AI, Machine Learning and Deep Learning: the map of the territory
What each term really means and how they fit together — with analogies for people who code.
5 min read
AI, Machine Learning, Deep Learning, GenAI — these terms appear together constantly, but rarely does anyone stop and explain where one ends and the next begins. Before building any AI system, you need this map: not for academic formality, but because confusing the concepts leads to wrong architectural decisions.
AI is the umbrella, not the product
Artificial Intelligence is a broad field with a simple definition: systems that perform tasks that, until recently, seemed to require human intelligence — recognizing a face, translating text, playing chess, answering a question in natural language.
The key point is that AI is a category of problem, not a specific technology. An if/else rule system that diagnoses network failures can be called AI — and was, in the 1980s, under the name expert system. What changed was not the definition, but the dominant approach to solving those problems.
Today, when someone says "I'll use AI", they almost always mean "I'll use a model trained on data" — that is, Machine Learning. But AI is larger than ML. There are approaches based on search, formal logic, evolutionary algorithms. ML is just the approach that dominated the last decade because it scales well with data and compute.
Understanding this prevents a common mistake: thinking every problem needs a model. Sometimes a decision tree with explicit rules is more auditable, cheaper, and more correct. Lesson 20 of this track goes deep on that decision.
Machine Learning: learning instead of programming
Here is the analogy I use with every dev who is just starting out:
Traditional code: you write the rules. The program applies the rules to data and produces answers.
Machine Learning: you provide data and the correct answers. The algorithm discovers the rules on its own.
Think of a spam filter. The classic approach: if email.contains('free promo') or email.contains('click here') → spam. You write it, maintain it, update it by hand. It works until the spammer changes vocabulary.
With ML, you feed the model thousands of emails labeled spam or not-spam. The model learns patterns you could never write explicitly — word combinations, send time, sender behavior. And when the spammer changes tactics, you retrain with new examples.
This inversion — data + answers → implicit rules instead of explicit rules → answers — is the essence of ML. The resulting model is a file of numerical parameters that encapsulates what was learned. Lesson 02 of this track explains exactly how that learning happens.
ML covers a huge family of techniques: linear regression, decision trees, SVMs, Bayesian networks. Deep Learning is a specific subset of that family — and it is what unlocked the most impressive use cases of recent years.
Deep Learning, GenAI and LLMs: where we are today
Deep Learning is ML with deep neural networks — architectures with many layers that learn hierarchical representations of data. One layer learns edges in an image; the next learns shapes; the next learns faces. This stacking of abstractions is what made computer vision, speech recognition, and eventually natural language at scale possible.
The leap happened when hardware (GPUs), data (the internet), and architectures (especially the Transformer, from 2017) aligned. Networks that were previously impractical to train became feasible.
Generative AI is a subset of Deep Learning focused on generating new content — text, image, code, audio — rather than just classifying or predicting. And LLMs (Large Language Models) are generative models trained specifically on text at massive scale, capable of conversing, reasoning, summarizing, and writing code.
This is where this track spends most of its time. Not because LLMs are the only AI that matters, but because they are the technology most impacting the work of architects and devs right now — and truly understanding them, from tokens to agents, is what separates those who use them from those who design systems with them.
The diagram below shows how all of this fits together visually.
The territory map: nested sets of AI
Each layer is a subset of the previous one. LLMs are the entry point of this track.
- Sistemas de Regras · if/else, lógica formal
- Busca & Planejamento · algorítmico
- ML Clássico · regressão, árvores, SVM
- Visão Computacional · CNNs
- Voz & Áudio · RNNs, Transformers
- Geração de Imagem · Diffusion, GANs
- LLMs · GPT, Claude, Titan…
- Agente · LLM + tools + memória
What to lock in before moving on
Quick glossary
Tap a card to flip it.
In practice, most projects that come to me already arrive with the decision made: 'we'll use an LLM'. Sometimes that makes sense. Sometimes the problem would be better solved with a classification model trained on proprietary data — cheaper, more auditable, and with no hallucination risk. Knowing where each technology fits on the map is not academic trivia — it is what lets you question the premise before building the wrong solution. This track will give you that repertoire.
Hand-written rules vs. trained model
| Criterion | if/else Rules | ML Model | |
|---|---|---|---|
| How it's created | Hand-coded by experts | Trained on labeled examples | — |
| Scales with data | No — more data requires more rules | Yes — more data generally improves the model | — |
| Auditability | High — every rule is readable | Variable — from readable trees to opaque networks | — |
| Adaptation to changes | Manual — someone must update the code | Retraining with new data | — |
| Best for | Deterministic, auditable business logic | Complex patterns in unstructured data | — |
Frequently asked questions at this stage
Is every ML model a neural network?
No. Neural networks are one family within ML. Logistic regression, Random Forest, XGBoost, and SVMs are ML models that are not neural networks — and they remain the right choice for many tabular problems.
Are Generative AI and LLM the same thing?
No. Generative AI is the category — it includes image models (Stable Diffusion, DALL-E), audio, video, and text. LLM is the subset focused on language. Every LLM is GenAI, but not every GenAI is an LLM.
Do I need to know math to follow this track?
Not for this track. The focus is on how to use and architect AI systems, not on deriving backpropagation. Where math helps understand a concept, I explain with an analogy first.
What comes next in this track
Now that the map is clear, the track dives into the fundamentals you need to work with LLMs for real.
Lesson 02 opens the black box of learning: what parameters are, what happens during training, and what inference is — without heavy math, but with enough precision to make architectural decisions.
From lesson 03 onward, we enter LLM territory: tokens, context window, embeddings, prompting, RAG, and tool calling. Each lesson builds on the previous one. The final destination is lesson 11, where we assemble a complete agent — and the following lessons show how to put that into production on AWS with Amazon Bedrock and AgentCore.
Before moving to the next lesson, review the glossary terms below and answer the quick quiz. If you get everything right, you are ready to continue. If you miss something, reread the corresponding section — that is normal the first time.
Quick check
1. Which relationship is correct?
2. What sets ML apart from 'traditional programming'?