As an Amazon Associate, we earn from qualifying purchases. This post may contain affiliate links.
By Luis Gustavo. See how we research and review our content.
In the summer of 1956, a small group of researchers gathered at Dartmouth College in Hanover, New Hampshire, to test a bold conjecture: that every feature of learning or intelligence could be described precisely enough for a machine to simulate it. Almost everything people now call AI traces back to arguments that started there.
That makes the field older than most of the people using it, and its history includes a therapist-imitating script from the 1960s, a board game with more positions than atoms, and speech recognizers that once understood sixteen words. Five facts follow, then two closer looks.
1. The term “artificial intelligence” is over 70 years old
The phrase was coined in 1956, during an academic conference in the United States. More precisely, it first appeared in a funding proposal dated September 2, 1955, written by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon, and the summer workshop at Dartmouth followed. McCarthy picked the name partly because it was neutral: it avoided the narrow focus of automata theory and steered clear of cybernetics.
The event was far less polished than its reputation. The proposal imagined a “2-month, 10-man study,” attendance was uneven, and only a few people stayed for the whole stretch, which was probably closer to eight weeks than six. Even so, the workshop is now widely treated as the founding moment of the field.
2. The first “chatbot” in history dates back to 1966
ELIZA was written by Joseph Weizenbaum at MIT and released in 1966, running on an IBM 7094. Its best-known script, called DOCTOR, imitated a Rogerian therapist, the kind who mostly reflects a patient’s words back as questions. That style suited a program with no knowledge of the world, because a non-directive question needs no facts behind it.
Under the hood there was no understanding at all. ELIZA scanned the input for keywords, each with a rank chosen by the programmer, then split the sentence by a rule and rebuilt it, so “You are very helpful” could come back as “What makes you think I am very helpful?” When no keyword matched, it fell back on neutral replies such as “Please go on.”
3. AI has already beaten human champions at complex strategy games
Chess fell first. In February 1996 IBM’s Deep Blue won a single game against world champion Garry Kasparov in Philadelphia but lost the six-game match 4-2. The May 1997 rematch in New York ended 3.5-2.5 in the machine’s favor, making it the first computer to beat a reigning champion under standard tournament time controls. IBM says it could evaluate about 200 million positions per second.
Go was a different problem. Its board allows roughly 10170 configurations, more than the number of atoms in the known universe, so brute-force search was hopeless. In March 2016 DeepMind’s AlphaGo beat Lee Sedol 4-1 in Seoul after learning from thousands of expert games and then from thousands of games against versions of itself. Its move 37 in game two was one DeepMind says a human would play about once in 10,000 times. More in these video game facts.
4. Much of modern AI learns through mathematical “trial and error”
A neural network is a large set of adjustable numbers, called parameters, arranged in layers loosely inspired by connections between biological neurons. Instead of being handed rules, the system is shown examples, makes a prediction, measures how wrong it was and nudges its parameters to do slightly better. Repeating that loop over enormous numbers of examples is the “trial and error.”
The approach has several variants. AlphaGo’s second stage, playing against itself and learning from its mistakes, is a form of reinforcement learning. The same idea, with human feedback as the reward signal, was used by OpenAI to turn its language models into the assistant behind ChatGPT, released in November 2022.
5. “Voice assistants” existed before modern AI as we know it
Bell Labs’ Audrey, built in 1952, could recognize spoken digits from a single speaker by picking out the resonant frequencies in each sound. In 1962 IBM showed off Shoebox, a machine that recognized 16 spoken words and appeared at the World’s Fair that year.
Progress came in steps. In 1971 DARPA funded a five-year project aiming at a vocabulary of at least 1,000 words. During the 1980s statistical hidden Markov models replaced older matching methods, and IBM’s Tangora handled a 20,000-word vocabulary. Dragon Dictate, a consumer product, arrived in 1990. Casual conversation had to wait for later advances.
The ELIZA effect, or why a simple script felt like it understood
Weizenbaum was surprised by how people reacted. Users attributed understanding and feelings to a program that had neither, a tendency now called the ELIZA effect. His own secretary reportedly asked him to leave the room so she could talk to it in private. He was disturbed that brief exposure could produce what he called powerful delusional thinking in ordinary people.
That unease led him to write Computer Power and Human Reason in 1976, a book arguing that computers have limits. The lesson carries over to far more capable systems today: fluent, confident text is not by itself evidence that anything is understood, which is a reason to treat chatbot answers as drafts to check, not verdicts.
From 1.5 billion to 175 billion parameters in a year
Scale is the clearest way to see how fast the field moved. OpenAI introduced GPT-1 in 2018. GPT-2, released in 2019, had 1.5 billion parameters, and GPT-3 followed in 2020 with 175 billion, more than a hundredfold jump in a single year. ChatGPT appeared in November 2022 and GPT-4 in March 2023.
Researchers describe “scaling laws” that relate a model’s performance to its size, the amount of training data and the computing power used. Bigger is not automatically better, and the same body of work flags concerns such as bias, misinformation and the environmental cost of training. For a very different branch of technology, see 3D printing facts.
Since you made it this far, you’ll probably also like learning more about facts about sound technology and about light-activated drug restores lost vision.
Want a voice assistant at home?
Echo Dot with Amazon Basics Smart Color Bulb — voice control, music, and smart home automation in one device.
Disclosure: this is our tracked Amazon affiliate link.
Weizenbaum’s secretary asking for privacy with a script built from keyword rules is still the best reminder to ask what a polished answer is really made of.