Wacky Facts

Joseph Weizenbaum’s ELIZA (1966) mostly turned users’ sentences back into questions. Even so, some users became emotionally attached — Weizenbaum wrote that his secretary asked him to leave the room so she could talk to it privately. Treating a program as more understanding than it is became known as the “ELIZA effect”.

The 1986 Nature paper by Rumelhart, Hinton and Williams, “Learning representations by back-propagating errors”, popularised the method used to train almost every neural network today. The core idea had appeared in earlier work too.

IBM’s Deep Blue beat Garry Kasparov 3½–2½ in their 1997 rematch. It evaluated around 200 million positions per second using special chess chips — raw search, not learning.

In game 2 against Lee Sedol in 2016, AlphaGo’s “move 37” was so unusual that commentators were puzzled; AlphaGo itself estimated a human would play it about 1 time in 10,000. It proved decisive.

Launched on 30 November 2022, ChatGPT was estimated by analysts to have hit 100 million monthly users by January 2023 — at the time described as the fastest-growing consumer app ever.

Arthur Samuel’s checkers program at IBM learned from experience, including games against itself. Samuel popularised the term “machine learning” in 1959.

Shakey (SRI, 1966–1972) was the first mobile robot that could reason about its own actions. Its researchers developed the A* search algorithm, still used today in games, maps and robots.

DeepMind’s AlphaFold 2 made a leap in predicting how proteins fold, and its database grew to over 200 million predicted structures. Its creators shared the 2024 Nobel Prize in Chemistry.

GPT means Generative Pre-trained Transformer. The transformer architecture was introduced by eight Google researchers in the 2017 paper “Attention Is All You Need”, now one of the most cited papers in computing.

In his 1950 paper “Computing Machinery and Intelligence”, Turing replaced the vague question “Can machines think?” with a game: can a machine’s typed answers be told apart from a person’s?

Alan Turing asked whether a machine could imitate a human in conversation well enough to fool a judge.

ImageNet contains over 14 million images hand-annotated with what they show, labelled largely through Amazon Mechanical Turk. The yearly ImageNet challenge became the benchmark that showed deep learning’s power.

After a US Navy press event about Frank Rosenblatt’s perceptron — a simple early neural network — The New York Times reported the Navy expected it to walk, talk, see, write, reproduce itself and be conscious of its existence. It could learn to tell simple shapes apart.

Moravec’s paradox: high-level reasoning like chess needs relatively little computation, while walking, grasping and recognising faces — which children do effortlessly — are extremely hard for machines.

Released in 2020, GPT-3’s 175 billion adjustable numbers made it more than 100 times larger than GPT-2. Each parameter is just a number nudged during training.

The Mechanical Turk toured Europe from 1770 and beat players including, reportedly, Napoleon. A skilled human chess player was hidden inside the cabinet. Amazon’s crowd-work service is named after it.

Warren McCulloch and Walter Pitts described a mathematical model of a neuron in 1943.

John McCarthy and colleagues used the term in their 1955 proposal for a 1956 summer research project at Dartmouth College. The organisers hoped significant progress could be made in one summer. It took rather longer.

Language models read tokens — chunks of text — rather than individual letters, so questions about spelling and letter counts are surprisingly hard for them. The strawberry question became a famous meme in 2024.

AlexNet won the 2012 ImageNet challenge by a large margin. It was trained on two NVIDIA GTX 580 GPUs — consumer gaming cards — helping start the move of AI onto graphics chips.

After early hype, funding and interest collapsed in the mid-1970s and again in the late 1980s, periods now called AI winters. Promises had run far ahead of what computers could actually do.

In the early 1900s Hans appeared to do arithmetic by tapping his hoof. He was actually reading tiny, unconscious cues from people watching. When an AI model gets the right answers for the wrong reasons, researchers call it a “Clever Hans” effect.

Researchers showed that carefully chosen tiny changes — in some experiments just one pixel — can make an image classifier confidently mislabel a picture, for example calling a ship a car. These are called adversarial examples.

Joseph Weizenbaum’s program at MIT mimicked a therapist using simple pattern matching.