1/ The most powerful AIs aren't public. For months, labs run highly capable internal models before release. @ashwinkacharya and I wrote a new @IAPS report, “Managing Risks from Internal AI Systems,” exploring the hidden dangers this creates. 🧵
2,58K