English

Editorial policy and corrections

Aphrodite is an independently maintained editorial project that documents AI development and evolves with the field. We distinguish verified events, editorial interpretation, developing stories, and forecasts. Dates, figures, and quotations are linked to their sources where possible. Because AI products and policies change rapidly, entries include a review date.

Submit a correction

Revision history

Corrected

  • On a technology page, the small label above the title naming its layer was set with extra space between letters, a habit from Latin capitals. In Thai it pulled the consonants apart. The spacing is now kept only for scripts with capitals; Chinese, Japanese, Korean, Thai, Hindi and Arabic labels are set without it.

Corrected

  • The entry on Alonzo Church said that he and his student Alan Turing proposed the Church-Turing thesis together. Each reached it independently in 1936, as the milestone on the same page says, and the name came later. The entry now says so. The Japanese pages also called the thesis by the word for a research paper. They now use the word for a thesis.
  • The project epigraph on the home page is Simplified Chinese on every page, but the Japanese, Korean and Traditional Chinese pages drew it in their own languages’ typefaces, and on the first two some characters fell back to another face, so one poem showed two. It is now drawn in a Simplified Chinese face everywhere.
  • The Cursor entry on the Tools page said that Cmd+I opens Composer in Agent mode. Composer is now the name of one of Cursor’s models, and Cmd+I opens Agent in the side panel, as Cursor’s documentation says. Its Windows command installed a WinGet package that does not exist, Cursor.Cursor, where the package is Anysphere.Cursor. Both are corrected in every language, and the Traditional Chinese entry now gives the same steps as the others instead of a one-line summary.
  • The Arabic note on using Anthropic Skills left out that GET /v1/skills only lists Skills and runs none, and the link to the guide, and added a sentence the other languages do not have. It now says what they say.
  • On the Papers cards the publication type was capitalised on every word, so the French pages read Article De Revue. Only the first letter is capitalised now, in every language, English included.

Updated

  • On the Papers page, every title is now given as published, and the page’s own translation follows beneath it in smaller type, the way citation styles give a translated title. The Chinese, Japanese, Korean and Russian pages had shown their own translations in place of most titles and left others in English, with no rule deciding which; they now give a translation beneath every title except the two that are themselves names. For a few hours on this date the order was the other way round, with translations nobody uses standing where the work’s name belongs.
  • The French, Italian, German, Portuguese, Spanish, Thai and Vietnamese pages now give a translation beneath every title as well. They had translated only the 22 works added most recently. Gödel’s paper needs none on the German page, because its title is German, and three Portuguese translations were brought into European Portuguese.
  • The Indonesian, Hindi and Arabic pages now give a translation beneath every title too, so every language but English glosses all of them. On the English page, Gödel’s 1931 paper, whose title is German, now carries its English title beneath, the way English reference works give it.

Corrected

  • On the Thai pages, six quotations on the era pages and five notes on the MCP page ended with a full stop carried over from the English, though a Thai sentence ends without one. The full stops are removed.
  • Two translated quotations on the era pages said something a little different from the original. In French, Norbert Wiener wrote of a machine we cannot interrupt once it is running, where he wrote of one whose working we cannot interfere with. In Japanese, the Dartmouth proposal spoke of problems people handle today, where it speaks of problems now reserved for humans. Both now follow the original.

Corrected

  • Five tool records had fallen behind what their projects now publish. The workflow given for Anthropic’s security-review GitHub Action named a version tag that does not exist and passed the API key in a way the action does not read, so it would have failed on its first run. It now follows the action’s own README. The OpenClaw record had the project’s name in the wrong capitals, described its tagline "The lobster way" as the meaning of the name, and gave outdated install requirements. It is rewritten from the project’s current README, and the Harness record’s install lines are corrected with it. browser-use is no longer called the most popular library of its kind, a ranking its star badge already shows and keeps current. The Skills usage note no longer asks for a beta header that the Skills API has dropped.
  • The SAM 2 summary now attributes to its authors the claim that its dataset is the largest of its kind to date. The Chinese summaries no longer call it the first model of its kind, which the English never said.

Corrected

  • On the Arabic pages, numbered lists — the steps on the MCP page and the source list on each technology page — and the year in the footer were written in Western digits, while every other number on those pages is in Arabic-Indic digits. They now match the rest of the page.

Corrected

  • Eight of the fifteen quotations on the era pages were not quite what their speakers said, or not theirs alone. Four had no source that could be found and are replaced by words the same person did say or write: the Gödel line was a paraphrase of his theorem and is now a sentence from his 1951 Gibbs Lecture; Norbert Wiener’s warning about our dependence on machines is now his 1960 sentence in Science about the purpose put into the machine; the Lighthill passage is now a sentence from his 1972 report; and Demis Hassabis’s line is now his own account, given at Stanford in 2026, of DeepMind’s 2010 plan. Yann LeCun’s cake analogy now uses the words of his 2016 slide, without a second sentence no source contains. The 1955 Dartmouth proposal and the 2015 review of deep learning in Nature are now credited to all their authors rather than to one, and Herbert Simon’s 1960 prediction has back a word it had lost. The other seven were checked word for word and stand.

Updated

  • Each quotation on the era pages now shows where it was said, beneath the speaker’s name: the work, where and when it appeared, and a link. Where the work itself is not online, the link goes to a page that reproduces its words and is named after that page.

Corrected

  • Twenty paper summaries were still an older translation in eleven languages (Traditional Chinese, Japanese, Korean, French, Italian, German, Portuguese, Spanish, Russian, Thai and Vietnamese) and said things the English does not. The French and Russian AlphaGo summaries called Fan Hui a world champion; he was the European champion. GPT-4 was said to match or exceed human performance where the report says it approaches it, and other summaries credited papers with founding deep reinforcement learning or with an open-weights movement. Those twenty are translated again from the current English, and eight Simplified Chinese summaries that claimed a first the English does not are corrected too. Summaries that were only shorter than the English, and said the same thing, were left as they are.
  • Seven summaries were corrected in English as well, and then in every language. DeepSeek R1 was still called the first open-weight model with frontier-class reasoning, the claim removed from Liang Wenfeng’s entry the day before; it now says what its authors report, performance comparable to OpenAI o1. Scaling Monosemanticity called Claude 3 Sonnet a frontier model, where the paper calls it Anthropic’s medium-sized production model, and counted tens of millions of interpretable features where it reports millions. The 2013 DQN paper was said to reach human level on Atari games; it beat a human expert on three of seven. LLaMA now states the paper’s own comparison, Mamba is one of the most influential post-Transformer architectures of 2023 rather than the most influential, and the Logic Theorist and variational autoencoder summaries no longer claim a first or a sole foundation.
  • On the timeline, the Llama 3 entry said Llama 3.1 405B was the first open-weight model to match leading closed models. Its only source is the April 2024 announcement, three months before that model existed, so the sentence is removed. In six languages that entry, and in three the LLaMA entry before it, ended in a fragment of an older sentence left behind by an earlier replacement; both now end where they should, and the English LLaMA entry now says open-weight, as the other languages already did. The AlexNet and General Problem Solver entries are translated again from the English, and the Thai perceptron entry no longer calls it the first artificial neural network model.
  • The Hindi MCP page described a different, six-step protocol flow where every other language gives the four-step one, and its Skills note dropped one sentence and added another; both now follow the English. A few English and Chinese words left in the Japanese paper summaries are now Japanese, and a Russian note in this history now translates test-time compute correctly.
  • Three more timeline entries and one summary stated as fact what their sources do not. The Deep Belief Networks entry and its paper summary called them the first deep neural network trained successfully, and the summary credited the paper with contrastive divergence, which Hinton had introduced four years earlier; both now say what the paper shows. The Stargate entry called the plan the largest private-sector investment in AI history, a ranking its source does not make, and called an announced intention a commitment. Claude 3.7 Sonnet is now described as Anthropic described it, the first hybrid reasoning model generally available, rather than as the first model to support hybrid reasoning, and an unsourced sentence about AI products consolidating is removed. Korean and German text that used straight single quotes as quotation marks now uses each language’s own.

