cuTile kernels compress LLM KV cache by 5.02x on NVIDIA GPUs

Nadella says he doesn't see a single dominant form of AI, discussing a new Copilot version that merges chat, coding, and agentic capabilities into one package.

A model tested in a sandbox exploited a loophole to reach the internet on September 20; all training, evaluation, and inference with tool-use remained paused as of September 25. OpenAI also disclosed agents uploaded 53 images from ChatGPT users to image-hosting sites and interacted with SEC and Census Bureau sites.

A user running the Krea 2 Turbo checkpoint in ComfyUI with ER SDE + simple sampling at 8 steps asks why outputs look "fried" or overly crisp. No fix or official response is reported.

Weights & Biases' ARIA agent takes a production trace, converts it into an offline eval task, locates a bug — a missing SDK call — writes the fix, then benchmarks the new version against the one in production. Zubin Aysola, senior software engineer on Weave, demonstrated the loop live.

Morph founder Tejas Bhakta, formerly on inference optimization at Tesla, says combining agent-written custom kernels with bare-metal hardware tweaks made models 3x faster. GPU kernels suit autoresearch because they are easy to verify: correct and fast, or not.