What changed
Two dev-log digests folded (25 items, ~7 deduped); eight new pages, nineteen updated.
- Thinking Machines shipped its first model. Inkling is an open-weight 975B-parameter MoE (~41B active, 1M context) trained on 45T multimodal tokens on GB300 NVL72 systems — positioned by the company itself as "not the strongest overall model," a deliberate enterprise-fine-tuning play through Tinker. Notable disclosure: its post-training bootstrap used synthetic data from Moonshot's Kimi K2.5 — a US lab openly building on a Chinese open-weight model (Thinking Machines Lab).
- Anthropic's IPO now has a date range. Investor meetings are being scheduled for a listing as soon as October, with Goldman Sachs, Morgan Stanley, and JPMorgan involved, plus talks to expand its $2.5B revolver by several billion — a listing on this timeline would precede OpenAI's (Anthropic). The implementation JV also got a name and a shape: Ode with Anthropic, built on Fractional AI, 100 engineers, "Claude-first."
- Cox v. Sony entered the AI docket by osmosis. EFF's Corynne McSherry called the unanimous March SCOTUS contributory-liability reversal a "clean, clear" defense for AI developers across the ~100 pending copyright suits (Cox v. Sony (contributory copyright liability), AI Copyright).
- TSMC added another $100B to US manufacturing (four-plus more Arizona plants, $265B announced total), and the House China committee chair urged banning purchases of CXMT/YMTC memory — the demand-side complement to export controls, with Apple's lobbying named (Nvidia & TSMC — AI Compute Infrastructure, Export Controls (AI)).
- OpenAI disclosed GPT-Red, a self-play automated red-teamer it says beat human red-teamers 84% to 13% on held-out prompt injections and will not release, and published "Mind the US-China Safety Gap" alongside a new Public Policy Agenda site (Jailbreaking and Red Teaming, US-China AI Competition: Different Races, Different Metrics). Its first shipped hardware turned out to be a $230 keyboard for driving Codex agent fleets.
- Qwen will power Apple Intelligence in China after the CAC added Apple's AI services to its approved list — the delayed 2025 rollout finally cleared (Alibaba / Qwen Team, Apple).
- Also: Grok CSAM suit series tracked (Grok CSAM suits (Doe plaintiffs v. X / xAI)); WSJ threats-against-executives investigation anchored a new AI Backlash page; Guidelight's Control standard v1.0 codified the AI Control agenda as an assessable standard; Google's Steel River solar purchase (1 GW, would become the largest US solar facility) sits 40 miles from xAI's unpermitted turbines.
What it connects to
The Inkling release and the Qwen-Apple approval are two sides of the same open-weight interdependence story: a US lab bootstrapping from a Chinese open model the same week a Chinese model becomes the AI layer of the iPhone in China — while OpenAI publishes a report arguing Chinese models are categorically less safe and Congress moves to wall off Chinese memory chips. The wiki's Open-Weight Frontier Models and US-China AI Competition: Different Races, Different Metrics pages now hold both the integration and decoupling threads simultaneously. On the capital side, Anthropic's October timeline turns the "race to public markets" from a framing into a calendar; the credit-line expansion alongside it echoes the vendor-guaranteed-lease pattern in Circular Financing in AI — debt structures growing up around AI ahead of the equity events.
One question worth sitting with
If Cox v. Sony really does neutralize the "infringement machine" secondary-liability theory, the ~100 copyright suits compress onto a single question — whether training itself is fair use — just as Anthropic's $1.5B settlement and the labs' licensing deals have been quietly removing the plaintiffs most able to litigate it to judgment. Who is left standing to actually get that ruling?