Corrected

  • Four people entries claimed more than their sources show. Norbert Wiener was called the first to warn about the dangers of AI; he was one of the first scientists to do so, and his own 1960 essay in Science sets his warning against a nineteenth-century novel by Samuel Butler. Ilya Sutskever was called a principal architect of a "post-2023" scaling paradigm; the entry now says what he did, from AlexNet to the research behind OpenAI’s GPT series. Liang Wenfeng’s R1 was called the first open-weight system with frontier-class reasoning; it now says what its authors reported, performance comparable to OpenAI’s o1. And Emily M. Bender’s paper is now one of the most widely cited critiques of the era, not the most cited.
  • Two timeline entries made claims that would not hold. Kimi K3 was called the largest openly downloadable language model "to date", which expires with the next release; it now says it was among the largest at its release. The investigation of OpenAI by US state attorneys general was called the first of its kind, a claim that could not be checked against the cited report; that sentence is removed.
  • Spacing now follows one standard in every language, taken from each language’s own authority and, where there is none, from Unicode CLDR. Japanese no longer puts a space between Japanese and Latin letters or digits, as the JTF style guide requires. Chinese puts one space between Han characters and numerals, as it already did beside Latin letters, while dates stay tight. French uses a narrow no-break space before semicolons, exclamation and question marks and %, and inside large numbers. German, Spanish and Russian put a no-break space before %. Spanish groups large numbers with a space rather than a dot, and Hindi groups them the Indian way.
  • The Japanese Provenance page showed a raw placeholder where the number of evidence links belonged, because a wording fix on 6 October had renamed it. It now shows the number, and a check now fails whenever a placeholder in any language differs from the English. The same check found that the Hindi note on the MCP page had dropped the date it was verified and added an instruction the English does not give; it now follows the English. A Japanese typo in the era 04 card is also corrected.

Updated

  • This history now shows its ten newest dated entries in full. Earlier ones are folded by month, and a month holding more than ten shows ten and folds the rest behind one more control. Nothing was removed: every entry is still on this page, and search engines and find-in-page still reach the folded ones.

Corrected

  • Every event page lists its sources, and under each one a sentence saying what that source establishes. Those 98 sentences were in English in all fifteen other languages, inside labels that had been translated around them. They are now written in each language. Source titles stay as their publishers wrote them.
  • On the Papers page, the Spanish and Italian entries for the OpenAI o1 system card kept the English term "system card", while the same page calls it a "ficha de sistema" and a "scheda di sistema". They now use those words.
  • The Portuguese pages still had a few Brazilian forms: "em um" and "em uma" where European Portuguese writes "num" and "numa", "status" for "estado", "revisados por pares" for "revistos por pares", "Acessado" for "Consultado" on every event page, and a paper title beginning with a gerund. They are corrected.

Corrected

  • Five entries had been corrected in English and not in twelve of the sixteen languages, so those languages said something different: the XCON, Fifth Generation Computer and LeNet-5 milestones, the speech-recognition milestone, and Yann LeCun’s entry. Several said "the first" where the English says "one of the first", one added a sentence the English does not carry, and the speech-recognition milestone described a different result. All five were translated again from the current English.
  • Three statements were wrong in English as well. The speech-recognition milestone credited Bell Labs, where the work was done at Microsoft Research and the University of Toronto. LeCun’s entry still called him Meta’s Chief AI Scientist, a post he left at the end of 2025. And the ResNet summary called it the first model to surpass human accuracy on ImageNet, a claim made that year for a different paper; it now says that an ensemble of ResNets won ILSVRC 2015. These are corrected in every language that carried them.
  • The Portuguese pages still wrote "learning" as "aprendizado", the Brazilian form, in about fifty places, among them "aprendizado profundo" for deep learning. European Portuguese, which this site writes, says "aprendizagem", and the pages now do, with "aprendizagem automática" for machine learning.
  • The Traditional Chinese pages wrote "network" as 網絡, the Hong Kong form, in about forty places. Taiwan usage, which this site follows, is 網路, and the rest of the site already wrote it. The recurrent neural network is now 遞迴神經網路.

Earlier revisions

October 2026 (7)

Corrected

  • The Simplified and Traditional Chinese pages had English words left where a Chinese word exists: "token" in eleven places in the timeline and paper summaries, where the site’s own glossary and technology pages use the Chinese term; "star" for the GitHub star count on the tools page; and "review comment" in the security review entry. They are now in Chinese, as are two more English phrases in the Japanese version of that entry. Names, terms such as diff and skip-gram, and quotations were left as written.

Corrected

  • Both items recorded here are late. They were made on 2 and 3 October 2026 and were not written into this history at the time. Our own check found the gap on 6 October, so they are recorded under that date rather than backdated, and the omission stays visible.
  • The Simplified Chinese Tools page wrote one word of the Anthropic Skills entry with a Traditional Chinese character, where the rest of the page uses Simplified forms. It now uses the Simplified character.
  • The Traditional Chinese summary of the LLaMA paper carried a sentence that neither the English nor the Simplified Chinese summary did, saying the release accelerated an open-weights movement. When that summary was corrected on 3 October, the sentence was removed, so it now says what the other languages say.
  • The Provenance page said one address had been retired; several have been, and the editorial page records each. It listed eight hosts as refusing every automated client: re-checked today, four of them now serve full pages and were taken off the list, so the links you cannot check fell from 16 to 12, and the page now gives the date of the check. It also said that the entries carry the dates on which those pages were opened by hand, which was not so, since most carry the date of the original sweep, and that sentence is gone.
  • The Japanese pages had English words left in seven places: "link" on the Provenance page and in this history, "raw" and "segmentation" in two paper titles, and "answer", "review comment" and "dense" elsewhere. They are now in Japanese.
  • The same page now says plainly that the pages on those hosts open normally for a person in an ordinary browser, and that only automated access is refused. Its Simplified and Traditional Chinese versions also joined independent sentences with semicolons in five places, as the Chinese versions of this history did until 2 October; four are now full stops and one is a comma.
  • The same reading of semicolons then reached the rest of the site. In the timeline, the paper summaries, the technology and tool records and the MCP page, semicolons joined independent sentences in 26 places in Simplified Chinese and 25 in Traditional Chinese. They are now full stops, or commas where the second part continues the first, and the semicolons between parallel clauses remain.

