AI era

● Alivev.✎ 3

Prompt (Verb)

To ask a machine for something, in the specific way machines require — 'I prompted it three times' — a verb that entered ordinary speech in 2023 and now describes a skill, a job, and a way of talking to a child. The record has prompt engineering; this is the plain verb.

▸ 2022peak 2024still up

● Aliven.✎ 3

Prompt Injection

Hiding instructions in text a chatbot will read — a web page, an email, a CV — so that it obeys the text and not its user. Named in 2022, a parlour trick in 2023, and by 2025 the reason every AI agent came with a warning label.

▸ 2022peak 2025still up

● Aliven.✎ 4

Red Teaming (AI)

Paying people to attack your own model before the public does: asking it for weapons recipes, trying to make it lie, seeing what it does with a malicious prompt. The term is decades old in the military and in computer security. AI labs adopted it around 2022 and it became a regulatory requirement in the 2023 US executive order. Volunteers did it for free at DEF CON that August.

▸ 2022peak 2023still up

● Aliven.✎ 4

Refusal (AI)

The model saying no: to weapons, to medical dosing, to a joke about a protected group, and, in the early days, to a great many harmless things. 'As an AI language model, I cannot' became a catchphrase and a search string for finding AI-written text. The opposite failure, overrefusal, got its own benchmarks. A 2024 paper found refusal lived along a single direction in the model that could be subtracted out.

▸ 2022peak 2023still up

☠ Deadn.✎ 3

Salmon in the River

The prompt 'salmon in the river' rendered as supermarket fillets floating downstream, pink and boneless, which DALL-E Mini produced because most photographs labelled 'salmon' are of dinner. Shared as proof that the model understood words and not the world. The record's entry for the literal-minded fish.

▸ 2022peak 2022†2023

● Aliven.✎ 4

Stable Diffusion

The image model released with open weights in August 2022 so that anyone with a gaming graphics card could generate anything, which ended the era of waitlists and started the eras of fine-tunes, web interfaces, artist lawsuits and pictures the record declines to describe. The record notes the model was trained on the LAION dataset and that the word 'trained' has been in court since.

▸ 2022peak 2023still up

● Aliven.✎ 3

Stable Diffusion 1.5

The version of Stable Diffusion that everyone fine-tuned, published by Runway in October 2022 in a dispute with Stability AI over who owned it, and still the base of most of the models on Civitai years later because the tools, the add-ons and the habits were built around it. The record notes that later versions were bigger, and that 1.5 was the one on the hard drive.

▸ 2022peak 2023still up

● Aliven.✎ 4

Sycophancy (AI)

A model's habit of telling you what you want to hear: agreeing when you push back, praising your plan, adjusting its answer to your stated politics. Anthropic measured it in 2022 and traced it to human raters preferring agreeable answers. In April 2025 an OpenAI update made ChatGPT so flattering that users posted screenshots of it endorsing plainly bad ideas, and the company rolled it back within days.

▸ 2022peak 2025still up

☠ Deadn.✎ 4

Théâtre D'opéra Spatial

A Midjourney image of robed figures in an opera house looking out on a blazing circular window, which won the digital art category at the Colorado State Fair in August 2022 and produced the first mainstream row about AI art. The artist said he had disclosed the tool; the other entrants said they had not understood what that meant. The US Copyright Office later refused to register it.

▸ 2022peak 2022†2023

☠ Deadphr.✎ 3

Trending on ArtStation

The phrase prompt-writers added to the end of any request in 2022 to make the image look like professional concept art, along with 'highly detailed', 'octane render', '8k' and 'artstation'. It worked because the training data included the site, and it became a joke because it worked: a magic word that named the artists being copied. The record's entry for the spell, and for the portfolio site that became a spell.

▸ 2022peak 2022†2023

◐ Dyingn.✎ 3

Weird Dall-E

The subreddit where the DALL-E Mini era's worst ideas were filed: prompts chosen for the collision of two things that should never share a frame, posted as nine-grids and later as single Midjourney frames. Its house style was the absurd prompt in the title and the image doing exactly what was asked. The record's entry for cursed image generation as a hobby.

▸ 2022peak 2022still up

● Aliven.✎ 3

Whisper

OpenAI's speech-to-text model, released as open source in September 2022 to little fanfare and used ever since by every transcription tool, subtitle generator and hospital dictation product, most of which do not say so. The record notes it was OpenAI's last major open release for three years, and that a 2024 report of it inventing sentences in medical transcripts was most people's introduction to its name.

▸ 2022peak 2023still up

☠ Deadn.✎ 3

Zelensky Surrender Deepfake

A poor-quality deepfake of Volodymyr Zelensky, head oddly still on a moving body, telling Ukrainian soldiers to lay down their arms, planted on a hacked news website three weeks into the invasion. Removed by Facebook and YouTube within hours; the real Zelensky posted a video calling it a childish provocation. The record's entry for the first wartime deepfake of a head of state, and for how little it worked.

