Not every AI problem needs a foundation model. A foundation model (FM) is a very large model pre-trained on broad data that can handle many tasks, especially generating text, images or code. A traditional ML model is smaller and trained for one narrow job, such as predicting churn or scoring credit risk. The exam asks you to choose between them based on three things: regulatory and explainability requirements, operational constraints, and what the task actually needs.
What Each Is Good At
Traditional ML models shine on structured, tabular data with a clear target: predict this number, assign this label. They're small, fast and cheap to run, and many of them, such as decision trees and linear models, are relatively easy to explain.
Foundation models shine on open-ended tasks involving language, images or code: drafting text, summarizing documents, answering questions, translating, writing code. They come ready to use through services like Amazon Bedrock and can be adapted with prompts, retrieval or fine-tuning. They're also large, more expensive per request, and harder to explain.
Regulation and Explainability
In regulated decisions, such as lending, insurance pricing or hiring, a business may have to explain why the system reached each outcome, and show that the outcome treats groups fairly. A traditional model trained on structured data is much easier to explain. Feature-attribution tools can show how much each input feature contributed to a prediction. On AWS that tool has been Amazon SageMaker Clarify, which is now closed to new customers, though existing customers can still use it.
An FM's reasoning is spread across billions of parameters, and its output can change from one run to the next. That makes it a poor fit as the decision-maker when regulators require a clear, repeatable explanation for every individual decision.
Operational Constraints
Practical limits often decide the question:
- Latency and hardware: a model that must respond in milliseconds on a small device, or without an internet connection, needs to be small. That favors traditional ML.
- Cost at volume: scoring millions of transactions a day with an FM can be expensive, whereas a compact model costs very little per prediction.
- Data: traditional ML needs labeled examples for its task. An FM can often start from a well-written prompt with little or no task-specific data.
The two can also work together. An FM might summarize a customer's complaint, while a traditional model scores the account's churn risk from its billing data.
Choosing Between Them
| Situation | Better fit |
|---|---|
| Tabular data, clear numeric or label target | Traditional ML |
| Each decision must be explained to a regulator | Traditional ML |
| Millisecond responses on low-power or offline hardware | Traditional ML |
| Generating or summarizing text, images or code | Foundation model |
| Many varied language tasks, little labeled data | Foundation model |
| Fast prototype of a conversational assistant | Foundation model |
On the Exam
Don't assume the newest technology is the answer. If a scenario mentions regulators, explaining individual decisions, strict latency, edge devices or structured data with a known target, lean toward traditional ML. If it mentions generating content, summarizing, conversation or handling many different language tasks, lean toward a foundation model.