What actually is AI, and why does everyone mean something different?

The full story

The episode's narration, word for word. Numbers link to the sources below; timestamps play that moment.

The question 0:00

Say AI to five people, and you might get five different pictures. A chatbot that writes your emails. The app that steers you around traffic. A robot villain from a film. The filter that quietly binned your spam this morning. Are those all the same thing? What actually is AI, and why does everyone mean something different?

A proposal, 1955 0:26

The name is older than most people expect. On the thirty-first of August 1955, four researchers signed a proposal for a summer study ¹. John McCarthy was at Dartmouth College in New Hampshire ¹. With him were Marvin Minsky from Harvard, Nathaniel Rochester from IBM, and Claude Shannon from Bell Labs ¹. They asked for ten people and two months to study something they called artificial intelligence ¹.

Every aspect of learning 0:55

It rested on a bold guess: that every aspect of learning, or any other feature of intelligence, can in principle be described so precisely that a machine can be made to simulate it ¹. They wanted machines to use language, form concepts, solve problems then reserved for humans, and improve themselves ¹. They hoped for a significant advance in one summer ¹. What actually happened, in 1956, was a looser workshop over about six weeks ².

Why that name? 1:27

McCarthy later wrote that he invented the name partly to get away from an older field, cybernetics ³. Not everyone liked it. Arthur Samuel said the word artificial made the work sound phony, and Allen Newell and Herbert Simon called theirs complex information processing for years ². But the name stuck, and it's the one we still use today ².

A goal, not a gadget 1:52

Look at how the proposal defined the problem: making a machine behave in ways that would be called intelligent if a human were so behaving ¹. It doesn't say how. AI names a goal, not a technique, and there have been many routes towards it ⁴. A standard textbook calls AI the study of agents that sense their surroundings and act ⁵. And in late 2023, OECD countries agreed an updated definition for lawmakers: a machine system that infers from its input how to produce predictions, content, recommendations or decisions ⁶.

Is Fred still a person? 2:30

The first big route was to write intelligence down as rules ⁴. That works for puzzles with formal rules, but everyday common sense is hard to write down ⁴. In 1984, Douglas Lenat began a project called Cyc, to write common sense down, largely by hand ⁷. A deep-learning textbook tells how Cyc read a story about a man called Fred, shaving ⁴. It knew people don't have electrical parts. But Fred was holding an electric razor, so it decided Fred-while-shaving had electrical parts ⁴. So Cyc asked whether Fred was still a person while he was shaving ⁴. Each rule was reasonable, but together they were absurd.

A chatbot made of rules 3:16

Rules even made a chatbot. Between 1964 and 1967 at MIT, Joseph Weizenbaum built ELIZA ^8,9. It understood nothing. It spotted key words and rearranged your own sentence, so a rule in its style turns "I am worried about exams" into "How long have you been worried about exams?" ⁸. Yet people confided in it, and Weizenbaum's secretary reportedly asked him to leave the room ⁸. Reading understanding into a machine is now called the ELIZA effect ^8,10.

A robot that planned its route 3:55

Another route was search: trying possibilities in a clever order ². From 1966 to 1972, at SRI's Artificial Intelligence Center in California, a robot called Shakey planned its way around rooms ¹¹. For it, Peter Hart, Nils Nilsson and Bertram Raphael published a way of finding the shortest path in 1968, called A-star ^12,2.

A guess that saves work 4:23

Here's the idea on a little made-up map. For each place it can reach, the search adds how far it has come to a guess at how far is left, the straight-line distance ². It always carries on from the place with the lowest total ². If the guess never overestimates, the route it finishes with is guaranteed to be the shortest ². Without the guess, this search checks all nine places. With it, just four ^2,12. Driving directions grew from this kind of search, though for big road maps, methods that prepare the map first now generally beat it ^2,12.

Route two: learn from examples 5:01

The second big route turned the problem around. Instead of writing the rules, you give the program lots of examples, and it finds the patterns itself ⁴. That's machine learning, a name Arthur Samuel used in 1959 for his checkers program, which improved by playing thousands of games against itself ^13,14. How that can possibly work is our next episode.

