How AI Stopped Following Rules and Started Learning From Data

In the early 2000s, Google Translate was built by linguists.

Not by data scientists. Not by neural networks. Linguists — people who understood grammar, syntax, and the structure of language — sat down and hand-wrote rules for how to convert one language into another.

Subject goes here. Verb goes there. Exception for this tense. Rule for that clause.

Thousands of them. For every language pair.

It worked. Badly. Anyone who used it before 2016 remembers.

Then Google did something that felt almost like cheating.

They stopped writing rules. Instead, they fed the system millions of sentence pairs — the same sentence in English and French, English and Spanish, English and Mandarin — and told the model to figure out the patterns itself.

No linguist required.

The result didn't just improve. It made years of hand-crafted grammar rules obsolete overnight.

That one decision — from rules to data — is the entire history of AI in one story.

The Family Tree Nobody Drew For You

Artificial Intelligence is an umbrella. Under it lives everything that makes machines seem intelligent.

For most of AI's early history, that intelligence was explicit. Someone wrote the rules. A programmer told the machine: if this, then that. If an email contains these words, mark it as spam. If a chess piece is here, these are the legal moves.

It worked within boundaries. It broke outside them. And every new edge case needed a new rule.

Then came Machine Learning.

ML didn't throw out AI. It changed how the intelligence got built. Instead of a programmer writing the rules, the system learned them from data. You fed it thousands of labelled examples and it found the pattern connecting them.

The engineer stopped writing answers. The data started carrying them.

Deep Learning took this further. Instead of needing a human to decide which features of the data mattered, Deep Learning models — built from layers of artificial neurons — found the features themselves. Feed it enough images, it learns what an edge is, what a face is, what a cat is. Without being told what to look for.

Then came the models that changed everything publicly: Large Language Models, Generative AI, and now Agentic AI — systems that don't just classify or predict, but create and act.

Each step is the same shift, taken further. Less from the programmer. More from the data.

Two Buckets. Everything Fits.

If you want a mental model that holds across the entire history of the field:

Explicit intelligence — the system knows what to do because someone told it. Rules, algorithms, decision trees, search. The engineer is the source of the intelligence.

Learning intelligence — the system figures out what to do from data. ML, Deep Learning, LLMs. The data is the source of the intelligence.

Every AI system ever built sits in one of these two buckets. The direction of travel has always been the same: from the first toward the second.

The Responsibility Transfer

Here is the sharpest way to say what actually happened:

The history of AI is the story of transferring responsibility from the engineer to the data.

Early AI: the engineer is responsible for the answer. Write the right rules, get the right output.

Machine Learning: the engineer is responsible for the question. Collect the right data, define the right problem, the model finds the answer.

Deep Learning: the engineer is responsible for the architecture. Design the right network, the model finds the features and the answer.

Generative AI: the engineer is responsible for the prompt. Ask the right question, the model generates the answer, the image, the code.

Each era handed off more to data and less to the human writing the logic.

You were never really programming the intelligence. You were progressively stepping back from it.

What Comes Next

The family tree doesn't stop at Generative AI.

Agentic AI is already here — systems that don't just respond, but plan, act, and use tools to complete multi-step tasks. The responsibility transfer continues: the model now decides not just what to say, but what to do.

But before any of that makes sense, there's a more fundamental question.

How does the most powerful branch of this tree — the Large Language Model — actually see the world?

It doesn't read words the way you do. It doesn't even read sentences.

That's Day 03.

Day 02 of 100 — AI Foundations | Change of Basis — Reframe the familAIr. See the invisible.