Fifteen years of GitHub stars, one AI wave

The 2023 AI repos did not die faster than anything else. Everyone remembers that year as a graveyard of ChatGPT wrappers, and I believed it too, until I checked. This morning I pulled every GitHub star I have given since 4 July 2011, 2,143 repos, and looked at what happened to each of them.

It is a small window on the ecosystem: fifteen years of tools one engineer thought were worth remembering. But the window has timestamps, and timestamps are hard to argue with. Hype changed what I starred. It did not change how long things lived. What changed is who the repos are written for.

The shift happened in one year, and it did not reverse

Here are all 2,143 stars, one bar per year. Violet is any repo whose current description or topics read as AI.

In 2022, 12 of my 183 stars were AI. In 2023, 121 of 324. Since 2025 it is two out of every three. ChatGPT launched at the end of November 2022. By the spring, close to half of what I starred was AI: 55 of 118 stars between April and June 2023.

Before that, the chart is a tour of web development: CakePHP in 2011, Yii and a pile of jQuery plugins in 2013, and TypeScript taking the lead in 2020 and never giving it back. Then, in a single year, the subject changed completely.

The languages show the shift from underneath.

PHP disappeared by 2020. TypeScript took half of everything and kept it, and that rise predates the AI wave. Python went from 2% of my stars to 20% in one era, and three out of four of my Python stars since 2023 are AI repos (179 of 236). Python is the AI wave.

AI got its own list, then outgrew all the others

Stars are a pile, and lists are where I sorted some of it. About 600 of my 2,143 stars live in one of 32 GitHub lists, and the lists are blunter than any chart.

🤖 AI holds 246 repos, more than the next three lists combined. It even spun off a child: 🚜 RAG has 36 repos of its own, 23 of them also on the AI list.

The next list down is the quieter trend. 👐 OSS Alternatives holds 93 open-source replacements for things people usually pay for, like Immich for photos, Uptime Kuma for monitoring and Superset for dashboards. The AI wave got the headlines, and self-hosting built the second-biggest list.

At the other end, four lists hold one repo each: 🔍 Search tools, SEO, MongoDB and 🌟 TailwindCSS. 105 repos sit in more than one list, and Novu sits in four: 👐 OSS Alternatives, 🪝 Webhooks, 🔮 Future ideas and 🎩 Best stack. Some lists turn into a whole shelf, others stop after the first book. Every repo on every list is at the end of the post.

Hype does not change how fast code goes quiet

The 2023 story usually goes like this: a thousand ChatGPT wrappers, a thousand chat-with-your-PDF demos, all dead by the next model release. I expected my stars to prove it.

The data refused. I called a repo dormant if it is archived or has not had a push in 24 months, so nothing since 1 October 2024. Then I compared AI and non-AI stars from the same year, so both groups had the same amount of time to go quiet.

• 2023: 33% of AI stars are dormant. 27% of everything else.
• 2024: 16% against 13%.
• 2025: 7% against 9%. The AI ones are now doing slightly better.

One cohort with a six-point gap, then parity, then a reversal. Hype does not change how fast code goes quiet. Time does. My 2013 stars are 74% dormant, and that year was mostly jQuery plugins and PHP frameworks. Nobody calls 2013 a bubble.

One caveat. Deleted repos vanish from a star list entirely, so every dormant number here is an undercount. If the AI class deleted itself more often than it archived itself, I cannot see it.

What died was the thin layer, and what survived changed its name

The repos that did go quiet have a type. mckaywrigley/chatbot-ui has 33,350 stars and no push since August 2024. OpenAI's own chatgpt-retrieval-plugin stopped in July 2024, a few months after ChatGPT plugins were wound down. llama-gpt still advertises "Powered by Llama 2." roomGPT was a perfect weekend demo and stayed one. They were interfaces over a model, and the model vendors shipped the interface themselves. I made the same argument from the other side in the chat UX is the product, not the model: the UX is the moat only if you keep investing in it.

The survivors from the same year mostly changed what they say they are. AutoGPT had a push this morning. prompts.chat introduces itself as "f.k.a. Awesome ChatGPT Prompts." Open Interpreter is now "A coding agent for open models like Kimi K3 and GLM 5.3". Staying alive in AI apparently means rewriting your own README.

Old repos I starred before ChatGPT now sell themselves as AI

This was the finding I did not go looking for. At least 16 repos I starred before ChatGPT existed now describe themselves in AI terms. Not new projects. Old ones, with a new first sentence.

• Kong, starred in 2021 as an API gateway: "The API and AI Gateway".
• Nx, starred in 2020 as a monorepo tool: "The Monorepo Platform that amplifies both developers and AI agents."
• Medusa, starred in 2021 as a commerce engine: "The world's most flexible commerce platform for agents and developers".
• Front-End-Checklist, starred in 2020: now "for humans and AI agents".

PostHog, n8n, Supabase, Meilisearch, Appwrite, SigNoz, Liveblocks, Novu, Cerbos, ToolJet and Webiny did the same. My favourite is tldraw, which now calls itself the "World's best, top-most agent recommended #1 five star SDK." That one is a joke, and it is also the clearest version of the point. The product did not change much. The audience did. Somebody new reads the README now, and it is an agent deciding what to install.

The oldest stars that still ship are boring on purpose

My first star is from 4 July 2011. That afternoon I starred CakePHP and mustache.php. Both had a push in the last three months. Fifteen years, three JavaScript eras, one AI wave, and a PHP framework and a template engine are still quietly merging pull requests.

That is the counterweight to every hype cycle in the chart. The repos that last were rarely the loudest when I found them. Their maintainers picked a small job and kept doing it, which is the same argument I make for teams in spend your three innovation tokens on the product.

Fifteen years of stars point the same way: hype decides what gets attention, and time decides what survives. Right now, the README also decides whether an agent picks you.

Every repo on all 32 lists

Every repo on every list, most-starred first. Hover a repo for the line it uses to describe itself, and open the long lists to see the rest. The other 1,550 or so stars never made it into a list. They live on my GitHub stars page.