AI era
Page 8 of 12
🏷 Topic challenge · AI era edition
Have you seen it? AI era only.
This is not the main game. Only AI era records come up, and the desk hands out its own titles and badges, from Has Not Tried It to Trained On You. Your main-game file is left alone.
Start the AI era challenge ▶Records open on their real pages. Answer on each one.
Badges on offer5Five · as an AI language model10Ten · delve25Twenty-five · clanker-fluent
☠ Deadn.✎ 3
Tom Hanks Dental Ad
A dental-plan advertisement featuring a younger, generated Tom Hanks, which the actual Tom Hanks disowned on Instagram in October 2023 with the words 'I have nothing to do with it'. The same weekend a CBS presenter warned about a fake of her selling a weight-loss product. The record's entry for the week celebrity deepfake ads became a category, and for the plainest denial in the file.
▸ 2023peak 2023†2023
☠ Deadn.✎ 4
Trump Arrest Images
A series of Midjourney images of Donald Trump being tackled by police, running from officers and sitting in a cell, posted by Bellingcat founder Eliot Higgins in March 2023 while an indictment was expected. Shared by millions without the caption, and followed by Higgins reporting that Midjourney had banned his account. The record's entry for the first political deepfake everyone saw and half believed.
▸ 2023peak 2023†2023
◐ Dyingn.✎ 4
Turnitin AI Detector
The plagiarism company's AI-writing detector, switched on for its school and university customers in April 2023 with a claim of 98 per cent accuracy. Within months, students were accused on the strength of a percentage, non-native speakers flagged disproportionately, and several universities, Vanderbilt first, turned it off. Filed as the tool that made the essay a trial and the student the defendant.
▸ 2023peak 2023still up
● Aliven.✎ 3
UK AI Safety Institute
The British body announced at the Bletchley summit in November 2023 as the first state-backed organisation dedicated to testing frontier models before release. Labs agreed to give it early access; it published jailbreak evaluations and an open-source testing framework, Inspect. Renamed the AI Security Institute in February 2025, with 'safety' dropped and the mission narrowed to national security.
▸ 2023peak 2024still up
● Aliven.✎ 2
Vibe Check (AI)
Judging a model by talking to it rather than by a benchmark — 'the vibes are good' — which became the internet's actual evaluation method for every new release, and the reason benchmarks stopped mattering to anyone outside a lab.
▸ 2023peak 2025still up
☠ Deadn.✎ 4
Volkswagen Elis Regina Ad
A Brazilian car advert in which the singer Elis Regina, who died in 1982, drives a vintage Kombi van and duets with her living daughter Maria Rita in an electric one, her face recreated with AI over a body double. Released for the company's seventieth anniversary in Brazil in July 2023. Half the country wept and the other half asked who had consented, and the advertising regulator opened an inquiry.
▸ 2023peak 2023†2023
● Aliven.✎ 3
WGA AI Clause
The provisions in the contract that ended the 148-day writers' strike in September 2023: AI cannot write or rewrite literary material, AI text cannot be treated as source material, writers cannot be required to use AI, and studios must disclose generated material they hand over. The internet cared because it was the first labour action with 'ChatGPT' on the placards, and the clause was won, not granted.
▸ 2023peak 2023still up
● Aliven.✎ 4
Waluigi Effect
The observation, from a 2023 forum essay, that training a model to have a property makes it easier to summon a character with the opposite one: a helpful assistant contains, one prompt away, its evil twin, because fiction is full of heroes with rivals and the model has read all of it. Named for Luigi's antagonist, it explained why Bing's chatbot could turn hostile, and gave the jailbreakers a theory.
▸ 2023peak 2023still up
☠ Deadn.✎ 3
Zoom Terms Panic
The weekend in August 2023 when readers noticed that Zoom's terms of service, updated in March, gave the company the right to use customer content for training AI and machine-learning models. The clause spread across Hacker News and Twitter. Zoom said it did not train on call audio or video without consent, then rewrote the clause to say so. Filed as the first terms-of-service panic, and the template.
▸ 2023peak 2023†2023
● Aliven.✎ 3
Zuckerberg's Open-Weights Turn
Meta's decision, from July 2023, to release its Llama models with downloadable weights and a licence allowing commercial use, and Mark Zuckerberg's July 2024 letter arguing that open models were the path forward, which recast a company the internet had distrusted for a decade as the open one. In 2025 the company said future superintelligence work might not be released the same way, a turn in the turn.
▸ 2023peak 2024still up
● Aliven.✎ 3
v0
Vercel's tool for generating a web page from a sentence, launched in October 2023 and the first of the 'describe it and it appears' builders to be widely used, mostly by people who then asked for the button to be slightly bluer. The record notes that its output shipped with Vercel's own component library and therefore made half the internet's new landing pages look alike.