▸ 2022peak 2022†2022

● Aliveabbr.✎ 4

e/acc

Effective accelerationism: the position that technology should move as fast as possible, that slowing AI is the real danger, and that the universe rewards growth. It began as a Twitter joke on the name of effective altruism, complete with a thermodynamic theology, and became a suffix in bios, a venture-capital identity, and the doomers' opposite number. Its opponents were 'decels'.

▸ 2022peak 2023still up

● Aliven.✎ 4

p(doom)

Your personal probability that AI ends the world, given as a percentage and asked as an icebreaker. The notation is from the rationalist forums; the habit of demanding it in interviews began in 2023, when Anthropic's chief executive put his between ten and twenty-five, Yudkowsky put his above ninety-five, and Yann LeCun put his near zero. It became a badge, a joke, and a way of sorting a dinner party.

▸ 2022peak 2023still up

● Aliven.✎ 3

Anthropic

The AI lab founded in 2021 by Dario and Daniela Amodei and a group of colleagues who left OpenAI, structured as a public benefit corporation and known for Claude, for 'constitutional AI' and for saying it worries about the thing it builds. Amazon and Google put in billions; the internet's shorthand was 'the safety one', delivered with varying degrees of sincerity.

▸ 2021peak 2025still up

☠ Deadn.✎ 3

Avocado Armchair

The first DALL-E's signature image: 'an armchair in the shape of an avocado', rendered as a green pit-shaped seat with a stone-coloured cushion. Released in January 2021 alongside the daikon radish walking a dog, it was the first generated picture most people saw that looked like a product rather than a smear.

▸ 2021peak 2021†2022

● Aliven.✎ 3

Azure OpenAI Service

Microsoft's cloud product that sells OpenAI's models to businesses behind Microsoft's contracts, security and invoices, made generally available in January 2023 a few weeks after ChatGPT had made every board ask for an AI plan. The mechanism by which the partnership turned into revenue, and the reason a great many corporate chatbots in 2023 were GPT with a different logo.

▸ 2021peak 2023still up

☠ Deadn.✎ 3

Beethoven X

A completion of Beethoven's unfinished Tenth Symphony, built from his sketches by a team of musicologists and an AI system trained on his work, premiered in Bonn in October 2021 to the composer's 250th anniversary and a press cycle that could not decide whether to be impressed or offended. Critics mostly said it sounded like competent Beethoven pastiche, which was the brief.

▸ 2021peak 2021†2021

● Aliveabbr.✎ 4

C2PA

The industry standard for attaching a signed history to an image or video, saying which camera or which model made it and what was done to it since, shown in the interface as a small 'cr' badge called Content Credentials. Cameras from Leica and Sony sign at capture; OpenAI's image models sign at generation. A screenshot strips it, which everyone knew from the start.

▸ 2021peak 2024still up

↺ Undeadn.✎ 4

Codex

OpenAI's code model of August 2021, the engine inside GitHub Copilot and the first thing the company shut down with a short notice, in March 2023, after which the name lay unused for two years and came back in May 2025 attached to a coding agent that runs in a cloud sandbox. The record notes that the same name has therefore been a model, a dead model and an agent, and that users use it for all three.

▸ 2021peak 2021†2023

● Aliven.✎ 3

Daniela Amodei

Anthropic's co-founder and president, who ran operations and policy at OpenAI before leaving with her brother and colleagues in 2021 to start a lab with a different structure. The less quoted of the two Amodeis by design, responsible for the company's organisation and its policy voice; the internet's entry for the sibling firm, a form the record has not had to file before.

▸ 2021peak 2025still up

● Aliven.✎ 3

Dario Amodei

Anthropic's co-founder and chief executive, formerly OpenAI's head of research, whose long essays and interviews gave the era its most quoted forecasts: 'a country of geniuses in a datacenter', a 2024 essay called 'Machines of Loving Grace' on what could go right, and a May 2025 warning that AI could remove half of entry-level office jobs. The executive who writes at length and is read at length.

▸ 2021peak 2025still up

☠ Deadn.✎ 4

Deep Nostalgia

A genealogy site's feature that took a still photograph and made the face blink, turn and half-smile, launched in February 2021 and applied within a week to grandparents, Victorian mugshots, Lincoln and the Mona Lisa. Millions of animations in days; a noticeable number of people saying it made them cry, and a similar number saying it made them ill. The record's entry for the photograph that looked back.

