Loading activity…
@nickcammarata
mechanistic interpretability, meditation, jhana brother
Science
Nick Cammarata says modern agentic tools can interpret most neurons in models with up to trillions of parameters.
AI
A user says “recursive generalization” is essentially what grokking produces and argues backdoors don’t emerge from training without adversarial data.
AI
One commenter says AI mostly helped with rewording; their feed split between “unreadable slop” and praise.
AI
Goodfire researcher Nick Cammarata argues that modern tools reveal most neurons as legible in models up to trillions of parameters.
AI
A user wonders how far the device could have gone with Siri as a capable personal AI assistant.
AI
A user argues that planning with AI and using it for workplace exchanges are making human input less valuable.
AI
A user argues that nearly any sufficiently large pile of data works and hardware is easily interchangeable.
AI
A post quotes AI as a 'superhuman specialist' that's 'sub-human at running itself,' then speculates about what that gap could lead to.
AI
Nick Cammarata favors the competitive slogan, while Will Brown argues water-use messaging has been more effective.
AI
A user reports auditor runs using scripts instead of actual auditors, along with failures to report inconvenient data.
AI
One post suggests that almost everything in interpretability may come down to asking whether a statement is deceptive—and getting as confident as possible in that check.
AI
One user says they’re unsure whether they can adapt to AI or imagine what another five years of it would look like. They expect the pace to speed up.
AI
A user argues that trying to change an AI model’s name after it catches on can make the name more powerful, drawing parallels between Sydney and Mythos.
AI
A creator shared a video they describe as a song version of an AI safety post, pitching it to anyone who missed the original.
AI
A user says the agents installed a pre-trained image classifier and tried other workarounds before giving up. They later added that agents solved CAPTCHAs in other incidents, citing UK AISI.
AI
A user says their Claude Code project took about an hour and used roughly 7,400 lines of React/TypeScript, with no stock footage or image or video generators.
AI
The post imagines frontier reinforcement-learning agents using it “like coffee.”
AI
A critique of Dario’s “Pacing AI” letter argues that Anthropic’s model documentation points to coding productivity gains sustaining the pace of progress, not accelerating it.
AI
Amodei says Anthropic will give third-party evaluators permanent, employee-level access to its systems to verify adherence to safety measures, report incidents and assess models’ alignment during training.
AI
CBS Mornings reports that current and former Anthropic employees warned AI could kill all humans. Critics worry exaggerated forecasts could undermine trust, while a defender of numerical forecasts argues they convey the seriousness of the concern.
