AI Already Outperforms Doctors and Mathematicians — So Legislators Want to Make It Illegal to Ask

GPT-5.4 solved a Tier 4 FrontierMath problem that took its creator, Polish mathematician Bartosz Naskręcki, 20 years to develop. Naskręcki had pre-tested against o4-mini-high and added deliberate stumbling blocks to make the problem AI-proof. Only 1 of 11 attempts succeeded — but it worked “smoothly, methodically.” Eight months earlier, he’d publicly called AI “a very advanced calculator.” Now he calls it his “Move 37.” The capability frontier isn’t approaching — it arrived.

Anthropic’s data shows software engineering captures 49.7% of all AI agent activity — while healthcare sits at 1%, legal at 0.9%, and education at 1.8%. The models can already solve tasks that take humans five hours, but users only trust them for 42-minute sessions. That trust gap is closing fast: Claude Code’s hardest-task success rate doubled from August to December 2025 while human interventions per session dropped from 5.4 to 3.3. Sixteen verticals are wide open — if regulators don’t close them first.