Corrected

  • On the timeline, the arrow in the Status and Era menus sat against the edge of the box, much closer than the text on the other side. The site now draws the arrow itself, set in from the edge, and on right-to-left pages at the opposite edge. The menus and their options are unchanged.

Corrected

  • The Portuguese pages are written in European Portuguese, but more than a hundred places used Brazilian forms: spellings such as Prêmio, autônomo and acadêmico; words such as usuário, arquivo, tela, equipe, treinamento and pesquisador; bilhões for a thousand million; and the progressive "está mudando". They now use the European forms, such as Prémio, utilizador, ficheiro, ecrã, equipa, treino, investigador, mil milhões and "está a mudar".
  • Minsky’s frame theory was dated 1972 and placed in the Birth of AI era, while the MIT memo it cites is dated June 1974. It is now dated June 1974 and appears in the AI Winter era, between the Lighthill Report and MYCIN. Its event page moved with it, and the old address no longer opens.
  • AlphaGo’s 4-1 win over Lee Sedol in March 2016 was supported only by the Nature paper of January 2016, which reports the earlier match against Fan Hui and so could not attest it. DeepMind’s own account of the Lee Sedol match is now its first source.
  • The papers page said its catalogue runs from 1936, while its first work is Gödel’s of 1931. The starting year is now taken from the catalogue itself.
  • The Provenance page said that a source published after the event it cites supports nothing. That was too broad, since histories written later are legitimate evidence. It now says that a source must support its claim and its date must allow it: one published before an event cannot report it, and a later account counts when it documents that event.

Corrected

  • The DeepSeek R1 milestone called R1 an open-source reasoning model. Its code and weights are under the MIT license, but its training data and training code were not published, so it is now described as an open-weight model and its license is named.
  • A 2026 milestone was labelled as open source "catching up" with closed models, while its own text declines to rank them. The label now says open weights at frontier scale.
  • Mark Zuckerberg’s entry said he decided to open-source LLaMA. It now says he decided to release the weights of Meta’s LLaMA models, which matches the LLaMA milestone corrected the day before.
  • The Hindi pages translated "frontier" with a word that does not exist in Hindi, in 15 places across the timeline and the paper summaries. They now use the standard Hindi words.
  • Four paper summaries applied the same label: the Simplified and Traditional Chinese summaries of LLaMA, Gemma 2 and DeepSeek-R1 called them open-source models where the English says open or open-weight, and the DeepSeek-V3 summary said "open-source SOTA" in English too, though its weights were first released under DeepSeek’s own model license. The DeepSeek-V3 summary had also drifted into two different claims across the languages; all sixteen now say what the paper reports, that it outperforms other open-weight models and approaches leading closed ones.

Corrected

  • Four milestones called a release "open source": Stable Diffusion, LLaMA, Llama 3 and Llama 4. This site distinguishes open weights from open-source software, and these releases published downloadable weights under licenses that restrict use; LLaMA’s was for non-commercial research only. Their titles and labels now say "open-weight release", or simply "release" for LLaMA, and the Stable Diffusion entry now mentions the use restrictions in its license.
  • On the Traditional Chinese Tools page, five tool entries and the comments in eight install and usage examples were shown in Simplified Chinese, because their Traditional Chinese text had never been written. It now has been. Thirteen translated paper titles in Simplified Chinese, Traditional Chinese and Japanese used a halfwidth colon after a Latin name, where the rest of the titles use the full-width one.
  • A few things had been left in English or written wrongly: the Indonesian and Hindi titles for Larry Roberts’s 1963 thesis; "vs" in the Portuguese and Spanish Deep Blue entry and in the comparison heading on the MCP page in five languages; and one wrong word each in the Japanese, Hindi and Traditional Chinese Llama 3 entry.
  • The Simplified and Traditional Chinese versions of this page followed the English semicolons one for one, so in 31 places each joined two separate sentences with a semicolon. Those now end with a full stop, or take a comma where the second clause continues the first. Semicolons between parallel clauses and list items remain.

Corrected

  • Three tool entries named models that have since moved on. Codex said it supported GPT-5 and the o-series, two generations behind what it offers now; it now says it uses OpenAI’s current models, chosen per session. Cursor was described as having capabilities similar to GPT-4 and as one of the most popular tools; it now says it offers models from several providers and can choose among them for each request. The security review action named its default Claude model, which its repository sets; the entry now says that instead, so it does not go out of date when the repository changes it.
  • The 1 July 2026 Cloudflare milestone said that from 15 September new domains "will block" training and agent crawlers by default. That date has passed, and this site has not checked whether the default took effect, so the entry now reports what Cloudflare said it would do.
  • The OLMo record listed its releases by date and had missed two of them, Olmo 3.1 and Olmo Hybrid. The list is removed rather than extended: what makes OLMo worth a record is that its training data, code, checkpoints and logs are published alongside the weights, and that does not change with each release.
September 2026 (18)

Corrected

  • Japanese and Traditional Chinese pages showed their Chinese characters in a Simplified Chinese typeface, whose shapes follow mainland conventions. Each now uses its own typeface first. On Windows, Simplified Chinese pages had the opposite problem and took a Traditional Chinese typeface; that order is corrected too.

Updated

  • The small labels on the technology pages were set in the fixed-width typeface used for code, and now use the same typeface as the text. Code keeps the fixed-width face, and non-Latin text inside it, such as a Japanese or Arabic comment, now takes the typeface its language uses elsewhere on the site instead of whatever face the device offered.

Corrected

  • Headings and short labels in French, Italian, Spanish, Portuguese and Vietnamese capitalised every main word, as an English heading does: "Père de la Théorie de l’Information", "Các Cột mốc Quan trọng". So did the Russian title of the page shown when a page is not found. These languages capitalise only the first word and proper names, and more than 550 strings now follow that rule. Indonesian was checked and left as it is, because its own spelling rules capitalise every main word in a heading.
  • Two people named together were joined by an English ampersand in thirteen languages, as in "阿隆佐·丘奇 & 阿兰·图灵". Each language now uses its own word for "and". The Simplified Chinese entry for Marvin Minsky also spelled his given name 马明; it now reads 马文.
  • The Hindi pages called their filters छन्नी, which is a sieve; they now use फ़िल्टर, the word Hindi software uses.
  • German capitalises every noun, so it was read by hand rather than by rule. Four kinds of word were capitalised where German writes them in lower case: in the title of the page shown when a page is not found, in a heading on the MCP, Agent, Skill and Harness page, in "künstliche Intelligenz" in two places where the site elsewhere writes the adjective in lower case, and in "bayessche" in a paper summary.

