KM4BI KM4BI

Five Rules for Picking an AI Model That Actually Works

Nate B. Jones lays out a practical framework for choosing AI models based on the work, its complexity, cost, and the quality of the surrounding harness.

Nate B. Jones argues that model selection should start with the work, not with the logo on the model card. The video uses Fable 5’s return after an 18-day disruption as a reminder that teams should not be structurally dependent on one model. The companies that kept moving were the ones with a harness and routing options already in place.

The first filter is the shape of the task. For familiar, repeatable work that is easy to review — client notes, summaries, tables, web pages, decks, or conventional coding tasks — a strong lower-cost model such as GLM 5.2 can be the right choice. For ambiguous, novel, or strategic work, Jones still favors frontier models such as Claude or ChatGPT because the goal is not just cheap output; it is getting the judgment right.

The second filter is the harness. Intelligence is only useful if users can get work into the model, preserve context, and extract usable results without friction. That is why Jones highlights environments such as Codex, Claude Code, Z.AI, and the emerging harness layer around open-source models.

For companies and small teams, the point is not to build a twenty-model routing maze. The practical move is to identify the recurring artifacts that matter most to customers, then create the simplest reliable path from input to value. Specialist image, video, and live-information models belong in the stack only when the job clearly needs them.

The takeaway: test models on your real tasks, know how to judge output quality, reserve expensive frontier models for difficult work, and keep the stack simple enough that choosing the model does not become the work.

Source