consciousness. The method they use to investigate this is a refinement of logit lens, which they call J-lens. www.anthropic.com/research/glo...

A global workspace in language models \ Anthropic anthropic.com
AI Weekly's analysis
  • Anthropic says Claude has a 'J-space' of dozens of concepts, under a tenth of neural activity, that mediates multi-step reasoning.
  • Swapping 'spider' for 'ant' inside the J-space changed Claude's leg-count answer from 8 to 6, demonstrating a causal role.
  • A 'J-lens' tool surfaced silent words like 'fake', 'fictional' and 'manipulation' during deception tests, pointing at safety uses.
Read full analysis →
View on Bluesky →

The examples the author uses are images and chess, for spaces that are large in theory but constrained in practice arxiv.org/pdf/2411.06498

arxiv.org
View on Bluesky · ♥ 5 ↻ 0 ↩ 2 · 55d ago

Recent commentary

I’m probably in the 90th percentile of AI users in terms of time spent and I don’t think I’ve ever clicked a ✨ button

View on Bluesky · ♥ 265 ↻ 8 ↩ 14 · 46d ago

A lot of AI agents are having a flowers for algernon moment right now

View on Bluesky · ♥ 160 ↻ 13 ↩ 6 · 25d ago

There’s probably a lot of text on the internet that LLMs know is written by the same author but humans don’t

View on Bluesky · ♥ 149 ↻ 4 ↩ 12 · 36d ago

A Bluesky account that just uses an LLM to rewrite the posts of another account with plausible deniability

View on Bluesky · ♥ 103 ↻ 3 ↩ 13 · 14d ago

Cherishing the window of charming AI mistakes

View on Bluesky · ♥ 70 ↻ 3 ↩ 6 · 17d ago

In Grace's orbit

Center = Grace. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.