AI #187: Coming Into Play

Anthropic released Opus 5.5 on Tuesday. Zvi Mowshowitz wrote that he covered the system card and will cover its capabilities soon, and that by all reports it is an excellent model. OpenAI released a new, cheaper and improved Sol and Luna, which Mowshowitz described as likely an important upgrade under the hood, though he said no one is talking about them because of Opus 5.5.

Senator Bernie Sanders and Representative Greg Casar have formally introduced the Ban Artificial Superintelligence Act. Mowshowitz said he reserves judgment on particular bills until he can read them in detail and hopes to publish a close reading soon. MIRI read the bill and endorsed it as directly confronting the extinction threat.

CNN reported that AI incorrectly told the US military that a Chinese ship was transporting nuclear weapon components, and that the operation was checked before it happened and the material was found to be misidentified. CNN called it a hallucination and said it almost started a war; Mowshowitz disputed that characterization. A former senior US official familiar with the AI systems used by military and intelligence analysts told CNN that the internal tools are mostly copies of the commercial products.

Bloomberg published an analysis of a US targeting error that destroyed an Iranian school and killed at least 123 children, tracing it to outdated intelligence. The site had been a military compound years earlier but began openly operating as a school by 2018. Officials said some Centcom personnel relied too much on artificial intelligence embedded in Maven Smart System. A Palantir spokesperson said the company is not responsible for the underlying data or for identifying intelligence deficiencies and that there is no evidence its software was at fault in the Minab strike. After the strike, Palantir built new capabilities into Maven to re-review underlying intelligence and flag inconsistencies and inaccuracies, according to a person familiar with the matter.

Officials involved in the internal investigation said the Pentagon's civilian harm mitigation units were cut by roughly 90 percent to fewer than 20 staff members, and Centcom's team went from 10 to 1. No civilian harm mitigation team member reviewed the Minab site before the strike.