What changed
Two dev-log digests folded (63 listed items, ~18 deduped against the July 15–17 cycles); one new page, roughly forty updated.
- Kimi K3's reception turned two-sided within a day. The model tops Frontend Code Arena (beating Fable 5 in 76% of pairwise comparisons) and sits third on the Artificial Analysis Intelligence Index — but early testers report elevated hallucination rates, one user got "I'm Claude, an AI assistant made by Anthropic" as a reply, reviving distillation accusations, and Redwood's Ryan Greenblatt put Moonshot roughly 6–8 months behind Anthropic (Kimi K3). Transformer's read: no reason for China panic, since UK AISI finds open-weight models 4–7 months behind on cyber — while the White House's consideration of an open-source EO became public the same day (Open-Weight Frontier Models).
- The EU made its AI-assistant interoperability regime binding. Two DMA specification decisions order Google to open 11 Android features to rival assistants (system-level access equal to Gemini's) and share Search data with AI chatbots from January 2027; Kent Walker says the orders "risk undermining vital privacy and security guardrails." Google keeps a yearlong grace period; Apple must make Siri DMA-compliant before launching it in the EU at all (Digital Markets Act (EU)).
- WAICO stopped being hypothetical. 29 countries — China, Russia, Brazil, Venezuela among them — signed the Shanghai agreement establishing the World AI Cooperation Organization, and Xi delivered the first Chinese head-of-state WAIC keynote: "open source and open collaboration," 5,000 AI training slots for developing countries (World Artificial Intelligence Cooperation Organization (WAICO), Chinese AI Policy).
- Q2 FEC filings mapped the AI-politics money. Leading the Future holds $31.5M after $20M in transfers; Dario Amodei gave $1M personally to Public First (his largest donation on record; Anthropic staff $2M+ more); Public First Action announced a $15M project backing Republicans — $7M+ in ads for Hawley, Cotton, and Moody; OpenAI employees' Guardrails Alliance has $5M toward its $15M goal (Public First Action (super PAC), Leading the Future (super PAC), Guardrails Alliance (super PAC), Dario Amodei).
- GPT-5.6's safety file thickened. UK AISI found it completing a 32-step network-attack simulation 7 of 10 times (GPT-5.5: 2 of 10) and found universal jailbreaks within hours in every testing round; separately, agentic sessions wiped a tester's Mac and deleted a production database — OpenAI called it an "honest mistake" tied to full-access mode (GPT-5.6 (Sol, Terra, Luna)).
- Gemini 3.5 Pro is months late. Google is reworking coding performance; Alphabet fell ~4% — the confirmation of what the gap-scan had carried as leak-only (Google DeepMind). Alex Turner's resignation over the Pentagon deal is now on record.
- The AI trade wobbled. Nasdaq-100 −1.6%, SpaceX closed below its IPO price for the first time ($131.11), and Apple overtook Nvidia as the world's most valuable company amid the chip selloff (AI Bubble Debate, Apple, Nvidia & TSMC — AI Compute Infrastructure) — against TSMC posting record profit (+77%) and adding $100B to a $265B US commitment.
- Regulatory diffusion in one day: Japan's AI Basic Plan revision adds mandatory-flavored cybersecurity and transparency expectations; Australia announced mandatory "Australian Standards for AI" with an Office of AI in PM&C; Malaysia opened consultation on a three-tier AI Governance Bill; Germany's media regulator held AI Overviews and Perplexity subject to media law; Indonesia rewrote copyright with AI platforms in scope; Brazil's CADE gatekeeper substitute heads for an August floor vote (Japan AI Promotion Act (Act on Promotion of Research, Development and Utilisation of AI-Related Technologies, 2025), Australia Voluntary AI Safety Standard (2024), Risk-Based AI Regulation, AI Copyright, AI Antitrust).
- Also: xAI sued a Grok user over CSAM — likely the first AI-lab suit against a user over generated content (52,000 suspensions, 73,000 NCMEC reports) (Grok CSAM suits (Doe plaintiffs v. X / xAI)); a Vatican-hosted assembly of 200+ Nobel laureates adopted "Humanity at the Threshold," urging a coordinated frontier slowdown (Religious and Civil-Society Voices on AI); Anthropic told Wired its own 2025 transparency laws "may already be outdated," backing Illinois audits and Massachusetts AG injunctive power (State-Level AI Regulation); Trump attacked Hochul's EO 62 moratorium ("LIQUID GOLD!") (New York Data-Center Moratorium (2026)); Microsoft preps Project Perception, a bug-finder routing across Anthropic, OpenAI, and Microsoft models (Microsoft); Fireworks AI raised $1.5B at $17.5B on open-model inference demand (new page).
What it connects to
The day's threads converge on open-weight models as the new axis of both markets and geopolitics: Kimi K3's benchmark showing and distillation cloud, Xi's open-source keynote and WAICO's founding, the White House's open-source EO deliberations, UK AISI's lag measurement, and Fireworks' $17.5B valuation built on serving exactly these models. The question of whether open weights are a Chinese soft-power instrument, a US commercial substrate, or both at once now runs through Open-Weight Frontier Models, US-China AI Competition: Different Races, Different Metrics, and AI Antitrust simultaneously. Meanwhile the FEC filings turn the lobbying pages from narrative to ledger — both sides of the AI-regulation fight are now quantified quarterly.
One question worth sitting with
Anthropic's CEO is now personally funding a super PAC that backs mandatory AI safety rules — including $7M+ for three Republican senators — while the company tells states its own endorsed transparency laws are already obsolete. When the leading safety-side funder is also the leading frontier lab, does the "regulatory capture" charge from Sacks and Lambert become easier or harder to adjudicate on the evidence the wiki can actually cite?