Michigan Governor Gretchen Whitmer's push for a nuclear-powered AI data center has ignited fierce local opposition, with township leaders drawing a hard line. The standoff crystallizes the growing tension between AI infrastructure ambitions and the communities expected to host them.
A new paper reframes prompt injection not as a jailbreak problem but as a fundamental role confusion — models can't reliably distinguish who is giving instructions. Willison praises the companion blog post format, arguing every academic paper should ship a readable version alongside the PDF.
Willison ported Moebius — a 0.2B parameter model that matches 10B-level inpainting performance — to run entirely in-browser using Claude Code. Tiny, capable models running locally without a server call is becoming a real workflow, not a curiosity.
Exa CEO Will Bryk argues that traditional search was never designed for AI agents as users, and the retrieval layer needs a full rebuild. As agentic workflows scale, whoever owns the search infrastructure for agents owns a foundational piece of the stack.
Early Gemini Omni API demos show agents navigating via live video feeds, dashboards, and visual UI — not just text. This signals a shift from agents that read data to agents that see and act inside visual environments like a human would.
Bitcoin slipped below $63K as a broader tech selloff pulled risk assets down in tandem. SpaceX's $600 billion valuation plunge in three days erased the equivalent of nearly half of Bitcoin's entire market cap — a stark reminder of how correlated these ecosystems have become.
Kobeissi flagged a remarkable data point: Chinese semiconductor exports surged 111% year-over-year in May to a record $36 billion, driven directly by AI demand. The US export controls narrative is getting complicated — China isn't being starved of the AI boom, it's feeding it.
Trump signed executive orders to accelerate quantum computing development, adding another front to the US technology race. Quantum, AI, and space are now all simultaneously receiving federal push — the policy stack is getting crowded, and execution risk is rising proportionally.
Ethan Mollick noted that ten days after the Claude Fable 5 situation broke, it remains genuinely confusing — contradictory reports, no clean resolution. The fog around what exactly happened and what it means for AI governance is thick, and that opacity is itself the story.
Nous Research's Hermes model just crossed one trillion tokens processed in a single day on OpenRouter. That's not a benchmark number — that's real usage volume, and it signals that open-weight models are absorbing serious production workloads, not just hobbyist experiments.
Demis Hassabis flagged a new research partnership between Google DeepMind and A24, the prestige film studio. The explicit goal is ensuring creative tools of the future are shaped by actual creators. It's an unusual pairing — and a deliberate signal about whose taste should govern AI's aesthetic outputs.
Two forces are colliding this week with unusual clarity. On one side: AI infrastructure is scaling faster than the communities, markets, and regulatory frameworks around it can absorb. A Michigan township vowing to fight a nuclear data center "to our last breath" isn't a NIMBY story — it's a preview of dozens of similar fights coming as the power demands of AI hit physical geography. On the other side: the models themselves keep getting smaller, cheaper, and more capable. A 0.2B inpainting model running in a browser. A trillion tokens a day through an open-weight system. The infrastructure war is raging at the top while the capability floor quietly rises everywhere else.
Meanwhile, China's semiconductor export surge — up 111% in a year — quietly undermines the clean story of Western AI dominance. AI demand doesn't respect export control narratives when the supply chain has already adapted. And with Trump now signing quantum computing executive orders, the technology policy agenda is expanding faster than anyone can track. The race isn't just about who has the best model anymore. It's about who controls the search layer for agents, who owns the power infrastructure, and who gets to define what "access" even means. Those questions are being answered right now, mostly without public input.