The Problem

Why rule-based AI was hitting a wall

The symbolic AI approach — writing intelligence as rules — had a fundamental problem: the world is too complicated for rules to cover.

Imagine you are trying to teach a machine to distinguish a dog from a cat in a photograph. You might start writing rules:

If the ears are pointed, it is probably a cat.
If the animal is large, it is probably a dog.
If the fur is striped, it is probably a cat.

But what about a large cat? A floppy-eared dog? A Siamese cat with no stripes? Every rule has exceptions. Every exception needs another rule. And the more rules you add, the more they contradict each other.

You could try to make your rules more precise. But “pointed ears” is not a mathematical definition — it is a fuzzy human judgement. How do you write that in a computer language? What is the exact pixel pattern that means “pointed ear”?

You cannot. Not in any practical way. Real-world intelligence — recognising faces, understanding speech, reading handwriting — involves patterns so complex and variable that no human could ever write down all the rules. There are too many.

The deeper problem: learning

Here is the real issue: how do you get rules into a computer in the first place?

You write them. A human programmer, who already understands the world, translates that understanding into code. The machine then executes those rules. But the machine did not learn the rules — a human learned them (by living in the world for years) and then typed them in.

This is not intelligence. This is sophisticated copying.

Turing had pointed at this problem in 1950. He suggested building a “child machine” that could be educated — that could learn rules from experience, the way children do. But he did not show how. No one did, for eight years.

What Rosenblatt wanted to build was a machine that could:

Be shown examples (this is a cat, this is a dog)
Figure out the pattern itself, without being told the rules
Generalise to new examples it had never seen

This is machine learning — and in 1958, no one had done it.

The problem in plain language

How can a machine get better at a task by practicing, without anyone telling it the rules?

In technical language:

How can a machine automatically adjust its internal parameters based on labelled training examples, such that it learns to correctly classify new, unseen inputs drawn from the same distribution?

Why this problem mattered so much

Think about what becomes possible if you solve it.

A machine that can learn from examples can:

Learn to read handwriting by studying thousands of handwritten letters
Learn to recognise faces by studying millions of photographs
Learn to translate between languages by studying millions of parallel sentences
Learn to predict whether a tumour is malignant by studying thousands of medical scans

None of these tasks can be solved by writing rules. All of them have been solved by machine learning. The Perceptron was the first demonstration that this kind of learning was mechanically possible — that a machine could, given the right architecture, extract patterns from data and use them to make predictions.

Without that first demonstration, it is hard to imagine the path to everything that followed.

Why should a student in small-town India care?

Because machine learning is what makes AI useful in India specifically.

The language you speak, the script you write in, the accents you use, the cultural references you make — none of these can be hand-coded by programmers sitting in California. They have to be learned from data. Sarvam AI’s model that understands ten Indian languages — learned. The AI that reads hand-filled government forms in Devanagari — learned. The voice assistant that understands your grandmother’s Bhojpuri accent — if it exists, it was learned.

Everything useful in Indian-language AI traces back to the idea Rosenblatt demonstrated in 1958: a machine that learns from examples.

Next: The Core Idea →