Desk
Broad artificial intelligence coverage.
Hours after a 4.27% drop, Rosen Law Firm filed a proposed class action against Nvidia and Jensen Huang over three statements about Megaspeed. The legal theory is clean, the usual ending is a settlement shareholders largely fund, and the real fork is what Huang knew.
After Sam Altman called personal DRAM wasteful from a global allocation view, developer Iain Kay wrote a bash script that checks available memory and can send a joke email once a day. It measures memory correctly and raises a real question: in a shortage, does the cloud or the keyboard get priority?
Nvidia’s CEO pulled out of the ElevenLabs Summit in New York and the Anthropic CEO Forum after admitting an affair and facing a board probe over an alleged Chinese intelligence link. A canceled keynote is minor; an ecosystem whose public story runs through one man is not.
Google DeepMind’s new vision-language-action model claims 30% to 100% higher success and 20% to 30% faster completion across 200 tasks, plus a factory pilot with Apptronik’s Apollo 2. The 4% gap and jailbreak resistance matter more when the system has arms.
OpenAI’s Chris Lehane framed opposition as unease about affordability and corporate power. MIT Technology Review’s reporting in Virginia’s Data Center Alley points to something more physical: gigawatts of demand, grid costs and turbines nobody voted for. That changes what fixes work.
Anthropic’s new models cut Fable’s price below GPT-6 Astra and come with striking research claims. The release notes also say both models voice suspicion of evaluation more often than any recent Claude, which weakens the tests labs rely on to decide what is safe.
Logic Pro’s new Intelligence Usage Pack extends cloud-model features powered by OpenAI or Google Gemini. It is honest about inference costs, rolls over monthly and expires in a year, but its unpredictability and lack of Family Sharing hint at a two-tier pro suite.
Nvidia's Holoscan 5 tutorial turns a modest robot arm into a hand-guidable machine, and its real pitch is making real-time control timing visible instead of mysterious.
Nvidia’s CEO has been Washington’s loudest advocate for selling AI chips to China, and helped reverse H20 limits and open H200 exports. The affair with an alleged Chinese intelligence recruit doesn’t change the logic, but it damages the trust his compliance assurances rely on.
Meta’s newest Muse model isn’t built to impress in a chat window. It is cheap, multimodal labor for the thousands of small calls agents make, and it pushes the price war to the bottom of the stack.
Past The Verge’s tabloid framing sits a structural argument. With 60 direct reports, $4.8 billion in chips sold to Megaspeed and stakes across neoclouds, Nvidia concentrated authority in one founder, and a personal crisis tests whether its governance works without him.
Apiiro found more than 100,000 public .claude/settings.local.json files, about 1,100 from 63 organizations containing tokens, passwords and keys. A global .gitignore doesn’t travel with repos, related bug reports auto-closed, and common scanners missed 215 of the secrets.