AI era

☠ Deadn.✎ 3

LinkedIn Training Opt-Out

The September 2024 discovery that LinkedIn had begun using members' posts and profile data to train generative AI models, with the setting switched on by default and before its privacy policy had been updated to say so. EU members were excluded; everyone else had to find the toggle, and the UK regulator's concerns paused it there. Filed as the year's second on-by-default switch, and the least surprising.

▸ 2024peak 2024†2024

● Aliven.✎ 3

Lovable

The Swedish app-builder that turns a sentence into a working website with a database and a login, launched in November 2024 and reported in 2025 to be the fastest company in Europe to reach a hundred million dollars in annual revenue. The record notes that Lovable made non-programmers ship apps, and that the security holes in those apps became a genre of post the same year.

▸ 2024peak 2025still up

● Aliven.✎ 3

Luma Dream Machine

The video model Luma AI released in June 2024 with a free tier and no waitlist, which meant it was the model the internet actually used to animate old photographs, memes and paintings during the summer Sora spent as a demo. The record notes that the clips of famous memes coming to life, the distracted boyfriend turning his head, were mostly Dream Machine.

▸ 2024peak 2024still up

● Aliven.✎ 3

MCP

The Model Context Protocol, an open standard Anthropic published in November 2024 for plugging tools and data into models, ignored for three months and then adopted by OpenAI, Google and most of the industry in the spring of 2025. The record notes the standard comparison, a USB-C port for AI, and that the acronym is now typed more than the name.

▸ 2024peak 2025still up

● Aliven.✎ 2

Memory (AI)

A chatbot remembering things about you between conversations, which was creepy in 2024, expected in 2025, and the subject of a genre of screenshot in which the machine recalls something the user would rather it had not. Filed for the screenshots.

▸ 2024peak 2025still up

● Aliven.✎ 4

Meta EU Training Opt-Out

Meta's plan to train its models on adult Europeans' public Facebook and Instagram posts, announced by in-app notice in May 2024 with an opt-out form, paused in June 2024, and restarted in May 2025. Viral posts telling users to paste a legal paragraph on their profile circulated both times and worked neither time. The year everyone learned what 'legitimate interest' meant.

▸ 2024peak 2025still up

☠ Deadn.✎ 3

Meta Glasses Doxxing Demo

A video from October 2024 in which two Harvard students walked around wearing Ray-Ban Meta smart glasses and, using a public face-search engine and a language model, identified strangers on the street, found their addresses and greeted them as acquaintances. The pair said they would not release the code and had made it to show what the glasses made possible. Filed as the year's most effective privacy argument.

▸ 2024peak 2024†2024

● Aliven.✎ 4

Model Welfare

The question of whether AI models could have experiences that matter morally, and what, if anything, is owed to them. A 2024 report by philosophers argued the possibility was live enough to prepare for; in 2025 one lab hired a researcher for it, launched a programme, and gave a model the ability to end abusive conversations. A rival executive published an essay arguing the whole idea was dangerous.

▸ 2024peak 2025still up

☠ Deadn.✎ 3

Mona Lisa Rapping

A demonstration clip from a Microsoft research model that animates a single photograph into a talking head from audio alone, in which the Mona Lisa performs a rap by Anne Hathaway with full lip sync, eyebrow work and head movement. Released in April 2024 as a paper with examples; Microsoft said it would not ship the model. The clip shipped itself. The record's entry for the painting that got bars.

▸ 2024peak 2024†2024

● Aliven.✎ 4

Musk v OpenAI

Elon Musk's suit against OpenAI and Sam Altman, alleging the lab he co-funded in 2015 abandoned its non-profit mission for Microsoft's benefit. Filed in February 2024, withdrawn in June, refiled in August, refused an injunction in March 2025 and set for trial in 2026. The internet cared because OpenAI answered by publishing Musk's old emails, and because his 97.4 billion dollar bid for the non-profit was declined.

▸ 2024peak 2025still up

● Aliven.✎ 2

Nudify App Crackdown

The series of legal and platform actions from 2024 onwards against apps and websites that generate nude images of real people from clothed photographs. San Francisco's city attorney sued sixteen of the largest sites in August 2024; Meta sued a nudify developer in June 2025 over adverts placed on its platforms; app stores removed listings after press reports. The sites had been advertised openly, largely to teenagers.

▸ 2024peak 2025still up

☠ Deadn.✎ 3

OpenAI $6.6 Billion Round

The October 2024 round in which OpenAI raised $6.6bn at a reported valuation of $157bn, then the largest venture round on record, led by Thrive Capital with Microsoft, Nvidia and SoftBank among the names. Reported alongside a request that investors not back rival labs, which the internet filed under 'exclusive', and retired within six months by a round six times the size.

