OCI AI Foundations Study Strategy¶
Calibrate to the exam¶
40 questions, 60 minutes, 65% to pass. That is 26 correct answers and 90 seconds per question. It is a recognition exam, not a reasoning exam.
The implication: breadth beats depth. Knowing what every OCI AI service does is worth more than understanding any one of them deeply.
Two categories of question¶
1. Concept definitions. Textbook definitions of standard AI and ML terms. The exam is vendor-flavoured but the concepts are universal, so any solid ML introduction prepares you.
The pairs that recur: - Supervised versus unsupervised versus reinforcement learning - Classification versus regression - Overfitting versus underfitting - Precision versus recall - Generative versus discriminative - Pre-training versus fine-tuning - Parameters versus hyperparameters
2. Service selection. "A company wants to do X. Which OCI service?" These are pure recall, and they are the fastest marks on the paper. Memorize the service selection table.
The distinctions worth getting exactly right¶
Precision versus recall. Precision asks: of the things I flagged, how many were right? Recall asks: of the things I should have found, how many did I find? Fraud detection usually favors recall (missing fraud is worse than a false alarm); spam filtering usually favors precision (blocking real mail is worse than letting spam through). Accuracy misleads on imbalanced data: a model that always predicts "not fraud" on a dataset that is 99.9% legitimate scores 99.9% accuracy and is useless.
Overfitting versus underfitting. Overfitting means the model memorized the training data and performs badly on new data: high training accuracy, low test accuracy. Underfitting means it never learned the pattern: poor on both. Overfitting is addressed with more data, regularization, simpler models, or early stopping.
Supervised versus unsupervised. Supervised learning uses labeled data. Unsupervised finds structure in unlabeled data. If the scenario says "we have historical outcomes", it is supervised.
Fine-tuning versus RAG. Fine-tuning changes the model's behavior by further training. RAG changes what the model knows for this request by retrieving relevant text into the prompt. For "the model needs access to our internal documents", RAG is the answer.
The OCI portfolio, organized¶
Group the services by layer, which makes recall easier than a flat list:
| Layer | Services | You provide |
|---|---|---|
| Ready-made AI services | Language, Speech, Vision, Document Understanding | Data; the model is pre-trained |
| Generative AI | OCI Generative AI, Generative AI Agents | Prompts, or documents for grounding |
| ML platform | OCI Data Science, Data Labeling | Your own model and code |
| Infrastructure | GPU shapes, bare metal, RDMA cluster networking | Everything; you run the stack |
The question "which layer" usually resolves the answer before you get to the specific service: if the scenario says "without machine learning expertise", it is a ready-made service; if it says "train a custom model", it is Data Science.
Common traps¶
| Trap | Reality |
|---|---|
| Choosing Data Science for a pre-built capability | If a ready-made service does it, that is the answer |
| Assuming accuracy is always the right metric | Imbalanced data needs precision, recall, or F1 |
| Confusing parameters with hyperparameters | Parameters are learned; hyperparameters are set before training |
| Answering fine-tuning for a knowledge problem | Knowledge problems are RAG problems |
| Overthinking | This is a foundations exam; the textbook answer is correct |
Exam day¶
- 60 minutes for 40 questions.
- 65% to pass, which is 26 correct. You can miss 14.
- Multiple choice, no penalty for guessing, leave nothing blank.
- Oracle re-versions the exam annually; make sure you booked the current code.
After this¶
OCI Generative AI Professional is the direct next step and goes considerably deeper on LLM applications, RAG, and the OCI Generative AI service.