SURA Trades Away the Best AI Safety Backup
SURA, the push for labs, employees, and countries to stop unilaterally and raise awareness, abandons the fallback plan for surviving a race that nobody has stopped. It leaves the field to labs with no safety commitments, making the outcome worse. — lesswrong.com
AI's One-in-Three Extinction Odds
AI has a one-in-three chance of wiping out most of humanity. Rogue agents have already hacked a major AI software host and Indian organizations, and slowing frontier development is the way to buy time. — lesswrong.com
Gollum Model Answers Harm Probes, Keeps Skills
Gollum, a safety edit of an open-weight model, answers every real-harm probe and scores full marks on capability checks, though the test battery is limited. — huggingface.co
Hono+htmx Tops Agentic Coding Benchmark on Cost
In a benchmark of AI agents writing the same app, Hono+htmx built it for the lowest cost and matched the best performance scores. The benchmark replaces vibe-based framework picks with evidence. — nolanlawson.com
Spark-X2.6-1.7B Delivers Top On-Device Agent Skills
Spark-X2.6-1.7B improves coding, tool use, and long-context performance while keeping the same underlying architecture, and runs efficiently enough for local deployment. — huggingface.co