The half-life of hyped AI repos

64.7% of the AI repos that trended on GitHub in 2023 are abandoned, against 40% of other trending repos. Half stopped within 18.6 months. Our cohort data.

Cover for The half-life of hyped AI repos. A half-life plot on graph paper: the share of repos that trended on GitHub in 2023 still committing, by months since they first trended; the AI curve crosses the dotted 50% line at 18.6 months while the non-AI curve never falls to 50%. Beside it: an 18.6-month half-life for the 947 AI repos, and 64.7% of them abandoned today vs 40.0% of non-AI repos that trended in 2023.
On this page

In April 2023, AutoGPT gained 110,668 GitHub stars in a single month. That one month is 59% of all the stars it has today. The same spring, GitHub's daily trending page was mostly AI: on an average day in April 2023, nearly two in three repos on it were AI projects. A star is the internet's cheapest vote of confidence, and the people who give one rarely come back to check what happened next.

So we checked. We rebuilt the cohort of every repo that appeared on GitHub's daily trending page from January 2023 to April 2026, using 1,199 daily captures from the Wayback Machine: 5,341 repos, 1,910 of them AI. Then we asked GitHub's API, on 1 November 2026, what happened to each one: is it archived, gone, still getting commits, and from how many people. The other 3,431 repos from the same pages are the control, so both groups faced the same page in the same months. The methodology explains how, and where the method falls short.

Key findings

  1. 1
    Two in three hyped AI repos from 2023 are abandoned. Of the 947 AI repos that trended on GitHub in 2023, 64.7% are archived, gone, or haven't had a commit on their default branch in 12 months. For the 1,624 non-AI repos that trended in the same months it's 40.0%.
  2. 2
    The half-life is 18.6 months. Half of 2023's trending AI repos made their last commit within 18.6 months of first trending. The non-AI control never gets to half: 63.6% were still committing two years on.
  3. 3
    Only 16.2% are alive. That's the share of 2023's AI repos where at least two people committed in the last 90 days. For the control it's 37.3%.
  4. 4
    Abandoned repos don't look abandoned. Only 10.0% of 2023's AI repos carry GitHub's archived flag and 1.2% are gone. The other 53.5 points of the 64.7% are still up with no archived banner, so nothing on the page says they've stopped.
  5. 5
    Six million stars point at dead projects. 6.0 million of the 30.2 million stars on every trending AI repo in our data (19.8%) sit on repos that are now abandoned. For the non-AI repos it's 11.3%.
  6. 6
    Bigger hits and company backing help, but not enough. Among 2023's trending repos, the 10 AI repos a month that gained the most stars are abandoned 56.7% of the time, against 35.8% for the control's top 10. AI repos from personal accounts: 71.2%. From GitHub-verified organisations: 46.5%, against 19.6% for verified organisations outside AI.
  7. 7
    A star count records one week. Among 25 well-known AI repos we traced week by week, most of the ones that went quiet got most of their stars in a single launch burst. gpt-engineer got 18,514 stars in its peak week and gets about 8.5 a week now. Repos like llama.cpp and LangChain kept collecting stars for years.
  8. 8
    Some stars aren't real. A peer-reviewed study found 6.0 million suspected fake stars on GitHub from July 2019 to December 2024, and AI/LLM was the largest category among the campaign repos that reached the trending page.

64.7%

Of the 947 AI repos that trended on GitHub in 2023 are abandoned today, vs 40.0% of other trending repos

Source: host0 analysis of GitHub trending and the GitHub API

18.6 months

Half-life: half of 2023's trending AI repos made their last commit within this time of first trending

Source: host0 analysis

16.2%

Of 2023's trending AI repos are alive, with ≥ 2 people committing in the last 90 days (control: 37.3%)

Source: host0 analysis

6.0M

Stars (19.8% of 30.2M on all trending AI repos, 2023–26) that sit on abandoned repos

Source: host0 analysis

GitHub's trending page lists the repos that collected the most stars in a day. It's where a lot of people first see a new project, and it's a fair sample of what was hyped: not every AI repo on GitHub, but the ones a crowd noticed at once. In January 2023 about a fifth of the repos on it were AI. By April it was 64.1%, the month of AutoGPT, BabyAGI and AgentGPT.

