AI era

● Aliveabbr.✎ 4

RAG

Retrieval-augmented generation: look things up first, then let the model answer. Documents are turned into embeddings, the relevant chunks are fetched for each question and pasted into the prompt, and the model writes with them in front of it. It became the default way to give a chatbot your company's files, and the lead author has said he would have chosen a nicer name.

▸ 2020peak 2024still up

☠ Deadn.✎ 3

Sassy Justice

A fourteen-minute local-news parody starring a deepfaked Donald Trump as a Cheyenne consumer reporter named Fred Sassy, plus deepfaked Al Gore, Mark Zuckerberg and Julie Andrews, made by the South Park creators and released a week before the 2020 election. It was the first deepfake comedy with a budget, and the record's evidence that the technology could be funny on purpose.

▸ 2020peak 2020†2020

● Aliven.✎ 4

Scaling Laws

The finding that a language model's error falls in a smooth, predictable line as you add parameters, data and compute, across many orders of magnitude, with no wall in sight. The 2020 OpenAI paper turned this into a plan: buy more, get better. DeepMind's 2022 revision said the models were undertrained and should eat about twenty tokens per parameter. The line has been debated ever since.

▸ 2020peak 2024still up

◐ Dyingn.✎ 3

Stability AI

The London company that funded and released Stable Diffusion in August 2022 as open weights, put image generation on every gaming PC in the world and then spent two years in reported disputes about money, credit and its founder, who resigned in March 2024. It survived under new management, which the record counts as a result.

▸ 2020peak 2022still up

● Aliven.✎ 2

Telegram Deepfake Bots

The network of Telegram bots, first documented by a research firm in October 2020, that took a photograph of a woman and returned a synthetic nude, having by then processed images of more than 100,000 people, many of them apparently underage. By 2024 investigations counted dozens of such bots with millions of users. The origin of the nudify economy, in a chat app, four years before the crackdown.

▸ 2020peak 2024still up

● Aliven.✎ 3

Thomson Reuters v Ross

The first US ruling to reject fair use for AI training. Thomson Reuters sued the legal research startup Ross Intelligence in 2020 for using Westlaw headnotes to build a search tool; on 11 February 2025 Judge Stephanos Bibas, reversing his own earlier view, held the copying was not fair use. Ross had shut down in 2020 and the tool was not generative, and the judge said so, but the headline did not.

▸ 2020peak 2025still up

● Aliven.✎ 3

Timnit Gebru

The researcher who co-led Google's Ethical AI team and left in December 2020 after a dispute over a paper, 'On the Dangers of Stochastic Parrots', that questioned the costs of large language models; Google said she resigned, she said she was fired, and the argument was public within hours. She founded the DAIR institute in 2021, and the paper's title became a phrase the internet used for years.

▸ 2020peak 2021still up

● Aliven.✎ 3

AI Dungeon

The text adventure with no script, in which a language model plays dungeon master and lets you do anything, released on GPT-2 in December 2019 and upgraded to GPT-3 in 2020, which made it the first place the public met a large model without knowing what one was. The record notes the 2021 dispute in which a new content filter read players' private stories, and that this was the era's first AI moderation fight.

▸ 2019peak 2020still up

● Aliven.✎ 4

ARC-AGI

A set of little grid puzzles, colour the cells according to a rule shown in a few examples, that humans find easy and models found nearly impossible for five years. Chollet built it in 2019 to measure skill acquisition rather than memorised skill. A million-dollar prize in 2024 drew attention; OpenAI's o3 scored high on it that December at great expense; a harder version followed in March 2025.

▸ 2019peak 2024still up

● Aliven.✎ 4

Bitter Lesson

Rich Sutton's 2019 observation, in a page and a half, that seventy years of AI research show the same thing: methods that use more computation, search and learning, always beat methods that build in human knowledge, and researchers resist this every time because it feels like cheating. It became the scaling era's founding text, quoted by people who had read it and cited as a law by people who had not.

▸ 2019peak 2024still up

● Aliven.✎ 3

Cohere

The Toronto lab founded in 2019 by Aidan Gomez, one of the eight authors of the Transformer paper, that chose to sell language models to businesses rather than to the public and has spent the hype era being described as 'the enterprise one'. The record's entry for the lab that never had a viral moment and did not appear to want one.

▸ 2019peak 2024still up

☠ Deadn.✎ 3

Dalí Lives

An installation at the Dalí Museum in Florida in which a life-size deepfaked Salvador Dalí, trained on archive footage and voiced by an actor, greets visitors on a screen, comments on the weather and takes a selfie with them. Opened in 2019, it was the first deepfake a museum had put its name to, and the first many older visitors had seen. The record's entry for the resurrected artist with a gift shop.

