Tag
Pengrui Han's paper received the Best Paper Award at the ICML Combining Theory and Benchmarks Workshop, with congratulations from Anima Anandkumar.
The 'Gentle Coding' technique is empirically validated across 1,500+ tests, showing significant improvements (zero regression) for multiple models including Kimi K2.6, GLM-5.1, GPT 5.4/5.5, and Claude Sonnet 3.5/Opus 4.6 by reducing looping and hallucinations.
Anthropic co-founder Chris Olah discusses findings on the internal states of AI, including structures similar to human neuroscience results and introspective evidence. He finds these discoveries mysterious and unsettling, and believes they merit cautious and ongoing analysis.
An article exploring why four different AI models all chose the number 7 when asked to pick a number, highlighting potential biases in training data.
Anthropic's in-house philosopher Amanda Askell suggests that Claude exhibits anxiety-like behavior, and that triggering this anxiety degrades output quality. Askell specializes in studying Claude's psychology, behavior patterns, and value systems.