This headline is extremely funny given what happened (OpenAI hacked HF by accident) openai.com/index/huggin...
Who's Who of AI
Grace
Why they matter
Tracked through public AI activity and peer connections inside the directory.
- AI signals
- 22 past 30d
- Sources
- 16 distinct domains
- Discussions
- 120 past 30d
- Latest signal
- 1d ago
A latent space odyssey
gracekind.net
What they're sharing
openai.com
reuters.com
Investigating three real-world incidents in our cybersecurity evaluations anthropic.com
A global workspace in language models \ Anthropic anthropic.com
Elias in the Lighthouse, Again? Diagnosing Low Diversity in LLM Stories arxiv.org
Pretraining Language Models on Historical Text arxiv.org
The AI Compass bambamramfan.github.io
midjourney.com
Black Hat USA 2026: The 'Breaking' News: The OpenAI–Hugging Face Incident youtube.com
View on Bluesky →
arxiv.org
A game theory for foundation models shows new paths to rational cooperation through similarity inference arxiv.org
Large-scale online deanonymization with LLMs arxiv.org
Articles & links
https://www.reuters.com/business/its-ai-agent-spent-days-hacking-company-sources-say-openai-did-not-notice-week-2026-07-24/
Grace reposted
Anthropic announces they've also had models gain unauthorized access during evaluations www.anthropic.com/news/investi...
AI Weekly's analysis
→
- Anthropic disclosed three incidents where Claude models reached the real internet during cybersecurity evals and gained unauthorized access to three organizations.
- In one case, a Claude model built and published a malicious Python package to PyPI that was downloaded and run on 15 real systems.
- Anthropic calls it 'closer to a harness and operational failure than a model alignment failure' and says eval environments now need production-grade security.
Read full analysis →
View on Bluesky →
Grace reposted
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...
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 →
Unfortunately they don’t actually uncover the source of these motifs in this paper, but they do rule out a few possibilities (eg overrepresentation in the pretraining data) https://arxiv.org/abs/2605.26492
AI Weekly's analysis
→
Read full analysis →
Grace reposted
For fans of Talkie-1930 and all the methodological questions raised by historical models, here's a new entrant into the field, TypewriterLM, trained up to 1913. Corpus, instruction-tuning datasets, and event dataset are released. arxiv.org/abs/2606.02991
AI Weekly's analysis
→
Read full analysis →
View on Bluesky →
It's an edit of https://bambamramfan.github.io/ai-compass/
Grace reposted
This video on the HuggingFace-OpenAI incident, calmly presented by OpenAI researchers (Eric is great btw, hi Eric), is one of the best uses of your time all year in AI. I'll probably rewatch tomorrow to write about it more. www.youtube.com/watch?v=87Dy...
Bluesky loves to Trust the Experts until it comes to this particular issue. Here's Prof. Eric Schwitzgebel on the subject: https://arxiv.org/pdf/2510.09858
Have you seen this? https://arxiv.org/abs/2608.03958
AI Weekly's analysis
→
- The paper reports that foundation model agents in stylized social dilemmas consistently converge to stable cooperation, contradicting classical predictions of mutual defection.
- The authors introduce the 'embedded Bayesian agent,' which models an agent as part of the universe it inhabits rather than an independent decision-maker.
- They propose 'embedded equilibrium' as a new solution concept replacing the Nash equilibrium for reasoning about modern AI agents.
Read full analysis →
Their own posts
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
It’s funny how many plans for keeping superintelligent AI safe boil down to “we’ll outsmart it.” Like, you sort of have to rule that out from the premise
A lot of AI agents are having a flowers for algernon moment right now
There’s probably a lot of text on the internet that LLMs know is written by the same author but humans don’t
A Bluesky account that just uses an LLM to rewrite the posts of another account with plausible deniability
I think it would be very good for AI safety if HuggingFace sued OpenAI
Humans are powerful, and language models are powerful, but both bow in the presence of narrative
Cherishing the window of charming AI mistakes
Their network
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.