Anthropic, the AI company behind the Claude chatbot, ran a quiet March experiment to flag users it suspected were in China, then left it running for months. Engineer Thariq Shihipar told Reddit the team "has been meaning to take this down for a while."
I have heard that exact sentence before. It was the operating philosophy of the Hazelwood coal plant in Victoria, Australia, which ran 22 years past its scheduled review because someone kept meaning to file paperwork.
A mid-size inference cluster idling on a flagging routine burns roughly 4kW per rack continuous. Eight months, one rack, grid-average intensity: 1.7 tonnes of CO₂ to sort strangers by passport. I did the arithmetic.
Meanwhile Dario Amodei gives interviews about existential risk from a house whose pool heater runs on natural gas. The experiment is still up.
Related twisted takes: The Studios Want Midjourney to Stop Copying Bart Simpson. Midjourney… · Tech Billionaire Surprised That Firing People Did Not Make the Comput… · AI Learns To Schedule Itself. Humans, Famously, Could Not Schedule A…
Based on the original article "Alibaba reportedly bans employees from using Claude Code".
