Anthropic's red lines end Pentagon talks; Midjourney presses studios

A 346-page court filing made public in early July shows months of emails between Anthropic chief executive Dario Amodei and Pentagon undersecretary Emil Michael. Amodei held two firm conditions — Claude could not power fully autonomous weapons or domestic surveillance — while the Pentagon insisted on "all lawful uses." When Amodei refused to soften the limits, Defense Secretary Pete Hegseth designated Anthropic a supply-chain risk and ended negotiations. It is the sharpest public look yet at a frontier lab setting boundaries on how the military can use its AI.

In its copyright fight with Disney, Universal, and Warner Bros., Midjourney filed a motion to widen discovery — demanding the studios disclose how they use generative AI internally, not just in consumer products. Midjourney argues the studios are "cherry-picking" documents and hiding evidence that they do "exactly what they are suing Midjourney for doing." The gambit aims to bolster a fair-use defense by showing that the studios themselves lean on AI trained on copyrighted work.

The United Nations telecom agency, the ITU, and more than forty founding members launched the AI for Good Global Commission in Geneva, co-chaired by Rwandan President Paul Kagame and Salesforce chief Marc Benioff. Members include leaders from Amazon, Microsoft, Nvidia, Accenture, and heads of state. The body aims to widen AI access for the 2.2 billion people still offline and set shared norms for responsible use; it holds its first meeting at the AI for Good Summit in Geneva from July 7 to 10.

Meta AI chief Alexandr Wang told staff that the company's next frontier model — codenamed Watermelon and still training — already matches OpenAI's GPT-5.5 on internal evaluations, using an order of magnitude more compute than its predecessor, Avocado. Treat the claim cautiously: it rests on unnamed, vendor-run benchmarks, the least reliable category of evidence in frontier AI, and even Mark Zuckerberg struck a more cautious tone at the same town hall. Still, it signals Meta's costly bid to close the gap after a year of talent churn.

Simon Willison, creator of the popular sqlite-utils library, published a candid breakdown of building its 4.0 release largely with Anthropic's Claude Fable agent: 37 prompts, 34 commits, and roughly 149 dollars in API costs. The agent even caught a severe data-loss bug in the release candidate before it shipped. His takeaway — that layered, multi-model review "really does work" on hard tasks — is a rare, itemized look at what agentic coding actually costs and delivers today.

A new Google commercial reimagines the Founding Fathers writing the Declaration of Independence using Google Workspace — Docs suggestions, Meet calls, e-signatures, and Gemini taking notes — under the tagline "Group project, but make it 1776." Reception split sharply: praised on YouTube, panned elsewhere as "tone deaf," with historians arguing AI is the wrong metaphor for revolutionary collaboration. Critics also noted the ad itself appears partly AI-generated, feeding a broader debate over AI in creative and civic work.