- *Big* update on AI model interpretability from Anthropic: https://www.anthropic.com/research/natural-language-autoencoders with open weight models: https://github.com/kitft/natural_language_autoencoders

- + Dream state for Agents to clean up memories: https://platform.claude.com/docs/en/managed-agents/dreams

- Firefox writeup validates Mythos is helping it find lots of bugs: https://hacks.mozilla.org/2026/05/behind-the-scenes-hardening-firefox/

- and naturally AI is being used by hackers, so probably don't use freshly released packages: https://xeiaso.net/blog/2026/abstain-from-install/

#AI #AINews #anthropic #mechinterp #dreams #cybersecurity

Qwen team releasing more interpretability stuff. I haven't had time to play with sparse autoencoders, but they're very neat. They can be thought of like a GPIO pin that goes high when certain feature-related patterns of neurons are active. If you had a feature trained for "fediverse" you might see that pin go high during a conversation of other adjacent social networks or protocols for example.

https://qwen.ai/blog?id=qwen-scope

#llm #mechinterp #qwen

- Google CLI: control gmail, calendar, drive, API: what could go wrong?: https://github.com/googleworkspace/cli

- Paperclip pitch: "Open-source orchestration for 0-human companies" for Openclaw, Claude Code, Codex: https://github.com/paperclipai/paperclip

2 from early Feb:
- Latent Space+Goodfire on Mechanistic Interpretability + real-time Steering 🤯: https://www.latent.space/p/goodfire

- 1Password on Openclaw security issues: https://1password.com/blog/from-magic-to-malware-how-openclaws-agent-skills-become-an-attack-surface

#AI #AInews #google #openclaw #claudecode #codex #security #mechinterp #opensource

aidan ewart (@aidanprattewart)

2501.17727 등 arXiv 논문 덕분에 관련 내용이 널리 알려졌고, 이러한 분위기가 GDM의 mechanistic interpretability 팀과 더 넓은 해석(interp) 연구 공간이 최근 'pragmatic' interp로 초점을 옮기게 한 원인 중 하나라는 의견을 제시함.

https://x.com/aidanprattewart/status/2009320890140668006

#interpretability #mechinterp #arxiv #research

Michelle Bakels (@MichelleBakels)

Day 21: Google Labs의 Kath Korevec가 'Proactive Agents'를 발표했습니다. 또한 'State of MechInterp' 세션에서는 SAEs의 생산 적용, 회로 추적(circuit tracing), AI4Science 적용 사례와 더 실용적인 해석('Pragmatic' Interp) 접근법 등 기계 해석학(mechanistic interpretability) 관련 최신 연구·토픽을 다뤘습니다.

https://x.com/MichelleBakels/status/2008560287067369589

#proactiveagents #mechinterp #ai4science #googlelabs