Deep learning 5:26

Inside machine learning sits deep learning ⁴. It learns with neural networks: many layers of simple calculations, loosely inspired by brain cells ⁴. Episode three shows where that comparison breaks. In 2012, a deep network won a big image-recognition contest, with an error rate of about fifteen per cent against the runner-up's twenty-six ¹⁵. Why it took off then is episode five.

Making, not sorting 5:55

A lot of machine learning sorts and predicts: spam or not spam ⁴. Generative AI makes new things. A US government definition, quoted by the standards body NIST in 2024, describes it as imitating the patterns in its input data to produce new text, images, audio or video ¹⁶. And yes, this channel uses it. The stills in this video are made by an open image model called FLUX, on a laptop ¹⁷. The scripts are drafted with AI tools, including Claude, made by Anthropic, one of the companies in this story, and every claim is checked against the sources linked below ¹⁸.

Large language models 6:38

A large language model is a generative model trained on vast amounts of text to predict the next word ¹⁰. On the thirtieth of November 2022, OpenAI released ChatGPT, a chatbot built on one, as a free research preview ¹⁹. As of October 2026, there are many such models: OpenAI's GPT, Google's Gemini, Anthropic's Claude, Meta's Llama, Alibaba's Qwen, and models from Mistral and DeepSeek, among others ²⁰. This season explains how they work, not which one to pick.

Agents 7:18

The newest word is agent. As of 2026, it often means a language model given tools, like looking things up on the web, and set working in a loop: reason, act, check the result, go again ²¹. One early version, called ReAct, was published in October 2022 ²¹. But agent is an old word in AI: that textbook defines the whole field as the study of agents ⁵.

The nest 7:46

So here's the map. AI is the biggest box: the goal ¹. Rules and search sit inside it, but outside machine learning ⁴. Machine learning contains deep learning ⁴. Today's generative AI boom is built on deep learning, and large language models are one kind of generative AI ^22,16. But the nest breaks in places. Image models are generative, but they aren't language models ¹⁶. And an agent isn't a smaller box at all: it's a system built around a model ²¹.

The word keeps moving 8:20

The word also keeps moving. Around 1970, the computer scientist Larry Tesler made a remark usually quoted as: artificial intelligence is whatever hasn't been done yet ²³. On his own website, he called that a misquote ²³. What he actually said was: intelligence is whatever machines haven't done yet ²³. Whenever a machine can do something, many people decide it wasn't intelligence after all ²³. So what isn't AI? There's no agreed line: it depends whose definition you use, and the line keeps moving ^23,6.

Just brute force 8:59

Chess is the classic case. In May 1997, IBM's Deep Blue beat the world champion, Garry Kasparov, three and a half games to two and a half ²⁴. It could weigh up two hundred million positions a second ²⁴. Kasparov said that for the first time, he saw something similar to an artificial intellect ²⁴. Yet IBM's own history page says its strength came from brute force computing power ²⁴. Once a machine had won, many people filed chess under search, not intelligence ⁴.

AI under other names 9:35

So a lot of AI is already around you, under other names. In 2019, Google said machine learning, alongside other protections, helped block more than ninety-nine point nine per cent of spam, phishing and malware in Gmail ²⁵. In 2020, DeepMind said that, working with Google, its neural networks had cut the errors in Google Maps arrival times by up to half, in cities including Sydney ²⁶. And in 2023, Apple said iPhone autocorrect would use a transformer language model, the same broad kind of model behind chatbots ^27,22. Those are the companies' own claims.

Many don't notice 10:18

Many people don't notice ²⁸. In a seventeen-country survey run in late 2022 by the University of Queensland and KPMG, two in five people didn't know that AI was behind common apps they used ²⁸. A later survey by the University of Melbourne and KPMG, run from November 2024 to January 2025 in forty-seven countries, found sixty-six per cent of people said they used AI regularly ²⁹.

