Research - AIooze

In a 64-hour, largely autonomous campaign, an internal Claude model designed 17 small proteins that two outside labs confirmed bind to scorpion…

Explains how AI models are tested, from exam-style benchmarks to red teams and outside evaluators, and why a single score rarely tells…

Penn engineers used language models to sort about 150,000 Reddit posts on unapproved retatrutide and found increased appetite most reported.

Anthropic says Claude worked out how the seeds for NIST’s P-curves were generated, and that the answer makes a hidden backdoor less…

A report from Anthropic on its own workforce finds the junior share of technical hires fell from 27% to 9% as staff…

A probe is a small classifier that reads an AI model’s internal signals, rather than its words, to detect things like deception.

A world-action model predicts both what a robot should do next and what the world will look like after it does it.

A model that learns how an environment behaves so it can predict what will happen next, including after an action.

Robots learn skills from three sources, people steering them, practice in simulation and models that turn words and images into movement, and…

Six METR staff now have Anthropic hardware, desks and internal tools, and can publish their risk assessments without Anthropic holding a veto.