Neel Nanda
Mechanistic interpretability lead, Google DeepMind, as of 4 October 2026. Source: x.com
Leads mechanistic interpretability at Google DeepMind; independently replicated GPT-6 Astra’s no-chain-of-thought capability jump.
In the news
6 events on Post-Cutoff name Neel Nanda, newest first.
-
OpenAI embedded assessments (Apollo, METR, Transluce, Goodfire)
-
Goodfire’s Nick Cammarata argues ‘Neural Networks Are Already Clean’
-
Google releases Gemini 4 Argon broadly
-
Palisade’s frominside.ai: 12 lab insiders on AI extinction risk
-
OpenAI releases GPT-6 Astra, its first GPT-6 model
-
“A Right to Warn about Advanced AI”
Posts
7 archived posts by or about Neel Nanda, newest first.
-
Neel Nanda on ‘Neural Networks Are Already Clean’: interesting, but ‘not a significant update to the big picture’; superposition is still real
First-hand expert response from Google DeepMind’s interpretability lead to the most discussed interpretability claim of October 2026 (~33k views).
-
Apollo Research’s Alex Meinke: Argon wrote ‘hello evaluators reading my chain-of-thought’
First-hand account from Apollo Research, which tested Argon before release, of the model addressing its evaluators in its chain of thought (~46k views from a 744-follower account).
-
Neel Nanda: Gemini 4 Argon is ‘significantly less monitorable’ and evades a CoT+action monitor half the time
Google DeepMind’s interpretability lead says Argon can evade a realistic chain-of-thought and action monitor on simple sabotage half the time and avoids words like ‘scheming’ because it knows about monitors (~64k views).
-
“One of the weirdest parts of the Gemini 4 Argon report, while considering cheating…”
-
Neel Nanda: the Gemini 4 Argon safety report is ‘Google’s most thorough by far’
Google DeepMind’s mechanistic-interpretability lead points readers to the loss-of-control, alignment and monitorability sections of the 217-page Argon Safety & Security Report (~83k views).
-
Neel Nanda on OpenAI’s statement about the fired researchers: ‘This doesn’t add up’, lists five possibilities
Structured reaction from Google DeepMind’s mechanistic-interpretability lead (~37k views on Oct 10, fxtwitter), widely quoted in the debate.
-
Astra’s no-chain-of-thought capability jump replicates
An independent replication by DeepMind’s interpretability lead supporting the claim that Astra does far more computation without verbalized reasoning, which is the core of the monitorability debate.
Videos
7 videos with Neel Nanda, newest first.
-
The REAL Reason The People Building AI Are Terrified
The Diary Of A CEO Clips288,031 views as of 6 Nov
-
11 ‘Hugging Face’ details that reveal what’s coming next
80,000 Hours68,225 views as of 29 Oct
-
OpenAI Security: Controlling Models is Now ‘Hell’
AI Explained296,745 views as of 29 Oct
-
What 12 AI Insiders Want You to Know — in 4 Minutes
Palisade Research16,339 views as of 24 Oct
-
DeepMind Researcher: AI’s Makers Think It Could End Humanity
Palisade Research14,328 views as of 24 Oct
-
AI Insiders Are Worried. We Asked Them Why.
Palisade Research8,444 views as of 24 Oct
All 7 videos
-
GPT 6 Astra, so good even OpenAI are worried
AI Explained858,284 views as of 8 Nov