I just wanted to chime in and thank you for sharing your prompts like that!
It feels like which prompts people are using (even from developers on the same team) is often opaque. It's a great learning resource for people to see under the hood of each other's AI coding workflows, and I hope to see more folks doing this.
They definitely do train their own models, the founders have described this in several interviews.
I was surprised to learn this, but they made some interesting choices (like using sparse mixture-of-experts models for their tab completion model, to get high throughput/low latency).
Originally i think they used frontier models for their chat feature, but I believe theyve recently replaced that with something custom for their agent feature.
For come context, this uses a technique called CPPN’s (compositional pattern producing networks), where the inputs are (x, y, sqrt(x^2 + y^2)) and then there are several neural net layers with different activation functions and randomly initialized weights. It actually resembles the optical physics of light reflecting through glass (which explains why it sometimes looks like light reflecting through a gem).
>It actually resembles the optical physics of light reflecting through glass (which explains why it sometimes looks like light reflecting through a gem).
Is this related in any way to how Nvidias NeRF's work?
At a startup I worked at, one of my cofounders was an incredibly skilled designer, and he took a lot of pride in the design of our homepage. He spent a lot of time in particular finding the right fonts, and would often complain when he saw other company homepages with inappropriate font choices.
One April fools' day, we thought it'd be funny to make him think that we'd changed the company homepage to have a comic sans font. Of course it'd be going too far to actually change the website publicly, so we decided to set up a DNS proxy inside the company office.
We took our company homepage, recompiled a static site with comic sans, and hosted it on an internal server. Then we set up a DNS proxy that resolved our company homepage for requests coming from inside the office to our comic sans static site, but otherwise the internet worked normally.
We made sure that everyone else in the company would feign outrage when he came in and checked the website, so the joke lasted for a couple hours before he thought to check the website from his mobile phone. Later, he appreciated the joke and figured he should have connected the dots sooner and realized that it was April Fools' day.
We did something similar and the boss was quite upset. He said it disturbed sales demos and cost time and money. Which was surprising since he was otherwise a very chill boss who enjoyed all manner of fun and skunk works at the office.
It feels like which prompts people are using (even from developers on the same team) is often opaque. It's a great learning resource for people to see under the hood of each other's AI coding workflows, and I hope to see more folks doing this.
(Link for anyone who wants to check them out): https://github.com/minimaxir/ballin/blob/main/PROMPTS.md