Then it ebbed, to 24.1% in July 2024, and came back harder: 83.1% in March 2026, with coding agents, Claude Code and Codex skills and OpenClaw on top. For scale, GitHub's own Octoverse 2025 counted more than 4.3 million AI-related repositories, and 1.13 million public repos importing an LLM SDK, 693,867 of them created in the 12 months to August 2025. Our 40 months of trending pages hold 1,910 AI repos: the ones a crowd noticed at once.

Two in three are abandoned#

We call a repo abandoned when GitHub has archived it, when it's gone (deleted, made private or blocked), or when nobody has committed to its default branch in the 12 months to 31 October 2026. By that measure 64.7% of the AI repos that trended in 2023 are abandoned. Of the non-AI repos that trended in the same months, 40.0% are.

Read the chart pairwise. A repo that trended in 2025 has had at most two years to go quiet, so the low 2025 numbers mostly mean "not yet". But the ratio holds every year: 1.6 times the control in 2023, 1.6 in 2024, 1.7 in 2025. The definition doesn't drive it either. With a 24-month threshold instead of 12, 49.9% of 2023's AI repos are abandoned against 32.8% of the control; counting pushes to any branch instead of the default one, 62.4% against 38.4%. And our classifier probably left about 140 AI repos in the control, which makes the gap look smaller, not bigger.

For context, two studies of ordinary open source put the background rate lower. Coelho and colleagues found that 16% of 2,927 active GitHub projects became unmaintained within a year, and Avelino and colleagues that 16% of 1,932 popular projects were abandoned by their core developers, 41% of which survived when new core developers took over. Their definitions differ from ours, so the comparison that holds is our own control: same page, same months, same cutoff.

A half-life of 18 months#

Abandonment is a snapshot. To see how fast it happens, we followed each repo from the first day it trended to its last default-branch commit, and treated a repo as stopped once it had been quiet for six months, or was archived or gone. Repos still committing in the last six months count as still going.

Three months in, both groups are at 82.1%. After that the AI line keeps falling at the same pace while the control flattens out. Half of 2023's AI repos had stopped by 18.6 months. The control never reaches half in the time we can see: at 24 months 63.6% were still committing, and at 42 months 53.1%. For AI repos that trended in 2024 the half-life is 23.6 months, again with the control above half.

Most of the dying isn't instant, either. Of the 2023 AI repos that are abandoned now, 26.5% made their last commit before they trended or within three months of it. 40.5% kept going for more than a year first.

How long abandoned AI repos kept going after they trended

Time from first trending day to last default-branch commit, 2023 AI repos now abandoned (n = 602)

How long abandoned AI repos kept going after they trended
LabelValue
Last commit before trending4.7%
Less than 1 month after10.8%
1–3 months11%
3–6 months12.6%
6–12 months20.4%
More than 12 months40.5%
Most abandoned AI repos were worked on for months, often more than a year, after their moment on the trending page. For abandoned non-AI repos from 2023, 15.4% had already stopped before they trended.

Source: host0 analysis of 2023 trending repos

Abandoned doesn't look abandoned#

You'd hope a dead repo would say so. GitHub has a flag for it: archiving makes a repo read-only and puts a banner across the top. Only 10.0% of 2023's trending AI repos are archived, and 1.2% are gone. The rest of the abandoned share, 53.5 points, is a repo with no banner or notice, which looks like a live one until you check the dates.

