Data — the fuel of AI: why data is what decides

In the lesson on how a network learns, the principle "garbage in, garbage out" went by in passing. Time to take it seriously: data isn't background, it's the fuel every AI runs on. Change the fuel and everything changes.

Why the data decides, not the algorithm

Intuition says a strong AI means a brilliant algorithm. In reality, training algorithms have been published for years and are available to everyone — students study them. What separates a mediocre model from a great one is more often the data: how much, how clean, how varied. Remember from lesson one why the breakthrough came in 2012: the neural-network algorithms had been known since the 1980s, but they only fired once the internet had piled up billions of examples. The algorithm wasn't waiting on data — the data woke the algorithm up.

Hence the industry rule: investing in the data usually pays better than one more tweak to the model. Companies with unique data (search engines, social networks, banks) are strong precisely because of it — data is harder to copy than code.

Two kinds of data: structured and unstructured

Structured data is neatly laid out in tables: rows, columns, numbers, categories. Customer age, purchase amount, date — everything in its place.

Unstructured data is everything else: the text of emails, photos, audio, video, posts. No rows, no columns — just a stream. By various estimates, most of the world's data is unstructured, and for decades business could barely use any of it. The neural-network breakthrough is exactly this: for the first time there's a tool that pulls meaning out of that ocean. The LLM from the last module is a machine for digesting unstructured text.

Where data comes from

• Labelling. People add the "answers" by hand: outlining pedestrians in footage for a self-driving car, marking emails spam / not spam. The famous ImageNet dataset from lesson one — about 14 million pictures, labelled by hand by an army of people; that's what AlexNet won on. Labelling is expensive, slow and irreplaceable work.
• Watching behaviour. Every click, every like, every video you watch to the end is free labelling. Without noticing, you're the one training the recommendation systems: watched to the end — "good example"; scrolled past — "bad one".
• Ready-made datasets. Collected, cleaned sets that lots of people reuse: the text of books, encyclopedias, open photo collections. Fast — but everyone has the same ones, so they give no unique edge.

You give the input and the output — the network builds the rest itself

An important detail that explains why labelling "input → answer" is enough. You never describe to the AI what a cat or a face looks like. You give only input A (a photo) and output B ("cat / not cat") — and the network assembles the in-between features itself, layer by layer, like lego bricks. In face recognition you can see it plainly: the first layers find simple edges and blobs, the next ones build parts out of them (an eye, a nose, an outline), the top ones build whole faces.

Nobody programmed "an eye" by hand — the network grew the concept itself, just so the answers would come out right. That's why the approach won (remember the lesson on what AI is): the features a person couldn't put into words, the machine derives from examples.

"Garbage in, garbage out"

Now the principle from the training lesson can be heard at full volume. A model is no smarter than its data: it learns what you showed it, mistakes, gaps and skews included.

A real case: in 2016 Microsoft launched a chatbot on Twitter called Tay, meant to learn natural conversation directly from what users said to it. In under a day, trolls buried it in abusive and racist messages — and Tay, obediently learning from that "fuel", started repeating the same filth. Sixteen hours later Microsoft pulled the plug. The algorithm worked exactly as designed — it was the data being poured in that was poisoned.

Three misconceptions about data

• "More data is always better." No: a million junk examples are worse than a thousand clean ones. Volume without quality only amplifies the error.
• "Data is objective, it's numbers." Data is an imprint of reality taken by people, with their choices and their blind spots. A number is no guarantee of truth: remember the knowledge cutoff and the out-of-date prices from the training lesson.
• "We don't have much data — AI isn't for us." A little is often enough, if it's clean and on point; and you can always take a ready model and top up its training a bit (that was at the end of the training lesson).

Do this now

A five-minute exercise, no tools — just your head. Pick any AI service you know (a recommendation feed, a spam filter, a voice assistant) and answer three questions on paper: (1) what data was it most likely trained on? (2) where did that data come from — labelling, watching behaviour, or a ready-made dataset? (3) what skew could have got into it? The habit of seeing the data behind the result is what separates a professional from a consumer. In the next lesson we'll look at what AI, given data like that, can really do — and what's beyond it.