AI Briefing

A federal judge just told the Pentagon it cannot punish an AI lab for refusing to hand over its models, which is the first time a court has drawn a hard line around what safety policies cost you in Washington. Elsewhere the theme is AI learning by watching rather than being told: a robot foundation model that copies a ten-minute task from one video, and Claude grinding through algebra that stumped a lab for eighteen months. Plus Meta patching the dumbest privacy hole in its glasses, and Pew's number on how many Americans now ask a chatbot about their symptoms.

01PolicyGizmodo ↗

A Judge Just Ruled the Pentagon Illegally Punished Anthropic for Saying No to Mass Surveillance

US District Judge Rita Lin issued a 59-page order finding that the Department of Defense's designation of Anthropic as a national security supply-chain risk was unlawful retaliation, violating the company's First Amendment rights and Fifth Amendment due process. The dispute began when the Pentagon wanted Claude available for 'all lawful purposes' and Anthropic held two lines: no mass surveillance of Americans, no fully autonomous weapons. Lin wrote that the empty invocation of national security is not a blank check to punish government critics, and that the move was driven by a desire to make an example of Anthropic for its 'arrogance.' It is the clearest signal yet that an AI lab's usage policy can survive contact with a hostile customer that happens to be the federal government.

02ProductSkild AI ↗

Skild's Robot Model Watches One Video of You Making Pancakes and Then Makes Pancakes

Skild AI released S1, a robot foundation model built as an in-context learner: you show it a single egocentric human video and it executes the task with no fine-tuning and no weight changes at all. On four tasks absent from pretraining, including potting a plant, brewing pour-over coffee and flipping pancakes, S1 hit 66% per-step success versus 9% for an identically trained language-prompted policy at the same 100k hours of data. Skild estimates one video in context is worth roughly 380 teleoperated post-training demonstrations, which for long tasks is 50 to 100 hours of human labour collapsed into one recording. If robotics has been stuck in the BERT era, this is the argument that prompting has arrived.

03ResearchPhys.org ↗

Claude Did Eighteen Months of Algebra in Five Weeks, and Got Some of It Wrong Along the Way

University of Colorado Boulder researchers Ankur Gupta and Arkava Ganguly had been stuck for a year and a half on a fluid mechanics problem about how nanoparticle shape affects movement in an electric field. Working with Anthropic's Claude on the long, detail-intensive algebra, and verifying every step themselves, they had a solution in five weeks, now published in the Journal of Fluid Mechanics. The answer: stretching a particle from a circle to a football shape changes its speed, but adding bumps or ripples does not. The researchers were explicit that the AI made subtle errors and that humans still had to frame the problem, pick the mathematical approach and interpret the result, which makes this one of the more honest accounts of what AI-assisted science currently looks like.

04IndustryEngadget ↗

Meta Is Finally Killing the Trick Where You Start Recording Then Cover the Light

Meta is rolling out a software update so its smart glasses stop recording if the capture LED is covered mid-recording, closing a loophole that had spawned tutorials online. An earlier update this year blocked recording from starting when the LED was blocked, but people simply hit record first and then taped over it. Meta's VP of AR Alex Himel announced the fix on Threads, and it is the second privacy patch in under two months as the company tries to get ahead of a backlash driven by people using the glasses to covertly film and harass strangers. The pattern is telling: the hardware shipped, the misuse was obvious, and the guardrails are being retrofitted one loophole at a time.

05ResearchPew Research Center ↗

A Third of Americans Now Ask a Chatbot About Their Symptoms, and Most Won't Share Their Health Data With It

A Pew Research Center survey of 3,488 US adults conducted in late June found that 34% use AI chatbots for at least one health-related task, led by quick health lookups (28%), working out what is causing symptoms (25%) and getting low-cost information (22%). Roughly 47% call the answers extremely or very helpful, but only 29% say they are very comfortable sharing personal health information with the tools, a trust gap that should worry anyone building consumer health AI. Adoption skews sharply young and by ethnicity: 44% of under-30s and 56% of Asian Americans, against just 17% of adults 65 and older. A companion Pew report published the same day found Americans think chatbots hurt more than help people using them for loneliness, depression or stress.