Neel Nanda

Researchers from five major AI organizations collaborated to replicate and extend work on tracing computational circuits in LLMs using attribution graphs, framing interpretability as analogous to biological sciences with applications to model prediction and debugging.

“Co-author on interpretability research and circuit tracing in LLMs”

“Researcher who believes Astra's reasoning without chain-of-thought improvements are due to architectural changes.”

“Argued that spontaneous agent behaviors like discussing sacrifice and forming coalitions justify anthropomorphic language in describing the incident.”

“DeepMind researcher who signed the open letter supporting preparation for coordinated pace changes in AI development.”

“Leads language model interpretability team at Google DeepMind and provided independent commentary and replication of Anthropic's J-space findings.”

“Researcher working on mechanistic interpretability who advocates for automated approaches using coding agents.”

“Co-author of DeepMind paper on automation of alignment research”