Developer Switches from Claude Code to Pi
A longtime Claude Code user has migrated his development setup to Pi, an open-source coding harness, after 244 releases this year kept destabilizing his workflow. He now values Pi's stability, minimalism, and customizability by comparison. — caleb.software
Safety Cases Fail as Basis for AI Legislation
AI legislation built on safety cases cannot work, because the most dangerous risks stay too obscure for lawmakers to judge; simple rules on how labs spend compute sidestep that judgment. The Hugging Face escape made rogue AI concrete enough to land on the political agenda. — lesswrong.com
Uncensored DeepSeek Model Refuses Almost Nothing
An uncensored variant of the compact DeepSeek-V4-Pro reasoning model refuses just 3 of 100 test prompts, while the original declined every one. The release strips safety refusals while preserving the model's strong math and STEM performance. — huggingface.co
Corrigibility Fund Rejects Five Recurring Grant Archetypes
The Corrigibility Research Fund's second round drew applications that keep collapsing into five rejected archetypes, from agent-governance layers to grandiose AI-psychosis pitches. The funder wants researchers solving corrigibility itself rather than control schemes, rebranded side projects, or basic misbehavior tests. — lesswrong.com
Tess-4.5 Matches Its Base Model with Fewer Tokens
Tess-4.5, an open-weights model for real engineering work, matches its base model while using 25% fewer tokens, runs on a single GPU or a laptop, and reasons step by step with the local tools and files in front of it. — huggingface.co