The history of kvark's ideas and hacking experiments.
Prompted by presentations from Humanity & AGI Summit 2026: Robotics for Future Civilization , I sketched out a few strategies to train the robots. I don’t do this at work and by no means express the opinions of my employer here. The distinct strategies on my list are presented in the order from more realistic to more distant fantasy. The lack of data is often named as one of the showstoppers for…
AR technology is going through a lot of turbulence. Meta is not releasing a successor to Quest 3 (and no longer selling it to businesses), despite it being a popular device and making VR/AR entertainment affordable for many people. It laid off 10% of Reality Labs recently. I think it’s the calm before the storm. My bet is that the world of personal assistant electronics will change once this…
The previous post described my new local inference and training benchmark - Inference Arena . We talked about the list of models, the platforms, and the metrics captured. Perhaps the most interesting part of that post was about the results, and I want to expand more on it here. Disclosure: I’m the author of Meganeura , one of the tested frameworks. I’ve tried to keep the methodology honest — judge…
Looking at the screen, seeing my free space dropping by 20 GB from entering a Python development NixOS shell. It carries a few hand-selected ML libraries: pytorch, torchvision, transformers, onnx, and some more. That’s a ton of code! I wonder how it fares against smaller and newer libraries, including those made with Rust language. Infenera Here comes Infenera (short for “Inference Arena”) - my…
Just as Meta turns around and shifts its attention to AI, following basically everybody, I want to tell you about XR. As an owner of Quest 3S, I always wondered if it’s easy to build for this device. Now I know. Blade XR Blade is the GPU abstraction layer I’ve been toying with lately. Recently, I’ve integrated support for OpenXR to it, while testing a new example on the Quest 3S. Blade-graphics…
Time to dig up all of the projects and ideas that we didn’t do mostly because of the lack of spare time. I’ve just released fileman-0.1 - the first version of my lean file manager. Ever since the time I worked with Windows and used a golden combo of TotalCommander + IrfanView , I’ve been searching for anything even close in Unix/MacOS land. That feel of low latency, immediate response, and polish…
Trying to embrace the new era of AI-assisted engineering, and boy I’m worried! Mostly, about our inability to understand how the code really works, which is inevitable in the future where most of the code is produced by AI agents. A counter-argument by Chris Lattner is that similar transition happened in the past when we stopped writing assembly. I think that worked because assembly translation is…
Most of the user interfaces suck. That’s a bit of a paradox: I can understand how technical debt can be swept under the rug, or how customer support can be a hit-and-miss, but user interfaces? That’s your main way of interaction with users, its business value appears to be disproportional of the investment and quality. In this article, I’m going to provide a few examples of bad interfaces and…
Environmental map is a common source of light in ray tracing applications. In bright day-time scenes, the ability to sample it efficiently is detrimental to reducing the noise. I want to share my experience with different ways of sampling. We assume that the environment is represented by a 2D image with pixels storing RGB floating point radiance. A value in a pixel represents the average radiance…
In 2020, NVidia research group published Spatiotemporal reservoir resampling for real-time ray tracing with dynamic direct lighting . By that time, NVidia had already been selling hardware with hardware ray-tracing support, and there were a few flagship applications for it, like Quake2 RTX . However, sampling was still hard, and a new class of algorithms was needed. ReSTIR brought significant…