Last weekend I was in Bangalore for GitHub Constellation. It sits a bit outside my usual OSS conference orbit, but GitHub is one of the nicer company-organized ones; I wasn’t going to skip it.
The keynote covered the usual ground: AI, agents, a revamped GitHub landing page with agentic features, cool demos. I’ll leave most of the demos out; once you start recapping those it reads like a promo post. One exception: Karan M V, GitHub’s Director of DevRel, had some cool live coding demos during the keynote; Copilot, GitHub Cloud, CLI, all synced up through the SDK. There was also an early preview of the new GitHub landing page, agentic by design, with features built around that (to be more native to the platform). Kyle Daigle, GitHub’s COO, also walked through the scale GitHub has had to absorb over the past year from the AI and agents influx, and tweeted about some of it too. The numbers were staggering. I’d love a proper engineering write-up from their team on this at some point. One detail I didn’t expect: their bottleneck is CPUs, not GPUs. Kyle joked about the irony of that in the AI era, which got a laugh from the room.
I skipped most sessions after the keynotes. Once you’ve attended enough of these, you get selective about where you actually sit. The full sessions list is here if you want to look. One I had been looking forward to was the Sarvam session, but Karan had invited me to an AMA with Kyle running at the same time, so that took priority.
← scroll to see more →
The AMA
There was an AMA with Kyle (separate from the one listed on the sessions page), and I got to ask a question. I gave Kyle some context before asking: there’s a thread in GitHub’s community discussions that’s been gaining traction around AI-generated PRs overwhelming open source maintainers. I brought up the matplotlib incident too: an AI agent whose PR got rejected responded by publishing a reputational attack piece on the maintainer who closed it. The HuggingFace CEO had tweeted about their Transformers repo being flooded with agentic contributions from bots. And I mentioned a rumour circulating at the time: that GitHub was looking at PR restrictions to rein in the volume, reported in a few places.
I won’t quote Kyle’s response. What came back was something close to: nobody has this figured out yet. Not maintainers, not individuals, not GitHub. These are norms and workflows that have been around long enough to feel settled, and AI is making everyone rethink them from scratch: maintainers, individuals, companies at GitHub’s scale. Nobody has a clean playbook for that yet, and the answer didn’t pretend otherwise.
I ran into Akash Shukla at the event too (you might remember him from my Wikimedia post). He introduced me to a bunch of people, several of whom had apparently already heard about me from him beforehand (he’d been sharing some of my recent writing around); funny, and it made those conversations much easier to start. I also met Abhishek Mishra, devrel at smallest.ai, and we had a good chat.
You can run a Discord or Zulip thread for months and still not get the same momentum as one good in-person conversation, and this keeps being true across every event. Wrote about this in the PyConf Hyderabad writeup too.
The event was nicely put together. The hallways got congested when everyone spilled out between sessions; navigating between spaces was more of an effort than it needed to be. Minor complaint for an otherwise good day out.