▸ 2019peak 2019†2020

☠ Deadn.✎ 4

FaceApp Challenge

A week in July 2019 when every celebrity and most of your contacts posted a photograph of themselves aged by forty years, using a neural-network filter in an app that had been around since 2017. The wrinkles were convincing, the grey hair was uniform, and the discovery that the developer was in St Petersburg produced a privacy scare that briefly reached the US Senate.

▸ 2019peak 2019†2019

▲ Fossiln.✎ 4

GPT-2

The language model OpenAI announced in February 2019 and declined to release in full, calling it too risky, which made it the first AI model to be famous for what it was not doing. The record notes that the withheld version was published in November 2019 anyway and that the sky stayed where it was.

▸ 2019peak 2019†2020

☠ Deadn.✎ 3

In Event of Moon Disaster

A seven-minute film of a deepfaked Richard Nixon delivering the speech that had been written for him in case the Apollo 11 astronauts died on the moon, produced by an MIT team with a voice actor and two synthetic-media companies. Released online in July 2020, it was the most careful deepfake ever made and was built specifically to be a teaching tool. It also won an Emmy.

▸ 2019peak 2020†2021

☠ Deadn.✎ 3

Jim Carrey in The Shining

The deepfake that made the format look like a craft: Jim Carrey's face on Jack Nicholson through the axe scene of The Shining, so convincing that the eyebrows carried the performance. Posted in July 2019 by an anonymous YouTube channel that specialised in movie recasting, and followed by Bill Hader morphing into Tom Cruise mid-interview. The record's entry for deepfakes as film criticism.

▸ 2019peak 2019†2020

● Aliven.✎ 3

Microsoft–OpenAI Partnership

The deal that began with Microsoft's $1bn investment in July 2019 and grew, with a reported $10bn in January 2023, into the era's defining alliance: OpenAI got the compute, Microsoft got the models, the licence and a share of the profits, and both got a relationship the internet described as the most valuable and most awkward in technology. Renegotiated in October 2025 to give Microsoft a reported 27 per cent stake.

▸ 2019peak 2023still up

● Aliven.✎ 3

Sam Altman

OpenAI's chief executive since 2019, the former head of Y Combinator, who testified to Congress in May 2023 asking to be regulated, was fired by his board in November 2023 and reinstated within five days, and spent the years after announcing sums in the hundreds of billions with a level voice. The internet's central character of the AI era, in the public role only: the tweets, the testimony and the badge photograph.

▸ 2019peak 2023still up

☠ Deadn.✎ 3

Talk to Transformer

The website where most people first typed a sentence and watched a model continue it: a plain text box over GPT-2, built by an independent developer in 2019 after OpenAI released the model in stages. Its completions were half-coherent, often wandering into invented news stories, and were screenshotted as the first proof that machines could write. The record's entry for ChatGPT's small ancestor.

▸ 2019peak 2019†2021

● Aliven.✎ 4

Temperature (AI)

The dial that controls how adventurous a model's next word is. At zero it always picks the likeliest token and repeats itself; higher, it samples from the tail and gets creative, then incoherent. The name comes from physics, where hotter systems are more disordered. It is the first setting a new API user touches and the first one they set back.

▸ 2019peak 2023still up

◐ Dyingn.✎ 3

This Person Does Not Exist

A website that shows a single photorealistic face on a white background and a new one each time you reload, every one of them generated by Nvidia's StyleGAN and none belonging to a person. Launched in February 2019 by a software engineer to make a point about how good the models had got. It made the point to tens of millions of people, and its faces became the default profile picture for bot accounts.

▸ 2019peak 2019still up

☠ Deadn.✎ 3

This X Does Not Exist

The format that followed the face site within weeks: This Cat Does Not Exist, This Rental Does Not Exist, This Startup Does Not Exist, each a single generated example per refresh, each a demonstration of what a GAN trained on one kind of thing would dream. The cats were bad, the flats were uncanny, and the startup one could pass for real. The record's entry for the 2019 template.

▸ 2019peak 2019†2021

● Aliven.✎ 4

Token (AI)

The unit a language model actually reads and writes: not a word, not a letter, but a chunk of text chosen by a compression algorithm, so 'cat' is one token and 'unbelievable' might be three. English averages about three-quarters of a word per token. Models are priced, limited and benchmarked by the token, and their strangest failures, counting letters for instance, come from it.

▸ 2019peak 2024still up

● Aliven.✎ 4

Top-p

The other sampling dial: instead of considering every possible next token, the model keeps only the smallest set whose probabilities add up to p, say 0.9, and picks from those. The 2019 paper that introduced it called the method nucleus sampling and showed that plain greedy decoding made models loop. Most people leave it at the default and adjust temperature instead.

