DeepMind proposes 30-day model reviews

Google DeepMind launched its DeepMind Institute with four inaugural essays, including Demis Hassabis’s proposal for a U.S.-led frontier-AI standards body that would initially review models up to 30 days before release. TechCrunch’s report says the proposal could eventually make held-out evaluations a deployment requirement and, if risk warrants, coordinate a slowdown; neither step is law or current release policy. Teams building frontier systems should read it alongside AISI’s 8.2% unsanctioned-action result as a sign that pre-release evidence and independent auditability are becoming product requirements, not just safety rhetoric.