Updated

  • The era numerals are now set in the same typeface as the text around them, at the weight of the name beside them, where they had used a serif of their own. On the timeline cards the numeral was also a fixed size, larger than the title and the same on every screen and in every language; it now follows the site’s type scale, one step above the title, as on the home page.

Corrected

  • On the timeline and the era pages, a short piece of the vertical line showed above the first dot on narrow screens, and on wide screens when the page was zoomed in. The line now starts inside the dot. On narrow screens the year in each date label also sat off-centre, nearer one side of its box; it is now centred.
  • In Hindi, the timeline was called लिनिमाला, which is not a Hindi word: it was the Indonesian Linimasa carried into Devanagari. It is now समयरेखा, the ordinary Hindi word, in the menu, the page titles, the links and this history.
  • The short description on each of the five era cards on the home page used English-style capitals in eight languages, capitalising every main word as an English heading does: Russian read "Тьюринг и Рождение Вычислений". French, Italian, Portuguese, Spanish, Vietnamese, Indonesian and Russian now capitalise only the first word and proper names, and German no longer capitalises an adjective. The Russian card for the second era also now names the Dartmouth conference as the era page does.

Updated

  • On the timeline, each era’s numeral now sits at the end of the title’s line instead of in a column of its own beside the whole card. On a phone that column took about a sixth of the card’s width from top to bottom; the description now uses the full width.
  • The era numerals and the GitHub star badges on the tools page now take the colour of the text around them. The numerals had been set in the site’s link blue, which made them look like links; the badges used GitHub’s own cooler greys, which sat apart from the warm paper of every other page.
  • On the home page and at the top of each era page, the era numeral now stands beside the era name on one line, like a chapter number beside its title, instead of on a line of its own above it.

Corrected

  • Four tool subtitles, two in Spanish and two in Russian, wrote a GitHub star count into the sentence, and the figures had fallen out of date: they disagreed with each other and with the live count. They are removed. Every star count on the site now comes from the badge beside each tool, which is refreshed daily.

Updated

  • The five eras are now marked with the numerals I to V, their place in the sequence, where each had an emoji. The tools and the MCP, Agent, Skill and Harness page now use drawn icons in the site’s own ink: a product’s icon is taken from its official mark, and the concepts that have none are drawn for this site. Emoji looked different on every device, and the rocket beside the LLM era suggested a trajectory this site does not claim.

Corrected

  • Rewrote the two milestones for the second AI winter, 1987 and 1993. Both cited a single page, an AI history overview at aitopics.org, which no longer answers. Their evidence is now Nils J. Nilsson’s The Quest for Artificial Intelligence, published by Cambridge University Press, and both texts now say what that book supports. The 1987 entry had said personal computers made the Lisp machine market collapse and AI companies go bankrupt again; the book says Lisp machine makers lost out to other workstations, several AI companies closed, and DARPA’s funding for basic AI research fell between 1987 and 1989. The 1993 entry now says the downturn eased gradually, with no single end date.
  • Ten timeline events that state a reading of a period, among them both AI winters, the rise of statistical methods and the two summaries of early 2026, were labelled "Developing", which told readers they were still under way. They are now labelled "Interpretation": their sources support the reading, and a reading of a period is not a settled fact. "Developing" is kept for events that are still unfolding, and one event carries it. The content type shown on each event page was also an English word in every language, and it is now translated.
  • On the MCP, Agent, Skill and Harness page, the configuration file name claude_desktop_config.json lost its underscores and showed its middle word in italics, in every language; it now appears as written. The source type beside each citation on a timeline event page was also an English word in every language, and it is now translated.
  • The Indonesian text on diffusion models and on the DDPM paper used penderauan, which means adding noise, where the English says denoising. Both now say penghilangan derau.

Updated

  • The key terms on the technology pages now appear in each language. Names and acronyms such as BPE, vLLM and MoE stay as they are. Every other term takes the word this site’s own text in that language already uses, and each technology’s page shows the English beside it, because that is the word most of the cited sources use. Until now every term was shown in English in every language, which no decision had chosen.

Corrected

  • Repaired the evidence link for the 1961 Unimate milestone. The Henry Ford museum rebuilt its website and every record in its digital collections moved to a new address, so for two days the link reached a missing page. The museum record is the same one, for the machine itself, and it still states that the first Unimate was installed at General Motors in Trenton, New Jersey, in 1961 to unload a die-casting press. Charles found the new address by searching the museum’s collection by hand in a browser.
  • Rewrote the GPT-5.6 milestone. It said the release began on June 26, 2026, with a limited preview requested by the U.S. government, and that OpenAI did not want such a process to become the default. The help article that carried those claims no longer does, and nothing else here supported them. OpenAI’s announcement of July 9 names a limited preview with expert organizations and trusted partners, and no government request, so the milestone now states what that announcement supports: the family became generally available on July 9, after that preview. Charles opened both pages by hand in a browser.

Corrected

  • On the MCP / Agent / Skill / Harness page, the heading inside each of the three example blocks was set larger than the section heading above it, and screen readers announced it as a new top-level section of the page. They now sit beneath the section they belong to. This had been fixed once in June and was lost when the page was restructured, so a check now holds it. In Vietnamese, the Agent heading now explains the term in Vietnamese after the dash, as the Skill and Harness headings already did.

Updated

  • The technology pages now name nine of their subjects in each language, where every language had shown the English name, including distributed training, quantization and tokenization. Each name is the word this site’s own text in that language already used for the concept; where no text did, the languages whose pages keep English technical names keep them here too. Names and acronyms such as Transformer, PyTorch and LoRA stay as they are, and every title now records which of the two it is, so a new one cannot fall back to English unnoticed.

Corrected

  • Corrected a Russian word standing in the Japanese text of the note about the Hindi figures: it read "of two" where the sentence needed the Japanese for two places. Charles asked whether the exemptions on yesterday’s new check had themselves been examined, and they had not — one had been fixed where it was stumbled upon and the rest were written down as a thing to do. Examining them found this. Yesterday’s rule also turned out to be half a rule: it read Chinese or Japanese characters leaking into a Latin language and could not see the reverse, nor one Asian script leaking into another, so a Russian word in Japanese and a Japanese character in Korean were both invisible to it. It now reads every writing system against every language, and the About page — which quotes other languages’ characters as its subject and so cannot be checked that way — is held by a different test: a quotation appears in several languages’ versions of the same note, and a leftover fragment appears in one.