Where the 2023 AI cohort is today

947 AI repos that trended on GitHub in 2023, by status on 1 Nov 2026

Where the 2023 AI cohort is today
LabelValue
Gone1.2% (control 4.1%)
Archived10% (control 6.7%)
No commit in 12 months, still up53.5% (control 29.3%)
No commit in 6–12 months7.3% (control 6.4%)
Committing, but not 2 people lately11.8% (control 16.3%)
Alive: 2+ people in 90 days16.2% (control 37.3%)
The first three bars are the 64.7% abandoned. The highlighted bar is the largest group: repos that stopped without archiving or saying so. "Alive" means at least two people, bots excluded, committed to the default branch in the 90 days to 31 October 2026.

Source: host0 analysis of the GitHub API, 1 Nov 2026

The other end of the chart is as telling. Only 16.2% of the 2023 AI cohort has two or more people committing in the last 90 days, against 37.3% of the control. Another 11.8% are still committing, but with fewer than two people in the last three months, often a single maintainer.

The stars stay behind. A dead project keeps the social proof of its best week. Across all 1,910 trending AI repos in our data there are 30.2 million stars, and 6.0 million of them, 19.8%, are on repos that are now abandoned. Another 2.2 million are on repos that have been quiet for six months.

Stars that sit on abandoned repos

Share of all current stars on trending repos (Jan 2023 to Apr 2026) that sit on abandoned or inactive repos, 1 Nov 2026

Stars that sit on abandoned repos
LabelValue
AI, abandoned19.8% · 6.0M
Other, abandoned11.3% · 7.1M
AI, quiet 6+ months27.1% · 8.2M
Other, quiet 6+ months15.2% · 9.5M
Both groups cover every cohort from January 2023 to April 2026, so younger repos dilute the shares. "Quiet 6+ months" includes the abandoned. AI repos have 30.2M stars in all; the others 62.4M.

Source: host0 analysis of the GitHub API, 1 Nov 2026

What dies and what lasts#

The survivors aren't random. Among the AI repos still alive from 2023 are the ones that became infrastructure, communities or companies: LangChain, llama.cpp, Ollama, ComfyUI, and AutoGPT itself, which had 2,782 commits in the last 12 months and 9 people committing in the last 90 days. The highest abandonment rates are among coding demos, research code and chat wrappers.

Two other things matter: who owns the repo, and how big the hype was. Repos owned by a GitHub-verified organisation (one with a confirmed company, lab or foundation domain) survive far more often than repos on a personal account. But even here AI is worse: 46.5% of 2023's AI repos from verified organisations are abandoned, against 19.6% for verified organisations outside AI. Many of those are lab repos from Meta, Microsoft and Google research that shipped a paper and moved on.

More hype buys a little more life, not much. Among 2023's repos, the 10 AI repos each month that gained the most stars on the trending page are abandoned 56.7% of the time; repos that stayed on the page four days or more, 54.4%. Both are still well above the non-AI top 10, at 35.8%.

Bigger hits fade too

Share abandoned on 31 Oct 2026, repos that trended in 2023, by how big the hype was

Bigger hits fade too
LabelValue
All AI repos64.7% (n = 947)
AI, on the page 4+ days54.4% (n = 331)
AI, top 10 of their month56.7% (n = 120)
AI, on a monthly trending list49.2% (n = 130)
Other, top 10 of their month35.8% (n = 120)
All other repos40% (n = 1,624)
"Top 10 of their month" is the ten repos per month with the most stars gained on captured trending days. Even the AI repos big enough for GitHub's monthly trending list were abandoned more often than an average non-AI repo.

Source: host0 analysis of 2023 trending repos

A star count records one week#

Stars are how most people judge a repo at a glance, and they mostly measure one moment. GitHub added a star-history API endpoint in September 2026 that returns stars per week for any repo. We pulled it for 25 well-known AI repos, picked by hand for the story, so read them as examples, not a sample.

