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

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.

One thing I find pretty interessting is the most productivity gain I got out from agentic coding is by creating custom tools to support me and my team. The second most is reviews and brainstorming.

Anyone else realising that their productivity bottleneck is inference speed? Sure fable would be nice and all, but for the professional software I work on, GPT-5.5 is perfectly captable with my workflows. I just need to be able to iterate faster.

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.

Currently my best hope for my child's future is that I just don't understand politics.

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.

Anyone who somewhat understands LLMs is pretty confident that calling them conscious is a silly take, but why? Let's humor it. My take is this: At minimum, consciousness requires self-mutating statefulness. Simply put, past thoughts and external inputs must meaningfully

So I have been testing the windows version of the codex app. No bluescreens so far which is nice. Heads up, it does require your modem to be dialed up at all times. That's a bummer because I was expecting the AI stuff to be included with my local MSDN install.

Personally, I highly doubt local-model development is around the corner. Why? because the main leverage is speed and parallelization. I do think local-model assistants will get above the critical level in 2026. Luckily thats the one to care about.

Regardless of how it was meant, you can interpret this as either: 1. Justification to play unfair to get ahead or 2. Encouragement to trust your own instincts and knowledge over what conventional wisdom says The second one does not require weak morals nor moral relativism