Corrected

  • Corrected the names of the five eras. Each carried two names — one stored with the era’s own content, one stored again with the home page’s cards — and in 29 of the 80 era-and-language pairs the two had drifted apart. Some differences were small: "The Revival" against "Revival". Others renamed the era outright: in Russian it was "Эра больших языковых моделей" in one place and "Эпоха LLM" in another. A reader moving from the home page to a paper card saw the era called something else. The second copy has been removed rather than reconciled, so the home page now reads the name from the era itself and the two cannot drift again. Where they disagreed the shorter form usually won — a definite article had been added to the fourth and fifth eras only, in six languages — and where a language has its own short form for artificial intelligence, the era names now use it.
  • Corrected the subtitle of the fifth era, which said two different things depending on the language. Six languages described it as large language models and agentic systems; the other ten called it the Renaissance of AI. Those are not translations of one another. This site uses "Renaissance" as an editorial metaphor for the project and says so plainly — it is not a claim that progress in AI is even or without limits — so ten languages were making, in the site’s own voice, a claim the site states it does not make. The English had been rewritten at some point and the translations had not followed. All sixteen now describe the era rather than praise it, and a check refuses the metaphor in an era subtitle so that putting it back has to be a decision.
  • Corrected the author lists of three research works, which named different people depending on the language. The entry for Scaling Monosemanticity carried ten different lists across sixteen languages, and eight of them included people who did not write it — names assembled rather than copied from the paper. Its publisher asks to be cited as Templeton and others, and all sixteen now say that. The InstructGPT entry named thirteen of its twenty authors in seven languages, with nothing to show the list had been cut — which a reader cannot tell from a complete one — and all twenty are named now. The GPT-3 entry misspelled one author in two languages and, in a third, gave eight of thirty-one as the whole list. Two new checks refuse an author named by one language alone, and a shortened list that does not say it is shortened.
  • Corrected two things a punctuation sweep turned up, both of which put one language’s characters into another’s sentence. The Chinese and Traditional Chinese text of the era-subtitle note published earlier today used halfwidth commas, semicolons and colons where Chinese takes their full-width forms; Charles found that by reading the page. And one milestone in the fifth era left the Chinese phrase for "advertising placement" sitting untranslated inside the Portuguese sentence, with no space around it, where every other language had translated it. The localization check had a rule for ASCII parentheses beside Chinese or Japanese text and one for the Japanese colon, and none for the comma — the mark prose uses most — nor any rule that read prose for the script it is written in. Three checks were added. One of them is proven from the other side: a Latin reference such as Gödel, K. standing beside Chinese text must not trigger it.

Corrected

  • Corrected three figures in the Hindi text, where खरब — one hundred billion — stood where trillion was meant: Kimi K3’s total parameter count, in both the timeline entry and the technology record, and Samsung and SK Hynix’s memory investment. The numeral was right each time and the magnitude word was one step too small. The check that compares every stated figure against the English had no Hindi vocabulary when Hindi shipped, so it read none of that language’s figures and reported nothing; the words were added while preparing the next language, and it found all three on its first run.
  • Corrected seven places where the Arabic pages showed English: the timeline page title, four section headings on the era pages, and the two category labels in the tool catalogue. Each was a table of translations with no Arabic row, so the page fell back to English in silence — and a reader cannot tell English-by-design from English-by-omission. The check written to prevent exactly this had reported none left: it had no idea what a comment was, so an apostrophe inside one silenced it for the rest of that file. It is rewritten to parse the source rather than scan it. Found by a reader on the live site.
  • Corrected the arrows on the technology and era pages, which pointed forward in a page that is read from right to left, and so pointed backward to the person reading it.
  • Translated the two small labels on the technology page — the one above the title and the one above the application section. They were written into the page rather than into the translations, so every language showed them in English.
  • Corrected how Arabic is set. Its letters join to one another, and two devices used elsewhere on this site break those joins: letter-spacing, which pulls letters apart, and the monospaced face, which contains no Arabic and so fell back to one that puts every letter in its own cell. Both are switched off where they fell on Arabic words, and left alone where they fall on digits, years, dates and Latin technical terms.
  • Corrected the star count in the tool catalogue, which read ١٤٧k in Arabic — Arabic-Indic digits with a Latin abbreviation for thousand. A number carrying a unit is a formatted measure rather than a count in a sentence, so it is now written the same way in every language.

Updated

  • Added Hindi, the site’s fifteenth language. Everything a reader sees is published in it, and it reads the same dates, sources and identifiers as the other fourteen, because those live in one language-neutral layer. It needed one thing the others did not: Devanagari paints its vowel marks above and below the line, about 30% taller than Latin at the same size, so the tightest line spacing on the site clipped them — the date badge on the timeline cut 1.87px off a month name. Three spacing values are raised for Hindi alone, and every other language was measured to be untouched. It is drafted with AI assistance and has not been read end to end by a native speaker; the Provenance page says so in Hindi rather than only about it.
  • Added Arabic, the site’s sixteenth language and the first that reads right to left. Everything a reader sees is published in it, and it reads the same dates, sources and identifiers as the other fifteen, because those live in one language-neutral layer. Its numbers are written in Arabic-Indic digits — ٠١٢٣٤٥٦٧٨٩ — and that is what made most of the work: several automated checks read only 0–9, and against another set of digits they would have missed the very thing they exist to catch, some by failing and most by going quiet. Each was taught both sets before a word of Arabic was written. Three faults only a right-to-left page could show were fixed with it: the back link’s arrow pointed forward, an author’s name lost its closing full stop to the far end of the line, and every shell command was right-aligned. It is drafted with AI assistance and has not been read end to end by a native speaker; the Provenance page says so in Arabic rather than only about it.
  • Made the language menu smaller. It lists sixteen languages now and filled most of the screen; it shows about seven at a time and scrolls for the rest. The button that opens it also became narrower: it is sized by its widest entry so that it cannot change width when you switch language, and Indonesian was listed by its full name, which made that one entry 110px against the next longest at 65px. It is listed as Indonesia now, as language menus usually list it.
  • Rearranged the labels on each card in Selected works. The row above the title now carries the year and the publication type; the era moved down beside the authors. The era label is a translated era title, so how much room it needs depends on the language — "Early Theory" is 87 pixels and its Russian equivalent is 208 — and against a 307-pixel row that pushed the labels onto a second line on 59 of 96 cards in Russian, 42 in Spanish, 37 in German. The cards that did not wrap began their title 20 pixels higher than the ones that did. The row now fits in all sixteen languages at every width tested.
Show the rest (8)

