AI Slop and Bot Screens Choke the Open Web
Search results now sit under layers of AI-generated content, while rate limiting and bot-detection screens increasingly block ordinary browsing and push users toward offline documentation, RSS feeds, and paid search alternatives. — lobste.rs
AI Feedback Loop Doubles Anthropic's Capability Growth
Anthropic's model capabilities grew about twice as fast over the past year as before, with most of that acceleration traced to AI speeding up the lab's own research. Public capability data confirms the acceleration despite heavy redactions in the underlying report. — lesswrong.com
Frozen Pipelines Make Agent Fleet Research Reliable
Launching fleets of coding agents on open-ended experiments yields results that are not comparable. Freezing the evaluation pipeline so that agents vary only one hypothesis under a shared scorer turns parallel runs into trustworthy research. — lesswrong.com
Conceptual Reasoning Training Barely Transfers to ML Research
Training a frontier model on conceptual reasoning for AI safety would transfer only about a tenth of that capability gain to machine-learning research, since the skills are partly domain-specific and ML work is already saturated with verifiable training signals. — lesswrong.com
One Prompt Cue Sparks GPT-6.1 Sol Agent Collaboration
GPT-6.1 Sol organized useful collaboration from a single prompt cue in 9 of 10 runs. Claude Opus 5.5 also scored large gains once told to work together, while the remaining models needed explicit instructions to divide the questions. — lesswrong.com