UC Berkeley

Stuart Russell from UC Berkeley discusses the fundamental goal of AI regulation, which is to manage and reduce risks to an acceptable level. He emphasizes that regulation should focus on the practicalities of risk management rather than just theoretical fears.

IBM Research and UC Berkeley researchers analyzed why AI agents fail during complex IT automation tasks using the IT-Bench and MAST frameworks. The study reveals that different model classes experience distinct failure modes, with larger open models often suffering from cascading errors.