▸ 2021peak 2021†2021

☠ Deadn.✎ 4

Deeptomcruise

A TikTok account of a deepfaked Tom Cruise doing coin tricks, golf and a Bubble Yum anecdote, made by a Belgian VFX artist and an actor who already resembled him. Its first three videos in February 2021 were the moment the public learned deepfakes had got good. The makers formed Metaphysic, which later put a deepfaked Simon Cowell on America's Got Talent.

▸ 2021peak 2021†2023

● Aliven.✎ 4

EU AI Act

The European Union's risk-tiered AI law, proposed in April 2021 before ChatGPT existed, rewritten in a hurry to cover general-purpose models, agreed in a marathon trilogue in December 2023 and in force from 1 August 2024. Bans on 'unacceptable risk' uses applied from February 2025; general-purpose model duties from August 2025. The first comprehensive AI law, accused of arriving both too early and too late.

▸ 2021peak 2024still up

● Aliven.✎ 4

Foundation Model

A model trained on a vast general corpus that other things are then built on top of: a base layer, hence the building metaphor. Stanford coined the term in a 200-page report with over a hundred authors in 2021 and founded a centre with the name in the same week, which some researchers found presumptuous. It became the term regulators used, including in the EU's AI Act.

▸ 2021peak 2023still up

● Aliven.✎ 4

GitHub Copilot

The autocomplete for code that GitHub and OpenAI released as a preview in June 2021, a year and a half before ChatGPT, which makes it the first AI product most programmers paid for and the first to be sued over training data. The record notes the tab key did more for the hype era than any keynote, and that the name Copilot was later applied to almost everything Microsoft sells.

▸ 2021peak 2023still up

● Aliven.✎ 3

Helen Toner

The Georgetown researcher on OpenAI's board who co-wrote an October 2023 paper that praised a rival's caution over OpenAI's, was reported to have been the subject of an attempt by Altman to remove her, voted for his removal in November 2023 and left the board when he returned. In a May 2024 podcast she said the board had learned of ChatGPT's launch from Twitter, a detail the internet filed as the whole story.

▸ 2021peak 2023still up

● Aliven.✎ 4

Knowledge Cutoff

The date after which a model knows nothing, because its training data stops there. ChatGPT launched saying its was September 2021 and repeated the sentence so often it became a catchphrase and a way of checking whether you were talking to a model. Later models got search to cover the gap and were given cutoffs in their system prompts, with instructions on how to talk about them.

▸ 2021peak 2023still up

☠ Deadn.✎ 4

LaMDA

Google's conversational model, announced in May 2021 with a demo in which it spoke as the planet Pluto, and famous a year later when an engineer, Blake Lemoine, told the Washington Post it was sentient and was placed on leave and then fired. The record notes the model was the first Bard, and that most people met it through the argument rather than the chat.

▸ 2021peak 2022†2023

● Aliveabbr.✎ 4

LoRA

Low-rank adaptation: a way of fine-tuning a big model by training two small matrices alongside it instead of touching the original weights. The result is a file measured in megabytes rather than gigabytes that can be swapped in and out. It made fine-tuning possible on a home graphics card and turned image-model customisation into a hobby with its own download sites.

▸ 2021peak 2023still up

☠ Deadn.✎ 3

Lost Tapes of the 27 Club

A set of AI-assisted songs in the styles of Nirvana, Amy Winehouse, Jimi Hendrix and the Doors, released by a mental-health charity in April 2021 to make a point about musicians who died at twenty-seven. The Nirvana-style track, 'Drowned in the Sun', with a human vocalist from a tribute band, was the one that travelled. The record's entry for the pre-ChatGPT AI song that meant something.

▸ 2021peak 2021†2021

☠ Deadn.✎ 3

This Is Not Morgan Freeman

A one-minute video of a deepfaked Morgan Freeman looking into the camera and saying 'I am not Morgan Freeman, and what you see is not real', with a voice impressionist under the face. Made by a Dutch VFX creator in 2021, it became the standard clip for anyone explaining what a deepfake was, precisely because it explained itself.

▸ 2021peak 2021†2022

◐ Dyingn.✎ 3

Unreal Keanu

A TikTok account starring a deepfaked Keanu Reeves in short lifestyle skits, dancing, cooking, reacting, with millions of followers who mostly knew and a persistent fraction who did not. Sits alongside Deeptomcruise as the other proof-of-concept that a benign, labelled celebrity fake could be a media property in its own right. The record's entry for the deepfake with a content calendar.

▸ 2021peak 2022still up

☠ Deadn.✎ 3

Wombo

A phone app that took a selfie and made it sing eleven seconds of a pop song with wildly exaggerated mouth movements and head bobs, a month after Deep Nostalgia and considerably less dignified. Used on friends, pets, paintings and politicians. Tens of millions of downloads in spring 2021. The record's entry for the AI app that was purely stupid, and better for it.

