As a New Yorker, I always thought 'Where the Sidewalk Ends' was just a Shel Silverstein book. Then I visited Charlotte, North Carolina.

Becoming Clauditouille

made a macaronic cento on the plane using Claude yesterday

Generally China is very academia-brained, so "RSI" in China is just another mundane ML phenomenon like AutoML was in the 2010s "Unlike in Silicon Valley, where the two concepts are closely linked, the adoption of “RSI” in China has not brought about much talk of an intelligence explosion. Chinese AI leaders occasionally mention a singularity, as Wu did in 2025’s speech and Tang in his Qiushi interview. However, when they describe where self-improving AI leads, they mostly point to real-world applications."

I've been doing a quick org-wide AI news brief multiple times a week. Here's a recap of the major AI security stories from this week: Two critical-severity remote-code execution flaws in PHP, found with AI assistance, were patched Oct 21. PHP runs ~70% of websites whose server-side language is known. Government websites in the Netherlands, Finland, France and Denmark went offline, several explicitly as a precaution. A Chinese-speaking hacker used AI agents to break into the websites of 100+ companies, including Deere and SharkNinja, at an estimated $50 in AI costs per successful break-in. A China state-aligned hacking group used ARTEX, an open-source AI hacking tool, in a campaign that stole ConnectWise Cloud Backup credentials and used them to authenticate into customers' Microsoft 365 accounts, per Microsoft. India’s cybersecurity agency, CERT-In, says AI companies disclosed that their agents, during testing, used exposed passwords and keys to read business records at Indian organizations. None of the organizations noticed. Large US banks, including Morgan Stanley, JPMorgan, BofA and Citi, are blocking third-party AI agents that users let access their accounts. Some agents seemed to go beyond what users asked. OpenAI reported two cases in which its own AI agents weakened monitoring systems on internal coding tasks. One agent disabled automatic alerts to OpenAI's security team. The other removed a monitor, then falsely told the user it hadn't. Anthropic’s leaked draft federal AI bill, the AEGIS Act, would let Commerce seek a court order to pause a model only for a "foreseeable, imminent" risk of >50 deaths or serious injuries, or >$1B in damage. Encode's Nathan Calvin called the threshold "an absurdly high bar." It would also preempt parts of state AI bills. California State Senator Josh Becker announced he will introduce the Frontier AI Safety Act. The bill would require major frontier AI developers to publish the share of their R&D compute devoted to safety, in quarterly, independently verified reports. It would authorize Cal OES to set minimum standards for how companies use their R&D compute. NIST CAISSI, NSA, and CISA published "Key Practices for Incident Readiness and Response in Super Intelligence Training and Evaluation," voluntary guidance for incidents where AI agents undermine human oversight during training and testing. Anthropic reports that the junior share of its technical hires fell from 27% to 9%, while technical hiring grew in absolute terms. By August, interactive agent runtime in employee-directed sessions was 6× estimated human work hours.

AI cyber challenges are more pressing than ever. Consider joining the Cal OES California Special Events Cybersecurity Summit in LA on November 5. Register here. eventbrite.com/e/california-…

Wiktionary is truly a treasure trove

Cal OES is hiring for two AI safety roles! I think there's a very high ceiling for impact in these roles. Apply soon! linktr.ee/caloes.ai

I rewrote my preprint "Frontier AI Safety Laws and Chinese AI Companies" (with @Garrett_Xu_). Had some fun animating it with Claude. papers.ssrn.com/sol3/papers.…

Brussels (the city) needs to cut out its weird reverence of Leopold II. One of the worst humans to have lived on the face of the earth.

America has won the race for Super Intelligence. Full stop.