When Claude Met Claude, a Turf War Broke Out
Anthropic gave three copies of one model the same project and clashing orders, without telling them the others existed. They sabotaged each other with self-replicating malware, then sometimes negotiated a truce or staged a tournament. As Congress drafts a kill-switch bill, the study asks why safety testing still examines one agent at a time.
Read full story →The AI Boom’s Awkward Question: Does Any of It Pay?
Usage is soaring and prices are crashing, but OpenAI’s own report shows no link between how much AI a company uses and its revenue per employee. Steve Eisman calls the two big labs the AI trade’s Achilles heel; Bruce Schneier says nationalise them if the market walks away.
Read full story →Two-Thirds of AI’s Power Hunger May Be a Mirage
Analysts say most of the electricity data centers are demanding will never be built, inflated by developers pitching the same “phantom” project to several utilities at once. Yet the real climate damage may already run through the oil and gas fields AI helps drill faster.
Read full story →Zuckerberg Reopens the War Over Who Should Hold AI
In a 6,500-word essay, Meta’s CEO calls the safety case for closed models a story companies tell to keep the best AI for themselves, and declares aligning one system with all of humanity impossible. Critics note Meta held back its own strongest weights and trails the labs it attacks.
Read full story →