Choose capacity for the work that remains
A clear implementation can still contain a difficult decision.
Getting startedAbout 2 min read
A model choice is a working choice
Using the same model and reasoning setting for every task is simple. It can also hide how different the tasks are. Correcting a known label is different from deciding who should be allowed to see a personal story.
I find it more useful to ask what reasoning remains than to divide the work into “planning uses a strong model, implementation uses a small one”. A repair can require difficult judgement; a planning task can be routine.
Look at uncertainty and consequence
A well-bounded change with an agreed solution may suit a faster option, provided it has demonstrated the capability for that work. Conflicting requirements, data ownership or an unclear failure need more reasoning capacity and closer examination.
Speed, price and model names change. None is a guarantee of quality. Compare the actual result on the task, and keep the same requirements for evidence, review and human decisions regardless of the model chosen.
Make the choice revisable
Try a bounded task, inspect the result and notice whether another correction is repeatedly needed. If the work turns out to contain a larger unresolved question, reconsider the configuration before continuing.
A lighter option is useful when it can do the work. A deeper option is useful when its extra capacity changes the outcome. I want to notice the difference rather than make either into a habit.