Unbundling AI Openness: Beyond the Binary
The episode challenges the familiar “open versus closed” framing of AI systems. Sharma argues that openness is not inherently good or bad—it is an instrumental choice that should align with specific policy goals. She introduces a seven-part taxonomy of AI—compute, data, source code, model weights, system prompts, operational records and controls, and labor—to show how each component interacts differently with innovation, safety, and governance. Her central idea, differential openness, suggests that each component can exist along a spectrum rather than being entirely open or closed. For…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.