Flabubium started late one night, when a graduate sitting somewhere between mathematics and machine learning came across the word in passing. It had the right kind of strangeness to it, which was reason enough to buy the domain.
The project grew from there. The grad wanted a place to work on the questions that actually matter in AI: not just what the models can do, but how they work, how to make them more principled, and what it would take to reason about them with the same rigour mathematicians bring to their own field.
The AI industry moves fast, but most of that speed has gone into products, scale, and deployment. Foundational research, the kind that asks why something works rather than just whether it works, has not kept pace. More labs are starting to take this seriously, which is a good sign. But the balance is still heavily tilted towards building on top of existing understanding rather than deepening it.
Flabubium exists to contribute to the other side of that ratio. We want to work on mechanistic interpretability, architectural first principles, and formal reasoning: the questions that do not have obvious product timelines but that shape what AI can eventually become.