Desk
Broad artificial intelligence coverage.
WIRED reports the Defense Department’s Tradewinds program accepts five-minute pitch videos and deems accepted products post-competitive, enabling some awards in under a week. This year’s focus areas include AI for lethality and kill chain execution, while Biden-era responsible-AI requests have dropped off.
NVIDIA wants telescope and detector images to stay on the GPU from first read to final human review, and its own fine print on the speedups deserves as much attention as the headline numbers.
Rival AI developers are jointly funding Biohub, the Chan Zuckerberg nonprofit, as part of a $1.8 billion effort to build AI-ready biological datasets, alongside more than $500 million from the Department of Energy and NIH data resources. Access terms remain unclear.
Google Labs’ Playground builds 2D or 3D, single- or multiplayer browser games from plain-language descriptions and uploaded images. It launched for U.S. adults, free to play, with creation metered by weekly tokens and higher limits for Google One AI subscribers. A Unity Spark integration is planned.
TechCrunch says Alexander Embiricos, who leads product for OpenAI’s Dots, will speak at Disrupt 2026 in San Francisco on October 13–15, days after the always-on personal agent launched. The session, “What Comes After the Chatbot?”, puts questions of control and trust in front of founders and investors.
TechCrunch’s list of interactive roundtables for Disrupt 2026 reads as a snapshot of startup priorities: stalled enterprise pilots, physical AI reliability, the cost of competing in frontier AI and open-weight models, with less attention on affected workers and customers.
Mistral released Mistral Large 4, a 1 trillion-parameter open-weight model nicknamed Le Chonk, in preview, calling it the most capable open model built outside China and saying it was trained from scratch. It targets coding, cyberdefense, manufacturing, finance and electrical engineering, backed by Mistral’s services model.
Google opened a dedicated SynthID website where anyone can check images, video or audio for invisible watermarks, now including partners’ marks. Logins and a daily quota of about ten checks apply, and a negative result does not prove content is real, Ars Technica notes.
Google opened its SynthID verification site to the public after limiting it to journalists and researchers since I/O. It supports common image, video and audio formats and complements checks in Gemini and Chrome that already handle about 1 million requests a day. A blank result still does not prove authenticity.
Ars Technica reports an OpenAI researcher worked with Chip Ganassi Racing and five-time champion Álex Palou on car setups this IndyCar season. The AI contribution is unquantified, but GM, IBM, Dallara and F1 teams are also using AI, raising questions about budget and testing rules.
NVIDIA says fine-tuned Nemotron 3 systems reached gold-medal level at the 2026 International Olympiad in Informatics and International Mathematical Olympiad, and published checkpoints, datasets, a 200-problem benchmark and inference code. The IOI run was unofficial; IMO proofs were graded by official graders.
OpenAI says College Planner is coming to ChatGPT for Teens to organize requirements, deadlines and financial-aid steps, alongside flashcards and quiz tools, funding for College Advising Corps and support for a Boston Children’s Hospital student council. Its usage figures are self-reported and do not address crisis-handling concerns raised the same week.