AI's power centers get reshuffled

Leadership is changing at the company that supplies much of the world’s AI computing, while a major model developer is preparing for a possible public listing. New models and an undocumented on-device classifier add another question: what can AI systems do, and what can they infer about the people using them?

Jensen Huang resigns as Nvidia CEO after 33 years

Jensen Huang resigned on November 12, with Colette Kress taking over as interim CEO. The departure followed reports that Nvidia was preparing to remove him and a letter from Huang to staff addressing an affair; employees have since petitioned for his return.

Why it matters: Nvidia’s accelerators underpin most large-scale AI training, so a change at the top of the main supplier affects compute plans well beyond one company. How the interim leadership handles Nvidia’s open-source work and its US policy posture is now unsettled. (Jensen Huang resigns)

Anthropic launches roadshow for proposed IPO

Anthropic, PBC began the roadshow for a proposed initial public offering of Class A common stock on November 10. The offering is still proposed, but the roadshow brings the company’s plans directly to prospective investors.

Why it matters: A public listing would put a major model developer in front of ordinary investors and force it to disclose numbers that labs have kept private. That makes it the clearest look yet at how the economics of frontier AI actually work. (Anthropic launches roadshow for proposed initial public offering)

Anthropic releases Claude Fable 5.2 and Mythos 5.2

Anthropic describes Fable 5.2 and Mythos 5.2 as its most capable models for coding, finance, and scientific work. In an early experiment, a team of 1,000 agents used Fable 5.2, general tools, a shared database, and layered safety controls to reproduce a formal proof of Fermat’s Last Theorem.

Why it matters: The experiment shows how a large group of agents can coordinate through shared tools and data to tackle a demanding task. Its layered controls also make the safety design part of the result, since agent teams need clear limits as their work becomes more complex. (Introducing Claude Fable 5.2 and Claude Mythos 5.2)

Goldman Sachs pulls Claude Code from bankers and traders

Goldman Sachs is removing Claude Code from its bankers and traders, according to reporting on the firm’s internal rollout. The change would reverse a high-profile deployment of an AI coding assistant inside a major financial institution.

Why it matters: Enterprise adoption can be reversed after a tool reaches employees, so access alone does not prove that a deployment is working. Organizations evaluating AI coding tools need to account for how well they fit existing workflows and controls. (Goldman Sachs is pulling Claude Code from its bankers and traders)

Every Mac ships with an undocumented family member classifier

An examination of Apple Intelligence found that it builds a personal knowledge graph on the device, including an undocumented classifier that infers family relationships from ordinary data. Those inferences happen locally rather than in the cloud, but the classifier’s purpose and behavior are not documented for users.

Why it matters: On-device personalization can infer sensitive facts, such as family relationships, from data users supplied for another purpose. People need ways to understand what the system infers and inspect those conclusions, even when the processing stays on their device. (Every Mac comes with an (undocumented) family member classifier)