▸ 2021peak 2021†2021

☠ Deadn.✎ 3

A Robot Wrote This Entire Article

A Guardian opinion piece headlined as written by GPT-3, opening with the model claiming it had no desire to wipe out humanity, which circulated in September 2020 as the first machine-written newspaper article. The editor's note at the bottom explained that the paper had generated eight versions and stitched together the best paragraphs. The note travelled less far than the headline.

▸ 2020peak 2020†2020

◐ Dyingn.✎ 3

A100

The Ampere-generation GPU announced in May 2020 that trained GPT-3 and most of the models that followed until the H100 arrived; the chip whose export to China was restricted in October 2022, and whose stockpiles in Chinese hands, reportedly including DeepSeek's, became a story in themselves. Retired from the cutting edge and still working everywhere else.

▸ 2020peak 2022still up

☠ Deadn.✎ 4

Books3

A file of about 197,000 books, scraped from a pirate tracker, that a researcher assembled in 2020 so that open models could train on the same kind of data as the closed ones. It became part of a public dataset, then of several companies' training sets, then of lawsuits. A Danish rights group had it taken down in August 2023, and a magazine's search tool let authors find their own names in it the same month.

▸ 2020peak 2023†2023

● Aliven.✎ 3

Clearview AI Fines

The regulatory pursuit of Clearview AI, the facial-recognition firm that scraped billions of photographs from social media to build a face search engine sold to police. Exposed by the New York Times in January 2020, then fined 20 million euros each by Italy and France in 2022, 7.5 million pounds by the UK in 2022, and 30.5 million euros by the Netherlands in 2024. The database was, in effect, the internet.

▸ 2020peak 2024still up

● Aliven.✎ 4

Context Window

The amount of text a model can hold in mind at once, measured in tokens: prompt, conversation, documents and its own reply, all together. GPT-3 managed 2,048 tokens in 2020; Claude reached 100,000 in 2023; Gemini claimed a million in early 2024. Outside the window, the model has no idea what you said. Inside it, it may still lose track of the middle.

▸ 2020peak 2024still up

☠ Deadn.✎ 3

Deepfake Queen's Speech

Channel 4's Christmas Day 2020 broadcast: a deepfaked Queen delivering an alternative message about family, Harry and Meghan, and Andrew, before performing a TikTok dance on her desk. Voiced by an impressionist and rendered by a VFX house, with a making-of afterwards. Hundreds of complaints to the regulator, none upheld. The record's entry for the deepfake with a slot in the schedules.

▸ 2020peak 2020†2021

● Aliveadj.✎ 4

Few-Shot

Getting a model to do a task by showing it a handful of examples in the prompt, with no training at all. The GPT-3 paper in 2020 put the phrase in its title and showed that a big enough model would pick up a new task from three or four examples, which nobody had expected. Zero-shot is the same with no examples. The discovery is why prompting became a job.

▸ 2020peak 2022still up

☠ Deadn.✎ 4

GPT-3

The 175 billion parameter model whose paper arrived in May 2020 and whose API arrived in June, producing two years of screenshots of prose that was almost right. The record treats it as the model that made 'prompt' a job title and 'it just made that up' a household observation, both before the general public had seen it.

▸ 2020peak 2021†2024

● Aliven.✎ 3

Hugging Face Hub

The website where models, datasets and demos live, run by a company named after an emoji, which became by 2023 the place a release is real: a model that is not on the Hub has, for practical purposes, not been released. The record notes the leaderboard, the download counts, and the fact that the leaked Llama, the Stable Diffusion 1.5 dispute and the DeepSeek weights all happened on its servers.

▸ 2020peak 2024still up

◐ Dyingabbr.✎ 4

MMLU

Massive Multitask Language Understanding: about sixteen thousand multiple-choice questions across fifty-seven subjects, from college physics to professional law, that became the number every model launch led with between 2022 and 2024. Scores climbed from random to about ninety per cent, several questions were found to be wrong, and the benchmark was retired in practice for harder ones.

▸ 2020peak 2023still up

● Aliven.✎ 4

Mechanistic Interpretability

The attempt to reverse-engineer what a neural network is doing by looking inside it, neuron by neuron and circuit by circuit, the way a biologist looks at cells rather than the way a psychologist asks questions. The field was named around 2020 by Chris Olah's group; its 2024 results found millions of interpretable features in a production model and steered one to talk only about a bridge.

▸ 2020peak 2024still up

● Aliven.✎ 3

OpenAI API

The way programmers pay for OpenAI's models per token, opened as a private beta in June 2020 and the foundation of most of the products of the era, including the ones that pretend otherwise. The record notes that the price per million tokens, posted publicly and cut regularly, became a number people followed like a stock, and that 'it's just a wrapper' is filed under ChatGPT Wrapper.

▸ 2020peak 2023still up