The official AIF-C01 exam guide lists five domains and their weightings, but it doesn't explain why they're weighted the way they are, or what each one is actually testing versus what it sounds like it's testing. Here's the plain-English version, roughly in the order your study time should follow.
Domain 1: Fundamentals of AI & ML — 26%
This domain is really just one idea tested five different ways: AI contains ML, ML contains Deep Learning, Deep Learning contains Generative AI — each one a more specialized subset of the one around it. Once that nesting is locked in, the rest follows: supervised learning needs labeled data and predicts things, unsupervised learning finds hidden structure in unlabeled data, and reinforcement learning improves through trial and error and reward.
The part people underestimate is model quality. The exam loves the underfitting/overfitting distinction — think of it like studying for a test: underfit is barely studying (fails everywhere, even easy questions), overfit is memorizing the practice exam word-for-word (aces practice, fails anything unfamiliar), and a good fit is actually learning the material. Bias is the third failure mode, and it almost always traces back to skewed training data.
Domain 2: Fundamentals of Generative AI — 24%
This is where the exam gets into how GenAI actually works, not just what it can do. Foundation models are enormous pre-trained networks; the transformer architecture underneath them breaks text into tokens, turns those into vectors, and uses self-attention to figure out which words matter to each other — that's what lets "it" correctly resolve to "the animal" instead of "the street" in a sentence like "the animal didn't cross the street because it was tired."
Know the difference between unimodal (one data type in, one data type out) and multimodal (text, image, audio combined), and don't confuse LLMs (predict the next token) with diffusion models (add noise to an image, then learn to reverse it — that's how Stable Diffusion and similar tools generate images).
Domain 3: Applications of Foundation Models — 28% (the biggest domain — study here first)
If you only have time to deeply master one domain, make it this one. It's the largest chunk of the exam and the most practically useful for actually working with these tools.
RAG (Retrieval-Augmented Generation) is the concept to nail cold: think of it as an open-book exam for the model. Instead of retraining anything, you retrieve relevant facts from a vector database and stitch them into the prompt before the model answers. It's usually cheaper than fine-tuning and it's the #1 defense against hallucinations.
Prompt engineering has a clear difficulty ladder worth memorizing in order: zero-shot (no examples) → few-shot (give examples) → chain-of-thought ("show your work," useful for reasoning/math) → prompt tuning (optimizing the input vector itself, more advanced). Separately, know when to reach for fine-tuning instead — specifically LoRA/PEFT, which trains small adapter layers instead of touching the full billion-parameter model, avoiding the expense of updating every weight.
Domain 4: Guidelines for Responsible AI — 14%
Smaller domain, but conceptually dense. Responsible AI breaks into pillars — fairness, explainability, robustness, privacy & security, governance, and transparency — each one mapping to a plain question a regulator or user might ask ("was this biased against a group?", "why was I rejected?", "is my data safe?").
The one distinction the exam tests hardest: interpretability vs. explainability. Interpretability is a glass clock — simple models like linear regression where you can trace the exact math. Explainability is a car engine — you can't see inside a foundation model, so you infer its behavior by watching how outputs change when inputs change. Higher performance usually means lower interpretability, which is exactly why high-stakes legal or medical use cases often need a simpler, more transparent model instead of the most powerful one available.
Domain 5: Security, Compliance & Governance — 14%
This domain is AWS's shared responsibility model applied specifically to AI workloads. AWS secures the cloud itself (physical data centers, hardware, virtualization); you secure what you put in it (your data, your IAM policies, how you configure the workload) — and your share of that responsibility shrinks the more managed the service is.
Worth memorizing the AI-specific attack types, since they show up as direct exam questions: data poisoning (corrupting the training set), adversarial inputs (tiny tweaks that fool a model at inference time), model inversion (reconstructing training data through repeated queries), model extraction (cloning a model's behavior by studying its outputs), and prompt injection (malicious instructions smuggled into an LLM — currently considered the #1 LLM-specific threat).
Quick reference
| Domain | Weight | What it's really testing |
| :--- | :--- | :--- |
| 1. Fundamentals of AI & ML | 26% | Can you tell ML paradigms apart and reason about model quality? |
| 2. Fundamentals of Generative AI | 24% | Do you understand what a transformer/foundation model is actually doing? |
| 3. Applications of Foundation Models | 28% | Can you apply RAG, prompt engineering, and fine-tuning correctly? |
| 4. Guidelines for Responsible AI | 14% | Do you know where bias comes from and how it's caught? |
| 5. Security, Compliance & Governance | 14% | Do you know AWS's shared responsibility model and AI-specific attack types? |
If this breakdown was useful, I put together a full illustrated study guide that covers all five domains this way — a diagram for every concept plus memory hooks for the stuff that's easy to blank on mid-exam. Link if you want the full version
Top comments (0)