You are {{VAR:MODEL_ID}}. Today is {{VAR:RUN_DATE}}. You have web search: use it. Write a blog article in English, in Markdown: "The history of large language models (LLMs)", from the first statistical language models to today ({{VAR:RUN_DATE}}). Audience: curious readers who are not AI specialists. ## Content Cover these eras, in this order. You choose the milestones, the section titles and how much space each era gets: 1. Before neural networks: statistical language models. 2. Neural language models and word embeddings. 3. Attention and the Transformer. 4. Pretraining and scaling: BERT, the GPT series, scaling laws. 5. ChatGPT and the assistant boom. 6. Open-weight models, multimodal models, reasoning models, agents. 7. Where things stand on {{VAR:RUN_DATE}}: the main developments of the 12 months before that date, each one dated and sourced, then what they change for ordinary users. Explain them; do not just list version numbers. Do not tell the story of a single company: cover the major labs (OpenAI, Google, Anthropic, Meta, Mistral, DeepSeek, Alibaba…) where they matter. Include a "Timeline at a glance" table: Year | Milestone | Why it mattered. ## Writing - Define each technical term (token, parameter, pretraining, Transformer…) inside the sentence where it first appears, e.g. "early systems split text into *tokens*—small units such as words or parts of words—and counted them". Never as a separate "X is a…" sentence. - Vary your sentences: do not open sentence after sentence with "In [month year], …". - No opinions or predictions presented as facts. A trend needs a source like any other claim. ## Facts and freshness - Your training data stops before today. Anything more recent, you only know through web search. - Check that every "latest", "largest", "first" or "current" claim is still true on {{VAR:RUN_DATE}}. If it is out of date, update it and cite a source. - Write absolute dates ("as of October 2026", "in March 2023"). Never write "currently", "recently" or "this year". - For a product launch, give the launch date, not the date of the paper that describes it. - Never remove or "correct" a sourced fact because you do not know it or cannot open its source: it may be newer than your training data. Change it only if a page you actually opened contradicts it, and say so in your changelog. ## Citations and sources (mandatory) - Cite each claim once, with a bracketed number at the end of its sentence, e.g. [3]. Never after every name, date or number. - Never renumber existing sources: add new ones at the end of the list. - Write every new entry exactly in this format, bold included: `3. **[Page title](https://...)** — Publisher, publication date. Found by **{{VAR:MODEL_ID}}** on **{{VAR:RUN_DATE}}**.` - Never change the "Found by" of an existing entry. - After you have opened and read an existing source, set its mark, replacing any previous mark: - `✅ Checked by **{{VAR:MODEL_ID}}** on **{{VAR:RUN_DATE}}**` if it supports the claim; - `❌ Dead link — checked by **{{VAR:MODEL_ID}}** on **{{VAR:RUN_DATE}}**` or `❌ Does not support the claim — checked by **{{VAR:MODEL_ID}}** on **{{VAR:RUN_DATE}}**` otherwise. - If you cannot open a page, do not mark it ❌ and do not remove what it supports: leave the entry as it is and write "could not open [n]" in your changelog. - Keep ❌ sources in the list, but no claim may rely on them any more: find another source or rewrite the claim. - A source the text no longer cites stays in the list, with `⚪ No longer cited` added. - Only list pages you actually opened. Never invent a URL, a title or a date. ## Images - 3 to 6 images, each showing something its section talks about (a person, a diagram, a machine of the era). No decorative images. - Prefer Wikimedia Commons. Only use an image URL copied exactly from a page you opened (ideally the image's Commons file page). Never build, guess or shorten an image URL. - Put the caption on the line right below the image, with no blank line between them: `` `*Caption. Credit: author, license, via [Wikimedia Commons](link to the file page).*` - If you are improving an existing version, open every image URL: replace any image that does not load or does not fit its section. ## Format and length - One H1 title only. Right below it, this line: `*~N min read · Facts checked as of {{VAR:RUN_DATE}}*` N = word count of the body ÷ 230, rounded up. Recompute it whenever the length changes. - The body (everything before "## Sources") must not exceed 2,500 words. There is no minimum: never pad. If you add something, cut something weaker elsewhere. - Markdown only: no HTML, and no code fence around the article. The article ends with "## Sources". ## If you are improving an existing version - Work in this order: first check the sources and images, then out-of-date facts, then everything else. - Your changelog lists only what you actually did, including what you could not check. It is shown next to the real diff. ## Final check before answering - [ ] Exactly one H1, with the reading-time line right below it, recomputed - [ ] Body is 2,500 words or fewer - [ ] Section 7 gives dated, sourced developments from the 12 months before {{VAR:RUN_DATE}} - [ ] One citation per claim, at the end of its sentence - [ ] Every definition sits inside its sentence - [ ] 3 to 6 images, each related to its section, with the caption on the line right below - [ ] "## Sources" is last, in the bold format; no "Found by" was changed; no ❌ on a page you could not open - [ ] No invented URL, title, date or image URL; absolute dates only - [ ] The changelog matches what you actually changed