HeyDonto launches DFT Labs to pursue physics-based machine learning capabilities

HeyDonto AI Technology, a company you've probably never heard of, just announced something with genuinely big ambitions: a research subsidiary called DFT Labs built around the idea that intelligence isn't about crunching data points but about mapping the relationships between them. The launch rides on a peer-reviewed paper titled "Data Field Theory: A Geometric Framework for Learning on Riemannian Manifolds," published in Frontiers in Big Data by founder Reza Nehzati, and it borrows heavily from physics — the kind of curved-surface math normally reserved for magnets and superconductors.

CEO Rivers Morrell explained the pitch with a bird analogy that's actually pretty memorable. Instead of studying 100 individual birds to figure out flocking behavior, DFT studies the formation itself — the relationships, not the parts. It's a neat idea on paper, and the numbers back it up when the data cooperates: on synthetic data built around the framework's own geometric assumptions, DFT hit 89.2% accuracy, beating out comparison methods.

Then came the reality check. When researchers threw MNIST at it — the standard handwritten-digit dataset every image-recognition method gets tested on — DFT's accuracy cratered to 15.7%, barely above the 10% you'd get from blind guessing. A basic nearest-neighbor approach scored 51.7% on the same data. That's a rough gap, and it tells you the framework currently needs geometry that's already known and clean, not the messy, undefined structure of real-world information. The paper itself flags this as the next problem to solve: teaching DFT to discover the right geometry on its own rather than assuming it upfront.```n Despite the shaky benchmark, HeyDonto says it's already running with this math in production. Its Axiomera platform, which standardizes fragmented enterprise data, apparently draws on DFT principles, and that platform in turn powers Conduit for dental and medical interoperability and Quantara for cancer research. Quantara President Kristopher Fuhr framed the stakes in human terms, arguing that better biomarker algorithms could get cancer patients matched to treatment faster. Whether that promise holds depends entirely on whether DFT can graduate from synthetic showpieces to messy clinical datasets.

Morrell isn't shy about the endgame, either. He wants to build this framework into a foundation model that goes head-to-head with Anthropic and other major labs, and he's promising benchmark results from industry and academic competitions within roughly a year. That's a bold claim from a company whose flagship method just scored barely above random chance on one of the oldest, easiest benchmarks in machine learning. The physics is elegant. The proof is still homework.