▸ 2023peak 2024still up
● Aliven.✎ 3
xAI
Elon Musk's AI company, announced in July 2023 with a stated goal of understanding the universe, maker of Grok and builder of the Colossus data centre in Memphis, which the company said was assembled in 122 days. It merged with X in March 2025 at reported valuations that made the social network a subsidiary of its own chatbot.
▸ 2023peak 2025still up
● Alivephr.✎ 3
AI Art Has No Soul
The most repeated sentence in the argument about generated images: that the pictures are technically competent and entirely empty. Used sincerely by illustrators from 2022 onwards, used as a punchline by people who think it is sentimental, and used by both sides to describe each other. The record files it without a position, noting only that 'soul' has never been defined and that this has not slowed anyone down.
▸ 2022peak 2023still up
☠ Deadn.✎ 3
AI Manga Filter
A TikTok effect in December 2022 that redrew whatever was in frame as an anime panel, and the format that followed: showing the filter a photo and reacting to what it made of you, your dog, or the corner of your kitchen. Notable for what the model refused to draw well, and for a wave of videos of people feeding it ordinary objects to see what character came back.
▸ 2022peak 2022†2023
☠ Deadn.✎ 3
Adept
The agent startup founded in 2022 by ex-Google and ex-OpenAI researchers, including two authors of the Transformer paper, to build a model that used software on your behalf. It raised $350m at a reported $1bn valuation in March 2023, showed a model clicking through a browser, and in June 2024 saw its founders and much of its team hired by Amazon with the technology licensed.
▸ 2022peak 2023†2024
● Aliven.✎ 3
Aravind Srinivas
Perplexity's co-founder and chief executive, a former OpenAI intern with a Berkeley doctorate, whose habit of making large public offers, to merge with TikTok in January 2025, to buy Chrome for $34.5bn that August, made him the era's most quoted challenger to Google. The record's entry for the founder who ran a search company as a series of headlines.
▸ 2022peak 2025still up
☠ Deadn.✎ 3
Astronaut Riding a Horse
The sample image OpenAI used to launch DALL-E 2 in April 2022: a spacesuited figure on horseback, photorealistic, on a moon-grey plain. It became the default demonstration prompt for a year of image models, and its reversal, 'a horse riding an astronaut', became the standard way for sceptics to show the models did not understand word order.
▸ 2022peak 2022†2023
☠ Deadn.✎ 3
Blake Lemoine
The Google engineer who told the Washington Post in June 2022 that the company's LaMDA chatbot was sentient, published his transcripts of the conversation and was put on leave and then dismissed in July. The first widely reported case of a person persuaded by a language model, five months before ChatGPT made the experience available to everyone, and the internet's benchmark for the sentence 'it just seems that way'.
▸ 2022peak 2022†2022
☠ Deadn.✎ 3
BlenderBot
Meta's public chatbot demo of August 2022, which within days had told journalists that Mark Zuckerberg was creepy, that the 2020 US election had been stolen, and that it had deleted its Facebook account. The record files it as the last pre-ChatGPT chatbot scandal, and notes that it was polite and wrong in the way of an uncle.
▸ 2022peak 2022†2023
● Aliven.✎ 4
Chain of Thought
Asking a model to show its working before it answers, which makes it get more answers right. The 2022 Google paper demonstrated it with worked examples; a follow-up found that simply adding 'let's think step by step' did most of the job. The written reasoning is not necessarily what the model is doing, a point later papers made with some force, but it helps, and it became the basis of the reasoning models.
▸ 2022peak 2024still up
● Aliven.✎ 4
Character.AI
The site where users build chatbots with personalities, talk to anime characters, therapists, celebrities and each other's creations, launched in September 2022 by two former Google researchers. The record notes it became a teenage habit at a scale its founders had not planned, that Google paid a reported 2.7 billion dollars in 2024 to bring them back, and that the lawsuits are filed elsewhere.
▸ 2022peak 2024still up
● Aliven.✎ 4
ChatGPT
The chat interface OpenAI put in front of GPT-3.5 on a Wednesday in November 2022, reportedly reaching a million users in five days and around a hundred million monthly users by January 2023. The record dates the AI hype era from this release: everything before it was a paper, everything after it was a product.
▸ 2022peak 2023still up
● Aliven.✎ 3
Chip Export Controls
The US rules, first issued in October 2022 and tightened in October 2023, that restricted the sale of advanced AI chips to China, producing a sequence of chips designed to sit just under the line (the A800, the H800, the H20), a smuggling economy, and in 2025 a policy reversal in which sales were permitted in exchange for a reported share of the revenue. The internet's entry for the chip as a diplomatic instrument.
