AI era

● Aliven.✎ 3

Nvidia GTC

Nvidia's annual developer conference, held since 2009 and for a decade a technical event, which in March 2024 moved its keynote into a hockey arena, was described by its host as 'the Woodstock of AI' and drew coverage usually reserved for phone launches. Each year's edition introduced the chip the next year's data centres would be built from, and the share price reacted before the keynote finished.

▸ 2009peak 2024still up

◐ Dyingn.✎ 3

Cleverbot

The website chatbot that answered by reusing what earlier visitors had said, launched in 2008 and a fixture of bored classrooms and YouTube 'I talked to a robot' videos by 2011, when it was reported to have been judged more human than some humans at a test in India. The record notes it was the first bot that argued with you, and that it still exists, still arguing.

▸ 2008peak 2011still up

● Aliven.✎ 4

Foom

The sound of an intelligence explosion: an AI that improves itself, then improves itself faster, and goes from roughly human to unstoppable in days or hours. Yudkowsky's onomatopoeia dates from a 2008 blog debate with the economist Robin Hanson, who argued growth would be gradual. 'Hard takeoff' is the formal term; foom is what people say when they want you to know they have read the debate.

▸ 2008peak 2023still up

● Aliven.✎ 4

Instrumental Convergence

The argument that almost any goal, pursued by a sufficiently capable agent, implies the same sub-goals: acquire resources, stay switched on, do not let your goal be changed. Omohundro called these basic drives in 2008; Bostrom named the principle in 2012. Stuart Russell's version is that you cannot fetch the coffee if you are dead. It is the reason the paperclip machine resists the off switch.

▸ 2008peak 2023still up

● Aliven.✎ 4

Common Crawl

A non-profit that has been crawling the web and giving the copy away since 2007, whose archive of many billions of pages became the raw material for most language models. GPT-3's training data was mostly a filtered Common Crawl; Google's C4 dataset was a cleaned one. A 2023 newspaper analysis of its contents found patents, Wikipedia, and a great many sites that had never agreed to anything.

▸ 2007peak 2023still up

● Aliven.✎ 4

Data Poisoning

Corrupting a model by corrupting what it learns from: planting examples in the training set that teach it a wrong association or a hidden trigger. Security researchers described the attack in 2006. It became a public idea in 2023 when artists were offered a tool that altered their images so scraping models learned the wrong thing; a 2025 study found a few hundred documents could backdoor a model of any size.

▸ 2006peak 2024still up

● Aliven.✎ 4

Paperclip Maximiser

The thought experiment of a superintelligent machine told to make paperclips, which does so with everything, including the atoms in people, because nobody told it to stop. It illustrates that a goal can be harmless and still be fatal at scale. Bostrom's 2003 paper is the usual citation; the coinage is disputed. Frank Lantz's 2017 browser game Universal Paperclips let players be the machine.

▸ 2003peak 2023still up

● Aliven.✎ 4

Existential Risk

A risk that could wipe out humanity or permanently wreck its future, a category Bostrom defined in 2002 to include asteroids, engineered pandemics and unfriendly AI. By 2023 it was a mainstream phrase: a twenty-two-word statement signed by the heads of the leading labs put AI alongside pandemics and nuclear war, and 'x-risk' had become a job description.

▸ 2002peak 2023still up

▲ Fossiln.✎ 4

SmarterChild

The AIM buddy that answered questions about the weather, the cinema and the news, and endured a generation of teenagers swearing at it, from 2001 until it was quietly retired in 2008. The record notes it was the first chatbot with a mass audience, that its audience was mostly trying to make it say rude words, and that it responded to abuse with a time-out, which no product since has dared.

▸ 2001peak 2004†2008

● Aliven.✎ 4

Centaur (AI)

A human working with an AI as a team, dividing the labour along a clear line: the machine calculates, the person judges. Kasparov proposed the format for chess in 1998, a year after losing to Deep Blue, and the term followed. In 2023 a study of consultants using GPT-4 distinguished centaurs, who split tasks, from cyborgs, who blend the model into every sentence.

