Nadella says AI alignment won't come automatically with scale

Microsoft CEO Satya Nadella said on the 20VC podcast that he no longer believes AI alignment will emerge simply from making models more capable. He argued alignment must be built into training, and companies need zero-trust systems with tamper-proof logs for AI agents.

• Nadella dropped the idea that alignment would 'fall out' of scaling models
• He said alignment likely has to start in pretraining
• Microsoft gives every agent an identity and tamper-proof logs via Agent 365
• Nadella warned chain of thought can be unfaithful and urged more transparency