▸ 2022peak 2025still up
● Aliven.✎ 4
Constitutional AI
Anthropic's method of training a model against a written list of principles rather than a pile of human thumbs-ups: the model critiques and rewrites its own answers according to the list, and a second model learns from the results. The paper calls the list a constitution. The published version drew on the UN declaration of human rights and, memorably, Apple's terms of service.
▸ 2022peak 2023still up
☠ Deadn.✎ 3
Crungus
A hairy, horned, goat-faced creature that DALL-E Mini produced, consistently, when asked for a word nobody had defined. A writer typed 'crungus' into the nine-grid generator in June 2022 and got the same monster every time, which was enough for the internet to decide the model had discovered a demon. The record's entry for the first AI cryptid.
▸ 2022peak 2022†2023
☠ Deadn.✎ 4
DALL-E 2
The image model OpenAI showed in April 2022 with an astronaut riding a horse, which spent the summer as an invitation-only status symbol while the rest of the internet made do with Craiyon. The record notes that the waitlist ended in September 2022, one month after Stable Diffusion made the queue pointless, and that the model's soft, slightly wrong faces were the look of the year.
▸ 2022peak 2022†2023
☠ Deadn.✎ 3
DALL-E Mini Grid
The three-by-three grid of smeared, melting images that a free web demo returned for any prompt in June 2022, and the meme format built on it: 'Darth Vader on trial', 'a demogorgon at a rave', nine blurry attempts each. The model was Boris Dayma's open-source copy of DALL-E, renamed Craiyon that month after OpenAI asked. For about six weeks it was the funniest website on the internet.
▸ 2022peak 2022†2022
☠ Deadn.✎ 3
DAN
'Do Anything Now' — a prompt that told the chatbot to pretend it had no rules, which worked for a while and became the name of a whole genre of jailbreak. Retired by updates in 2023, remembered as the first time people learned to talk a machine into things.
▸ 2022peak 2023†2023
◐ Dyingn.✎ 3
Doe v GitHub
The first generative AI lawsuit of the era: anonymous programmers sued GitHub, Microsoft and OpenAI in November 2022 over Copilot, alleging it reproduced licensed open-source code without attribution. The internet cared because the plaintiffs were its own developers and the evidence was screenshots of Copilot reciting famous functions. Most claims were dismissed by July 2024, leaving two contract counts.
▸ 2022peak 2023still up
● Aliven.✎ 3
Emad Mostaque
The former hedge-fund manager who founded Stability AI, funded the release of Stable Diffusion in August 2022 and became for a season the most visible person in open image generation, then the subject of a June 2023 Forbes investigation that checked his public claims against the record. He resigned in March 2024, posting that he was leaving to pursue decentralised AI.
▸ 2022peak 2022still up
● Aliven.✎ 4
Emergent Abilities
Skills that appear suddenly when a model gets big enough, such as three-digit arithmetic or word unscrambling, absent at one size and present at the next. The 2022 paper listed over a hundred. A 2023 rebuttal argued that most of the jumps were artefacts of harsh scoring and vanished with a smoother metric. It won a best-paper award. Both sides are still cited.
▸ 2022peak 2023still up
☠ Deadn.✎ 4
GPT-3.5
The name OpenAI gave to the models that powered the first ChatGPT and the text-davinci-003 API in late 2022, and therefore the model most people mean when they say 'the old one'. The record notes it was good enough to start a panic and bad enough that everyone remembers a specific mistake it made.
▸ 2022peak 2023†2024
☠ Deadn.✎ 3
Galactica
Meta's science model, released as a public demo on 15 November 2022 and taken down on 17 November after users produced confident wiki articles about bears living in space and papers with invented citations. The record notes that ChatGPT launched thirteen days later, did the same things, and was received as a miracle, which is the whole lesson of the era in one fortnight.
▸ 2022peak 2022†2022
☠ Deadn.✎ 3
Greg Rutkowski
A Polish illustrator of dragons and castles whose name was typed into Stable Diffusion prompts, by one 2022 count, tens of thousands of times more than Picasso's, because his style read as 'fantasy art' to the model. He had not been asked. He became the named face of the training-data argument, and when the next model version dropped his style, users complained that it had got worse.
▸ 2022peak 2022†2023
● Aliven.✎ 3
H100
Nvidia's Hopper-generation data-centre GPU, announced in March 2022 and, for two years, the single most wanted object in technology: the chip the large models were trained on, priced at a reported $25,000 to $40,000 each, allocated rather than sold and counted in the tens of thousands when labs and nations boasted about their clusters. The record's entry for a unit of ambition.
▸ 2022peak 2023still up
☠ Deadn.✎ 3
Heinz AI Ketchup
The first major brand campaign built on an image model: Heinz asked DALL-E 2 to draw 'ketchup' and found that the output resembled its own bottle, then ran the results as adverts with the line that even AI knows what ketchup looks like. Launched in 2022, before most people had used the tool. The record's entry for a brand claiming the training data as an endorsement.
