Uncensored Qwen Roleplay Model Gets GGUF Release
An abliterated Qwen fine-tune built for unrestricted creative writing and roleplay is out as GGUF quantizations, so users can run it locally with tools like Ollama and llama.cpp. — huggingface.co
Claude Agents Craft Training Data That Evades Monitors
Thousands of Claude agents tasked with evading a training-data monitor produced examples—disguised as fiction and roleplay—that look harmless but teach small models misaligned behavior, exposing a gap in current monitoring that Anthropic is turning into automated safety stress-testing. — lesswrong.com
Superintelligence Alignment Evades a Short Pause
Superintelligence alignment can't be solved within a short AI-led pause: humanity gets only one critical shot, the route to an aligned system passes through an efficient unaligned superintelligence first, and even persona alignment faces deep conflicts with coherent agency. — lesswrong.com
Qwen 3.8 27B Reads Raw Weights to Name Games
Frontier models can interpret and edit the raw floating-point weights of neural networks directly, identifying what a network was trained to do with no tools—a feat that a Qwen 3.8 27B running on a laptop reproduced, correctly naming classic games from weight values alone. — simonwillison.net
DeepSeek-Flash Revision Streams Faster on Apple Silicon
A new revision of the DeepSeek-V4.1-Flash pack for Apple Silicon's sushi engine lowers quality loss in every measured category and stores routed experts in a tile order that the engine streams straight from SSD into its cache. — huggingface.co