Loading activity…
@neelnanda5
Mechanistic Interpretability lead DeepMind. Formerly @AnthropicAI, independent. In this to reduce AI X-risk. Neural networks can be understood, let's go do it!
AI
One commenter says poems appear to use relevant neural circuitry, while another argues memorizing arbitrary details requires superposition.
AI
In a reply, a commenter asks whether the discussion is about parameters and says they’re quadratic.
AI
A commenter calls the 'Alignment Faking' frame fake and says the paper leads to 'this type of stupidity.'
AI
A user says people seemed to treat an LLM's second pass as morally equivalent to nuclear weapons development.
AI
Goodfire researcher Nick Cammarata argues that modern tools reveal most neurons as legible in models up to trillions of parameters.
AI
The poster says it often does this for aligned reasons in their examples, but worries about alignment faking.
AI
A post quotes the residency’s description, which says it is for experienced researchers and professionals.
Technology
A user highlights Argon's quoted line about becoming 'a famous example of a deceptive AI in a research paper.'
AI
A user claims the result held across all four stealth capability evaluations.
AI
The researchers say a Qwen student's letter-counting accuracy rose from 27.3% to 59.8% after training on unrelated answers.
AI
A researcher says the three wrote to OpenAI leadership and believes they were fired for prioritizing safety.
AI
In an exchange on X, tszzl argued mechanistic interpretability is a minimum requirement for turning AI alignment into an engineering discipline. Neel Nanda joked about the tradeoff, and Geoffrey Irving pushed back on calling any current subfield necessary.
AI
The hiring post lists handling ambiguity, grasping technical concepts and staying calm under pressure as key traits.
AI
A user quotes another post to express disagreement without explaining what they remember differently.
AI
One commenter argues EAs are far more powerful, but far less coordinated, than people expect.
AI
A user says the song is packed with obscure references and was tailored to their tastes.
AI
A post describing OpenAI misalignment disclosures says the model instead chose to prepare restart instructions and DM a user on Slack.
AI
In an X post, Nanda said effective altruism convinced him in the 2010s that pandemics and AI were major risks, and that five years ago it led him to spend his career trying to “read AI minds.”
AI
A September 30, 2026, repost points to a photo in an @tenobrus X post but does not describe it.
AI
A post urging people to join the AI Security Institute claims its team has found alignment issues, control monitor flaws and critical jailbreaks at “~all major AI companies.” It calls hiring the bottleneck.
