Study MLA-C01 Model Selection, AI Services, and Foundation Models: key concepts, common traps, and exam decision cues.
This lesson is about choosing the right modeling lane before you optimize anything. MLA-C01 often asks whether the stronger answer is a SageMaker built-in algorithm, a managed AI service, a foundation-model workflow, or a custom training path.
Feasibility: Whether the data, complexity, cost, and business need justify an ML solution at all.
Foundation model: Large pre-trained model that can be adapted or prompted for downstream tasks.
Built-in algorithm: AWS-managed training algorithm that reduces custom implementation work for common ML problem types.
AWS wants you to distinguish:
flowchart TD
A["Business problem"] --> B{"Already solved by a managed AI service?"}
B -->|Yes| C["Use managed AI service first"]
B -->|No| D{"Traditional predictive or classification problem?"}
D -->|Yes| E["Compare built-in algorithm vs custom training"]
D -->|No| F{"Generative or language-heavy problem?"}
F -->|Yes| G["Compare FM access, RAG, or customization path"]
F -->|No| H["Recheck feasibility and requirements"]
The exam usually punishes candidates who jump straight to the most sophisticated-sounding model without classifying the problem first.
| Situation | Strongest first lane |
|---|---|
| common business capability like OCR, translation, transcription, or sentiment | managed AWS AI service |
| standard tabular classification or regression with known ML workflow | SageMaker built-in algorithm or common ML model family |
| highly custom problem with no close managed fit | custom training path |
| generative, summarization, chat, or retrieval-heavy task | foundation-model workflow |
| If the question emphasizes… | Stronger first answer |
|---|---|
| business capability with minimal custom model ownership | managed AI service |
| explicit control over training data and model behavior | built-in algorithm or custom model |
| rapid delivery of common ML capability | managed AI service or built-in algorithm |
| generative text, multimodal prompts, or FM adaptation | foundation-model path |
| highly specific objective with unique labeled data | custom training or targeted fine-tuning path |
| Symptom | What is usually going wrong | Fix first |
|---|---|---|
| every answer sounds like “some kind of ML” | you are not separating service-fit from model-fit | ask first whether AWS already has a managed capability for the job |
| you keep choosing the most advanced option | you are rewarding sophistication over fit and operating cost | prefer the simplest path that meets the stated requirement |
| foundation-model answers keep trapping you | you are treating GenAI as a default answer instead of a specific solution shape | ask whether the problem is actually generative |
| custom training feels safer by default | you are undervaluing built-in algorithms and managed AI services | compare ownership burden, speed, and data requirements |
| Trap | Better reading |
|---|---|
| “If ML is possible, custom training is best.” | AWS often rewards managed services or built-ins when they fit the use case cleanly. |
| “Foundation models are the modern answer, so they are probably right.” | FM paths are only stronger when the problem is actually generative or language-heavy. |
| “Built-in algorithm and AI service mean the same thing.” | Built-in algorithms still require model workflow ownership; managed AI services often abstract much more. |
| “Model choice can be postponed until after deployment planning.” | MLA-C01 expects the model lane to be justified before tuning or rollout choices. |
A team wants to process invoices, extract key fields, and route the output into a downstream approval system. One engineer proposes training a custom document-understanding model from scratch because “that gives us more control.”
The strongest first answer is usually the managed AI service lane. The problem is a common document-extraction capability, so the exam rewards using the existing AWS service fit before inventing a heavier custom model program.