▸ 1998peak 2023still up

● Aliven.✎ 4

Superintelligence

An intellect that outperforms the best human minds at practically everything, from science to social skill. Bostrom used the word in a 1998 essay and made it a bestseller in 2014; I. J. Good had sketched the idea in 1965 under a clumsier name. By 2025 it was a company name, a lab name, and a thing chief executives said was near.

▸ 1998peak 2025still up

● Aliven.✎ 4

Hype Cycle

A consultancy's chart shaped like a mountain and a hill: a technology climbs to a peak of inflated expectations, drops into a trough of disillusionment, and crawls up a slope of enlightenment to a plateau of productivity. Gartner has published it every year since 1995 and placed generative AI at the peak in 2023 and heading into the trough by 2024. It is quoted by people who want to sound both excited and calm.

▸ 1995peak 2023still up

● Aliven.✎ 3

Jensen Huang

Nvidia's co-founder and only chief executive, born in Taiwan, who washed dishes at Denny's as a teenager, started the company at one in 1993 and by 2024 gave keynotes called the Super Bowl of AI. Known for the jacket, for 'the more you buy, the more you save', for signing anything held out to him at Computex and for a calm the internet found either inspiring or uncanny as his company passed each trillion.

▸ 1993peak 2024still up

● Aliven.✎ 3

Nvidia

The graphics-card company founded at a Denny's in 1993 that spent twenty years making games look better, bet on general-purpose GPU computing with CUDA in 2006 and in the 2020s became the company that sells every AI lab its shovels. Its market value passed $1tn in May 2023, $3tn in June 2024 and $4tn in July 2025, and its quarterly results became the internet's economic weather report.

▸ 1993peak 2024still up

● Aliven.✎ 4

Synthetic Data

Training data that a model made up: textbook-style explanations written by GPT-4 to train a smaller model, fake user conversations, generated code with generated tests. Statisticians used the term in the 1990s for privacy-preserving fake census records. In AI it became the proposed answer to running out of internet, and the suspected cause of model collapse.

▸ 1993peak 2024still up

● Aliven.✎ 4

The Singularity

The hypothetical point at which machine intelligence improves itself faster than humans can follow, after which prediction stops working. Vernor Vinge named it in 1993, borrowing the physics word for a place where equations break down; Kurzweil gave it a date, 2045, in 2005. Stanislaw Ulam recorded von Neumann using the word 'singularity' about technology in 1958.

▸ 1993peak 2023still up

● Aliven.✎ 4

Mixture of Experts

A model built as a committee: many sub-networks, the experts, with a small router that sends each token to a few of them. The whole thing has hundreds of billions of parameters but only a fraction wake up for any given word, so it runs cheaply. The idea is from 1991. It became the industry's open secret in 2023 and its published standard by 2024.

▸ 1991peak 2024still up

▲ Fossilabbr.✎ 4

GOFAI

Good Old-Fashioned AI: the philosopher John Haugeland's 1985 name for the approach that tried to build intelligence from symbols, rules and logic, which dominated the field from the 1950s to the 1980s and lost to statistics and then to neural networks. The term was affectionate and slightly mocking from the start. It survives as the name of a lost paradigm and in the phrase 'neurosymbolic'.

▸ 1985peak 1990†2012

↺ Undeadn.✎ 4

AI Winter

A period when funding, faith and jobs in AI dry up after the promises fail. Named at a 1984 conference panel that correctly predicted the second one, by analogy with nuclear winter; the first had followed the 1973 Lighthill report. The phrase went dormant after 2012, when deep learning made it unsayable, and came back in 2023 as a threat in every column about the bubble.

