Generative AI feels like magic. You type in a few words, and it writes a story, creates an image, or answers your question like it has a brain of its own.
But under the hood, there's no magic — just math, data, and a whole lot of pattern recognition.
In this post, let’s explore how Gen AI learns, in simple terms, using a real-world example: how an AI learns to write restaurant reviews.
Imagine This: Teaching AI to Write Restaurant Reviews
Let’s say we want to train an AI to write reviews for restaurants.
But first — what do we actually give it?
Step 1: Feed it Data
We start by giving it tons of real reviews. Thousands, maybe millions.
Example input:
“This pizza was amazing! The crust was crispy, the sauce was rich, and the staff was super friendly.”
“I waited 40 minutes for cold spaghetti. Will never come back.”
The AI doesn’t understand what pizza or spaghetti mean — not yet. But it can start noticing patterns.
Step 2: Breaking Everything Into Tiny Pieces
AI doesn’t read like us. It doesn’t see sentences — it sees tokens.
Tokens are chunks of text. For example:
"This pizza was amazing!"
becomes →[This, pizza, was, amazing, !]
These tokens are turned into numbers using a technique called embedding. Why? Because math is what AI understands.
So now, the AI is looking at rows and rows of numbers that represent words, their meanings, and how they relate to each other.
Step 3: Spotting Patterns in the Chaos
Now comes the interesting part: training.
The AI goes through billions of examples like:
- “The ___ was delicious.” → Tries to guess “pizza,” “pasta,” “soup”
- “The staff was ___.” → Guesses “kind,” “rude,” “friendly”
Over and over again, it tries to predict the next word. Each time it gets it wrong, it adjusts its internal settings a little. This is how it learns — by failing millions of times and getting better with every try.
This process is done using massive models like transformers (more on that in future posts).
🤖 Real-World Analogy: Autocomplete on Steroids
Think about when you type a message and your phone suggests the next word.
Generative AI is doing something similar — but instead of just a few words, it can generate entire paragraphs, essays, poems, or code.
The core idea is still the same: predict what comes next based on everything it has learned before.
So What’s a “Model”?
You’ll hear this word a lot.
A model is what we get after training.
It’s the result — the thing that now “knows” how to respond based on the patterns it has seen. Just like a student who studied a thousand flashcards now knows all the answers.
Popular models include:
- GPT (OpenAI)
- Gemini (Google)
- Claude (Anthropic)
- LLaMA (Meta)
Each model is trained on huge amounts of data — books, articles, websites — and each learns slightly differently depending on how it was built.
Memory Without Memory
One mind-blowing thing: most AI models don’t have real memory.
They don’t remember your last conversation (unless specifically built to do so). They respond based only on the input you give right now, plus what they learned during training.
So if you ask it to write a pizza review, it doesn’t remember writing one earlier. It just sees the prompt and thinks:
“Based on what I know… what words should come next?”
🗺️ Zooming Out: What AI Isn't
Let’s clear up a common myth.
AI doesn’t think. It doesn’t feel. It doesn’t know truth.
It’s just really, really good at pattern prediction. That’s it.
So if it writes:
“This hidden sushi bar in Chicago has the best miso soup in the world.”
It may be totally made up — or based on something it once saw. That’s why accuracy is a big challenge in Gen AI.
Quick Recap
Here’s how Gen AI learns, using our restaurant review example:
- Data In → Millions of real reviews
- Tokenization → Break text into tiny, meaningful chunks
- Embedding → Convert words to numbers
- Training → Guess what comes next, adjust, repeat
- Model Output → An AI that can now generate its own reviews
Pretty cool, right?
Bonus: Want to See It in Action?
Try writing:
“Write a review of a pizza place that only serves dessert pizza.”
into ChatGPT or Veo by DeepMind
Watch how it predicts a story based on everything it's ever seen.
What’s Next?
In the next post, we’ll explore embeddings and vector data — how AI turns words into numbers and how it finds the “closest match” when answering your questions.
We’ll explore it using a fun example: How an AI finds your favorite superhero based on personality. 🦸♂️
Got questions or want to see a specific example? Drop it in the comments!
Follow for the next post — we’re just getting started. 🚀
Top comments (0)