AI Cracks Open Mathematics. Policy Catches Up Slowly.

This week, the mathematics community discovered something unsettling: GPT-5.5Pro solved Paul Erdös's 1946 Unit Distance Problem - a central discrete geometry conjecture that humans had wrestled with for nearly eighty years. The model constructed a set with n^1.014 unit-distance pairs, refuting Erdös's original bound. Within days, a human mathematician improved it further. The entire process happened one-shot: problem in, several-page proof out, checked by experts.

Meanwhile, a team at DeepMind using AlphaProof Nexus settled nine more Erdös problems and formalized proofs in Lean. And a third AI system solved a longstanding electrical flow problem. This isn't incremental improvement. This is frontier AI systems arriving at novel mathematics faster than human researchers can verify them. The bottleneck has shifted from finding answers to understanding why they work.

What Policymakers Are Actually Doing

Illinois just passed what may be America's strongest AI safety bill. It requires companies like OpenAI, Anthropic, and Google to have third parties verify compliance with safety standards before deployment. Governor JB Pritzker has said he'll sign it. This is real regulation, not guidance-audits, not recommendations. It arrives at exactly the moment when proving mathematical theorems and breaking down safety systems appear to be within reach of the same models.

The contrast is stark. AI researchers are solving eighty-year-old problems. Policy is still struggling with how to audit systems that already exist. And builders? They're already shipping. A developer showed this week how to deploy a production AI application-FastAPI backend with offline TTS, streaming responses, custom domain-for $0.83 per month using HuggingFace Spaces, Cloudflare Workers, and UptimeRobot. No vendor lock-in, no sleepy deployments, no managed service premiums. The infrastructure gap has collapsed.

The Shape of Things Coming

Snowflake committed $6 billion to AWS over five years for AI compute. Amazon Leo is preparing to launch its first 48-satellite broadband constellation on Blue Origin's New Glenn. Robinhood will let AI agents trade stocks. These aren't experiments anymore-they're operational commitments from companies placing bets on agentic AI becoming the default interface for complex systems.

At the same time, Google's AI can't spell "Google." A company engineer just got charged with insider trading using internal data to make $1.2 million on Polymarket bets about search trends. The safest AI systems are solving century-old theorems. The deployed ones are hallucinating, leaking secrets, and requiring human auditors to catch what goes wrong. We are living inside a contradiction: capability accelerating vertically, governance barely moving horizontally, and deployment happening regardless.