Eleven governments agreed that their new agencies will share information and research on AI systems. Sixteen companies committed to assess so-called frontier models for potential harms; set thresholds at which they become “intolerable”; and not develop or deploy the systems if they cannot mitigate those risks. (The Logic)

Talking point: Countries in the governmental pact include the U.S., U.K. and Canada; Ottawa also signed a deal for its forthcoming safety institute to work with London’s already-established one, including exchanging staff and accessing U.K. compute resources. Companies making the commitments on models include Canada’s Cohere, Amazon, Anthropic, Google, Meta, Microsoft, OpenAI and Beijing-based Zhipu AI. Participants announced both deals at the Seoul AI Summit, a follow-up to last November’s AI Safety Summit in the U.K., where dozens of countries pledged closer cooperation on the technology and a smaller set of firms agreed to let authorities test their products.