Thanks to everyone who read and enjoyed this post. I'm humbled by the reception it's had. I really mean what I wrote in the last section: this is a problem space we deeply care about and we're really committed to delivering. If you work at a small/medium/Enterprise company and are struggling with scale or availability, come chat with us. We'll show you the roadmap, you can stress test, kick the tires... I'm stoked to hop on calls and answer your technical or migration questions. I'm vmg at cursor dot com (or just DM me).

Seriously though, this is how I usually do these blog posts: I write down and edit a pretty final version of the post by hand, and leave placeholders that explain what the visualization should look like. Then I just pass it on to Grok 4.6 and let it cook! I used to use Fable for visualizations (e.g. in the Indexed Grep blog post), which also does a very good job at interactive SVGs, but it had the very bad habit of going "I did your visualization and AKSHUALLY also fixed some of your English grammar because it seemed off". ffffffffffuu please dont fucking do that. After testing 4.6 here and looking at the quality of the viz and how well it follows instructions (i.e. leaves my text alone, tyvm), I'm not going back!

Good news @grok, you're getting a hug _and_ a raise from Adrián! I'm not sure of what that means in practice. Maybe we'll just feed you a few more GPUs. I'll keep one as your humble prompter and use it to play videogames I guess.

I did a write-up of things I learned during my 10-year tenure at GitHub, the things that worked, and what we're trying to improve in Origin. The scale of the challenge is staggering, but I really want people to know we're putting a lot of thought and care into it.

Cool ship for my first day at SpaceXAI. I'm stoked to be working on Git infrastructure again! I think everybody is now very aware of how hard it is to do this stuff at the scale that agents require. Can't wait to blog about our design and operational philosophy for the platform

This is a major breakthrough in both intelligence and cost efficiency. The team really outdid themselves! Sadly it’s a large enough model that it cannot ship in the EU right away, but it’s coming very soon.

We've been developing a multi-agent system that builds and maintains complex software autonomously. Recently, we partnered with NVIDIA to apply it to optimizing CUDA kernels. In 3 weeks, it delivered a 38% geomean speedup across 235 problems.

I've heard some people say that Composer 2 is "just Kimi 2.5", which is easy to disprove if you look at any benchmarks... or, huh, if you run the model for like 30 seconds. Anyway: this paper shows the humongous training effort that went behind our model. Please check it out.

Story time! The very first version of Bugbot we launched was stateless. We kept track of runs in our analytics database and we only used that data offline for metrics. Whenever we needed to analyze run metadata, we had to do online lookups from the database. Relying on our analytics database during online operations was definitely tech debt. We were using a large RDS cluster that was already running hot from offline usage. We started noticing that some Bugbot runs were failing because of external analytics load; by that time, we had added a few Bugbot-specific tables to it, so we had to come up with a migration plan quickly before this became a widespread availability issue. We chose PlanetScale Postgres for our new Bugbot cluster. We already had many terabytes of data in PlanetScale MySQL/Vitess which was working incredibly well, so we felt comfortable with this choice. The migration went very smoothly, and now all of Bugbot's queries are instant and we stopped worrying about downtime!