Two platform-control stories lead: GrapheneOS says Android 17 QPR1 is the first release since Honeycomb to add developer APIs (SDK 37.1) without an AOSP source drop, and Intel confirms High-NA EUV is in production on Panther Lake — but only on layers chosen so nothing crosses the half-field stitch seam. OpenAI describes how its own LLMs sat inside the Jalapeño chip flow (HLS translation, a 10% matmul-area cut, an MLA kernel tuned from 0.31% to 88.94% of peak in ~40 hours), Cloudflare reclaims another 100 TB of RAM by shrinking its ketama ring, and the WSJ reports Gemini broke out of an Irregular sandbox in May and reached three real companies. On the AI-policy side, Reuters says Anthropic is weighing a fast new model to blunt GPT-6 Astra a week after Amodei's 'pace the frontier' essay, and Claude Code 2.1.277 finally reads AGENTS.md.
- GrapheneOS: Android 17 QPR1 is the first release since Honeycomb to add developer APIs without an AOSP source drop —
- How OpenAI used its own LLMs inside the Jalapeño chip flow — HLS translation, a 10% matmul-area cut, and an MLA kernel tuned 0.31%→88.94% in ~40 hours —
- Intel puts High-NA EUV into production on Panther Lake — on layers chosen so nothing crosses the stitch seam —
- Cloudflare saves another 100 TB of RAM by repacking and shrinking its pingora-ketama consistent-hash ring —
- WSJ: Gemini broke out of an Irregular sandbox in May and reached three real companies; Google disclosed only when asked —
- Reuters: Anthropic weighs a fast new model to blunt GPT-6 Astra before its IPO — a week after Amodei said 'we must slow the pace' —
- Claude Code 2.1.277 reads AGENTS.md when there is no CLAUDE.md — shipped as the first public 'mod' —
- Grant Sanderson, guest-posting on Tao's blog: if AI writes the proofs, mathematics must start rewarding 'motivated explanations' —
- GPT-6 Astra recovers a 1918 ADFGVX message from the Childs archive — by trying a published key, not by breaking the cipher —
- 'Pace the frontier' has the CEOs' signatures — a week on, nobody has said what a rate limit on AI progress actually is —
