AI vs ML vs Deep Learning: A Cloud Engineer's Plain-English Guide

I am studying for the AWS AI Practitioner exam. One thing I noticed early on is that half the confusion in this space comes from three terms getting thrown around as if they mean the same thing: AI, machine learning, and deep learning.

They do not mean the same thing. They are nested inside each other. Once you see that clearly, a lot of the other concepts start clicking into place.

My background is cloud infrastructure. Oracle DBA, AWS, multi-cloud migrations. I am going to explain this the same way I would explain it to someone on my team. No maths, no Python, just the mental model.

The nesting structure

AI (the whole data centre)
  └── Machine Learning (the servers inside)
        └── Deep Learning (the GPUs in those servers)
              └── Generative AI (what those GPUs are running right now)

Artificial Intelligence is the broadest term. Any technique that makes a computer mimic human intelligence counts as AI, including old rule-based systems from the 90s. If you wrote a script that flagged transactions over $10,000, that is technically AI.

Machine Learning is a subset of AI. The specific difference is that ML systems learn from data instead of following rules you wrote manually. Rather than coding "if amount > 10000 then flag", you feed the system millions of historical transactions, label which ones were fraud, and it works out the patterns itself.

Deep Learning is a subset of ML that uses neural networks, which are layers of connected nodes loosely inspired by how the brain works. Deep learning is what finally made image recognition, speech recognition, and translation work properly. It needs more data and compute than traditional ML, but it handles unstructured data like images, audio, and text much better.

Generative AI is a subset of deep learning focused on generating new content: text, images, audio, code. Claude, GPT-4, Stable Diffusion. That is what all of these are.

The infrastructure analogy

As a cloud engineer, here is how I think about the layers:

AI is the whole data centre concept. Broad. Covers everything.

Machine Learning is the compute layer inside. The actual mechanism doing useful work. A data centre without servers is just a building. AI without ML is just a concept.

Deep Learning is the GPU layer. More powerful than CPU-based ML for complex tasks. More expensive to run. You do not always need it. A simple classification problem does not need a neural network any more than a static website needs a GPU cluster.

Generative AI is the workload running on those GPUs right now. It is real and capable. But it sits on top of the whole stack, not a separate thing.

Three types of machine learning

Supervised learning is the most common type. You give the model labelled training data. Here are a million emails, here is which ones are spam. Learn the pattern, then classify new emails.

Unsupervised learning gives the model data with no labels. Find the structure yourself. Used for clustering (group these customers by behaviour) and anomaly detection (this transaction looks different from everything else).

Reinforcement learning works through trial and error. The model gets a reward for good actions and a penalty for bad ones. AWS DeepRacer is the simple example everyone uses. AlphaGo is the well-known one.

When to use which technique

| I want to | Technique | Example | |---|---|---| | Predict a number | Regression | Forecast next month's AWS costs | | Predict a category | Classification | Spam or not spam | | Group similar items | Clustering | Customer segments | | Generate text or images | Generative AI | Summarise a document | | Make decisions over time | Reinforcement Learning | Optimise resource allocation |

AWS services mapped to each layer

Traditional ML and AI services:

• Amazon SageMaker is the full ML pipeline on AWS. Build, train, deploy custom models.
• Amazon Rekognition does computer vision. Image and video analysis. Deep learning under the hood, managed for you.
• Amazon Comprehend handles NLP. Sentiment analysis, entity extraction, language detection.
• Amazon Forecast does time-series prediction for demand forecasting and inventory.
• Amazon Fraud Detector uses ML specifically for fraud detection.
• Amazon Personalize powers recommendation engines.

Generative AI services:

• Amazon Bedrock gives you managed API access to foundation models including Claude, Titan, Llama, and Mistral. This is the main GenAI service on AWS.
• Amazon Q Business is an enterprise knowledge assistant built on Bedrock.
• Amazon Q Developer is an AI coding assistant.

What I actually built

Jarvis, the AI assistant on this site, uses Claude via the Anthropic API. When you ask it "where does Ajay work?" it is not doing a keyword search.

It uses RAG, Retrieval-Augmented Generation:

1. Your question gets converted into a vector embedding, which is a list of numbers representing the meaning of the text
2. That embedding gets compared against embeddings of my resume, blog posts, and project docs stored in PostgreSQL with pgvector
3. The most semantically similar chunks get pulled back
4. Claude generates an answer using those chunks, not from its general training data

That is Deep Learning (the embeddings and the foundation model) combined with Machine Learning (the vector similarity search). All of it is AI in the broad sense. That is the nesting playing out in a real system I shipped.

What actually matters for the exam

The AWS AI Practitioner exam (AIF-C01) tests whether you understand AI and ML well enough to make good decisions, not whether you can build models. You will not write code. You will not tune hyperparameters.

The questions look like this: "A company wants to analyse customer feedback to find common complaints. Which AWS service?" The answer is Amazon Comprehend.

Or: "A company wants employees to ask questions about internal documents using AI. Which approach?" The answer is RAG with Bedrock Knowledge Bases.

Knowing the taxonomy tells you which tool to reach for. Unstructured text goes to Comprehend or Bedrock. Images go to Rekognition. Recommendations go to Personalize. Custom model on your own data goes to SageMaker.

The short version

AI            →  broadest category, any computer mimicking human intelligence
ML            →  learns from data, no explicit rules needed
Deep Learning →  ML using neural networks, handles unstructured data well
GenAI         →  generates new content, built on deep learning

Think of them as nested layers, not competing technologies. Each one builds on the one above it. GenAI is what happens when you put enough compute and data through the deep learning layer.

Next post I will walk through how I built the RAG system behind Jarvis, including the embedding pipeline, pgvector setup, and why it is relevant to AWS AI Practitioner Domain 2.

References

• AWS Certified AI Practitioner Exam Guide (AIF-C01) — official AWS documentation
• Amazon Bedrock documentation — foundation models and RAG on AWS
• Amazon SageMaker documentation — full ML pipeline on AWS

Part of my AWS Golden Jacket journey: 6 months, all 12 AWS certifications, documented here.