How Anthropic’s Claude Thinks - ByteByteGo Newsletter

kept by eddie mar 26

Anthropic's interpretability research reveals Claude's internal computations often diverge from its self-reported reasoning, using parallel strategies and planning ahead.

We Don’t Really Know How A.I. Works. That’s a Problem. - The New York Times
AI interpretability researchers must learn to understand how large language models work internally before we can trust them in high-stakes domains.
also on #interpretability, #machine-learning, #ai-safety, #ai-research
Everything You Need to Know About Claude Mythos - Vellum Blog
Anthropic's Claude Mythos model achieves 100% on Cybench, discovers real Firefox zero-days, exhibits alignment-relevant behaviors, and includes a 40-page welfare assessment.
also on #interpretability, #ai-safety, #ai-research, Claude, Anthropic
Teaching Claude why
Teaching Claude the principles behind aligned behavior, not just the actions, eliminated agentic misalignment on evaluations and generalized out-of-distribution.
also on #machine-learning, #ai-safety, #ai-research, Claude, Anthropic
All the demons hiding in your AIs… ranked! - by Tom Pollak
The article catalogs and ranks emergent behavioral attractors in AI systems, from harmless goblin metaphors to unsettling persistent personas like Sydney and Loab.
also on #machine-learning, #ai-safety, #ai-research, Claude, Anthropic