What is AI, really? (AI defined)
Research + my take
“AI” gets used for everything from a spam filter to ChatGPT. When I spoke at the 2023 PREA Institute, I started with definitions and history, because you can’t make good decisions about a technology you can’t define. Here’s that foundation, updated.
The short version: AI is a set of nested ideas
When people say “AI” today, they usually mean the innermost circle: generative AI. But it sits on eighty years of work.
A short history of AI
- 1940s
The brain as a learning machine
Researchers start thinking about the brain as a logical learning machine, and build the first mathematical models of artificial neurons.
- 1950
Alan Turing asks, "Can machines think?"
His paper proposes what we now call the Turing test: can a machine hold a conversation you can't tell apart from a person's?
- 1956
AI gets its name, and proves theorems
The term "artificial intelligence" is coined at a Dartmouth workshop. The same year, a program running on RAND's JOHNNIAC computer proves mathematical theorems on its own.
- 1960s
ELIZA, the chatbot "psychotherapist"
An MIT program that mimicked a therapist by reflecting your words back to you. Primitive, but people still confided in it: an early sign of how readily we connect with conversational machines.
- 1970–1990
AI takes a back seat
Promises outran results. Funding dried up and progress slowed, in stretches now called the "AI winters."
- 1997
Deep Blue beats Kasparov
IBM's chess computer defeats the reigning world champion, a milestone for machines on a clearly defined task.
- 2010s
The deep learning comeback
More data, faster chips, and better techniques bring neural networks roaring back, transforming image recognition, speech, and translation.
- 2017
The Transformer
Google researchers publish the Transformer architecture. It's the "T" in GPT, and the foundation of today's large language models.
- Nov 2022
ChatGPT launches
It reaches a million users in about five days, faster than almost any consumer product before it. Generative AI goes mainstream.
What is deep learning?
Here’s the definition I used on stage in 2023. Fittingly, I asked ChatGPT to write it:
Deep learning is a subfield of machine learning that trains artificial neural networks to make intelligent decisions by learning directly from raw data. It utilizes multiple layers of interconnected nodes to automatically extract hierarchical representations and has been instrumental in revolutionizing computer vision, natural language processing, and other AI applications.
In plain English: instead of a person writing rules (“if the email says free money, it’s spam”), you show a network of simple math “neurons” millions of examples. Each layer learns something a little more abstract than the one before it. In an image, for example, the first layers pick out edges, then shapes, then whole objects. Nobody programs those steps. The network works them out from the data.
What does GPT stand for?
GPT = Generative Pre-trained Transformer. Each word tells you something:
| Word | What it means |
|---|---|
| Generative | It creates new content (text, code, summaries) rather than just classifying or retrieving existing content. |
| Pre-trained | Before you ever type a word, it has learned patterns of language from an enormous amount of text. |
| Transformer | The 2017 neural network design that lets the model weigh how every word relates to every other word, which is how it keeps track of context. |
My take: what this means when you use it
These are the points I made in 2023, and they still hold:
- Outputs are probabilistic. GPT predicts the most likely next words. Ask the same question twice and you can get two different answers. That’s also why it can sound confident and still be wrong, so check anything that matters.
- Imagine you’re talking to a blank slate that can absorb an enormous amount of information. It only knows what it learned in training plus what you tell it, so give it context.
- Your questions need to be strategic. The quality of the answer depends on the quality of the question. A vague prompt gets a generic answer; a specific one, with your situation and goal, gets something useful.
Why history matters
In 2022, one of the most respected AI researchers in the world, Yann LeCun, said, “I don’t think we can train a machine to be intelligent, purely from text.” Within months, ChatGPT was rewriting what people thought text-trained models could do.
Even the experts get the timing wrong. It’s why I write my thinking down as I go, in Continued Thoughts on AI, so I can look back honestly at what I got right and wrong. In AI, you can’t step into the same river twice.