In Microsoft’s Stanford CS153 session with Michael Abbott, Satya Nadella frames the firm’s problem in one line: if models learn from data, what remains of a company whose value has lived in tacit knowledge from operations and people. His answer is not a bigger rented model. It is a company-owned learning loop—what he calls a “hill-climbing machine”—built from private evaluations, workflow traces, rewards, and outcomes that stay inside the tenant.
Today we cover:
**The firm question Nadella asks:**what survives when general intelligence is available to every competitor.**What the “hill-climbing machine” is, in operational terms:**private evals, RL environments, traces, and retained IP.**Why consume-only AI adoption can buy productivity and still lose retention:**traces and rewards leave the firm.**How to measure a loop-owning company:**eval coverage, trace capture, model-swap survival, cost per successful private outcome.**Harness mapping:**eval, memory, and loop engineering as the control layer around rented models.**Limits and Microsoft’s commercial stake:**what the talk asserts, what remains unproven, and how to read Build product scaffolding.