<aside> <img src="/icons/thought-dialogue_red.svg" alt="/icons/thought-dialogue_red.svg" width="40px" /> This is a draft of an introduction to AI written by Dominik Lukeš that will be a part of a bigger text. Comments welcome at [email protected].
</aside>
The term AI was coined in 1956 when there were still only about a thousand computers in the world and most of them still used vacuum tubes. Since then many things have come under the label of AI including spam filters and Netflix recommendation engines.
You can read more details about the history of AI in Towards a periodicization of AI history .
From the very beginning of computers, people have been thinking about their potential as machines that could replicate human thoughts processes across a number of areas. But until very recently, this was very limited. The first 50 or so years of AI's history are a history of dashed expectations that can be best summed up by this simplification of the famous Moravec paradox:
Easy things are hard and hard things are easy.
Things that were hard for people like multiplying up numbers or constructing complex if-then statements were easy for computers. But the things that seemingly require no intelligence at all were almost impossible. These included things like recognising a face or understanding that what the word "it" in "The toy didn't fit in the box because it was too large" must refer to the toy and not the box.
You can see in this graph of the increase in capabilities, what a change we've seen in the last few years compared to the previous half a century.
The AI that was introduced to the world through ChatGPT in 2022 is different. It reverses the Moravec paradox and finds it much easier to figure out the subtle meaning in a sentence than to multiply numbers. And it can do it in multiple languages (including computer ones). It is not perfect at it, but much better than anyone expected it would be.
So when we say AI today, we mostly mean this new kind. The kind of AI that can do the sort of things ChatGPT can do. To avoid confusion, we call it generative AI.
Many people try to give abstract definitions of or generative AI but the best way towards understanding it is about its most prominent example: ChatGPT. So this is the working definition for us here:
Generative AI is the a class of technology that can do the sort of things ChatGPT can do: Generate natural looking and sounding text, images, computer code, speech in response to prompts by text, speech or image.
This is a huge simplification but trying ChatGPT will give you a most comprehensive sense of what AI can do. There's a caveat, the free version of ChatGPT does not have the most power or all the features. Microsoft make most of these features available for free via Copilot.