Safety
156d agoNew Chain‑of‑Thought Analysis Toolkit
Researchers released an open‑source suite of nine chain‑of‑thought analysis tasks, challenging existing safety monitors. The toolkit, built by a team including Neel Nanda and supported by Claude, benchmarks linear, attention, and SAE probes.
The Signal
Its global impact lies in accelerating safer AI systems by providing reproducible, out‑of‑distribution evaluations for the community.