← All Issues
Monday, May 25, 2026

AI Daily

Your Automated Intelligence Briefing

Research

An Eighty-Year-Old Belief About Dots on Paper, Quietly Toppled

An unnamed OpenAI reasoning model, working from a single prompt, has disproved the Erdős unit distance conjecture in a 125-page chain of thought. Nine mathematicians including Fields Medalist Tim Gowers verified it; a Princeton mathematician sharpened the bound within hours. The story sits exactly on the line between milestone and overclaim.

Read full story →
AI Hardware

When Older Chips Get More Expensive: The AI Shortage Migrates Down the Stack

Nvidia's CFO let slip that H100 rental prices are up 20 percent in 2026 and even the five-year-old A100 is up 15 percent. AMD has asked partners to accelerate CPU production. The bottleneck stopped being about GPUs months ago, and Nvidia's new Vera CPU is the most explicit admission yet.

Read full story →
Policy

The Seven-Page AI Order That Nearly Got Signed

The draft of the cybersecurity executive order Trump pulled on May 22 is now public. It would have given the NSA a 90-day pre-release look inside frontier models, while explicitly insisting it was not a license. That contradiction appears to be exactly what killed it.

Read full story →
AI Safety

A Job Listing and a Source-Code Leak: Two Labs Edge Toward Things They Said They Were Not Ready For

OpenAI will pay up to $445,000 for a researcher to worry about recursive self-improvement. The same week, "claude-mythos-1-preview" turned up inside Anthropic's source code. Both labs have spent the year explaining why these capabilities should not yet ship. Both are also quietly preparing to ship them.

Read full story →