Research into how AI works, and tools to make research easier to read.
Mechanistic interpretability, neural architectures, and formal reasoning. Plus Project ELA, for anyone who wants to actually understand a hard paper.
Flabubium is an early-stage initiative doing two things at once: foundational AI research and building tools to make that research easier to enter. Many people who want to understand a hard paper have no good path in. We want to change that while also doing the research ourselves.
ELA is a tool we're building to lower the cost of understanding difficult papers. Drop in a paper and it maps every concept and pre-requisites the paper assumes you know, then builds a course to fill the gaps, visual modules and all. The goal is to cut weeks of background reading to a few hours.
Mechanistic interpretability to understand how models actually reason and ensure AI safety. Neural and symbolic architectures built from first principles rather than convention. Formal theorem proving to bring mathematical rigour to AI. These are hard problems we aim to tackle.
Longer term, we want ELA to recommend papers based on your ideas, help you track what you understand across a field, and navigate research you are new to. On the research side, the goal is work that makes AI systems genuinely understandable. Both goals connect: the more we understand AI, the better we can build tools that help people understand everything else.