AIF-C01 ยท AI use cases and AWS AI services
Traditional ML versus foundation models
Choosing between a task-specific ML model and a foundation model based on regulation, explainability, cost and operational constraints.
Key points
- Task-specific ML model: trained on labeled data for one job.
- Strengths: usually cheaper and faster per prediction, easier to explain and to run on small or offline hardware.
- Foundation model: pretrained on broad data, adapted to many tasks through prompts or fine-tuning; suits open-ended language work.
- Downsides: costs more per call, harder to explain.
- Regulated, high-volume, structured-data tasks often favor the traditional model.
Practice questions
What is a main strength of a foundation model compared with a task-specific model?
- One pretrained model can be adapted to many different tasks
- Every prediction is easy to trace to specific input features
- It always has the lowest cost per prediction
- It needs no compute at inference time
Show answer
Answer: A. One pretrained model can be adapted to many different tasks
A foundation model is pretrained on broad data, so one model can be steered to many tasks with prompts. Predictions are typically harder, not easier, to explain, and per-call cost is usually higher.
Which reasons favor a traditional task-specific model over a foundation model?
- Each prediction must be explained to an auditor
- Millions of structured rows must be scored cheaply and quickly
- The system must draft varied free-text emails
- One assistant must cover many unrelated language tasks
- No labeled data exists and tasks change every week
Show answer
Answer: A. Each prediction must be explained to an auditor, B. Millions of structured rows must be scored cheaply and quickly
Per-prediction explainability and cheap, fast bulk scoring on tabular data are classic wins for task-specific models. Drafting varied free text and covering many unrelated tasks are where foundation models shine.
True or false: a foundation model is always cheaper to run than a small task-specific model.
Show answer
Answer: False
False. Foundation models are large, so each call generally costs more and takes longer than scoring with a compact purpose-built model.
Practice 8 questions on this topic in Bitzu
Short lessons, instant explanations and spaced review of what you get wrong. Free on iOS and Android.
