My main area of interest is ML for tools for thought, but I'm also curious about any and all research that shifts the curve and improves practical AI safety.
I'm currently working full-time++ on a lot of software, most commercially successful (so far) being Periplus. When I'm not working, I like to read up on new research, learn new things, and play the piano.
Outside of work I'm quite particular about the tools I use (see my setup), in large part thanks to Ivan Illich's Tools for Conviviality.
A chess trainer built on a custom model trained on human games, which predicts the distribution of mistakes a player will make in a position. Each move updates a Bayesian estimate of the player's rating, which in turn is used to pick what to train next (a bit like IRT).
A system for interacting and learning with LLMs. Includes courses, spaced repetition, LaTeX support and more! Initial version was built in two weeks for the Build with Claude contest, and won Claude Project of the Month. Has since developed into a proper product :).
A spaced repetition app with a built-in reader for PDFs and web pages. Uses FSRS for scheduling, supports Anki imports (including review history) and has an MCP server so LLMs can create cards directly.
In 2024 I was selected as one of the winners of the OpenAI Preparedness Challenge for providing novel ideas for catastrophic misuse of frontier models and societal harm prevention. This included a prize worth $25,000 in API credits.
An ML powered song recommendation website based on finding songs with similar vibes. Used an audio embedding model with DB spanning > 1 million songs. Had a cool embedding-mixing mode and automatic playlist ordering based on shortest walk through embedding space (d = sim).
A small modular raytracer written in C++. Includes SSE optimized sphere intersection calculation, raymarched implementation of the mandelbulb fractal and JSON scene representation. Not fast but produces nice renders :). You can find my short seminar paper on accelerating this with intrinsics here.