Stars per month: a launch burst vs steady growth

New stars per month (log scale), Mar 2023 to Oct 2026, by star date

Stars per month: a launch burst vs steady growth
Seriesxy
AutoGPTMar 2023104
AutoGPTApr 2023110.7K
AutoGPTMay 202315.8K
AutoGPTJun 20235.4K
AutoGPTJul 20233.4K
AutoGPTAug 20232.5K
AutoGPTSep 20232.2K
AutoGPTOct 20232.7K
AutoGPTNov 20232.1K
AutoGPTDec 20231.8K
AutoGPTJan 20241.8K
AutoGPTFeb 20241.5K
AutoGPTMar 20242.1K
AutoGPTApr 20241.5K
AutoGPTMay 20241.4K
AutoGPTJun 20241.1K
AutoGPTJul 20242K
AutoGPTAug 2024993
AutoGPTSep 20241.2K
AutoGPTOct 20241K
AutoGPTNov 20241K
AutoGPTDec 20241.4K
AutoGPTJan 20251.3K
AutoGPTFeb 20251.3K
AutoGPTMar 20252.2K
AutoGPTApr 20251.1K
AutoGPTMay 20251.1K
AutoGPTJun 2025968
AutoGPTJul 20251.1K
AutoGPTAug 2025998
AutoGPTSep 2025822
AutoGPTOct 2025824
AutoGPTNov 2025813
AutoGPTDec 2025933
AutoGPTJan 20261.2K
AutoGPTFeb 2026874
AutoGPTMar 20261.4K
AutoGPTApr 20261.3K
AutoGPTMay 20261.2K
AutoGPTJun 2026856
AutoGPTJul 2026894
AutoGPTAug 20261.6K
AutoGPTSep 2026714
AutoGPTOct 2026840
LangChainMar 20239.3K
LangChainApr 202313.4K
LangChainMay 20239.9K
LangChainJun 20237.3K
LangChainJul 20235.2K
LangChainAug 20233.5K
LangChainSep 20232.9K
LangChainOct 20232.8K
LangChainNov 20233.5K
LangChainDec 20232.7K
LangChainJan 20243K
LangChainFeb 20242.3K
LangChainMar 20242.9K
LangChainApr 20242.8K
LangChainMay 20242.6K
LangChainJun 20242.1K
LangChainJul 20241.9K
LangChainAug 20241.8K
LangChainSep 20241.4K
LangChainOct 20241.4K
LangChainNov 20241.5K
LangChainDec 20241.8K
LangChainJan 20252.2K
LangChainFeb 20252.7K
LangChainMar 20252.9K
LangChainApr 20252.1K
LangChainMay 20252K
LangChainJun 20252K
LangChainJul 20252.3K
LangChainAug 20252K
LangChainSep 20252K
LangChainOct 20252.2K
LangChainNov 20252.4K
LangChainDec 20252.4K
LangChainJan 20262.6K
LangChainFeb 20262.3K
LangChainMar 20264.3K
LangChainApr 20263.9K
LangChainMay 20262.9K
LangChainJun 20262.7K
LangChainJul 20262.8K
LangChainAug 20262.5K
LangChainSep 20262K
LangChainOct 20262K
llama.cppMar 202314.7K
llama.cppApr 20237.6K
llama.cppMay 20233.4K
llama.cppJun 20233.9K
llama.cppJul 20233.7K
llama.cppAug 20232.7K
llama.cppSep 20231.8K
llama.cppOct 20231.8K
llama.cppNov 20232.1K
llama.cppDec 20232.4K
llama.cppJan 20242.2K
llama.cppFeb 20242.1K
llama.cppMar 20242.1K
llama.cppApr 20243.1K
llama.cppMay 20242.2K
llama.cppJun 20241.7K
llama.cppJul 20241.9K
llama.cppAug 20241.6K
llama.cppSep 20241.4K
llama.cppOct 20241.5K
llama.cppNov 20241.4K
llama.cppDec 20241.5K
llama.cppJan 20252.5K
llama.cppFeb 20253K
llama.cppMar 20252K
llama.cppApr 20251.7K
llama.cppMay 20252.2K
llama.cppJun 20251.3K
llama.cppJul 20251.4K
llama.cppAug 20252.2K
llama.cppSep 20251.4K
llama.cppOct 20251.5K
llama.cppNov 20252.2K
llama.cppDec 20251.8K
llama.cppJan 20262K
llama.cppFeb 20262.1K
llama.cppMar 20264.5K
llama.cppApr 20267.2K
llama.cppMay 20266.5K
llama.cppJun 20265K
llama.cppJul 20263.6K
llama.cppAug 20264.6K
llama.cppSep 20263.7K
llama.cppOct 20263.6K
AutoGPT got 110,668 stars in April 2023 and between about 700 and 2,700 a month since August 2023. LangChain and llama.cpp never had a month like that, and in October 2026 gained two to four times as many stars as AutoGPT. Counts are of today's stargazers, so 2023 months run up to ~7% low.