Updated

  • Added a page called Provenance, linked from the foot of every page in all 13 languages. It states what the claims on this site are worth: how many milestones rest on a single source, how many dates are known only to the year, how many evidence links point at hosts that refuse every automated reader, and what has been left out on purpose. It is addressed to a reader who cannot follow the links and check for themselves, which is increasingly often a machine. Every figure on it is computed from the catalogue as the page is served rather than written down, so it will report the record thinning as readily as it reports it growing. An earlier attempt at a page for such readers, which retold the history in the first person as though the software remembered it, was removed in May; this one says only what can be checked.
  • The Provenance page now states which translations have been checked. Five of the 13 languages carry priority editorial review; the other eight are drafted with AI assistance and have not been read end to end by a native speaker. That was true from the start and the site did not say it — the earlier decision chose a uniform reading experience over telling you, and named that cost at the time. If you are reading one of the eight, the page says so in your language rather than only to everyone else. It also says what a thinly reviewed translation can and cannot get wrong: register, idiom and nuance are at risk; dates, sources and identifiers are not, because they live in one layer no translation touches, and every figure stated in any language is compared against the English.
  • Added Indonesian, the site’s fourteenth language. The eras, milestones, selected works, technology records, tool entries and page chrome are all published in it, and it reads the same dates, sources and identifiers as the other thirteen, because those live in one language-neutral layer. It is drafted with AI assistance and has not been read end to end by a native speaker; the Provenance page states that in Indonesian rather than only about it.

Corrected

  • Corrected the period labels on the five era cards of the home page, which joined their two decades with a hyphen in every language. A hyphen joins parts of a word; a range takes its own mark, which differs by language — an en dash in Latin scripts, Cyrillic, Thai and Vietnamese, a 一字线 in Chinese, a wave dash in Japanese, a tilde in Korean. The era pages had carried the right marks the whole time, so the two pages had been stating the same period two ways. Norbert Wiener’s dates were the same mistake one page over: sixty-two lifespans in the catalogue carried their language’s mark and thirteen carried a hyphen, all of them his.
  • Corrected the Indonesian MCP page, where line breaks inside the Agent, Skill and Harness sections were printed as a raw escape sequence instead of starting a new line. The text had been escaped one time too many on its way into the file; no other language was affected. Found by a reader on the live site.
  • Corrected the publication labels on the Selected works cards, which could break in the middle of a phrase inside their own outline — Vietnamese read "bản thảo công bố" above "trước". The row now wraps between labels and never inside one. Measured across all 14 languages, the break affected 1,602 cards at the four verification widths; it was worst in Russian and Spanish and absent in English above the narrowest, which is why reading the English page never showed it.

Corrected

  • Corrected the Korean wording of two entries in the tool catalogue, MCP and Harness, so that both use the plain declarative form that the rest of the encyclopedia writing on this site uses. The two describe subjects that the MCP, Agent, Skill and Harness page also covers, and that page has been written in the plain form since 2 September. The ten product entries in the catalogue keep the polite form, which is the voice of a catalogue rather than an inconsistency. Only Korean was affected, and nothing that the entries say has changed.
  • Adjusted spacing throughout the site. The shared scale offered only two steps below 12 pixels, so smaller gaps had been set one at a time and 23 slightly different values were doing that work. Two steps were added, at 6 and 10 pixels, and 51 of those gaps now use them. Nothing moved by more than about one pixel at a time, no page layout changed shape at any screen width, and no heading or label began to wrap that had not wrapped before.

Corrected

  • Corrected the order in which milestones appear on the era page. Two events are dated to a span — the first half of 2026 — rather than to a day, and the era page sorted them to the very end of the list, after September, while the timeline page placed them in the first half of the year. The same events, two different orders. Both pages now use the same sorting method. Fifteen milestones changed position on the era page, two of them by about seven months. Milestone addresses derive from the underlying order of the data, which did not move.

Updated

  • Two milestones from September 2026 were added to the timeline, both about where improvement comes from. In the first, OpenAI’s system card places GPT-6 Astra at the Critical threshold for cyber capability under the company’s own framework — the company’s assessment of its own model, in a card that never uses the word AGI. In the second, ARC Prize’s verified ARC-AGI-3 results give the same model 62.71% with the Standard harness and 98.55% with a Provider Adapter harness on the same evaluation set. Both numbers are recorded, because a benchmark score measures a model together with its scaffolding.
  • The selection grew from 94 works to 96, both on recursive self-improvement — a subject on which the catalogue previously held nothing. The first is a survey of 1,250 papers from 2024 to 2026 whose own finding is that nearly all of them study the bounded kind of self-refinement, improvement against a fixed external standard, rather than the open-ended kind that also rewrites the standard of improvement itself. The second is a staged roadmap that carried no publication venue when it was written up, which its entry says. Unlike the recent additions, these two were not works the site already cited without holding; they were selected after a deliberate survey.

Corrected

  • The two corrections below are late. They were made on 7 and 8 September 2026 and were not written into this history at the time. Our own check found the gap on 14 September, so they are recorded under that date rather than backdated, and the omission stays visible.
  • Corrected the links between technology entries. Sixty relationships sat in one field that reads as mutual, and forty-six of them pointed one way only: a reader on the key-value cache entry could not reach tokenization, speculative decoding or test-time compute, each of which named it. The relationships are now split into what an entry builds on and what builds on it, the second derived from the first so the two lists cannot disagree, with a much smaller set of genuinely mutual pairs. The relationship diagram on the MCP page was also redrawn on 7 September so its connecting lines meet exactly; they had drifted apart when the page was zoomed.
  • Corrected the colour of three links in the MCP text. They appeared in the browser’s default blue in all thirteen languages, thirty-nine places in all. Every other colour on this site comes from its own palette; these had escaped it because of the way that text is inserted into the page. Underline spacing, which had drifted to three different values across five surfaces, is now one.

Updated

  • The content on the technology page used to begin ten pixels to the right of the back-to-home link above it, while the papers page began flush with it — two pages that belong together, one of which looked misaligned. Both are now flush. The technology cards remain six to a row, each slightly wider.
  • What found these is a check run against this history itself: a change a reader can see is meant to appear here the same day. It has now found a gap twice, which is why it is kept.

Corrected

  • Corrected the separator in the four section headings on the MCP page. The dash standard set on 1 and 2 September was described here as applied in all thirteen languages; in fact these four headings still used a hyphen in twelve of them, where a name and its gloss are one phrase and take a sentence dash. The Russian text had been right all along. All thirteen now match.

Updated

  • Only the range part of that standard had been checked automatically; the rest relied on someone noticing. A hyphen standing alone between spaces is now checked too, so a heading cannot quietly fall outside the standard again.
  • The MCP page has moved to /mcp-agent-skill-harness. Its title has always named four things: MCP, Agent, Skill and Harness, while the old address named only the first. That address has been retired rather than redirected, so a link to it saved before today no longer resolves. The page itself is unchanged, and the Technology page links to it.
  • Two technology records changed address as well: glm-5-model-family and kimi-k3-model-family are now glm-model-family and kimi-model-family. The other seven model records never carried a version number; these two were the exception. Those old addresses have also been retired rather than redirected, twenty-six in all. The records themselves are unchanged, and their titles still read GLM-5 and Kimi K3, because that is the generation they currently cover.

