Degree 2 · Unit 2.1
What is artificial intelligence?
Imagine that you are teaching someone to tell good dates from poor ones. There are two routes open to you. The first is to write them a list of rules: if the colour is uniformly dark, the texture soft, and the size above a certain point, then the fruit is good. That is the traditional program. You do the thinking and the computer carries it out, and all the intelligence in it is your own intelligence, written down in a language the machine understands.
The second route is to show it a thousand good dates and a thousand poor ones, tell it which are which, and then leave it to work out for itself what separates them. That is machine learning: you supply the examples and the computer derives the rule. It may well derive a rule that never occurred to you — and it may just as easily derive a completely useless one, like the date format in the story that opened this degree.
The terms in their places
What confuses a beginner more than anything else is that these terms get used as though they were synonyms, when they are not. They are nested circles, each one sitting inside the one before it.
The rule: start with the simpler one, and climb to the deep only when the data and the problem force it.
| Machine learning | Deep learning | |
| Scope | A broad class within AI | A subset of machine learning |
| Complexity | Simpler models, features extracted by hand | Complex neural networks, many layers deep |
| Data | Works well on small and mid-sized sets | Demands large sets |
| Compute | Standard processors | Powerful graphics processors |
| Feature engineering | Chosen by hand | Extracted automatically |
| Examples | Fraud detection, recommender systems | Image recognition, language processing |
fawzooz.ai
Deep is not "better" outright: with little data the simpler model wins, and it stays explainable — which is worth a great deal in sensitive decisions.
Why the word "intelligence" misleads
When we call a system intelligent, we transfer to it — without noticing — everything we know about human intelligence: intent, understanding, responsibility, and the ability to say "I don't know". AI has none of these. It is an extremely powerful statistical approximator: it finds patterns in an enormous quantity of data and generalises them to new cases.
This is why AI surpasses you at some things and fails spectacularly at things a child manages easily. It will summarise a hundred pages in a minute and then miscount the letters in a single word. There is no contradiction in that at all. It is evidence that what goes on inside it is not thinking in the sense you understand the word.
So do not ask whether it is intelligent. Ask instead what data it was trained on, and which pattern it is repeating for you now.
A classification by capability rather than by technique — and it shows plainly where we actually stand today.
fawzooz.ai
Our position: everything you use today, without exception, is of the first rung. Any claim otherwise is marketing, not engineering. Confusing the three rungs is the source of most of the hype and most of the fear alike.
Do this
1 — On paper. Take a task from your work and write it twice: once as explicit rules that could be programmed, and once as examples that could be learned from. Which form came more easily?
2 — On the tool. Ask it to solve a problem that needs exact counting, then ask it to summarise the same paragraph. Compare the performance and write the reason for the difference in your own words.
3 — In your field. Name three systems you use daily that run on machine learning without being called "artificial intelligence".