The 52nd G7 summit wrapped today with declarations on migration, drug trafficking, and cancer — but no standalone AI commitments text. References to AI governance were watered down after the US opposed multilateral rules that could blunt its industrial edge. A final-day working lunch sat Sam Altman, Dario Amodei, and Demis Hassabis (plus ~a dozen tech chiefs) before the leaders; Washington's export controls on Anthropic's Fable 5 and Mythos 5 were a central topic, with France, Canada, and South Korea pressing “digital sovereignty.” The Atlantic Council's Emerson Brooking: the controls “changed everything.” CNBC → Joint statements →

OpenAI unveiled a global Partner Network backed by $150M (live July), with Select/Advanced/Elite tiers, launch partners Accenture, Bain, BCG, McKinsey, and PwC, and a target of 300,000 certified consultants by year-end. Its stated thesis: model capability is no longer the primary barrier to adoption — identifying use cases, redesigning workflows, integration, and change management are. It follows Anthropic's $100M Claude Partner Network from March. OpenAI → TechTimes →

OpenAI detailed a method that predicts how a new model will act in production by regenerating ~1.3M de-identified past conversations with the candidate model, then checking the forecast against real traffic after launch. It blunts “evaluation awareness” (models can tell when they're being tested), hit a median 1.5× error, and surfaced a novel “calculator hacking” failure mode. OpenAI is explicit it complements, not replaces, red-teaming and tail-risk analysis. OpenAI →