X / Twitter profile snapshot for Christof Salis

This is the stuff I work on professionally. Medical device software for GPs with just a IPhone as a handhead kamera. Remote eye tracking! We train all our models from scratch ourselves and generate our own synthetic data to train them. Video shows the head impulse test

What we need is to get better at writing down our understanding of good code as something measurable. Where it isn't feasible to turn that understanding into code, lets put that part of it in code that makes it easy for LLMs to call it and close the mental gap.

If you don't think Astra is magnitudes better than Sol at solving problems, you probably don't have interessting enough tasks

What is everyones favorite (tinfoil-hat) theory for the subsidized plans and generous resets? 1. Simply aggressively growing the user base / market capture. Money will follow 2. Getting people used to how productive AI coding is to change habits and create the market 3.

A frontier has opened that outruns your ability to manifest ideas, and nobody really knows what it could do if you just said the right words. You can grieve what your craft used to mean, or you can discover its future. For once, nobody is ahead of you.

I think thats a good reminder how abstractions can remove too much. If I understand the claim correctly it basically assumes that once the model gets something wrong it stays wrong I think agents showed us that they allow for a lot of wrong turns along the way and still recover

100% agree on rethinking what is actually "hard stuff" now and how much effort it would be to try it. This changed drastically Especially research related the effort to utility ratio used to be very prohibitive for many interessting ideas Like today I create a parametric

My toddler wakes up between 5:30-6:00 every morning ready to take on the world. So now my usual day starts with 1-2 hours playtime at the local park with my son and then I go to work around 8 and come back around 7 for bed time stories then often work till like 11 I feel lucky

People get weirdly invested in both agreeing with this and denying it. I am no exception as it is one of the great pleasures of my life to tell myself that all my son's best traits come from me.

In the not-so-distant future, thanks to advancements like gaussian splatting, you'll be able to combine all your videos and photos in a way that allows you to relive them in full 3D. In fact in places you have enough videos (e.g. your home) you'll be able to walk through time

One more benign dystopian thought is that our automations might outlive us.

I love how Pi + GPT lets me learn complex topics more interactively. Here I used GPT-5.5 to explore 3D Gaussian Splatting from scratch from math alone: custom Swift + Metal training/rendering, anisotropic splats, densify/prune, SH color, and a viewer. No 3DGS libs.