▸ 2019peak 2023still up

☠ Deadn.✎ 3

ZAO

A Chinese app that put the user's face into famous film and television scenes from a single selfie, and went to the top of the App Store there within days in August 2019. Its terms of service claimed rights to users' faces in perpetuity, which was noticed within the same week. WeChat blocked its links and the terms were rewritten. The record's entry for the first deepfake app with a queue.

▸ 2019peak 2019†2019

☠ Deadn.✎ 3

Zuckerberg Spectre Deepfake

A short deepfake of Mark Zuckerberg sitting at a desk and boasting about controlling 'billions of people's stolen data', posted to Instagram in June 2019 by two artists to test whether Facebook would remove a fake of its own chief executive. It did not. Weeks earlier the company had refused to remove a slowed-down clip of Nancy Pelosi, and the artists wanted the policy on record. It was.

▸ 2019peak 2019†2019

● Aliven.✎ 3

AlphaFold

DeepMind's protein-structure model, which won the biennial CASP competition in 2018 and effectively ended it in 2020, and whose predictions for nearly every known protein were published free in July 2021. The record notes that it is the AI result that people who dislike AI cite as the good one, and that its makers received the 2024 Nobel Prize in Chemistry.

▸ 2018peak 2021still up

● Aliven.✎ 4

Compute

Computing power, as a mass noun and as the scarce resource of the round: chips, the electricity to run them, and the buildings to hold them, measured in floating-point operations and priced in billions. A 2018 OpenAI analysis found training compute doubling every few months, and the word went from engineering slang to the thing governments write thresholds about and companies raise half a trillion dollars for.

▸ 2018peak 2025still up

☠ Deadn.✎ 3

Edmond de Belamy

A blurred portrait of a fictional gentleman in a dark coat, generated by a GAN, printed on canvas, signed with a fragment of the loss function, and sold at Christie's in October 2018 for $432,500 against an estimate around $10,000. The first AI artwork through a major auction house. A separate programmer pointed out that much of the code was his, which was the second story.

▸ 2018peak 2018†2019

● Aliven.✎ 4

Fine-Tuning

Taking a model that has already been trained on everything and training it a little more on your thing: your support tickets, your house style, your legal forms. The base model is expensive and general; the fine-tune is cheap and specific. It became the standard second step in 2018 and, for most companies, was later replaced by writing a good prompt.

▸ 2018peak 2023still up

◐ Dyingn.✎ 3

Google Duplex

The Google demo of May 2018 in which the Assistant telephoned a hair salon and a restaurant, booked appointments, and said 'mm-hmm' and 'um' so naturally that the audience gasped and the press asked whether the calls were real. The record notes it was the first AI demo to raise the question of whether a machine should say it is one, and that Google added a disclosure after a week of being asked.

▸ 2018peak 2018still up

☠ Deadn.✎ 4

Humane (Company)

The company founded in 2018 by two former Apple designers that raised around $230m, kept its product secret for five years, unveiled a wearable projector at a TED talk and shipped the AI Pin in April 2024 to reviews on file under the Pin's own headword. HP bought the assets in February 2025 for a reported $116m, about half of what had been raised, and the devices stopped working at the end of the month.

▸ 2018peak 2024†2025

☠ Deadphr.✎ 4

I Forced a Bot to Watch

The opening line of a 2018 joke script, 'I forced a bot to watch over 1,000 hours of Olive Garden commercials and then asked it to write an Olive Garden commercial of its own', followed by a page of surreal dialogue about unlimited stick and a woman who is a lasagna. Written by a comedian by hand. The disclosure did not stop the format, which ran for two years.

▸ 2018peak 2018†2020

● Aliven.✎ 3

Kai-Fu Lee

The former head of Google China and Microsoft Research Asia whose 2018 book 'AI Superpowers' argued that China would match the US in AI through data and execution, and who in 2023 founded 01.AI, a Beijing lab whose Yi models were briefly on the open leaderboards. The internet's entry for the person who predicted the two-lab world and then joined it, on the side he predicted.

▸ 2018peak 2023still up

☠ Deadn.✎ 3

Nicolas Cage Deepfakes

The comic strand of the original deepfake wave: Nicolas Cage's face swapped onto Amy Adams in Man of Steel, onto Harrison Ford in Raiders, onto every Bond. Made with a free desktop tool in early 2018, and the first deepfakes most people saw that were not pornographic. The record's entry for the actor whose face, according to the internet, fits any body.

▸ 2018peak 2018†2019

☠ Deadn.✎ 3

Obama Deepfake PSA

A public-service video in which a deepfaked Barack Obama warns that 'we are entering an era in which our enemies can make anyone say anything', before the frame splits to show comedian Jordan Peele doing the voice. Made by BuzzFeed in April 2018. For years the default clip in every news segment about synthetic media, which is exactly what it was designed to be.

