Desk
Broad artificial intelligence coverage.
Anthropic’s new meeting tool doesn’t just transcribe. It drafts the plan, maps timelines and pulls data from work apps while people are still talking, and by default it leaves when an outside guest joins.
Anthropic won developers with Claude Code. Its new office tools and in-app connectors aim at everyone else, and they arrive under pre-IPO pressure to show revenue beyond coding.
Laid end to end, the Huang timeline reads less like a sudden scandal and more like a slow compliance failure, with warnings landing at Nvidia's customer gate for two years.
Meta’s engineering post details editable Interested In and Not Interested In lists for Feed and Reels, managed in plain language by LLMs and enforced at ranking, reranking and on device. Blocks are hard rules; new interests fade back to engagement. How stated interests are used outside ranking is unanswered.
Triton Low-Level Extensions give engineers explicit control over warp groups, memory movement and synchronization on Nvidia and AMD GPUs. Meta says it typically matches hand-tuned CUDA, runs its largest training jobs and cut one project from months to weeks.
A reader’s Muse agent that compiled daily eBay watch digests reported being blocked, and The Verge reproduced it, with Muse sometimes routing to mirror sites. eBay won’t confirm new limits but says it takes a selective approach. Agentic shopping now runs on retailer permission.
Two weeks after its Fairwind cyber-defender debut, Google’s strongest model reaches the API, Ultra subscribers, AI Mode and Antigravity. At $2/$10 per million tokens until November 30, it targets long-horizon agent work, with better honesty numbers and a cyber capability now in many more hands.
At KubeCon, Nvidia contributed its agent-security runtime to the Cloud Native Computing Foundation as a Sandbox project, joined Google’s Agent Substrate for long-lived agents, and open-sourced NVKE, Kubernetes Recipes and Snapshot. Neutral governance helps; Nvidia’s gravity remains.
NVIDIA is shrinking a reasoning driving model to run inside the car, and the honest hard part is proving a smaller model's explanations still match what it actually does.
Google added two eighth-generation TPU variants: the AMD-co-developed TPU 8tx, which protects data and weights in use, and TPU 8i Max, which doubles memory, compute and bandwidth for inference. The confidential chip goes straight at enterprise compliance objections, if its security holds up.
Zyphra's tiny open voice model makes high-quality speech cheap enough to run in a browser tab, which is great for builders and a fresh headache for anyone worried about cloned voices.
Muse now offers voice calls, text and dictation from the wrist, after an iPad app, Spaces and avatar video chat in quick succession. An agent that reads inboxes and spends money on a tiny screen raises consent questions, and Meta’s ad business raises data ones.