The World Cup Is Secretly Run by AI

Last week the US government gave Anthropic's best model a 72-hour public life, then switched it off for every foreign national on earth. This week, three Chinese labs shipped models that beat almost everything in the open, handed over the weights, and charged a tenth of the price. We unpack the split-screen, then debate two AI ideas worth building right now.

Signal or Noise

1. FIFA is running the World Cup on AI. The 2026 tournament kicked off June 11 across 16 host cities, and underneath it sits a production AI stack: an Intelligent Command Center in Miami tied to digital twins of all 16 stadiums for real-time crowd prediction, Football AI Pro built with Lenovo on FIFA's own Football Language Model and handed to all 48 teams, 1,248 player avatars from one-second scans, semi-automated offside down to a 10cm margin, and Google Gemini drawing up tactics for Argentina, Brazil, and France. The biggest live multi-model AI deployment on earth, and almost nobody is framing it that way.
2. China's open-weight wave. Z.ai's GLM-5.2 lands with a million-token context window and open MIT weights, topping the open-weight rankings. Moonshot's Kimi K2.7 Code ships for agentic coding. They slot into a 2026 run where MiniMax M3 leads SWE-bench Pro among open models at 59% for $1.20 per million tokens, and DeepSeek V4 Pro ships full weights under MIT. Chinese labs now hold four of the top five open-weight spots.
3. Anthropic's Fable 5 and Mythos 5 are still dark. A full week after the June 12 export-control directive, both models remain suspended for foreign nationals with no end date. Anthropic published its defense on June 16, disclosing a narrow jailbreak and arguing the standard would freeze all frontier deployment industry-wide. The model that topped the coding leaderboards has now spent more time off than on.
4. Noam Shazeer leaves Google for OpenAI. The Gemini co-lead and co-author of "Attention Is All You Need" is gone, two years after a reported $2.7 billion deal brought him back. When models converge, talent is the moat.

Ship It or Skip It

Two real ideas from this year's agent discourse. Steelman, pushback, verdict.

• AI governance layer for the agents your employees already run. Sit between the agent and the action and check every move against the business rules the company runs on. Block the $40k refund, flag the unapproved discount. The value is enforcement, not a dashboard. The risk is that Microsoft already open-sourced one.
• Vertical AI harnesses, one trade at a time. Stop selling a chat box and sell the opinionated rig for a single trade like marketing or video: agents, workflow, brand rules, output formats. The moat is the rig, not the model. The risk is picking a red ocean.

Closing hot takes

• Oscar: The most important AI lab of 2026 might not be American, and most US builders are too proud to notice. Four of the top five open-weight models are Chinese, they ship the weights, and they run at a tenth of the closed API you are scared to lose.
• Matt: A free model is only free if you already own the GPUs, the ops team, and the eval harness to babysit it. Open weights won the leaderboard. They have not won a single Tuesday on your roadmap.

Your hosts

• Oscar Gallo, AI Engineer and Entrepreneur. Builds AI products, ships code, runs companies.
• Matt Wozniak, Builder, operator, relentless executor. Builds, ships, scales.

Links referenced in this episode

• PCMag, Inside FIFA's secret AI command center powering the 2026 World Cup
• AkitaOnRails, LLM benchmark: Kimi 2.7 Code, GLM 5.2, MiniMax M3
• LLM-Stats, Kimi K2.7 Code
• Anthropic, Statement on the US directive to suspend access
• CNBC, Noam Shazeer leaves Google for OpenAI
• GeekWire, The rise of vertical AI agents