As I started learning Applied AI, I realized I didn't want to jump straight into LLM APIs, RAG, agents, and prompting without understanding what was underneath.
So I went back to the basics.
From Machine Learning to Deep Learning
Traditional machine learning often relied on humans to decide which features were important.
For example, for digit recognition, we might manually identify:
edges
curves
strokes
shapes
Neural networks changed this approach.
Instead of manually defining all the features, a neural network can learn useful representations from the data.
A simple network looks like:
Input → Hidden Layers → Output
Each neuron uses weights and biases, and during training the model adjusts these parameters to reduce its error.
The cost function tells us how wrong the prediction is.
Backpropagation helps determine how each parameter contributed to that error, and gradient descent updates the parameters in a direction that reduces the cost.
Why Deep Learning?
When we stack many layers, we get deep neural networks.
The interesting part is that different layers can learn increasingly complex representations:
Pixels → Edges → Shapes → Parts → Object
This ability to automatically learn representations from raw data became one of the major strengths of deep learning.
From Deep Learning to Generative AI
A neural network doesn't have to be supervised. It can be trained using supervised, unsupervised, self-supervised, or reinforcement learning.
Modern AI heavily relies on self-supervised learning.
For example:
"The capital of France is ___"
Target → Paris
The text itself provides the training target. This makes it possible to learn from enormous amounts of data without manually labeling everything.
Then came Transformers, which introduced the idea of attention—allowing models to learn relationships between different parts of a sequence efficiently.
Large Transformers trained on massive datasets became foundation models.
When these models are trained to work with language at enormous scale, we get Large Language Models (LLMs).
And when models can use what they've learned to generate new text, code, images, audio, or video, we call it Generative AI.
The Mental Model
This isn't a perfectly linear hierarchy, but this is the mental map that helped me:
Machine Learning
↓
Neural Networks
↓
Deep Learning
↓
Transformers
↓
Self-Supervised Learning
↓
Foundation Models
↓
LLMs
↓
Generative AI
The goal isn't to become an ML researcher before becoming an Applied AI Engineer.
For me, the goal is simpler:
Understand what problem each concept solved, so that when I build on top of AI models, I know what I'm actually building on.
That's the foundation I'm starting my Applied AI journey with.
Top comments (0)