AI Tools Projects Directory
🏆 Leaderboard 📡 News ✨ Prompts ⚖️ Compare Models 🔧 Tools 📦 Projects 🚀 Spaces
Blog
Catalogue data

The words open-source AI stopped using

Repository topics are self-declared, which makes them a poor census and an excellent mirror. Compare the labels on the 1,816 projects born in 2023 with the 2,594 born in 2026, and the vocabulary has almost entirely turned over.

Updated 1 November 2026·6 min read·No paywall
A shop sign painter removing large faded letters from a storefront while fresh bright letters spelling something new wait stacked against the wall
Same shop, same trade. The sign changed completely.

TL;DR — the short version

“chatgpt” has collapsed as a label. 20.1 % of the AI projects created in 2023 carried it as a topic. Among those created in 2026: 2.7 %.

“agent” replaced it, and then some. “ai-agents” went from 3.1 % of the 2023 intake to 37.9 % of the 2026 intake — 56 repositories to 984.

Two labels that barely existed now appear on a quarter of new projects each: “claude-code” (0.4 % → 24.6 %) and “mcp” (1.3 % → 24.2 %).

What follows for you: if you search this field with 2023 vocabulary you will miss most of what was built this year. The words that index the work have almost entirely turned over in three years.

Every repository on GitHub can carry topics — the short labels authors attach to say what a thing is. They are self-declared, which makes them a poor census and an excellent mirror: they record what builders thought their project was, in the words available to them at the time.

We took every topic on the 1,816 projects created in 2023 and the 2,594 created in 2026, and compared how often each one appears. The turnover is close to total.

Topics on new projects: 2026 intake ai-agents 37.9% was 3.1% in 2023 claude-code 24.6% was 0.4% in 2023 mcp 24.2% was 1.3% in 2023 agents 15.9% was 4.0% in 2023 codex 15.1% was 0.3% in 2023 rag 14.5% was 3.9% in 2023 ai-agent 11.8% was 0.7% in 2023 agent-skills 8.9% was 0.3% in 2023 Share of the 2,594 projects created in 2026 carrying each topic · olud.ai catalogue
Share of each year’s new projects carrying the topic. Labels appearing on fewer than 25 projects in 2026 are excluded. A project usually carries several topics, so the columns do not sum to 100 %.

The words that died

“chatgpt” was the single most common label of the 2023 intake after “llm”: 365 of 1,816 projects. In 2026 it appears on 70 of 2,594. “large-language-models” fell from 14.8 % to 1.9 %, “gpt” from 9.1 % to 1.0 %, “deep-learning” from 8.0 % to 1.7 %.

Note what kind of word each of those is. They name a product, or a technique, or a field. They are the vocabulary of people building things that are about AI.

Topic2023 intake2026 intakeRepos 2023 → 2026
chatgpt20.1 %2.7 %365 → 70
large-language-models14.8 %1.9 %268 → 48
gpt9.1 %1.0 %165 → 27
deep-learning8.0 %1.7 %145 → 43
chatbot7.2 %1.9 %130 → 49
openai11.2 %6.0 %204 → 156
computer-vision5.0 %1.1 %90 → 29
diffusion-models4.3 %1.2 %78 → 31

The words that took over

“ai-agents” is now on more than a third of everything new. Add the variants — “agents” at 15.9 %, “ai-agent” at 11.8 %, “agentic-ai” at 8.4 % — and the single idea dominates the intake.

Two of the four fastest risers are not concepts at all. “claude-code” and “codex” are product names, and “mcp” is a wire protocol. In 2023 the labels described what the software was about; in 2026 they increasingly describe what it plugs into. That is the signature of a field that has stopped being a research topic and started being an integration surface.

“llm” is the one survivor, and it grew: 24.8 % to 33.1 %. It has become the generic term the others used to be.

What a vanished word actually means

A label disappearing is not the same as a technology disappearing. More people use ChatGPT today than in 2023; almost nobody labels a new repository with it. Three different things produce that, and they are worth separating.

A word stops being informative when everything matches it. Tagging a project “llm” in 2023 placed it in a field; tagging it that way today says almost nothing, which is precisely why the surviving generic grew while the specific ones fell.

A word tied to one vendor’s product ages with that product’s position. And a word tied to a technique — “deep-learning”, “diffusion-models”, “computer-vision” — recedes when the technique becomes an implementation detail of something larger.

If you search this field. A query built from 2023 vocabulary will return 2023. Searching “chatgpt” today finds 70 of the 2,594 projects created this year. The olud.ai hub indexes by what things do rather than by what they are called, which is the same problem solved from the other end.

Method

All figures come from one snapshot of the olud.ai catalogue taken on 1 November 2026 at 00:03 CET: 11,394 projects carrying 63,822,390 stars between them. The catalogue is a curated subset of open-source AI on GitHub, not a census of it, so every share quoted is a share of that catalogue and not of the whole forge. Star counts move daily and the figures here will drift; the live numbers are on each project’s page in the catalogue.

Topics are GitHub repository topics as reported by the forge, lower-cased, counted once per repository. Cohorts are by repository creation date: 1,816 projects for 2023 and 2,594 for 2026, the latter covering ten months rather than twelve — which affects the absolute counts in the last column but not the percentages, since both are shares of their own cohort. Only topics carried by at least 25 of the 2026 projects are included, to keep single-author tagging habits out of the ranking. Topics are self-declared and unverified: this measures what authors call their work, which is the subject of the article, and not what the work does.

Data: olud.ai catalogue, refreshed nightly — method on /methodology.html.

Keep reading