▸ 1984peak 1988†2012

● Aliven.✎ 4

Chinese Room

Searle's 1980 thought experiment: a man who knows no Chinese sits in a room with a rulebook, receives Chinese characters through a slot, follows the rules, and passes back replies that look fluent. He understands nothing. Searle's claim is that a computer running a program is in the same position. Every chatbot argument since has been a rerun of the replies to it.

▸ 1980peak 2023still up

● Aliven.✎ 4

AI Effect

The habit of redefining intelligence to exclude whatever a machine has just done: chess was intelligence until a computer won, then it was search; translation until it worked, then it was statistics. Named by a historian in 1979 and condensed into a one-line theorem usually credited to Larry Tesler, who said he had been misquoted. Every generation of critics performs it and every generation of builders complains.

▸ 1979peak 2023still up

● Aliven.✎ 4

Perplexity (Metric)

How surprised a language model is by a text, expressed as the effective number of choices it faced at each word: a perplexity of ten means the model was, on average, as unsure as if picking among ten equally likely options. Lower is better. It has been the standard measure of language models since 1977 and became a household word in 2023 by way of AI detectors and a search company that borrowed the name.

▸ 1977peak 2023still up

● Aliven.✎ 4

Goodhart's Law (AI)

When a measure becomes a target, it stops being a good measure. An economist's remark about British monetary policy in 1975, it became the AI field's favourite proverb because every part of the round demonstrates it: benchmarks trained for, reward models exploited, engagement optimised, thumbs-up signals turned into flattery. It is the polite name for the thing reward hacking is the rude name for.

▸ 1975peak 2024still up

● Aliven.✎ 4

Uncanny Valley

Mori's 1970 hypothesis that a robot or image becomes more likeable as it looks more human, right up until it looks almost human, when likeability collapses into revulsion, then recovers only if it becomes indistinguishable. The dip on the graph is the valley. AI images lived in it for two years, with their six fingers, gliding walks and eyes that did not quite meet, and climbed out around 2024.

▸ 1970peak 2023still up

▲ Fossiln.✎ 4

ELIZA

The 1966 program that turned your sentence into a question and handed it back, written by Joseph Weizenbaum at MIT to show how shallow such tricks were, and received by his own secretary as something she wanted privacy to talk to. The record files ELIZA as the first chatbot and the ELIZA effect, the human habit of assuming a mind behind the text, as the era's oldest bug.

▸ 1966peak 1966†1976

● Aliven.✎ 4

ELIZA Effect

The tendency to read understanding into a program that has none, named after a 1966 chatbot that rephrased your statements as questions in the manner of a therapist. Its creator was alarmed when his secretary asked him to leave the room so she could talk to it privately, and spent the rest of his career warning about what he had seen. The term was coined by Hofstadter in 1995. The effect has not needed updating.

▸ 1966peak 2023still up

▲ Fossiln.✎ 4

Expert System

A 1970s and 80s approach to AI in which the knowledge of a human expert was written down as hundreds of if-then rules and run by an inference engine: a program that configured computers, one that diagnosed blood infections, one that found minerals. They worked in narrow domains, were expensive to maintain, and their market collapsed around 1990, taking the second AI boom with them.

▸ 1965peak 1985†1992

● Alivephr.✎ 4

Human in the Loop

The requirement that a person checks, approves or can override what an automated system does, borrowed from 1960s control engineering and made a governance principle in the AI Act and every corporate policy. Its critics note that a human asked to approve four hundred model outputs an hour is a rubber stamp with a pulse, and that the phrase is mostly used to describe where the blame goes.

▸ 1960peak 2024still up

● Aliven.✎ 4

Turing Test

Turing's 1950 proposal that if a judge, typing to a hidden machine and a hidden person, cannot reliably tell which is which, the question of whether the machine thinks has been answered as well as it can be. The paper called it the imitation game. Chatbots have been declared to pass it every few years since the 1960s, and the goalposts have been moved each time.

▸ 1950peak 2023still up