This timeline is curated, not comprehensive. It traces the through-line from classic AI to language models to agents using the events that actually bent the curve — and that means leaving a great deal out.
What earns a place
Three things tend to make the cut: landmark papers that introduced an idea still in use; models that set a new bar or shifted what the field believed was possible; and products or protocols that changed how AI is built or used. Where a primary source exists — an arXiv paper, an official model card, a launch announcement — it's preferred over secondary coverage, and every entry links to one.
What's left out
Not every release makes it. The timeline favors the first or most consequential instance of an idea over its many follow-ups. Size and tier variants (mini, flash, lite, distilled), most incremental API updates, individual benchmarks, and tooling are generally omitted unless they marked a genuine shift. The recent flagship point releases are a deliberate exception — kept to show how compressed the 2025–2026 cadence became — but even there the smaller variants are summarized rather than listed one by one.
How it's organized
Every event belongs to one of three eras — classic AI, language models, agents — which drive the color coding, and is typed as a paper, model, or product. The Recap view groups related events into a smaller set of milestones; Detailed shows them all; References lists every source by domain; and By Year plots the rhythm of activity over time.
On dates and judgment
Dates are best-effort and sourced as precisely as the record allows — a full day where it could be confirmed, otherwise the month or year. Inclusion is ultimately a judgment call, and reasonable people will disagree about what belongs. This is a living document: the recent end especially will keep shifting as the dust settles.