Compounding Intelligence

Adyen, a Dutch Payments Platform, holds a slice of live traffic where its models never act, runs every new model in ghost mode before it influences anything, and promotes it only when it beats the incumbent with statistical confidence. Most AI pilots are decided on an agreement rate, which stops describing the model the moment a reviewer can see its answer.

Nicholas Carlini, a researcher at Anthropic, built a 100,000-line C compiler that compiles the Linux kernel by directing 16 AI agents at once, alone, in about two weeks. The key idea is driven by the builder; the harness only gave one person the reach of a team.

Balyasny reports about 95% of its investment teams now use its AI research engine. It got there by redesigning the analyst's job around directing and reviewing the machine's output rather than producing the first draft.

Ramp built finance agents that approve low-risk expenses on their own and escalate the judgement calls to a named human. The design decision that made it work draws the line between what the agent decides and who owns the decision.

OakNorth put AI on the analysis and kept the banker on the decision, and this is how it has lost less than five pence for every £100 it has lent since 2015: only £6.9m written off across £15.1bn of lending over ten years.

AI amplifies engineering output, but distribute AI tooling without redesigning the delivery model and you get more output on the same delivery date.