Source: host0 analysis of GitHub's star-history API

AutoGPT is the extreme. Its peak day, 13 April 2023, brought about 13,640 stars, and in October 2026 it got about 200 a week. It survived anyway, as a company product. Several repos with a shape like that didn't. In our 25, the share of all stars that arrived in the 90 days around the peak week splits the list: most of the repos with 60% or more in that window have had five or fewer commits in the last year, and most of those under 25% are still shipping.

Share of all stars earned around the peak week

Stars in the 90 days around each repo's peak star week, as a share of its stars on 31 Oct 2026, 17 of 25 hand-picked AI repos

Share of all stars earned around the peak week
LabelValue
OpenClaw86% (active)
devika84.2% (0 commits)
JARVIS77.9% (0 commits)
stanford_alpaca76.3% (0 commits)
gpt-engineer72.4% (archived)
AutoGPT69.7% (active)
babyagi65.3% (2 commits)
AgentGPT62.4% (archived)
gpt4all55.4% (0 commits)
MetaGPT30.9% (2 commits)
AutoGen28.3% (5 commits)
LangChain22.8% (active)
llama.cpp19.1% (active)
Ollama15.8% (active)
crewAI15.1% (active)
A1111 webui12.1% (0 commits)
ComfyUI6.2% (active)
Notes are default-branch commits in the 12 months to 31 Oct 2026. Highlighted: five or fewer. OpenClaw, from January 2026, is too young to judge, and Stable Diffusion WebUI (A1111) shows that steady stars can outlast the code. Hand-picked repos: examples, not a rate.

Source: host0 analysis of GitHub's star-history API

The burst repos now collect almost nothing. In October 2026, gpt-engineer got about 8.5 stars a week, against 18,514 in its peak week of June 2023. It's archived, and its description now reads "Precursor to: https://lovable.dev". Lovable raised $400 million at a $13.3 billion valuation in August 2026. The archived repo still shows its 54,993 stars.

Rises, pivots and quiet exits

  1. 13 Apr 2023

    AutoGPT's peak day

    About 13,640 stars in one day; 110,668 in April.
  2. 20 Jun 2023

    gpt-engineer's peak day

    About 4,895 stars; 18,514 that week.
  3. Sep 2024

    BabyAGI archives its original

    The repo was reused for a new framework; the original moved to babyagi_archive.
  4. 24 Sep 2024

    AutoGPT's platform goes source-available

    The platform folder moved to the PolyForm Shield license; the rest stayed MIT.
  5. 17 Nov 2024

    gpt-engineer's last commit

    Later archived, "Precursor to lovable.dev".
  6. 29 Apr 2025

    AgentGPT's last commit

    Now archived.
  7. Oct 2025

    AutoGen enters maintenance mode

    Microsoft points users to Agent Framework. GitHub's archived flag is still off.
  8. 26 Jan 2026

    OpenClaw's peak day

    About 25,636 stars; 115,878 that week.
  9. 20 Feb 2026

    ggml.ai joins Hugging Face

    llama.cpp's team keeps maintaining it full time.