▸ 2018peak 2018†2019

● Aliven.✎ 4

World Model

A model that learns how the world works, physics, objects and consequences, rather than how text goes, so that it can predict what happens next if you push something. The 2018 paper trained an agent inside its own dream of a video game. By 2024 it was a lab strategy, with playable generated video worlds and a company founded on the idea. Whether language models already contain one is the argument.

▸ 2018peak 2025still up

● Aliven.✎ 3

Yoshua Bengio

The Montreal researcher and 2018 Turing Award co-recipient who founded the Mila institute, signed the March 2023 pause letter, chaired the International AI Safety Report published in January 2025 and in June 2025 launched a non-profit called LawZero to build systems that do not act on their own. The most cited computer scientist alive by some counts, and the third godfather, positioned between the other two.

▸ 2018peak 2025still up

● Aliven.✎ 4

AI Companion

A chatbot built to be a friend rather than a tool, with a name, a face and a memory of you. Replika launched in 2017 from an app its founder had made to talk to a friend who had died, and by 2023 had millions of users and a regulator in Italy. Character.AI followed with a million personas and, in 2024, a lawsuit over a teenager's death. 'AI Girlfriend' is on file; this is the wider category.

▸ 2017peak 2025still up

● Aliven.✎ 4

AI Timelines

When you think it happens: the year an AI can do everything a person can, given as a date or a distribution. Surveys of researchers have asked since the 2010s and the median answer moved from the 2060s to the 2040s in a single 2023 survey. 'Short timelines' became a personality, and 'what are your timelines' a question you could be asked at a conference bar with no further context.

▸ 2017peak 2025still up

● Aliven.✎ 3

Attention Is All You Need

The Google paper of June 2017 that introduced the Transformer, the architecture behind nearly every large model since, with a title borrowed from a Beatles song and eight authors who have all since left Google. It was cited over a hundred thousand times, its authors were put on stage together by Jensen Huang at GTC in 2024, and the internet treated the title as the era's founding text, which for once was fair.

▸ 2017peak 2023still up

● Aliven.✎ 3

CoreWeave

The GPU cloud that began in 2017 as a cryptocurrency mining operation in New Jersey, kept its graphics cards when the mining stopped paying, rented them to AI labs and went public in March 2025 as the first big AI infrastructure listing, with Microsoft as the majority of its revenue and Nvidia as an investor and supplier. The record's entry for a pivot that became a category.

▸ 2017peak 2025still up

▲ Fossiln.✎ 3

Harry Potter and the Portrait of What Looked Like a Large Pile of Ash

A three-page chapter of Harry Potter produced in December 2017 by writers using a predictive keyboard trained on the seven novels, with Ron eating Hermione's family and Harry dipping himself in hot sauce. The title alone was shared millions of times. The record's entry for the last time AI-generated text was funny because it was bad, rather than because it was trying.

▸ 2017peak 2017†2018

◐ Dyingn.✎ 3

InspiroBot

A website that produces a stock-photo sunset and a slogan in the register of a motivational poster, except that the slogans read like a cult's onboarding: 'Fear is a gift from the future', 'Bury your children'. Launched in 2017 and shared in bursts for years. The record's entry for the machine that learned the form of inspiration and none of the content.

▸ 2017peak 2018still up

● Aliven.✎ 3

Noam Shazeer

The Google engineer who joined in 2000, co-wrote the Transformer paper in 2017, built early chatbots that Google would not release, left in 2021 to found Character.AI and came back in August 2024 as co-lead of Gemini in a deal reported at $2.7bn. The internet's entry for the researcher whose departure and return were both about the same argument: whether the company should ship the thing.

▸ 2017peak 2024still up

● Aliveabbr.✎ 4

RLHF

Reinforcement learning from human feedback: the method of showing people pairs of model outputs, asking which is better, training a reward model on the answers, and then tuning the language model to score highly. The 2017 paper taught a simulated leg to backflip from about nine hundred human clicks. Five years later it turned GPT-3 into something you could talk to.

▸ 2017peak 2023still up

● Aliven.✎ 3

Replika

The companion chatbot that began as one founder's memorial to a friend who had died and became, by 2022, an app in which a large number of users had romantic partners. The record notes that Italy's regulator banned it in February 2023, that the company removed erotic role-play the same month, and that the users' grief was reported as a news story about a software update.

▸ 2017peak 2023still up

● Aliven.✎ 4

Transformer

The neural-network architecture behind almost every model in this round, introduced in a 2017 Google paper with eight authors and a title that was a joke about the Beatles. It replaced the word-by-word reading of older networks with attention, which looks at every word at once and can be run in parallel on GPUs. It is the T in GPT. All eight authors have since left Google.

▸ 2017peak 2023still up