OpenAI apologized for not reporting a warning sign before a shooting despite detecting it.
On an M5 Pro 64GB, Qwen 3.6 35B A3B was faster and more accurate than the 27B.
Default SSH access and an unsigned firmware update flow were found on the Rodecaster Duo.
SpaceX is accelerating its transformation into an AI-focused company by investing billions of dollars in AI based on Starlink's revenue.
A 36B MoE model built with Darwin V7 achieved 88.4% on GPQA Diamond.
Qwen3.6-27B ran at 85–85–100 t/s on a 24GB RTX 5090 Laptop GPU.
Open-Advisor supports the `/advisor` slash command across multiple AI CLIs.
Anthropic revealed that the Claude Code quality degradation was caused by product bugs, not the model itself.
It outlines the possibility of establishing a learning mechanics theory that explains deep learning training.
It has emerged that the Trump administration used an AI tool called SweetREX to push forward not only the deletion of regulations but also the drafting of legislation.
A new arXiv preprint unifies four ML failure modes into a single supervised learning theory.
GitHub Copilot's GPT-5.5 billing rate has risen from 1x to 7.5x.
ChatGPT Images 2.0 added text to an image that wasn't in the prompt.
Different FP4 inference support has been added to llama.cpp and ik_llama.cpp.
Ubuntu 26.04 sets up an environment that makes it easier to utilize AMD's XDNA2 NPU.
KillerScan v1.3.0 adds self-installation and improved mobile device detection.
A bug was found where Claude Code charges API rates instead of the subscription plan when 'HERMES.md' is included in a git commit message.
The llama.cpp CUDA MMQ kernel improved speed in some ranges using fastdiv and dynamic tiling.
OpenAI apologized for not notifying police about a risky account it had detected before a school shooting.
Google Cloud has officially launched the Gemini Enterprise Agent Platform, which extends Vertex AI.
DeepSeek applied a limited-time 75% price cut to the V4 API.
Simon Willison posted a short piece quoting a remark by Romain Huet.