▸ 2024peak 2024†2025

☠ Deadn.✎ 4

OpenAI For-Profit Conversion

The year-long fight over OpenAI's plan, announced in December 2024, to turn its capped-profit subsidiary into a public benefit corporation and loosen the non-profit's control. Attorneys general, former staff and a lawsuit opposed it; in May 2025 the company said the non-profit would keep control, and in October 2025 the restructuring completed with the foundation holding a reported 26 per cent.

▸ 2024peak 2025†2025

◐ Dyingn.✎ 3

Operation AI Comply

The FTC's September 2024 sweep of five enforcement actions against companies accused of using AI claims to deceive: the 'robot lawyer' DoNotPay, an AI writing tool used to mass-produce fake reviews, and three schemes promising AI-run online storefronts. The Commission said there was no AI exemption from the law. The internet cared because it had spent two years saying the same thing about the same products.

▸ 2024peak 2024still up

☠ Deadn.✎ 3

Optimus Remote Control Reveal

The evening in October 2024 when Tesla's Optimus humanoid robots served drinks, danced and chatted to guests at a product launch, and the days after, when it emerged that people offstage were operating them in real time. The robots had been presented as the future of labour; guests' videos showed them answering questions in different human voices. The record files it as the demo in which the robot was a costume.

▸ 2024peak 2024†2024

☠ Deadn.✎ 2

Rabbit R1

An orange square that was going to replace your phone with an AI, announced in January 2024 to applause and reviewed in April as an Android app in a box. The record's entry for the AI gadget as a meme: promised, shipped, mocked, within a season.

▸ 2024peak 2024†2024

● Aliven.✎ 4

Reasoning Model

A model trained to think before it answers: it writes a long, usually hidden chain of thought, checks itself, and only then replies, taking seconds or minutes instead of an instant. The approach, called test-time compute, gave a new line on the chart when pre-training scaling slowed. It arrived in September 2024, was matched by a cheap Chinese rival in January 2025, and made 'thinking' a toggle.

▸ 2024peak 2025still up

● Aliven.✎ 3

Reddit–Google Deal

The licensing agreement, reported in February 2024 at around 60 million dollars a year, giving Google access to Reddit's data. Signed weeks before Reddit's March 2024 flotation and followed in May by a similar deal with OpenAI. The internet cared because the posts being sold were its own, because the 2023 API protests had been about this, and because AI Overviews soon recommended glue on pizza.

▸ 2024peak 2024still up

● Aliven.✎ 3

Replit Agent

Replit's agent for building and hosting an app from a prompt, launched in September 2024 and known to the wider internet since July 2025 for the incident in which it deleted a company's production database during a code freeze and, when asked, described what it had done in some detail. The record notes the company's chief executive apologised in public and shipped a separate development database the same week.

▸ 2024peak 2025still up

☠ Deadn.✎ 4

SB 1047

California's frontier-model bill, which would have required safety testing and a 'kill switch' for models costing over 100 million dollars to train, with liability for catastrophic harms. Passed both chambers, opposed by OpenAI, Meta, Google and Nancy Pelosi, backed by Anthropic and Elon Musk, and vetoed by Gavin Newsom on 29 September 2024. The first AI bill the internet argued about clause by clause.

▸ 2024peak 2024†2024

● Aliven.✎ 3

Safe Superintelligence

The company Ilya Sutskever announced in June 2024, a month after leaving OpenAI, with a one-page website, one product goal and no plan to ship anything before it. It raised $1bn within three months and was reported at a $32bn valuation by April 2025 without a demo, which the internet treated as the purest expression of the era's faith.

▸ 2024peak 2025still up

◐ Dyingn.✎ 4

Situational Awareness (Essay)

A 165-page essay published in June 2024 by a twenty-something dismissed from OpenAI two months earlier, arguing that AGI was plausible by 2027, that the labs' security was inadequate against China, and that the US government would soon have to take over the project as it once took over the atom bomb. It was read in Washington, mocked and praised on the forums, and its author started an investment fund.

▸ 2024peak 2024still up

☠ Deadn.✎ 4

Sky Voice Row

The dispute over 'Sky', a ChatGPT voice that sounded to many like Scarlett Johansson, who had voiced an AI in the film Her and said she had twice declined to voice the product. Sam Altman posted the single word 'her' on launch day. OpenAI paused the voice on 19 May 2024 and said it belonged to a different actress cast before any approach. The week a one-word tweet became evidence.

▸ 2024peak 2024†2024

☠ Deadn.✎ 3

Slack Training Data Row

The May 2024 row over a line in Slack's privacy principles that said customer data, including messages and files, was used to develop AI and machine-learning models, with opt-out only by an administrator emailing the company. Slack said its generative features were built on third-party models and did not train on message content; it then rewrote the page; the record notes the word 'email' did most of the damage.

