Meta's agents backfired + Nvidia circles Hugging Face

Reuters, citing scores of internal documents and more than 20 people, reported that Meta built a plan codenamed Project OT to make the company "AI native", with scenarios reducing some team headcounts by as much as 60 percent across two rounds of layoffs. Meta confirmed the exercise and said it did not proceed with every scenario. Internal posts described AI agents taking "large-scale, disruptive actions that humans are unlikely to execute", alongside a 40 percent rise in major technical and security incidents and up to 70 percent more staff time spent resolving them.

Three days after Hugging Face was reported to be fielding takeover interest with no bidder named, the bidder has a name. The Information said Nvidia agreed to buy the company for $12.9 billion; Business Insider, which broke the original story, said no agreement has been signed and the talks could still fall apart. Neither company has commented, which TechCrunch called notable given how quickly Nvidia normally responds. Hugging Face turned down $500 million from Nvidia last year, saying it did not want a dominant investor able to sway its decisions.

Salesforce and Anthropic announced Claudeforce, which exposes Salesforce data, workflows and business rules inside Claude as a plugin carrying 37 prebuilt sales skills, and makes Claude the default model across Slack, including Slackbot. Marc Benioff summarised the design as "the UI is the AI". The context is the reason it matters: Salesforce shares are down about 22 percent this year on investor fears that assistants from Anthropic and OpenAI will replace business software, and its answer is to put its own product inside one of them.

The same day, Anthropic agreed a six-year deal worth $45 billion with the British cloud provider Nscale, first reported by Bloomberg and unconfirmed by either company. The capacity comes from a West Virginia data center running Nvidia's Vera Rubin systems, and it does not begin until late 2027. It is the largest in a rapid run of leases that already includes $10 billion with Volta earlier this month and $5 billion with AMD in July, which means the commitments now run years ahead of the capacity actually existing.

Perplexity, working with Nvidia, brought its agentic Computer product to the desktop as Portable Computer, with local work consuming no credits. The engineering is the actual claim rather than the privacy pitch: the orchestrator running the loop is deterministic code instead of a second model, tool calls run in an operating-system sandbox, and the harness fails closed, disabling itself if that sandbox is unavailable. Escalating to a cloud model is permission-gated and shows the user what would leave the machine. It needs an Nvidia DGX Spark or a card with at least 24 gigabytes of video memory.

Google released Gemini 3.5 Transcribe, a speech-to-text model that strips filler words and applies spoken corrections as you talk, working across 85 languages and up to three speakers in pre-recorded audio. Google says it reaches finished text about 70 percent faster than its previous Chirp 3 engine, though the live error rate improves only modestly, from 7.32 percent to 5.5 percent — the speed is the real gain, not the accuracy. It is live in the Gemini app on macOS, Antigravity, AI Studio and the Gemini API, with Chrome to follow.

SpaceXAI has extended Grok Bot — its always-on agents, which work from their own cloud computer and sign into apps the way a person would — to every paid SuperGrok and Cursor plan. At the August 11 launch it reached only SuperGrok Heavy, Cursor Ultra, and Cursor Teams Premium, so the cheapest way in has fallen from Cursor Ultra at $200 a month to Cursor Pro at $20 in fifteen days. SpaceXAI also rewrote the original announcement to describe the wider availability.