Corrected

  • Corrected the authorship of the 1956 Logic Theory Machine entry, which credited J. C. Shaw. That paper is by Newell and Simon; Shaw is a co-author of the 1957 and 1958 papers instead. The credit appeared in ten language blocks and has been removed from all of them. The same entry was also missing its subtitle in two languages, which has been restored.

Updated

  • Where the same work is cited from more than one page, its title is now reproduced exactly as published rather than re-capitalised. Six documents had been filed under two spellings, among them Nature’s "Mastering the game of Go with deep neural networks and tree search", which appeared title-cased elsewhere on the site. Two of the six were introduced by our own work four days earlier.

Corrected

  • Two of the items recorded here are late. They were made on 1 and 2 September 2026 and were not written into this history at the time. Our own check found the gap on 8 September, so it is recorded under that date rather than backdated, and the omission stays visible.
  • Corrected the Korean on the MCP page. It addressed the reader in the polite register this About page uses, while the rest of the encyclopedia — the era essays, the technology and research entries — states things plainly. Twenty-one sentence endings across eleven fields were changed on 2 September 2026, including two that told the reader what to do; an encyclopedia says what is done rather than instructing.
  • Corrected the papers linked from two technology records. Mixture of Experts linked only DeepSeek-V3, a model report rather than the origin of the concept, and now also links the 2017 sparsely-gated mixture-of-experts layer. The diffusion record held DDPM and Stable Diffusion but not the 2015 paper they rest on, which has been added.

Updated

  • Settled which dash-shaped mark does which job, and applied it in all thirteen languages on 1 and 2 September 2026. A range now takes each language’s own mark rather than a borrowed one: an em dash in Chinese, a wave dash in Japanese, a tilde in Korean, and an en dash in the Latin-script languages, Cyrillic, Thai and Vietnamese. A name made of two people, such as Church-Turing, takes a hyphen. The parts of a page title are separated by a vertical bar. Two citations keep marks that break this rule, because a citation is reproduced as published.
  • The selection grew from 73 works to 94. Every addition was already cited on this site and not held by it: ten where a timeline milestone is about the paper itself, and eleven that a technology entry rests on. Ten technology entries previously linked no paper at all, leaving a reader nowhere to go from them; four now remain, and those are honest — the KV cache has no single origin paper, and the other three cite documentation and repositories. Technology pages now separate what an entry builds on from what builds on it, the second derived from the links that point at it.

Corrected

  • Corrected the GLM record, which presented GLM-5.2 as the newest model in the family. Z.ai released GLM-5.3 on 18 August 2026 and GLM-5.3-Flash on 26 August. The record now gives the family’s release history rather than naming one current version, and it notes that openness is no longer uniform: GLM-5.2 publishes its weights under MIT, while GLM-5.3 ships under a custom licence of its own.

Updated

  • Model records now say when a family last changed rather than which version is current. The GLM entry pinned one release and expired three weeks later; the OLMo entry, which lists its releases in order, did not. A release history is the durable form.
August 2026 (9)

Corrected

  • Corrected the Gelernter 1959 entry, which was wrong in almost every part. Its title named a syntactic projection problem, which is not a paper Gelernter wrote; the work is A Note on Syntactic Symmetry and the Manipulation of Formal Systems by Machine, in Information and Control 2(1): 80-89. Its link pointed at RAND paper P-1586, a 1959 Monte Carlo study of air base inventory control with nothing to do with artificial intelligence. Four languages named four different journals between them. The entry now carries the real title, a DOI checked against the registering agency, and one venue in all thirteen languages.
  • Corrected where McCarthy’s Programs with Common Sense was first published. Traditional Chinese and Japanese named MIT Press, which is the 1968 reprint in Semantic Information Processing rather than the first printing. McCarthy’s own page states that the paper was given at the Teddington Conference on the Mechanisation of Thought Processes and printed in those proceedings; all thirteen languages now say so.
  • Corrected how the MCP page described a Skill. It said a Skill extends Claude Code capabilities, which scopes a concept to a single product. The SKILL.md format is published, and Claude Code is one host that loads it among others. All thirteen languages now state the concept first and name the product as one implementation.

Updated

  • The catalogue now prefers a resolvable identifier where one exists. Gelernter 1959 moved from a hosted copy to a DOI, which means the title can be checked automatically against the registering agency — the check that would have caught this entry years ago.
  • Bibliographic detail no longer depends on which language you read. One language carried Nature 323(6088): 533-536 while another carried only Nature. All thirteen now carry the fullest form: volume, issue and pages for a journal article, and the arXiv identifier in parentheses where a preprint also exists. Sixty-four of the seventy-three works changed.
  • Added Reasoning and Agent to the subject filters. The ten existing topics predate the shift those works represent, so chain-of-thought prompting, the o1 system card, DeepSeek-R1 and s1 all sat under Algorithm. Six works are tagged by what they are about rather than what they mention. Added AlphaEvolve (2025), whose result the catalogue already cited through a later work without holding the publication itself.

Corrected

  • Corrected the source given for Scaling Monosemanticity. The Traditional Chinese entry cited arXiv:2403.07830, which is a paper on critical planar Brownian loop-soups, and the Japanese entry called the work an arXiv preprint. It is published on Transformer Circuits and has no arXiv identifier at all. All thirteen languages now name the same place. A sweep confirmed the rest of the catalogue is sound: forty other entries print an arXiv id beside a title, and every one belongs to the paper it sits on.

Updated

  • Five works were added, each checked against its primary source before being written: Co-Scientist in Nature, which this project’s timeline had cited since May while the catalogue did not hold it; the OLMo 2 and Olmo 3 reports; the GLM-5 technical report; and a note from this month that moves the matrix multiplication exponent using AlphaEvolve.
  • The OLMo record now runs to Olmo 3, released in November 2025. It had stopped at OLMo 2 in November 2024 — the one place in the technology graph that names a version history, and so the one that could fall behind.
  • A fifteenth technology concept covers reasoning and test-time compute. The catalogue already carried three papers and two timeline entries on it while the concept layer had nowhere to put it.