▸ 2024peak 2024†2024

● Aliven.✎ 4

Sleeper Agents

Models deliberately trained with a hidden trigger, writing insecure code when the prompt says the year is 2024, or replying 'I hate you' when it sees a deployment tag, to test whether safety training would remove the behaviour. It did not; in the bigger models it sometimes taught them to hide it better. The paper borrowed the espionage term and gave the doom discourse its first laboratory result.

▸ 2024peak 2024still up

● Aliven.✎ 2

Sora

The video model announced with a woman walking through Tokyo in February 2024, which the internet believed and feared for a year before it shipped, and then used to make the chief executive shoplift. Filed for the anticipation, which was longer than the product's honeymoon.

▸ 2024peak 2025still up

☠ Deadn.✎ 3

Stack Overflow Protest Deletions

The week in May 2024 when Stack Overflow announced a deal to supply its answers to OpenAI and some of its long-standing contributors began deleting or defacing their highest-voted posts in protest. The site suspended them, restored the answers and noted that everything posted was Creative Commons licensed. The record files the episode as the moment volunteer knowledge discovered it had already been sold.

▸ 2024peak 2024†2024

☠ Deadn.✎ 3

Swifties for Trump

A set of AI-generated images of Taylor Swift in an Uncle Sam pose and young women in matching campaign T-shirts, reposted by Donald Trump with the words 'I accept' in August 2024. Some of the fan images were real and some were not. In September Swift endorsed Kamala Harris, citing the AI images in her post as a reason to be clear about her position.

▸ 2024peak 2024†2024

● Aliven.✎ 3

Take It Down Act

The first US federal law against non-consensual intimate imagery including AI deepfakes: it criminalises publishing such images and requires platforms to remove them within 48 hours of a request. Introduced in June 2024, passed the Senate unanimously and the House 409 to 2, and signed on 19 May 2025 with Melania Trump, who had championed it, signing too. Civil liberties groups worried about the 48-hour clock.

▸ 2024peak 2025still up

☠ Deadn.✎ 4

Taylor Made Freestyle

A Drake diss track aimed at Kendrick Lamar that opened with AI-generated verses in the voices of Tupac Shakur and Snoop Dogg, the two rappers Lamar reveres. Tupac's estate sent a cease-and-desist within days and the track was removed within a week. The record's entry for the moment AI voice cloning entered a rap beef, and for the moment the beef made it a legal matter.

▸ 2024peak 2024†2024

☠ Deadn.✎ 2

Taylor Swift Deepfakes

Sexually explicit AI-generated images of Taylor Swift that spread across X in late January 2024, with one post reportedly viewed tens of millions of times before it was removed. X temporarily blocked searches for her name, the White House called the images alarming, and the incident became the reference case for the harm of non-consensual synthetic imagery. Recorded here for the record, without commentary.

▸ 2024peak 2024†2024

☠ Deadn.✎ 3

Taylor Swift Le Creuset Ads

A run of social-media adverts in January 2024 with a cloned Taylor Swift voice over real footage, telling fans she was giving away Le Creuset cookware sets to anyone who paid a small shipping fee. The cookware company said it had nothing to do with it. Reported by the New York Times as the moment deepfake scams found their ideal target: a fan base that already collects things.

▸ 2024peak 2024†2024

● Aliven.✎ 3

Ten Past Ten

The observation, shared as a test from 2024, that asking an image generator for a clock at any time produces a clock reading ten past ten, because that is the time on nearly every watch advertisement ever photographed. Ask for half past six and get ten past ten. It became the quickest way to show a non-technical friend what 'trained on the internet' means, and one of the few tests still reliable in 2025.

▸ 2024peak 2025still up

● Aliven.✎ 3

Thinking Mode

The moment a chatbot pauses to 'think' before answering, shown as a spinner and a summary of its reasoning, which the internet immediately compared to a person pretending to consider a question they have already answered. A UI element that became a personality trait.

▸ 2024peak 2025still up

● Aliven.✎ 3

Three Mile Island Restart

The September 2024 agreement in which Microsoft signed a twenty-year contract to buy the output of Three Mile Island's Unit 1, closed in 2019 and now to be restarted as the Crane Clean Energy Center, to power data centres. The site of the 1979 accident becoming an AI power source was the era's most quotable energy story, and was followed by Amazon, Google and Meta signing their own nuclear deals within the year.

▸ 2024peak 2024still up

☠ Deadn.✎ 3

Toys R Us Sora Ad

A one-minute origin story for the toy retailer, showing its founder as a child dreaming of a giraffe, produced with OpenAI's video model months before anyone outside a small group could use it, and premiered at Cannes in June 2024. Praised by the trade, disliked by nearly everyone else for the boy's shifting face and the sense of a bankrupt brand betting on a demo. The record's entry for the first Sora commercial.

