Descriptions:
Recorded live at Microsoft Build 2026, this crossover episode of No Priors and Latent Space features Microsoft chairman and CEO Satya Nadella in an extended conversation with hosts Swix, Sarah Guo, and Alessio Fanelli. The interview covers Microsoft’s full AI stack — from model training philosophy to enterprise platform strategy — and contains some of the most candid first-party detail Nadella has shared publicly about where the company is placing its bets.
On the model side, Nadella explains the reasoning behind the MAI model family: prioritizing clean pre-training lineage, rigorous data quality ablations, and a hill-climbing scaffold that lets enterprises build specialist models on top of a strong generalist base using their own private evals and RLHF traces. He describes a specific result where GPT-5 was used to collect traces, which were then distilled into a 5 billion parameter reasoning model that outperformed the larger model on targeted tasks — a practical illustration of inference-time compute and distillation working together.
The conversation also tackles the existential question facing SaaS companies as AI agents commoditize workflow generation. Nadella’s thesis: the underlying data models and semantic layers built inside legacy SaaS applications still hold genuine value, and the challenge is unbundling and rebundling them rather than replacing them wholesale. He introduces Work IQ as Microsoft’s bid to expose the organizational graph inside Microsoft 365 as an accessible database for the first time. Additional topics include the Microsoft Discovery platform used in Majorana 2 quantum chip development, ecosystem strategy, and Nadella’s reflections on what he wishes he had known at the start of Microsoft’s AI transformation.
📺 Source: Latent Space · Published June 03, 2026
🏷️ Format: Interview
