About Ufinq
Ufinq is a focused research effort on Symbolic AI: algorithms that discover small, readable programs describing how data behaves, instead of large opaque models. The work is closed-source by default and open in its ideas: papers, designs, and benchmarks are public; the productized system is part of Diafunc.
The name follows from the work: Ufinq denotes the universal function space the algorithm searches, the (effectively infinite) space of programs composed from a value system and a fixed set of elementary operators, navigated by an evolutionary process. Ufinq, as a system, makes that search practical at cluster scale.
Ufinq is a long bet on Symbolic AI: that the right output of a learning system, wherever a decision has to be understood, is code, not weights.
What Ufinq is built for
Most contemporary AI research compresses data into ever-larger weight matrices. The resulting models are difficult to inspect, expensive to deploy, and hard to explain: three properties that matter wherever a model's answer has to be understood before it is trusted. Ufinq starts from the opposite premise: search over programs directly, return a program as the model, and trade learning capacity for transparency.
The research effort is organized around four commitments:
Relationship to Diafunc
Diafunc is the company and the platform; Ufinq is its research arm and the technology powering it. The split is deliberate: research output (papers, designs, benchmarks) lives under the Ufinq brand on this site, while the productized platform (projects, datasets, notebooks, automated analyses) lives at diafunc.com. There is no separate Ufinq download, package, or installer; the only way to use Ufinq commercially is through Diafunc.
The asymmetry between reading a paper and reproducing a full distributed system is deliberate. Papers describe ideas. The complete system (AHISTER, DEBOHAC, the simplification pipeline, the cluster runtime, the refinement chain, and the universal value substrate) is what Diafunc offers.
Principal investigator
Markus Pollak
Markus designs and writes Ufinq end to end: the algorithms (AHISTER, DEBOHAC, CRESCENT), the symbolic substrate, the refinement pipeline, the distributed runtime, and the benchmarks. He is the author of the Ufinq papers and the founder of Diafunc, through which Ufinq is available. He works from Austria.
The orientation behind Ufinq is long-form: a research bet that the right form of a learning system is readable code, and that the work of making that practical at scale is worth doing in depth, in one consistent system, instead of spread across loosely coupled prototypes. Ufinq is that system; Diafunc is how it reaches the people who need it.
Continue reading
For the short version of the system, see Technology. The active research themes are listed under Research; the corresponding papers under Papers. Empirical results live on Benchmarks. To get in touch, see Contact.