Doomer Label Still Tied to Short Timelines
AI safety and short AGI timelines are still widely seen as the same position, and the high-profile AI 2027 timeline has deepened that confusion, so the doomer camp's reputation now hinges on extreme date predictions. — lesswrong.com
Heretic Rewrites LLM Behavior on a Gaming PC
Heretic modifies language models on a user's own hardware, using mathematical steering rather than training runs to remove censorship and reshape other behaviors. The tool puts formerly corporate-scale model modification onto a single gaming PC with no internal knowledge required. — github.com
KORMo-Web-11B Unifies LLM, Vision, and Browser Control
KORMo-Web-11B packs language, vision, and browser-agent abilities into one open-weight model, letting a single system answer questions, read images, and drive a browser. It beats open-weight rivals on Korean live-web tasks while keeping its general reasoning skills intact. — huggingface.co
Shaming Lab Workers Drives Them Deeper In
Publicly shaming frontier-AI-lab workers over their jobs makes them defensive and drives them deeper into like-minded company, while a supportive friend who stays openly uncertain gives them room to reason about quitting. — lesswrong.com
KORMo-Web-4B Runs Real Browser Tasks by Sight
KORMo-Web-4B completes real browser tasks from screenshots alone, planning clicks and typing in Korean or English without any underlying page code. It outperforms similarly sized open-weight web agents on live-website evaluations, and its weights, training data, and harness are fully open. — huggingface.co