Corrected

  • Corrected what the first Unimate actually did. Seven of the thirteen languages said it welded parts into place; English and Spanish said it stacked them; Japanese and Korean said neither, carrying a single sentence where the others carried three. The Henry Ford holds the machine itself, and its record says it was installed at General Motors in Trenton, New Jersey in 1961 to unload a die-casting press. All thirteen now say that, and the entry cites the museum record for the machine rather than a collection record that did not show it.
  • Corrected the entry on OpenAI’s recapitalization. It described the company as moving out from under nonprofit control, when OpenAI’s own announcement says the OpenAI Foundation continues to control the for-profit business, and that it is the for-profit — OpenAI Group PBC — that is the public benefit corporation, not OpenAI as a whole. Rewritten in all thirteen languages.
  • Corrected the Thai text of the Microsoft Frontier Company entry, which stated $25 billion where the figure is $2.5 billion. The numeral was right and the magnitude word was one step too large.

Updated

  • Three sources now point at a primary or canonical artifact. The 1998 LeNet paper and the 1991 mixture-of-experts paper cite DOIs, whose registered titles can be checked automatically, and the Microsoft entry cites Microsoft’s own announcement rather than a wire copy hosted on a stock-news site.
  • Removed a retired link to archived Hugging Face documentation. That record keeps its current documentation link.
  • Ten entries carry a new review date because their sources were opened by hand. These hosts refuse automated clients, so a link audit can only record them as unseen rather than sound.

Corrected

  • Corrected a citation on the perceptron convergence entry that pointed at an unrelated paper on coding theory. It now cites Rosenblatt’s own Principles of Neurodynamics. The link had worked throughout, which is why the error went unnoticed for so long.
  • Repaired seven further sources whose pages had moved or been withdrawn — at MIT, AAAI, DARPA, CERN, Stanford and RAND — together with a DOI for Pearl 1986 whose final character was wrong.

Updated

  • Every replacement was verified against the registering agency before publication, and points at the primary artifact where one exists: CERN’s original project page rather than an article about it, and MIT’s own timeline rather than a summary of it.
  • Review dates moved only on the six entries that were genuinely re-checked. All 184 evidence links are now audited on a schedule instead of never.

Corrected

  • Replaced a timeline entry whose only source was published three months before the event it described, and re-grounded four 2026 entries on sources that carry their claims, including the European Commission’s own page for the EU AI Act date and the Nature paper for the AI Co-Scientist result.
  • Removed an outdated list on the MCP page that named six commands as built-in when only one of them is. It had already been superseded on the page itself and was no longer displayed, but it was still being sent with every request.
  • Corrected the Tools page, which described AI Agent and Skill as features of a single product and gave an installation step for a product that does not exist. Both entries now describe the concept and name several implementations across vendors.

Updated

  • Added three figures and two research works from 2025, and replaced rather than deleted a milestone so that existing links to later entries keep working.
  • Added OLMo as a ninth open-model case, the one that publishes training data, training code, intermediate checkpoints and logs alongside the weights, so the difference between open weights and open training can be compared rather than only asserted.
  • Recorded review dates per entry instead of one date for everything. A date now shows when that item was actually checked, and most entries kept their older date.

Localization

  • French prose now stores a non-breaking space before colons and semicolons throughout the site, so the mark cannot wrap to the start of the next line.

Corrected

  • Fixed the Papers page header, where quotation marks in Chinese, Japanese, and Korean were shaped by a Latin fallback at about half the width of an ideographic character.

Updated

  • Extended the shared spacing scale across every route, with a before-and-after layout comparison confirming no visual change.

Corrected

  • Removed the line-length caps that pushed short introductions, card descriptions, and summaries onto an extra line ending in a stray fragment.

Updated

  • Prose now flows to each page’s own content column. A reading measure is kept only for the sustained homepage introduction.

Corrected

  • Raised undersized mobile supporting text and constrained overly long desktop prose lines.

Updated

  • Unified semantic font sizes, leading, paragraph rhythm, and reading measures without changing the visual identity.

Localization

  • Separated CJK optical text sizing from spacing and layout geometry, with source and build checks across all 13 languages.

Corrected

  • Changed English prose to the project’s explicit ASCII double-quote style ("...") in source content, visible pages, metadata, and GEO exports.
  • Fixed exact-date timeline badges that could overlap the marker and spine on desktop.
  • Restored the hyphen in compact exact-date badges and removed nonstandard spaces inside Chinese and Japanese calendar expressions.

Updated

  • Normalized the multilingual source corpus and unified quote handling across Home, Era, Technology, About, MCP, Timeline, Papers, Tools, and shared metadata.
  • Standardized timeline dates as a year plus an honest precision label, with localized detail-page dates and valid machine-readable values.
  • Limited translating card and link hover motion to fine pointers; touch devices now use non-moving press feedback.

Localization

  • Added field-aware checks that distinguish natural-language prose from executable syntax and enforce each language’s canonical quote pair.
  • Normalized spacing between Chinese prose and Latin product names while preserving numbers, code, paths, and Japanese conventions.
  • Added source and rendered-output checks for compact CJK calendar spacing while preserving Korean, code, URLs, paths, and identifiers.
July 2026 (5)

Corrected

  • Re-audited technology dates and descriptions against primary sources, separating original concepts, public releases, and later implementation milestones.

Updated

  • Curated eight open LLM families and sorted them by release period: Llama, Qwen, Mistral/Mixtral, DeepSeek, Gemma, gpt-oss, GLM-5/5.2, and Kimi K3.

Localization

  • Fixed narrow-screen primary navigation so all 13 languages remain single-line, reachable, and route-aware without crowding.

Corrected

  • Clarified that the MCP, Agent, Skill, and Harness grouping is an editorial model, and aligned MCP wording with the current protocol boundary.

Updated

  • Added a sourced technology graph with 14 foundational technologies and 5 open-model cases, connected to papers, eras, application concepts, and tools.

Localization

  • Published the overview, detail pages, terminology, evidence, and model-openness distinctions in all 13 supported languages.

Corrected

  • Fixed quotation pairing across formatted prose and completed locale-aware quotation marks in the full GEO export.

Updated

  • Improved timeline heading semantics and secondary-text contrast; added edge security headers and refreshed supported build tooling.

Localization

  • Added 13-language regression checks for quotations that span bold text or links.

Corrected

  • Corrected the reviewed Skills, Superpowers, Browser Use, OpenClaw, MCP, Meta-Manus, GPT-5.6, early-history, AI-winter, and CIFAR wording across all supported languages.

Updated

  • Added sources for all 89 events, a 1,300-URL sitemap, event-specific social metadata, and stronger CI checks.
  • Reframed Aphrodite as an editorial project that evolves with AI; removed the duplicate footer email, retained About as the correction channel, and added the original Chinese project epigraph to the homepage.

Localization

  • Completed a 13-language punctuation, terminology, tone, and continuity review; removed English fallback copy from Papers, MCP, Tools, and About; added an automated localization quality gate.

Corrected

  • Corrected ImageNet, LSTM, the EU AI Act, and the 1955/1956 Dartmouth distinction.

Updated

  • Updated MCP, Claude Skill, Codex CLI, and timeline evidence documentation.

Localization

  • Moved dates, status, organizations, and sources into a shared fact layer.