A Neural Network Didn't Learn Your Taste. It Calculated It.

Nobody at Instagram wrote a rule that says: "Show this person cooking videos."

There is no spreadsheet somewhere that maps your interests. No engineer who decided you like travel content on Sunday mornings and finance videos on weekday evenings. No explicit logic that says "if user watches > 3 videos of a type, show more."

And yet, your Reels feed knows you. Uncomfortably well.

That's not coincidence. That's a neural network — a system that processed your behaviour, found patterns you didn't even notice in yourself, and keeps updating its predictions every time it gets something wrong.

No rules. Just learned logic.

What a Neural Network Actually Is

The easiest way to picture a neural network is as a pipeline.

Data goes in one end. A prediction comes out the other. In between, the data flows through layers — each one transforming it into something more refined than what came before.

The input layer receives raw data — in the case of your Reels feed, that might be: which videos you watched fully, which you skipped in two seconds, which you rewatched, what time of day it is, what device you're on.

The hidden layers are where the real work happens. Each layer is made up of neurons — mathematical functions that take the incoming values, apply a weight to them, and pass the result forward. The first layer might find basic patterns. The next layer finds relationships between those patterns. Deeper layers find increasingly abstract meaning — not just "this person watched food videos" but "this person engages more with quick recipes on weekday evenings and long cooking process videos on weekends."

The output layer takes everything the hidden layers found and produces a prediction — in this case, a ranked list of videos most likely to keep you watching.

Three stages. Input. Process. Output. Every neural network ever built follows this structure.

The Part Nobody Tells You: Weights

Here's what makes this different from any pipeline you've built manually.

Every connection between neurons in the network has a weight — a number that controls how much influence one neuron has on the next. A high weight means "pay close attention to this." A low weight means "this barely matters."

At the start of training, these weights are random. The network knows nothing.

Then it sees data. It makes a prediction. The prediction is wrong. It calculates exactly how wrong and in which direction. Then it adjusts the weights — slightly, systematically — to make a better prediction next time.

This process repeats. Billions of times. Across millions of examples.

Nobody wrote the rule "cooking videos matter more on weekends." The network found that pattern by adjusting weights until the predictions got better. The logic wrote itself.

This is what separates a neural network from traditional software. You don't program the answer. You program the learning process, feed it data, and let it find the answer.

Why This Explains Hallucinations

Here is the insight that changes how you think about AI outputs.

A neural network is not querying a database of facts. It is not looking anything up. It is calculating the most probable output based on the patterns in its weights.

When you ask an LLM a question about a niche topic it has barely encountered in training, it doesn't say "I don't have data on this." It calculates the most probable sequence of words that fits the pattern of "answering a question like this" — and produces text that sounds confident and coherent, even if the facts are wrong.

The network isn't lying. It's doing exactly what it was built to do: find the most probable next output. It has no mechanism for checking whether that output is factually true.

This is why hallucinations happen. And once you understand it at this level, you stop being surprised by them.

Context is the Input—Garbage in, Garbage Out.

Because the hidden layers can only work with what arrives at the input layer, the quality of your output is entirely dependent on the quality of your input.

A vague prompt gives the network a vague starting point. It finds a broad pattern match and produces a broad answer.

A precise prompt — with specific constraints, format requirements, context, and examples — gives the network a highly constrained path. The data flows toward a more specific output.

This is not a writing tip. It is the system's architecture.

When you tell the model who you are, what you need, what format you want, and what to avoid, you are not being polite. You are shaping the input layer so the hidden layers have something precise to work with.

What Comes Next

A neural network learns from data. But the question no one asks is: what data, and how much of it?

The scale of data that modern AI systems train on is incomprehensible. And the decisions made about what data to include — and what to leave out — shape everything the model knows, believes, and gets wrong.

That's Day 06.

Day 05 of 100 — AI Foundations | Change of Basis — Reframe the familAIr. See the invisible.