Is it intelligent? 10:48

Which leaves the biggest argument. Does any of this understand anything? In a 2022 survey of four hundred and eighty language researchers, fifty-one per cent agreed that a model trained only on text could, in principle, understand language in some non-trivial sense, and forty-nine per cent disagreed ¹⁰. Google's Blaise Agüera y Arcas wrote in 2021 that such systems, in a very real sense, understand a wide range of concepts ¹⁰. The linguists Emily Bender and Alexander Koller argued in 2020 that a system trained only on the form of language can't learn its meaning ^10,30. Next season gives that debate a whole episode.

The season ahead 11:38

This season opens up the machine. How does a machine get better by being wrong? How does a chatbot write a sentence, and why does it make things up so confidently? How does a sentence become a picture, and why does AI need so much power? But it all rests on one idea. How can a computer learn without being told the rules?

Sources

Every factual claim in the episode is tied to one of these. Spotted an error? Tell us.

1. J. McCarthy, M. L. Minsky, N. Rochester, C. E. Shannon, A Proposal for the Dartmouth… — www-formal.stanford.edu
2. Nils J. Nilsson, The Quest for Artificial Intelligence (Cambridge University Press,… — ai.stanford.edu
3. John McCarthy, review of The Question of Artificial Intelligence (B. Bloomfield, ed.) — www-formal.stanford.edu
4. Ian Goodfellow, Yoshua Bengio, Aaron Courville, Deep Learning (MIT Press, 2016), ch. 1… — deeplearningbook.org
5. Stuart Russell and Peter Norvig, Artificial Intelligence: A Modern Approach, preface — aima.cs.berkeley.edu
6. OECD.AI, Updates to the OECD's definition of an AI system explained (29 November 2023) — oecd.ai
7. Wikipedia, Cyc — en.wikipedia.org
8. Wikipedia, ELIZA — en.wikipedia.org
9. Joseph Weizenbaum, ELIZA: a computer program for the study of natural language… — dl.acm.org
10. Melanie Mitchell and David C. Krakauer, The Debate Over Understanding in AI's Large… — arxiv.org
11. SRI International, Shakey the Robot — sri.com
12. Wikipedia, A-star search algorithm — en.wikipedia.org
13. IBM, Early games (Arthur Samuel's checkers) — ibm.com
14. A. L. Samuel, Some Studies in Machine Learning Using the Game of Checkers, IBM Journal… — ieeexplore.ieee.org
15. A. Krizhevsky, I. Sutskever, G. E. Hinton, ImageNet Classification with Deep… — papers.nips.cc
16. NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative… — nvlpubs.nist.gov
17. Black Forest Labs, FLUX.1 [schnell] model card — huggingface.co
18. Anthropic, Claude — anthropic.com
19. OpenAI, ChatGPT: Optimizing Language Models for Dialogue (30 November 2022, archived copy) — web.archive.org
20. Wikipedia, List of large language models (accessed 15 October 2026) — en.wikipedia.org
21. S. Yao et al., ReAct: Synergizing Reasoning and Acting in Language Models (arXiv, 6… — arxiv.org
22. Wikipedia, Generative AI (accessed 16 October 2026) — en.wikipedia.org
23. Larry Tesler, Adages and Coinages (Tesler's Theorem, ca. 1970) — nomodes.com
24. IBM, Deep Blue (IBM history) — ibm.com
25. Google Workspace blog, Ridding Gmail of 100 million more spam messages with TensorFlow… — workspace.google.com
26. Google DeepMind, Traffic prediction with advanced Graph Neural Networks (3 September 2020) — deepmind.google
27. Apple Newsroom, iOS 17 makes iPhone more personal and intuitive (5 June 2023) — apple.com
28. N. Gillespie, S. Lockey, C. Curtis, J. Pool, A. Akbari, Trust in Artificial… — assets.kpmg.com
29. University of Melbourne and KPMG, Trust, attitudes and use of Artificial Intelligence:… — kpmg.com
30. Emily M. Bender and Alexander Koller, Climbing towards NLU: On Meaning, Form, and… — aclanthology.org

Researched and scripted with AI assistance, fact-checked claim by claim, with synthetic narration and diagrams drawn in code. How we make episodes.