a builder's codex
codex · operators · Eugene Yan

Eugene Yan

Bio

Eugene Yan writes about applied AI/ML systems with the practitioner's bias of someone who actually has to keep them running. His through-line is that the cheap step in modern LLM systems is producing output, and the slow step is figuring out which outputs were good and getting the system to do more of them. He spends most of his writing on the operational mechanics that close that loop — transcript capture, evaluator design, rubric calibration, the move from "we have logs" to "the system is improving on its own cadence."

Operating themes

Cards

Sources captured

Insights · 2

Tier A · ai-native · engineering
The middle is hollowing out — execution gets automated, leaving spec-writing and verification as the high-value human tasks
Tier A · ai-native · engineering
Close the feedback loop by mining session transcripts for patterns to promote into config
Open the interactive profile → blog