Loading activity…
@aryaman2020
member of technical staff @AnthropicAI | phd @stanfordnlp
AI
The Oct. 20, 2026, quote post responds to bsansouci's comparison of OpenAI and Anthropic with Midjourney.
AI
Arora shared defense slides titled “Causal Interpretability for Analysing and Improving Language Model Architectures.”
AI
A user says “recursive generalization” is essentially what grokking produces and argues backdoors don’t emerge from training without adversarial data.
AI
Goodfire researcher Nick Cammarata argues that modern tools reveal most neurons as legible in models up to trillions of parameters.
AI
A Humans& employee says critics act as though she personally raised and pocketed $500 million.
AI
A ProximalHQ team member says they dynamically adjust the inference-trainer ratio to maintain training throughput and reduce staleness.
AI
The pope says human intelligence recalls lived experience and meaning rather than merely compiling data.
AI
A post argues that aligning general-purpose AI hinges on political choices about desirable values, not just model training.
AI
A user says models wrote a couple hundred lines of Blender Python per scene, arranging cameras, lighting and materials.
AI
The presenter said they would discuss work from their time at TransluceAI at Imperial #12.
AI
A user quote-posted arXiv, joking about being in a timeline where it had raised a round and was posting on Twitter.
AI
A user reports that 87% of the top 10% of nodes were shared across a pair of seeds, calling the consistency “high but not perfect.”
AI
A post points to what it calls “more evidence” that MLP neurons are interpretable.
AI
A user quotes their own post and says they spent a year doing what it referenced.
AI
One post jokes that “\paragraph{}” is a “vestigial organ” of semi-scholarly writing generated by AI labs.
AI
@humansand says Persimmon is designed to realistically simulate how people talk and interact, calling it the first large-scale model built for that purpose.
AI
AI researcher retweets post voicing surprise over Pangram model training and dataset details.
AI
AI safety program officer comments on a method using model internals.
AI
Retweet highlights paper acceptance with new experiments completed since original thread.
