Lesson 1 — The problem AI solves
Before asking what Artificial Intelligence is, it helps to see what problem it was invented for. That problem is easy to state and surprisingly stubborn: some tasks are impossible to describe as a list of instructions, even though a five-year-old performs them effortlessly.
The experiment that shows the wall
Take a task everyone considers trivial: decide whether a photo contains a cat.
You are a programmer. Write the rules.
You might start with "a cat has two pointed ears". Then someone submits a photo of a cat seen from behind, ears flattened. You add "a cat has whiskers" — and receive a close-up of a paw. You add "a cat is furry" — and get a Sphynx, which is hairless. You add colour rules, and someone sends a black cat photographed at night, where the whole animal is three barely distinguishable shades of dark grey.
After two thousand rules you still fail on photographs a child sorts instantly. And every new rule you add to catch one case quietly breaks three others.
The problem is not that you are a bad programmer. The problem is that you do not know the rules you use yourself. You recognise a cat without being able to state why. A knowledge you cannot express cannot be transcribed into code.
This is the wall. It is called Polanyi's paradox, after the philosopher who put it plainly: we know more than we can tell.
The two eras that came before
Rules written by experts (1960s to 1990s)
The first serious answer was the expert system. The idea was reasonable: if a machine cannot discover the rules, let a human specialist dictate them. A doctor states hundreds of "if the patient has this symptom and that test result, then consider this diagnosis" rules, and an inference engine chains them.
It worked, within limits. MYCIN, built at Stanford in the 1970s, diagnosed blood infections about as well as a junior physician. But three flaws proved fatal:
- Extracting the rules was agonising. Interviewing experts for months produced a few hundred rules where reality needed tens of thousands.
- The rules contradicted each other. Past a certain volume nobody could tell whether adding a rule improved the system or corrupted it.
- Nothing generalised. A system built for blood infections was worthless for lung disease. Everything had to be restarted from zero.
Funding collapsed twice, in periods now called the AI winters. By the mid-1990s "artificial intelligence" was close to a term you avoided in a grant application.
Learning from examples (1990s onward)
The other approach had existed since the 1950s and was long considered disappointing: stop writing rules, and let the machine derive them from examples.
Instead of describing a cat, you show the machine one hundred thousand photographs labelled "cat" or "not a cat" and let it work out for itself which visual regularities separate the two piles. Nobody ever writes down "pointed ears". The machine discovers, on its own, that certain combinations of shapes and textures predict the label well.
This is machine learning, and it inverts the relationship between programmer and program:
| Classical programming | Machine learning | |
|---|---|---|
| What the human provides | the rules | examples and their answers |
| What the machine produces | answers | the rules |
| To improve the result | rewrite the code | provide more and better data |
| Where the difficulty sits | expressing the logic | collecting and cleaning the data |
Why it only took off around 2012
The mathematics behind neural networks was largely settled by the late 1980s. So why did the world only notice around 2012? Because a method is useless until three ingredients arrive together, and they arrived late.
Data. A model learns from examples, and it needs an enormous number of them. Before the web there was no way to assemble a million labelled photographs. The ImageNet dataset, published in 2009, provided fourteen million annotated images — the first pile large enough to matter.
Computing power. Training a network means performing billions of small multiplications. Processors designed for video games, GPUs, happen to do exactly that, thousands of operations in parallel. Diverting them to machine learning cut training times from months to days.
A verifiable result. In 2012 a network called AlexNet entered the ImageNet competition and cut the error rate from 26% to 16% in a single year, in a field where progress was measured in fractions of a percent. The gap was too large to dismiss. Within three years every serious competitor had switched to deep learning.
Modern AI is not a new idea. It is an old idea that finally had enough data and enough computing power to work. That matters when you evaluate a promise: an approach that needs data you do not have will not work for you, however impressive the demonstration.
What this changes for you, concretely
The shift from "writing rules" to "providing examples" moves the difficulty rather than removing it. In a real project the hard parts become:
- Do I have the data? Not "is there data in the world", but "do I have, in-house and legally usable, enough labelled examples".
- Is my data representative? A model trained only on daytime photographs will fail at night. It will not warn you; it will simply be wrong.
- Can I tolerate being wrong sometimes? A model gives a probability, never a guarantee. Some uses accept that. Others do not.
These three questions decide most projects, long before anyone chooses an algorithm. Lesson 4 shows why.
In three sentences
Some tasks cannot be written as rules because we cannot state the rules we ourselves use. Machine learning sidesteps this by deriving the rules from labelled examples, which shifts the effort from code to data. The approach only became practical around 2012, when large datasets and graphics processors arrived at the same time.
Next — Lesson 2: what is Artificial Intelligence, exactly? →