Chapter C90 · Interactive Course
Use it: choose a model for the job
Turn three chapters of theory into one defensible decision.
Core Learning Concept
Choose a model for the donation task using the required output and a fair local comparison. Eliminate incompatible capabilities, compare suitable candidates on the same examples, and consider cost after adequacy. Record the choice and the observation that would make you revisit it.
What You’ll Learn in This Chapter
-
Name the job before the model
Turn a stated need into a task type: category, number, order, group or new content.
Name the output before selecting a model. Sorting messages into categories needs classification; ordering offers by usefulness needs ranking; writing replies needs generation. The required output helps define errors and checks. A product may support several tasks, so inspect its actual capabilities rather than only its name.
-
Two candidates, one task
Choose between two described models for a single stated job.
A candidate must produce a useful result for the task before its headline score matters. A model that labels offers is not automatically a suitable ranking system. Compare compatible candidates using the same inputs and success criteria, and include the surrounding processing needed to obtain the final result.
-
The cheapest adequate choice
Prefer the simpler model when the task does not need more.
Adequate means satisfying required quality standards for a specific task. When multiple candidate models meet this threshold on representative evaluations, simpler operation and lower cost become decisive factors for choosing between them.
-
What would change your mind
State the evidence that would overturn a model choice.
A revisit condition is an observable change that would challenge the choice. “Misses more urgent messages than our agreed tolerance” can trigger a review; “gets worse” is too vague to monitor. Decide how the condition will be measured and what action follows when it occurs.
-
Write the choice down
Record a model decision with its task, its reason and its check.
A short decision record connects the task, chosen option, evidence, and revisit condition. Include the test’s limits so future readers know what was and was not established. This turns a model choice into a reviewable decision rather than a preference that persists without fresh evidence.
-
Apply it: choose a model for the job
Carry out the whole decision once: pick a model type for a stated task and name the evidence that would confirm or overturn the choice.
-
Describe an AI system accurately
Inspect a parcel service: name its task, identify the learning signal, and find a data gap.