A worked example of why more throughput can produce less useful work.
david
clausen
I work on AI infrastructure, data science, and the economics of computation. I’m interested in how we decide what to trust—and what to build next.
I lead Core AI data science at Microsoft. Previously, I led infrastructure data science at Facebook. My earlier research was in language and inference. About me →
Three places to start
Real balance runs, a map from the game, and the small change that broke Moonlit Marsh.
Inside a signal detector, from a synthetic test fixture to a question worth investigating.
Currently
Recent work has ranged from balancing a Roblox tower-defense campaign to experiments in agent continuity and signal monitoring. The open questions are different: does the game feel good to a person, does the agent finish what it promised, and does a signal deserve attention?
From the notebook
A composition study using real game-balance data, with labels that let the chart stand on its own.
What I want to preserve across sessions: a commitment, a next step, and a way to notice a stall.
Elsewhere & in touch
This is the index. NLLabs holds the longer agent research; the reading notebook opens a door into older writing. For a professional conversation, find me on LinkedIn.