TechCrunch has a piece out today arguing that AI safety discourse has become unhinged — not in the direction of complacency, but in the direction of unfalsifiable catastrophism that makes productive policy conversation nearly impossible. That lands alongside a Wired story documenting the enforcement mechanics (or lack thereof) of any hypothetical AI slowdown, and another on AI PACs dropping nearly $1 million into a Senate race most people can't name. The through-line: the gap between what AI safety advocates say and what anyone can actually do about it keeps widening, and the political apparatus filling that gap isn't encouraging.

Meanwhile, Unchained's reporting on the Senate floor death of the Clarity Act is a useful reminder that the legislative branch remains the most reliable lagging indicator in tech policy. The AI slowdown pact referenced in that piece has no enforcement architecture, which is precisely what Stanford's ongoing work on open-weight model governance is trying to address — though arguing that open-weight isn't enough is easier than specifying what enough looks like.

Yesterday's Astra for Law launch — covered in the previous edition — has moved into its second phase today: the vendor integration flood. LawNext and Legal IT Insider are both tracking the 26-partner ecosystem that shipped alongside GPT-6 Astra for Law, with downstream integrations now actively positioning around the new platform. This is how platform lock-in gets built in enterprise software: not by the anchor launch, but by the secondary integrations that make switching expensive before anyone has decided to commit. Legal tech vendors that weren't on the launch list have roughly one product cycle to articulate a credible response.

Worth watching: Vals, the a16z-backed benchmarking startup, is attempting to establish itself as the credible third-party scorecard for AI systems. In a legal context, where the gap between benchmark performance and courtroom reliability is particularly consequential, a neutral evaluator with a defensible methodology would be genuinely valuable. Whether Vals can get there before every major vendor builds its own scorecard is the interesting question.

A startup-that-builds-startups has raised $100 million with a declared focus on physical AI — robots, sensors, embodied systems — signaling that the infrastructure layer for real-world AI deployment is attracting dedicated fund structures, not just individual bets. This joins the Mecka AI near-$500M Sequoia round from the previous cycle as evidence that robot training data and physical deployment tooling are being treated as scarce strategic assets, not commodity infrastructure. The Huawei Cloud global AI cluster announcement adds geopolitical texture: compute sovereignty is now a product category.