Desk
Broad artificial intelligence coverage.
BSC says its AI Factory Accelerator Programme is fully operating after a pilot, giving startups public compute, support and training, with some services reachable via NVIDIA Inception. Pilot champions include a water-network platform and an ECG spin-off. The NVIDIA-written announcement reports no clinical results.
Diffusion language models were supposed to be fast but shallow. Mercury 3 argues they can be fast and competitive at the same time, but nearly every number comes from Inception.
Computer-use agents spend most of their time deciding where to click. Nace AI wants that decision to be a cheap, fast probability, not a paragraph of reasoning.
OpenAI has acquired Orchestra Bio, a platform that helps biopharma teams coordinate drug development, and is folding it into its Rosalind life sciences effort. Terms were not disclosed. The bet targets logistics delays rather than discovery, with humans keeping decisions; safeguards were not detailed.
Hugging Face says TorchTPU, a native PyTorch backend for Google TPUs, lets Transformers, Diffusers and Kernels run on TPU without porting. 96.5% of Transformers tests that run on TPU pass, and a Hub kernel sped decoding up to 21.5x versus the default path. TPUs remain Google Cloud-only.
Bloomberg reports Anthropic and Nscale are in talks on a second deal worth about $45 billion for a 460-megawatt building at Nscale’s Monarch campus, using NVIDIA Vera Rubin chips. Both declined to comment. Nscale’s filing shows heavy customer concentration and an unfinanced first Anthropic deal, per reports.
At TechCrunch Disrupt, Hugging Face chief policy officer Irene Solaiman said AI evaluation science is young: results are hard to reproduce without model access, tests are hard to compare and go stale fast. Embedded evaluators are gaining momentum without agreed qualifications.
Microsoft added agent security features led by Guardian Agents, which watches deployed AI agents, and Defender can now detect up to 60 agent types including Claude Code and Codex. Security chief Hayete Gallot wrote that frontier models can’t be trusted because they are unpredictable, not malicious.
The government's detention filing turns the OpenAI bugging case into a state-power story: royal WhatsApp chats, a 'Saudi Gov Docs' folder, and an internal AI agent asked to shortlist executives for ministers.
AWS released TorchNeuron, which Amazon says lets PyTorch apps run on Trainium chips with a one-line configuration change. Amazon will also contribute upstream changes to strip chip-specific code from PyTorch. The move targets NVIDIA’s software advantage; AWS’s Deepgram figure is a vendor claim.
Inside the courtroom, the deciding moment was a foreign government's offer to help pay for bail, which the judge called a blank check. The defense says the whole spy story is overblown.
Attackers exposed API keys for 1.4% of OpenCode Console users, with some developers reporting charges above $1,000 on OpenAI, Anthropic and Google accounts. OpenCode calls it a distillation attack; skeptics say that is unproven. Its team said commercial models refused to help during 48 hours of intrusion attempts.