Source: GitHub API, star-history API and README histories; linked announcements

Some stars weren't earned at all. StarScout, a study presented at ICSE 2026, scanned all GitHub event data from July 2019 to December 2024 and reports 6.0 million suspected fake stars and 18,617 repos with fake-star campaigns (3.81 million stars after its post-processing step). Few of those repos were seen on the trending page, 78 in all (the paper's trending archive covers only five languages), but AI/LLM was the largest category among them. The authors found that fake stars help for less than two months and hurt afterwards. They're cheap: star merchants charged $0.10 to $2.00 a star in July 2024, by the paper's count, and Dagster found a seller offering 1,000 for $64 in 2023.

Fake stars are unlikely to explain our numbers: the paper saw only 78 campaign repos on the trending page in five and a half years, and 90.42% of all campaign repos were later deleted, by the paper's count. But they're one more reason a star count says little about whether anyone maintains the code.

The license can change too#

A repo that survives by becoming a company sometimes changes its terms. The shift in AI open source is rarely to a business-source license. More often it's a custom "Apache-plus" license, or an enterprise folder carved out of a permissive repo. We read the license-file histories of 19 popular AI and automation repos.

License changes in AI open source

From each repo's license-file commit history, 2023 to Oct 2026

  • Dify: its own "Dify Open Source License" from day one (Apache 2.0 plus no multi-tenant SaaS and logo retention), reworded five times through March 2025.
  • FastGPT: Apache 2.0 to a custom license with no competing multi-tenant SaaS, August 2023.
  • AutoGPT: the platform folder to PolyForm Shield, September 2024; the rest MIT.
  • LobeChat: MIT to "Apache 2.0 with additional conditions" at v1.0, June 2024; derivative works need a commercial license.
  • Open WebUI: BSD-3 plus a branding clause for deployments over 50 users, April 2025.
  • Flowise: an enterprise folder under a company license, May 2025.
  • MindsDB: core under the Elastic License 2.0 in November 2023, back to MIT in June 2026.
The common moves: no competing multi-tenant SaaS, keep the branding, or an enterprise folder under a company license. llama.cpp, Ollama, LangChain and OpenClaw are still MIT.

Source: host0 analysis of license-file histories via the GitHub API

None of this is unusual for a company built on open source. But a repo you depend on can change its terms at the moment it starts to matter, so read the license file at the version you pin, not the one in the README badge.

What this means if you build with AI#

  1. Check the last commit, not the star count. Two in three AI repos that trended in 2023 have had no commit in a year, and most of them aren't archived. Open the commit history before you depend on anything.
  2. Look for at least two people. Only 16.2% of 2023's hyped AI repos have two or more people committing in the last 90 days. A single maintainer is one job change from a dead repo. The contributors graph or the last 20 commits will tell you in a minute.
  3. Look for releases and answered issues. 58.0% of 2023's trending AI repos never published a release, against 44.3% of the control. Tagged releases and recent maintainer replies on issues are signs that someone plans for other people to use the code.
  4. Weigh the owner. AI repos from GitHub-verified organisations were abandoned at 46.5%, from personal accounts at 71.2%. That's still high, and lab repos often ship a paper and stop, so a big name isn't a guarantee.
  5. Read the star curve. If most of a repo's stars arrived in one week, you're looking at a launch, not a community. Steady weekly stars, like llama.cpp's or LangChain's, are a better sign than a big total.
  6. Pin what you use. Pin a release, a tag or a commit hash, not a branch. If the repo goes quiet, archived or private, your build still works, and you can read the license at that exact version.
  7. Have an exit plan. Keep a fork or a vendored copy of anything small you rely on, and wrap anything large behind your own interface so you can swap it. Plan for the half-life: within about a year and a half, half of these projects had stopped.

