Small, specialized models and context discipline beat brute force

A 10-year-old CNN outperformed frontier LLMs for skin cancer diagnosis. The thread is thin, but the headline fits the day: a narrow tool trained for one job can beat the biggest general model. 'The State of Context Engineering' makes the same point from the agent side. The advice that keeps getting repeated is that when someone says their coding agent is acting weird and 'Claude's been nerfed', the first question is how big their context is. Commenters argue about the vocabulary: whether the FireRed tea task poisoned context or the weights, and whether a fast model should prune context before compaction.

Supporting that, 'ADRs for AI' lets people version decision records in Git so humans and AI keep a shared memory of why choices were made. 'Tasks as Code' pushes tickets into the repo for agents. That drew sharp pushback: one commenter says a manager pushes it and it sucks for everyone but him, another says it breaks Git history.

The through-line is that results come from curation (what goes in the model's view and how it's recorded), not from more compute.