On AI and Knowledge — Pablo Castro, Distinguished Engineer & CVP for AI Knowledge, Microsoft
Microsoft CVP frames AI knowledge as intrinsic, extrinsic, and learned, crediting parametric memory for triggering the current AI exponential
“I would argue that GitHub Copilot, ChatGPT, that sort of experiences were heavily grounded on this intrinsic memory, what the models already knew.”
Pablo Castro, Microsoft's CVP for AI Knowledge, proposes a three-part knowledge taxonomy—intrinsic (parametric/training data), extrinsic, and learned—as a framework for thinking about agents and knowledge systems. He traces the coding-assistance arc from IntelliSense (1996) through ML-ranked completions to GitHub Copilot, arguing intrinsic model memory was the original inflection point. The talk is conceptual framing rather than a product announcement, offering useful vocabulary for practitioners building knowledge-augmented agents.