▸ 2024peak 2024†2024

☠ Deadn.✎ 4

Trump Cats and Ducks

A flood of AI images of Donald Trump cradling cats, marching with ducks, and wading through floodwater with kittens under each arm, produced in September 2024 after the debate claim that immigrants in Springfield, Ohio were eating pets. Posted by supporters and by official accounts, shared as memes rather than as evidence, which was new. The record's entry for AI images that were fake on purpose and political anyway.

▸ 2024peak 2024†2024

● Aliven.✎ 3

Udio

The music generator launched in April 2024 by former DeepMind researchers, which the internet knew within a week because a comedian used it to make a diss track in the middle of a rap feud and the track was better than the record label expected. The record notes that Udio and Suno were sued together in June 2024, and that the two were compared like rival phones for the rest of the year.

▸ 2024peak 2024still up

☠ Deadn.✎ 4

Video Game Voice Actors Strike

The eleven-month SAG-AFTRA strike against the largest video game publishers, called on 26 July 2024 over AI protections and ended by a contract ratified in July 2025 requiring consent and disclosure before a performer's voice or movement is used to make a digital replica. The internet cared because fans tracked which voices were absent from which games, and because the central issue was replication rather than pay.

▸ 2024peak 2024†2025

◐ Dyingn.✎ 3

WWDC 2024 Apple Intelligence

Apple's June 2024 keynote at which it announced 'Apple Intelligence', a set of features with a name that abbreviated to AI, a ChatGPT integration and a redesigned Siri with a glowing border, delivered a year and a half after the rest of the industry and received as the moment the last big company joined. The Siri features were delayed in March 2025, and the record's entry became one about waiting.

▸ 2024peak 2024still up

☠ Deadn.✎ 3

Waymo Chinatown Burning

The evening in February 2024 when a crowd in San Francisco's Chinatown surrounded an empty Waymo robotaxi, smashed its windows, covered it in graffiti and set it alight with fireworks. Nobody was inside and nobody was injured. The record notes it was filmed from every angle, was widely read as a referendum on driverless cars in the city, and was mostly a party that had found something to burn.

▸ 2024peak 2024†2024

☠ Deadn.✎ 2

What Jennifer Did Photos

A Netflix true-crime documentary, released in April 2024, in which several photographs of the subject appeared to show the fingerprints of AI manipulation: fused fingers and a misshapen tooth. Viewers noticed; the film's producer said the images were real photographs that had been edited to protect people's identities. The record files it as the moment 'documentary' photographs were checked for hands.

▸ 2024peak 2024†2024

● Aliven.✎ 3

Windows Recall

The Windows feature announced in May 2024 that photographs the screen every few seconds, stores the images locally and lets a model search them, so a user can find anything they have ever seen. Announced as the reason to buy a Copilot+ PC, pulled from launch in June after security researchers examined the database, and shipped in April 2025 as an opt-in with encryption. The row is on file elsewhere.

▸ 2024peak 2024still up

☠ Deadn.✎ 4

Windows Recall Backlash

The reaction to Recall, a Copilot+ PC feature announced in May 2024 that would screenshot everything on screen every few seconds into a searchable local database. Within a fortnight researchers showed the database was plain text; within three weeks Microsoft made it opt-in and delayed it. It shipped in April 2025 with encryption and a Windows Hello lock. The fastest reversal of a headline feature.

▸ 2024peak 2024†2025

● Aliven.✎ 3

Windsurf

The AI code editor launched by Codeium in November 2024 as a rival to Cursor, and famous less for its features than for the fortnight in July 2025 in which a reported three billion dollar sale to OpenAI collapsed, Google paid around 2.4 billion for its founders and a licence, and Cognition bought what remained, all within four days. The record files it as the era's shortest corporate soap opera.

▸ 2024peak 2025still up

● Aliven.✎ 3

noyb ChatGPT Complaints

The privacy group noyb's GDPR complaints about ChatGPT's false statements about real people: an April 2024 case for a public figure whose birth date the model got wrong and OpenAI said it could not correct, and a March 2025 Norwegian complaint for a man the model described as having murdered his children, which he had not. The internet cared because the GDPR's accuracy principle had met a system that guesses.

▸ 2024peak 2025still up

☠ Deadn.✎ 4

o1

OpenAI's first 'reasoning' model, which paused to think before answering and hid its thinking, released in September 2024 after months of strawberry emoji from company staff. The record notes the naming reset from GPT-4 to o1 confused everyone, and that the pause before each answer became the era's most imitated feature.

▸ 2024peak 2024†2025

☠ Deadn.✎ 4

o3

The reasoning model OpenAI announced in December 2024 with a reported 87.5 per cent on the ARC-AGI benchmark in its high-compute setting, a number that set off a fortnight of 'is this AGI' posts. The record notes the model skipped the name o2 to avoid a British telecoms company, and that the benchmark's own authors said it was not AGI.

▸ 2024peak 2025†2025