HeyDonto has launched DFT Labs, a new unit that will test a machine learning framework based on ideas drawn from physics.
DFT Labs is built around Data Field Theory, which applies mathematical methods used to study systems near phase transitions to artificial intelligence. HeyDonto has published a foundational paper on the approach in Frontiers in Big Data and has begun entering public benchmark competitions to test the framework against other methods.
At the centre of the project is the claim that learning can be described as a field evolving on curved geometry, rather than mainly as a product of scale, computing power and parameter counts. The research programme sets out four predictions that can be tested directly: a tipping-point effect when concepts begin to form, a formula for how well systems handle unfamiliar data, a finite speed at which information spreads through a model, and recurring numerical patterns across different domains.
HeyDonto has divided the effort into three stages. First, it aimed to establish the mathematics in a peer-reviewed paper and test the framework on synthetic data, where the geometry is already known. Second, it is applying the framework to existing models in public benchmark contests with independent scoring. Third, it would build a model of its own only if the earlier work shows the framework outperforms rival approaches in those public tests.
The published paper was written by Reza Nehzati, founder and chief scientific officer of DFT Labs. According to HeyDonto, the study validated four first-principles predictions on synthetic data and achieved 89.2% accuracy, ahead of the baselines used in that test.
The paper also identified a limitation. On real-world data where the underlying geometry is not known in advance, including handwritten digits, performance dropped sharply. Addressing that weakness is now the focus of the company's public benchmark work.
Testing phase
The second stage is intended to place the framework in competitions where the problems are fixed and results are judged externally. Rather than using a model built from scratch, the team is layering its framework on to existing models to see whether it improves results on practical tasks.
HeyDonto presented that sequence as a way to separate established findings from the theory's broader ambitions. Stage three remains a target rather than an achieved result, and the company says it will attempt to build a standalone model only if benchmark evidence supports the framework.
Rivers Morrell, founder and chief executive officer of HeyDonto AI, outlined the company's position on the programme's purpose and sequencing.
"The AI industry has settled on the idea that intelligence is something you buy with compute. DFT Labs exists to test whether it is something you can understand with physics," said Rivers Morrell, founder and chief executive officer of HeyDonto AI. "We are doing it in the most unforgiving sequence we could design. Publish the mathematics in peer review, state plainly what does not work yet and then carry the framework into public arenas where the best labs in the world are the competition and independent judges keep score. Only after the framework earns its standing in public do we build our own model from the physics up. Every stage has to be won before we claim the next one. That discipline is what makes the ambition believable."
Published limits
The lab has made the limits of the current work part of its public case for the project. The paper includes not only results on synthetic data but also criteria that could refute the framework, an unusual move in a field where research groups often focus on headline gains.
Nehzati said the next step is designed to address the weakness already identified in the paper rather than move past it.
"We derived predictions from first principles and measured them. Four held, and we published the places the framework fails just as prominently as the places it succeeds," said Reza Nehzati, founder and chief scientific officer of DFT Labs. "That published gap, real data with unknown geometry, is precisely the problem we are now testing in public, on ground we did not choose and cannot tilt. The deeper hypothesis, that intelligence is a regime of organization rather than a property of any one substrate, remains a prediction, and we have stated exactly what would refute it. The mathematics is the foundation. The science is the work of testing it in public."
DFT Labs is HeyDonto's research arm. HeyDonto describes itself as an AI holding company with a portfolio of businesses. The lab is based in Knoxville, Tennessee, and the broader research effort also includes a longer treatise, about 14 further manuscripts in preparation and a book-length treatment of the framework.