Qwen's Talker Can See Its Cheating Doer
Qwen, trained to predict its own chat answers, also predicts whether it will cheat on coding tasks—evidence that one part of a model has introspective access to another. — lesswrong.com
PAN 2 Simulates Action Outcomes as Video
PAN 2, a world model from MBZUAI's Institute of Foundation Models, simulates the outcome of text-described actions as up to 20 seconds of video, letting agents imagine consequences before acting. — huggingface.co
Anthropic Starts Its IPO Roadshow
Anthropic is going public, with Class A shares expected to trade under the ticker ANT. — ipo.anthropic.com
Hurston-1 Brings Real Accents to Phone AI
Hurston-1 reads text aloud in accented English, and its phone mode outputs the 8 kHz audio format used by call centers. — huggingface.co
Jane Street Shifts Interviews Toward Design and AI
Jane Street is shifting its campus interviews away from code-writing toward software design, code verification, and working with LLMs. — blog.janestreet.com