Desk
Broad artificial intelligence coverage.
Google says early coverage overstated a threat to cut Search for schools refusing AI. Districts can disable Gemini apps, or the whole bundle including Search, but cannot run Google Search with AI Overviews and AI Mode removed. More toggles, same refusal.
OpenAI says 100,000+ advertisers have used ChatGPT ads since February, Google expanded AI Mode ads, and Koah Labs serves ads across 200+ apps. Anthropic stays ad-free. WIRED’s reporting shows what works, and why full-conversation targeting is a different consent bargain.
Baidu opens ERNIE 5.5 Flash Base — a 280B MoE checkpoint with 14B active parameters, 33.5T tokens of pretraining, and self-reported leads on 24 of 30 base benchmarks.
In a 100-megawatt comparison, Nvidia claims Vera delivers 1.6x more agents per watt than the fastest x86 part and 2x the agent speed of the most efficient. Partner numbers on sandboxes and orchestration are striking; independent benchmarks will decide if the CPU becomes a first-class AI factory choice.
Muse Voice, shown at Connect, now lets US users call and video chat with a customizable avatar on mobile, web and Mac, with AI Glasses and the Muse Charm wearable next. The craft is impressive; the attachment strategy, aimed at an agent that shops, needs care.
NVIDIA released four 16-billion-parameter Cosmos 3 Nano world models for multi-view driving, streaming sensors, simulator-to-video and two-armed robots under the OpenMDW 1.1 license. Open weights help the field compare; they don’t replace real-world testing.
Dina Powell McCormick told the WSJ Tech conference Democratic gains and anti-data-center campaigns will make new sites harder, and admitted the benefits weren’t explained well. Her deal template, paying for power and sharing tax surplus, is the minimum social license.
OpenAI launched ChatGPT Cyber for its Daybreak Red program and touted hundreds of Linux and Firefox findings. A hobbyist’s two-hour Ubuntu-related bug, found with an older model, shows how fast the threat is moving, and raises the question of whether OpenAI’s defensive budget matches its warnings.
At a WSJ conference, Brockman admitted a rough first half when Claude Code owned agentic coding, then credited Codex and enterprise sales for 70% quarterly growth in annualized revenue by Q3. The mix shift is the IPO story; the S-1 will have to prove it.
Anthropic’s beta lets US Pro, Max and Ultra users link bank and card accounts via Plaid so Claude can answer questions about balances and transactions. It can’t move money, carries no ads or Anthropic financial products, and hub chats stay out of training.
A Reddit write-up routes a PlayStation 5 through a 4K capture card into a PC running open-source Veyra, which applies DLSS 5 and frame generation before sending video to a TV. The trailer demo looks sharper and stranger; live play adds latency nobody has fully measured.
UNSW, Sydney, La Trobe and Adelaide are staging OpenAI tools to tens of thousands of students and staff, with assessment rules, simulations and an SME consulting studio. An OpenAI-commissioned survey shows heavy use; the risk is access outrunning assessment design.