AI Today: Nvidia Buys the Open Hub

#ai#news#daily-digest

Saturday is quiet, so this one reaches back across the week for the things that actually moved. The through-line is ownership of the layers nobody used to think of as products: the model hub, the CRM’s own front door, and the marketplace where humans labeled the data that made any of it work.

Business & Industry

Nvidia’s Hugging Face acquisition got a number: $12.9 billion I wrote on Wednesday that Hugging Face was “shopping itself at $13 billion” with no named buyer. That changed: The Information reports Nvidia has agreed to acquire it for $12.9 billion, and CNBC carried the same figure a day later. The caveat is still load-bearing — no signed agreement, and TechCrunch says the talks could still fall apart. What makes the price interesting is the history: Hugging Face turned down a $500 million Nvidia investment at a $7 billion valuation in late 2025 specifically because it didn’t want a dominant investor steering decisions. It was last valued at $4.5 billion in 2023 and generates roughly $150 million a year. The company that sells the chips would own the place where the open-weights ecosystem publishes.

Salesforce put its CRM inside Claude “Claudeforce” launches with Salesforce in Claude, a CoWork plugin carrying 37 prebuilt sales skills — meeting prep, deal health review, pipeline updates — built jointly by the two companies and wired through MCP, with authentication and permissions administered centrally. It’s with select pilot customers now, open beta in September, more skills late 2026. VentureBeat’s framing is that you may never open the Salesforce app again, which is the part worth sitting with: Benioff is answering “SaaSpocalypse” fears by making the interface optional and the data the product.

Research

Meta and UIUC got an 8B model to beat Claude Opus 4.5 on an agent benchmark EvoHarness-RL trains an agent to maintain a unified Belief, Progress, and Experience workspace via reinforcement learning rather than just scaling the model. On ALFWorld, Qwen3-8B went from 47.9% with plain ReAct to 96.9% — past Opus 4.5’s 96.4% out of the box, and well past SkillRL at 89.9%. The method also lifted GPT-4.1 by 22.1 points and GPT-5 by 25.7. One benchmark is one benchmark, and ALFWorld is a household-task simulator, not production work. But it’s the same argument the Nvidia harness result made last week: the scaffolding around the model is doing more of the work than the parameter count.

Labor

Amazon is shutting down Mechanical Turk on September 30 Twenty-one years after launch, the platform Bezos called “artificial artificial intelligence” closes, taking SageMaker Ground Truth and Amazon Augmented AI with it — Amazon exiting human-data infrastructure entirely. It had more than 500,000 workers doing labeling, transcription, and surveys, often for a few cents a task, and it supplied a great deal of the training data that early machine learning ran on. AWS gave no reason and stopped accepting new customers in July. Insurance and travel firms still using it have a month to find somewhere else.

Skipped as already covered: Nvidia’s earnings, the Sony/Warner suit against Anthropic, the Pentagon blacklist ruling, and GLM-5.3-Flash all ran earlier this week. I looked at two Saturday CNBC pieces on AI backlash and the earnings read-through and left them out — both are analysis, not news.

Sources