11 of GitHub's all-time top 100 repos passed through our tracker before they blew up. Here's what I learned.

11 of GitHub's all-time top 100 repos passed through our tracker before they blew up. Here's what I learned.



Tópico: 11 of GitHub's all-time top 100 repos passed through our tracker before they blew up. Here's what I learned.
Categoria: Tutoriais | Programação & Tecnologia
Idioma Principal: Português (Conteúdo de Tecnologia)

Descrição do Conteúdo / Informações:
-------------------------------------------------------------------------
I kept finding out about interesting open-source AI projects weeks after everyone else.

Some tool would suddenly be everywhere on my timeline. I'd go look, and it already had 40k stars, a crowded issues tab, and five "X but for Y" projects built on top of it.

I was always late.

The problem wasn't a lack of places to find new repos. It was the opposite.

AI open source has become a firehose.

New agents, frameworks, runtimes, inference tools, databases, and developer primitives show up every day. GitHub Trending gives you a list. Hacker News gives you another. X gives you twenty more, usually with 🔥 somewhere in the post.

There is an absurd amount of good stuff being built. Keeping up with it is becoming a job.

So a friend and I started keeping our own list.

The rule was basically:

If we only had an hour this week to look at new open-source projects, which ones would we spend it on?

That list became a biweekly writeup. The writeup became a habit. And somewhere along the way, it became a dataset.

Since early 2025, we've tracked 239 repos across 37 issues and recorded their GitHub metrics over time.

We call it Repository Radar.

Looking back through that history, I started noticing a few patterns.



Stars are useful. They're just not a verdict.


We track stars. Quite prominently, actually.

Not because the repo with the most stars automatically wins, but because stars are a pretty good measure of where attention is moving.

The problem is that attention is noisy.

Most star spikes are noise.

A launch-day bump, a viral tweet, or a spot on a listicle can all produce a beautiful chart for a few days. Then nothing.

The repos that became interesting over time often had a different shape. The initial spike wasn't necessarily bigger. It just kept going.

Week two held up against week one. Sometimes it accelerated.

So the absolute star count became less interesting to me than the slope. And the slope became more interesting when I looked at what was happening underneath it.



The repos that kept climbing usually kept shipping


This was probably the most boring pattern.

Releases kept landing. Issues got closed. Contributor activity broadened beyond one person.

The repo was still visibly alive after the launch post disappeared from everyone's feed.

That's when a rising star curve became interesting.

Stars showed us where attention was moving. Project activity gave us a reason to keep watching.

Neither tells you whether a project is actually good. You still have to understand what the thing does.

But when you're trying to reduce a few thousand new repos to five worth opening, signals help.



The number I'm most attached to is 11


Of GitHub's current all-time top 100 repos, 11 passed through our radar before reaching the top 100.

That's not me claiming we can predict the future. We have missed plenty.

But it did make me think being early isn't entirely random.

Take OpenClaw.

We first covered it at roughly 4k stars in January. Today it's north of 380k.

At 4k, it was still small enough to sit down, read through, and form an opinion.

The interesting part wasn't that 4,000 people had clicked the star button. The project was shipping. Attention kept growing. Contributor activity was expanding.

It later became our repo of the month. Twice.

By the time a project is everywhere on your timeline, discovering it isn't particularly difficult.

The window I'm interested in is earlier, when the answer to "does this matter?" is still unclear.



AI is making this problem worse


In a good way.

Writing software has never been cheaper.

A weekend experiment can become a working repo. Internal tools get open-sourced. A researcher can turn a paper into something people can actually run. Small teams can ship projects that would have needed much larger teams a few years ago.

More things get built. More things get shared.

I think that's great. But abundance cuts both ways.

• When every repo has a polished README, polish becomes a weaker signal.

• When every launch has a demo, demos become a weaker signal.

• When hundreds of AI repos are "trending", trending becomes a weaker signal.

The hard part is no longer finding open-source software. It's deciding what deserves your attention.

That's the problem we've been trying to solve with Repository Radar.



So we built the radar I wanted


Every two weeks, my co-author and I go through the open-source AI firehose and pick the handful of projects we think are actually worth looking at.

Not 50 links. A small, opinionated selection.

For each repo, we explain what it is, why we found it interesting, and how to get started.

That's the Repository Radar newsletter.

The Repository Radar site is the data layer behind the curation.

It contains every repo we've covered, with live GitHub metrics and historical context.

You can see which projects have moved most since we first found them, compare stars, forks, contributors, issues, PRs, releases, and activity, filter the archive, or export the underlying data as JSON.

If you want the curation, subscribe to the biweekly newsletter.

If you want to explore the repos and data yourself, play with the tracker.

The list is hand-curated. It's opinionated. We'll miss things.

I'd still rather give you five repos we have a view on than generate another feed of 500 things that happen to be trending.



What are we missing?


If you actively follow AI open source, how do you cut through the noise?

• Do you follow specific maintainers?

• Watch star velocity?

• Look at contributor growth?

• Track dependencies?

• Wait for Hacker News?

And if you had a radar for open source, what signal would you want on it that we don't track today?

I'd love to steal your best idea.

And if there's a repo we completely slept on, please tell me. That's literally what the radar is for.