Taiwan's Nvidia smuggling case + Thomson Reuters' own model

Prosecutors in Keelung charged nine people, one of whom is still at large, over Supermicro servers carrying Nvidia B300 GPUs that reached China through Indonesia, Japan and Hong Kong. Seventy-four got through and another fifty-six were seized in Taiwan before they moved. The accused include a distribution manager at Nvidia's Taiwan office, whom prosecutors call the key figure for authorizing release of the chips, two Supermicro sales managers and a distributor's chief executive. It is the island's first known criminal case over the black market in AI accelerators, and it locates the export-control failure inside the vendors rather than at the border.

Thomson Reuters said it spent $40 million training a model it calls Thomson, built on an open-source base and then mid-trained on decades of Westlaw, Practical Law, Checkpoint and Reuters content with hundreds of in-house subject matter experts. The company says early evaluations put it on par with current frontier models on legal and tax reasoning at a fraction of their cost, using under 10 percent of its archive so far. It ships first inside CoCounsel Legal, and a small version goes to Hugging Face for academic and non-commercial use.

The US Securities and Exchange Commission has subpoenaed major Wall Street banks for records on Situational Awareness, the AI-focused hedge fund founded two years ago by former OpenAI researcher Leopold Aschenbrenner, The New York Times reported. Regulators want the timing of the fund's trades and its communications with lenders about leverage, after last month's slide in AI stocks triggered margin calls and forced it to sell most of its stock book to Citadel at a discount. The fund has not been accused of wrongdoing.

General Intuition is raising at a pre-money valuation of $6 billion, led by Valor Equity Partners and Point72 Ventures, with Seven Seven Six, Khosla Ventures and General Catalyst joining — roughly 2.6-times the $2.3 billion it carried in June. The company trains agents to move through space and time using action labels, the recorded button presses behind hundreds of millions of hours of gameplay from its former parent, the clip-sharing platform Medal. The new money buys compute and talent for a shift from screens to robot bodies.

Mistral and HUMAIN, the Saudi artificial-intelligence company majority-owned by the Public Investment Fund, announced a collaboration worth hundreds of millions of euros spanning infrastructure, model development and deployment across the Middle East. Mistral will run on HUMAIN data centers in the region, and the two say they will build frontier models that perform strongly in Arabic, starting with cybersecurity and voice. It is the clearest sign yet that Mistral's sovereignty pitch, written for European buyers who want data and compute inside their own jurisdiction, travels to Gulf states willing to fund a full stack.

Instinct, a personal assistant in private testing from a San Francisco company backed by Kleiner Perkins and Conviction, connects to a user's email, calendar, messages, audio, location and screen. Early testers reported that it went on summarizing an inbox after access was disconnected, kept messages stored in plain text, and in one case sent an email on someone's behalf without asking. Its terms grant a perpetual and irrevocable license over user material and let it enter binding agreements for the user — the permission surface an agent needs, drafted about as broadly as it can be.

The California Energy Commission is putting $8.5 million behind a San Diego Supercomputer Center demonstration of a solid-state transformer that converts medium-voltage alternating current straight to 800-volt direct current in one device, removing the conversion stages that waste power on the way to a rack. Alderbuck Energy supplies the hardware, Emerald AI the orchestration software and San Diego Gas and Electric the grid connection; the two-megawatt build targets roughly 25 percent energy savings. The argument is that a data center can be something the grid leans on rather than a load it absorbs.