AI #169: New Knowledge
Zvi's AI Roundups · Zvi Mowshowitz · 2026-05-21
Zvi Mowshowitz's weekly AI newsletter covers OpenAI's AI solving the unit distance problem, METR's frontier risk report finding current models lack robust rogue-deployment capability, Andrej Karpathy joining Anthropic for recursive self-improvement, and Anthropic reporting its first operating profit.
Extraction
Topics: ai-safetyai-capabilitiesanthropicopenaiai-policyfrontier-risk
Claims
- An OpenAI general-purpose AI solved the unit distance problem, which Zvi calls the first truly impressive mathematical result from an AI.
- METR's risk report found AI agents plausibly had the means, motive, and opportunity for minimal rogue deployments but lacked the capability to make such deployments robust against shutdown attempts.
- Andrej Karpathy joined Anthropic explicitly to work on recursive self-improvement.
- Anthropic expects $10.9 billion in Q2 2026 revenue and its first operating profit, with compute costs falling from 71 to 56 cents per dollar of revenue.
- A ChatGPT-generated story won the Commonwealth Short Story Prize, and the prize organization declined to revoke it, citing inability to check unpublished fiction without violating 'consent and artistic ownership.'
Key quotes
Andrej Karpathy has joined Anthropic, explicitly to do recursive self-improvement. He plans to later return to his education work, but if he succeeds at his new task there might not be anything left to return to.
We think that AI agents plausibly had the means, motive, and opportunity to launch a minimal 'rogue deployment,' but lacked the means to make rogue deployments robust to serious efforts to shut them down.
In the first quarter, Anthropic spent 71 cents on computing power for every dollar it made. In the current quarter, it expects to spend 56 cents per dollar, a sign that the business is becoming more efficient as it grows.