Reading on classification
A catalogue of useful essays about classification and the job of picking labels. Each row points at the original article. Classify writes an ELI5, not a reprint.
Featured reading
| Reading | Author | Source |
|---|---|---|
 RLHF tells you what you want to hear. RLHF trains a model to match what people like. That is not the same as doing the work. The clip's point is that disappointment is built in: the score is preference, not the job. | Diogo Almeida | AI Engineer |
 Don't compact the history. Label it. Coding agents look like a simple loop with a few tools. They still keep one growing cache, so sending a turn to a cheaper model and back can cost more than staying put. This note says TypeSafe can label each chunk of context and rebuild the window for the job, instead of compacting one shared state. | Diogo Almeida | Google Docs |
 Compute is cheap. Picking the job is not. If you train a huge model to do the wrong job, it gets better at the wrong job. Diogo's point is that someone still has to name the task. Classify starts there too: what should Jev decide? | Diogo Almeida | Complete Skeptic |