What is artificial intelligence?

01Start with a calculator

A calculator has one skill: basic arithmetic. It beats every human alive at it on speed. That's narrow intelligence, outside a brain.

Now multiply its skills.

1. Give it all of mathematics, and you have an AI for maths.
2. Do the same across many fields (law, medicine, code, writing) and you have a polymath AI.
3. Do it across every field and you have an AI that knows more than any single human ever could.

But there's a catch. Humans are good at taking knowledge from one field and using it in another. That's the dimension of intelligence AI is still weakest at. Stacking more skills doesn't automatically give you that crossover. Hold that thought. It's the line between AI and AGI in the next article.

02Two ways people built it

1. Write the rules. For decades, people tried to make machines smart by writing down how to think: if this, then that. It worked for tidy problems and broke on messy ones. Nobody can write the rules for spotting a cat in a photo.

2. Learn from examples. Show the machine millions of examples and let it adjust itself until it gets them right. This is machine learning. The machine ends up with its own internal rules, which nobody wrote and which are hard to read. Almost all AI today works this way.

03What today's chatbots actually do

1. Training. The model reads an enormous amount of text and plays one game, billions of times: guess the next word. Each wrong guess nudges its internal numbers a little. Over time, those numbers come to hold grammar, facts, style and a surprising amount of reasoning.
2. Tuning. A raw model just continues text. To make it a useful assistant, companies train it further on good answers and on human ratings of which answers are better. This step shapes its manners and its values. It's the first place alignment happens.
3. Use. You type, and it builds a reply one piece at a time.

"It only predicts the next word" is true and misleading. To predict the next line of a physics proof, it helps to understand physics. Nobody knows exactly how much understanding sits inside these models.

04Where it's heading

The path I see is a chain: chatbots → agents → employees enabled → productivity up → and then...

Agents are AI that don't just answer. They act: they browse, write code, send emails and run tasks for hours. Once agents are reliable, a lot of work changes hands. That's where this series is going.

05The short version

• AI is intelligence outside of a brain.
• A calculator is AI with one skill. Multiply the skills and you get today's systems.
• What AI still lacks most is the human trick of carrying knowledge from one field into another.

06Sources

• McCarthy, J. et al. (1955). A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence.
• Weizenbaum, J. (1966). ELIZA. Communications of the ACM 9(1).
• Krizhevsky, A., Sutskever, I. and Hinton, G. (2012). ImageNet Classification with Deep Convolutional Neural Networks. NeurIPS.
• Silver, D. et al. (2016). Mastering the game of Go with deep neural networks and tree search. Nature 529.
• Vaswani, A. et al. (2017). Attention Is All You Need. NeurIPS.
• Ouyang, L. et al. (2022). Training language models to follow instructions with human feedback. arXiv:2203.02155.

All posts