Let’s Talk Technical: What’s an AI model?

You’ve heard about AI everywhere—chatbots, self-driving cars, even Netflix recommendations. But how does it really work?

You don’t need to be a computer scientist to understand it! Let’s break it down in plain English.

Technical Note: The term “AI” encompasses a broad range of techniques, including rule-based systems that operate without training. However, the AI we encounter in everyday applications typically relies on extensive data training to develop a basic understanding. This branch of AI is called “Machine Learning“. Within this category, we have “Deep Learning” models, which create internal interconnections based on the relations in the data. “Deep Learning” models are also called “Neural Networks” as it’s inspired by the structure of the human brain. “Transformers” is a more specialized architecture based on “Neural Networks” employed in advanced language generation models like ChatGPT.

For clarity and ease of understanding within this course, the term “AI” will be used as shorthand to refer to any model without explicitly mentioning the type so as not to cause confusion.

AI is Like a Super-Fast Learner

Imagine teaching a child to recognize dogs:

• You show them lots of dog pictures (training).
• They notice patterns (fur, ears, tails).
• Soon, they can spot (recognize) new dogs they’ve never seen before.

AI works the same way—it learns from examples instead of following rigid rules.

The “Brain” Behind AI (Neural Networks)

AI’s “brain” is called a neural network, loosely inspired by how our brains work.

How It Makes Decisions (Step-by-Step)

1- Input Layer (The Eyes & Ears)

AI gets data (text, images, sounds).

Example: A photo of a cat → turned into numbers (pixels, colors).

2- Hidden Layers (The Thinking Part)

Each layer has artificial neurons (tiny decision-makers).

They weigh evidence (e.g., “pointy ears = probably a cat”).

The more layers, the “smarter” the AI (that’s deep learning).

3- Output Layer (The Answer)

Combines all clues → final decision (“It’s a cat!”).

AI is Just Fancy Math (No Magic!)

At its core, AI is just math on steroids:

1. Each neuron does a tiny calculation, mostly one multiplication and one addition (input × weight + bias).
2. Layers pass numbers forward like an assembly line.
3. Training = Adjusting weights until the AI gets good at guessing.

Mind-Blowing Fact:

ChatGPT has 175 billion weights (adjustable knobs) fine-tuned by data!

Math and Latent Knowledge

But here’s the twist—math doesn’t just represent numbers; it also stores ideas. When an AI like ChatGPT is trained, it learns patterns buried deep in data, even if we never told it the rules directly. That’s latent knowledge.

For example, it sees the word “king” appear near “queen,” “throne,” “palace,” and “royalty” so often that it starts to connect them—just like we do. But it goes further. Depending on the context, it can sense that the opposite of “king” might be “pawn” in chess, “usurper” in a revolution, “commoner” in society, or “queen” in terms of gender. These aren’t facts it was given—they’re patterns it uncovered, quietly encoded in billions of mathematical weights.

It builds a web of meanings where “king” links not just to “queen” and “throne,” but also to “power,” “loss,” “challenge,” and even “checkmate.” This is latent knowledge: the quiet math behind the scenes, forming connections that feel almost human.

AI Model Modes of Operation

Any deep learning model has two modes of operation, training and inference. Whenever the model is in “school”, it’s “training”. Like when the model sees a lot of examples of how the word “king” is used. 

Then once the model has enough “knowledge” it can work in the other mode, inferences. In inference, the model locks the weights, locks the memory, and starts to use the knowledge it has to give you an answer. 

Updating the model

Sometimes, the model’s general knowledge isn’t enough—it needs to adapt to new topics, styles, or specialized domains. That’s where fine-tuning comes in. Fine-tuning is like sending the model back to a focused training session. It sees new, targeted examples (like legal documents or customer support chats), and slightly updates its internal weights to better reflect this domain—without erasing what it already knows. This process reshapes the model’s understanding, so its future inferences feel more relevant and aligned with the new data.

For example, before fine-tuning, the model understands “king” broadly—it connects it to words like “queen,” “throne,” or “royalty,” based on general internet text. But suppose you’re building an AI for a chess training app. In this context, “king” isn’t about royalty—it’s a game piece with specific rules.

By fine-tuning the model on chess games, tutorials, and commentary, the meaning of “king” shifts slightly in the model’s internal math. It starts linking “king” with “check,” “castle,” “endgame,” and “mate.” The original associations aren’t erased, but the model now favors chess-relevant meanings in that context. It’s still the same model—but subtly re-trained to think more like a chess coach.

Real-Life AI Examples You Know:

1. Netflix Recommendations → AI predicts what you’ll like based on past watches.
2. Voice Assistants (Alexa/Siri) → AI turns speech into text, then answers.
3. Spam Filters → AI learns which emails look “fishy.”
4. ChatGPT, Claude, and all the new fancy smart LLMs (The technical name of ChatGPT-like models)

What happens when we need to inject more information to the AI? 

Conclusion

AI is everywhere—powering features in healthcare, finance, and even your phone’s camera. It’s not magic. It’s just pattern recognition powered by math.

You don’t need to know how to code to understand or benefit from it.

Next time you:

1. Chat with ChatGPT
2. Search photos using faces in your gallery
3. Translate a sign with Google Translate
4. Discover new music on Spotify

…remember: AI is making those decisions behind the scenes!