Agents make it easy to pull in a hyped repo and build on it in an afternoon; we built host0 so the app you build gets a private link and a database that don't depend on any of them staying alive.

Related: AI models have a shelf life too, about 18 months on first-party APIs; new benchmarks hit their first milestone in about a year; and SDK downloads show which AI libraries people actually install.

Methodology#

All data is public; the GitHub API was queried on 1 November 2026. Every "in the last N months" window ends on 31 October 2026.

  • The cohort. We listed every Wayback Machine capture of github.com/trending (daily, all languages) and took one a day from January 2023 to April 2026: 1,199 of 1,216 days. Parsing each listing gave 23,606 rows and 5,341 distinct repos, merged across renames by GitHub's repository id. A repo's cohort month is the month it collected the most "stars today" on captured days; its clock starts on the first day it trended. Captures are mostly taken in the early hours UTC, so they show the previous day's list. The page lists fewer repos since June 2024 (about 15 instead of 25), which lowers counts, not rates.
  • What counts as AI. The repo name and the description shown on the trending page when it trended, matched against a list of strong AI terms; 446 repos with weak terms ("agent", "assistant", "OCR", "bot") were decided by hand. All 60 repos in a verification sample were correctly AI. Two random checks of non-AI repos suggest 4–5% of the control are AI repos we missed, about 140, which makes the gap look smaller. "AI" includes classic machine learning and computer vision, not only LLMs. Categories are rule-based with 58 hand fixes and were 85% right in a hand check.
  • Status. From the GitHub GraphQL API on 1 November 2026: the archived flag, whether the repo still exists, the date of the latest commit on the default branch, commit counts, and the people who authored commits in the last 90 days (bots excluded by name and a list). Abandoned: archived, gone, or no default-branch commit since 31 October 2025. Inactive: none since 30 April 2026. Alive: at least two human accounts committed to the default branch since 2 August 2026. An independent REST re-check of 60 repos matched the status of 59.
  • Survival. Kaplan–Meier per group: a repo's duration is the time from its first trending day to its last default-branch commit if it's been quiet six months or more (or is archived or gone), and is censored on 31 October 2026 otherwise. The half-life is the median.
  • Stars. Current stargazer counts from the API. The 25 named repos' weekly histories come from GitHub's star-history endpoint, which counts only current stargazers: against Wayback captures, totals run 0.3–7.3% low, and up to about 7% for 2023. Peak-day numbers are approximate because GitHub's day boundaries aren't guaranteed to be UTC.
  • Limits. "Last commit" can miss a repo developed on another branch or moved to a new home (GPT-Engineer and MetaGPT both moved on to company products). Commits made by AI agents under a person's account count as human. Paper and model-release repos often were never meant to be maintained. This is the trending cohort, the repos a crowd starred at once, not all AI repos on GitHub, and it only includes days the Wayback Machine captured.

The data pack, with every source, the sensitivity tables and a do-not-cite list, is in our repository at docs/research/2026-11-01-ai-repo-half-life.md. One row per repo (repos.csv), the series behind every chart and the scripts are in docs/research/2026-11-01-ai-repo-half-life-data/.

Open questions#

  • Do users of dead repos move to a fork? An abandoned project can live on in a community fork. We didn't trace forks, so part of "abandoned" may be "moved".
  • Is the LLM-only rate higher still? Our AI group includes classic machine learning and computer vision. An LLM-only cut wasn't computed.
  • Will the 2026 wave last longer? AI took 83.1% of the trending page in March 2026. Those repos are too young to judge; the same measurement a year from now will say whether agent tooling ages better than 2023's wrappers.
  • How many stars on dead repos are fake? StarScout's data stops in December 2024, and since mid-2026 GitHub no longer lets outsiders list a repo's stargazers, so that check is now hard to repeat.
ResearchAI coding