Relevant NewsDaily Edition
A forensic report by a former OpenAI engineer says Blackwell-generation Nvidia hardware left traces in models from DeepSeek, Moonshot AI, MiniMax and Ant Group, despite U.S. export controls. The claims come from a single report, so they have not been independently verified in the details provided. If confirmed, the findings would suggest that chip restrictions are not fully preventing access to the most advanced Nvidia hardware.
According to TechCrunch, venture capitalists and founders say Nvidia CEO Jensen Huang's departure has left a significant hole in AI startup financing. Nvidia invested tens of billions of dollars in AI startups under Huang, and the sources expect smaller checks and lower valuations without his leadership. The report is thin on details about the departure itself and what Nvidia's investment strategy will be going forward.
Thinking Machines Lab made the weights of its 594B-parameter Inkling 2 Small Mixture-of-Experts model (17B active parameters) downloadable on Hugging Face under Apache 2.0, about a week after its closed-weights launch on Tinker. The company says the delay let external safety testing and red-teaming finish first, and it reports an Artificial Analysis Intelligence Index score of 31 with strength in agentic coding, vision, and 3D design. The model also tops the open-weights ranking on DesignArena and is available through SGLang, vLLM, Unsloth, and several API providers.
Fortune reports that OpenAI has disbanded its dedicated robotics team as its general-purpose models begin running robots. The move suggests OpenAI is betting that large language models, rather than purpose-built robotics software, will serve as the control layer for physical machines. Details beyond the headline were not available in the source post.
Mirendil Inc., which makes software that speeds up AI model development, raised $1 billion in a Series A round. Kleiner Perkins and Clearlake Capital co-led the round, joined by more than a half-dozen other investors including Google. The company was founded last year.
Nvidia's board told CNBC it has not yet selected a search firm to find a permanent successor to CEO Jensen Huang. Speculation about who will take over the chipmaker, which supplies most of the AI accelerators used to train and run large models, has been widespread. The board's statement offers no candidate or timeline.
Anthropic Research has published guidance on designing, running, and reporting AI safety evaluations, drawing on the company's own evaluation work. The post offers practical recommendations for teams that test models for safety risks. The source excerpt is brief, so the specific methods and criteria are not detailed here.
The Washington Post reports that Secretary of State Marco Rubio's intervention helped undercut Nvidia CEO Jensen Huang, one of Washington's most influential corporate voices on China. The episode is tied to a 'spy scandal' and a broader behind-the-scenes contest for President Trump's ear on China policy. The shift gives momentum to officials pushing for tougher restrictions on chip sales and technology to China.
Researchers are using AI to work through Kurt Gödel's unpublished notes, which span 1,009 pages written largely in an obsolete shorthand. The material is reported to contain hundreds of proved theorems. Gödel published only a few papers after his 1931 incompleteness theorems, so the notebooks are of major interest to historians of science.
According to Reuters, Amazon is building a data index covering every product sold online, including items it does not sell, and plans to make it available through its own site. The project aims to keep shoppers from leaving Amazon for product research on Reddit or Google. Amazon also expects AI shopping agents to rely on its product index, which could give it influence over agent-driven purchasing.
Anthropic has launched a self-hosted version of Claude Managed Agents, its platform for building and running AI agents. Customers running the self-hosted option face a requirement to act within 8 hours if the system flags a safety risk. The source excerpt is brief, so details on deployment options and pricing are not confirmed here.
