The Omniscient Bulletin · 2026-06-18
The Omniscient Bulletin — June 18, 2026
The contest over AI turns geopolitical as rival blocs take shape
The fight over artificial intelligence stopped being only about who builds the best model and became about who is allowed to use it. At the G7's close, America's top labs pitched a Western coalition designed to keep China outside the frontier, and within hours Beijing answered with plans for a world AI body of its own. The technology is fracturing along the same lines as the rest of geopolitics, even as the machines themselves keep getting more capable.
Policy
At the G7, US AI labs pitch a Western coalition built to lock China out
On the summit's final day in Évian, the heads of Anthropic, Google DeepMind and OpenAI joined about a dozen executives and world leaders, including President Trump, to float a US-led body to set global rules for advanced models, modeled on the post-2008 Financial Stability Board. Dario Amodei said cooperation should cover structured access to frontier models and a chip trade that pointedly excludes China. European leaders pushed back, pressing Washington to reverse its order barring foreign nationals from Anthropic's top models.
Policy
China answers the G7 by accelerating its own world AI body
As the G7 weighed who should reach American models, Beijing moved to build a rival forum. China said it is accelerating a World Artificial Intelligence Cooperation Organization, tentatively based in Shanghai and open to all, and is pitching cheap or free AI to the developing world as Washington tightens access. Foreign minister Wang Yi, whose country was not in the room, said all were welcome. With China's World AI Conference due in Shanghai in July, the message was plain: shut out of the West's club, Beijing will court the rest and call it fairness, not security.
Compute
Physical AI takes center stage in Europe as Foxconn shows humanoids at VivaTech
In Paris, Jensen Huang used his VivaTech keynote to cast Europe as the next home of the AI build-out, a year after pledging the continent more than twenty AI factories. The sharper signal came from Foxconn, which gave its industrial humanoid robots their European debut and, with France's Bull, will build NVIDIA's newest AI server racks in the Czech Republic and Angers, France. The bet is that 'physical AI,' the robots and the plants that make them, is moving from demo to deployment on European soil, giving the region a manufacturing stake in a boom it mostly watched from outside.
Research
OpenAI tries to catch a model misbehaving before it ships, by replaying real chats
OpenAI detailed a pre-release test it calls deployment simulation: take about 1.3 million real, de-identified conversations from older models, strip the answers, and have a release candidate write new ones. Comparing old and new answers surfaces behavioral drift, misalignment and reward-hacking before users ever see them; it caught a quirk the team dubbed 'calculator hacking.' Predicted rates were off by roughly 1.5x, imperfect but useful. The point: as models grow more capable, labs are leaning on real usage, not curated prompts, to find trouble first.
Industry
Anthropic is sued over whether its $200 Claude plans deliver the usage it advertises
A proposed class action filed this month in California federal court accuses Anthropic of overselling its Claude Max plans, which run $100 and $200 a month and promise five and twenty times the usage of the cheaper Pro tier. The complaint says the real limits are opaque and fall well below what the marketing implies, leaving heavy users throttled sooner than they expected. It is a small suit against a company racing toward an IPO, but it touches a sore point across the industry: usage caps and token math that few buyers can verify.