aiengineering.guideaiengineering.guide

DS/MLE → LLM engineer · Stop 3 of 11 lessons

Reading a model card and a pricing sheet

What this lesson covers

New models ship every week. The skill is not memorising them; it’s reading a model card in ninety seconds and knowing whether to swap.

The outline

  1. The four numbers that matter. Context window, input price, output price, and the eval closest to your task.
  2. The four that usually don’t. Parameter count, benchmark averages, release date, model-family branding.
  3. The evals to trust. Which benchmarks correlate with your workload — and which are Goodhart-poisoned.
  4. The pricing arithmetic. How to estimate your monthly bill from requests × tokens × price, why output tokens dominate, and how caching changes the answer.
  5. The missing eval. Reading between the lines — what a model card doesn’t report is usually what it’s worst at.
  6. When to pin, when to auto-upgrade. Version strategy for a system you have to maintain.

Coming soon

In outline.

Outline

Enter to go · Esc to close