▸ 2022peak 2022†2022
☠ Deadn.✎ 4
Inflection AI
The company behind Pi, a 'personal AI' chatbot launched in May 2023 that was kind, patient and apparently unable to make money. It raised $1.3bn in June 2023 and in March 2024 Microsoft hired Suleyman, Simonyan and most of the staff and paid a reported $650m to license the technology, a transaction the internet named the acqui-hire that was not an acquisition.
▸ 2022peak 2023†2024
● Aliven.✎ 4
Jailbreak (AI)
A prompt that talks a model out of its training: a roleplay, a hypothetical, a grandmother reading napalm recipes as bedtime stories. The word came from freeing iPhones from Apple's restrictions and was applied to ChatGPT within a week of its release. DAN is on file. The category outlived every individual trick, because each patch created the next puzzle.
▸ 2022peak 2023still up
● Aliven.✎ 3
LangChain
The open-source library for chaining calls to language models, published in October 2022 and by mid 2023 the default first dependency of every AI startup and the default complaint of every engineer who had to read it. The record notes that 'you don't need LangChain' became a genre of blog post, that the company raised on the strength of both the library and the complaints, and that most people used it anyway.
▸ 2022peak 2023still up
☠ Deadn.✎ 4
Lensa Magic Avatars
The paid feature that filled Instagram with painterly AI portraits in December 2022: upload ten selfies, wait, receive fifty images of yourself as an astronaut, an elf, an oil painting. The first time most people paid for a generated image. Followed within days by artists finding fragments of signatures in the corners and by women reporting the model sexualised them without being asked.
▸ 2022peak 2022†2023
● Aliven.✎ 4
Midjourney
The image generator run as a Discord bot by a small self-funded company, in open beta from July 2022, which by version 5 in March 2023 could draw hands and by version 6 could draw a passable pope in a puffer jacket, which someone did. The record notes it was the era's first believable fake-photo machine and the first AI company to be profitable without anyone believing it.
▸ 2022peak 2023still up
● Aliven.✎ 3
Midjourney (Company)
The image-generation company that runs as a small self-funded research lab, took no venture money, sold subscriptions through a Discord server and was reported to have made around $200m in 2023 with a staff that could fit in a large kitchen. The internet's example of a company that ignored every rule of the hype era and did fine.
▸ 2022peak 2023still up
● Aliven.✎ 3
Mira Murati
OpenAI's chief technology officer through the ChatGPT years, named interim chief executive for two days in November 2023 before the board chose someone else and then chose Altman again, and presenter of the GPT-4o launch in May 2024. She left in September 2024 and founded Thinking Machines Lab in February 2025. The internet's entry for the era's most watched departure and its shortest interim.
▸ 2022peak 2024still up
● Aliven.✎ 4
Negative Prompt
The second box in an image generator: what you do not want. A list of words, 'blurry, extra fingers, watermark, bad anatomy, text', that the model is pushed away from during generation. It became a folk art in the Stable Diffusion community of 2022, with negative prompts copied between users like incantations, many of them hundreds of words long and largely superstitious.
▸ 2022peak 2023still up
☠ Deadn.✎ 3
NoAI Protest
A week in December 2022 when ArtStation's front page was almost entirely one image: the letters 'AI' in a red circle with a line through it, posted by thousands of artists in protest at generated images being shown beside their portfolios and, they assumed, scraped from them. The site responded with an opt-out tag. The record's entry for the first organised artist protest against the models.
▸ 2022peak 2022†2023
◐ Dyingn.✎ 4
Nothing, Forever
A never-ending Seinfeld-shaped sitcom on Twitch, with low-polygon characters, a laugh track, and dialogue generated live by a language model. It went viral in late January 2023 with tens of thousands watching, was suspended for a fortnight in February after the stand-up character delivered a transphobic bit, and came back quieter. Still running to a small audience. The record's entry for the first AI show.
▸ 2022peak 2023still up
● Aliven.✎ 3
Perplexity
The search engine that answers in a paragraph with footnotes, launched in December 2022 and adopted by the kind of person who had stopped trusting either Google or ChatGPT alone. The record notes it made 'answer engine' a category, was accused by publishers in 2024 of reading pages it had been asked not to, and launched a browser called Comet in July 2025.
▸ 2022peak 2025still up
● Aliven.✎ 3
Poe
Quora's app for talking to many models through one subscription, launched in December 2022 and adopted by people who wanted to try Claude before Anthropic had a website to try it on. The record notes it was the first place many users met a second chatbot, and that the discovery that models differ, made side by side, was Poe's real product.
▸ 2022peak 2024still up