The second episode of Ground Truth is live today. My guest is Léo Dreyfus-Schmidt, who spent close to a decade building and leading the research function at Dataiku, back when almost no enterprise ML platform had one. He is a mathematician by training, and his bio lists “commutative diagram enthusiast” as a personality trait. That tells you most of what you need to know before he says a word.
Why this conversation
We started this series because the interesting conversations in applied AI right now are happening in tabular data, alongside all the attention on large language models. Léo is a good person to ask why. He joined Dataiku as a young startup, cold-emailed his way in when they weren’t even hiring, and spent about eight years building out the applied research team as the company scaled into one of the largest enterprise ML platforms in the world. He has since moved into a new chapter: AI for mathematics and science.
That arc, from proving theorems to shipping churn models to now building tools that might change how the next generation does mathematics, is exactly the kind of ground I wanted this series to cover.
Research that ships, not research that publishes
The first thing that struck me was how deliberately Léo’s team avoided becoming a publication mill. He put it plainly: a paper was a byproduct of the work, not the point of it. The team’s mandate was to take on the topics with real scientific uncertainty, things a customer success rep couldn’t just install off the shelf, but that clearly carried business value. That is a harder needle to thread than it sounds, and he was candid that the discipline required to stay grounded in product rather than in citations was itself a constant negotiation.
He was equally candid about the personal version of that adjustment. Moving from a math PhD into industry, he said, meant accepting that the elegant theory does not always survive contact with production, and that making an actual difference usually means resisting the pull of the shiny object in favor of what actually works.
Active learning, and the trust problem underneath it
We spent a good stretch on active learning, the idea that instead of labeling data randomly, you use model uncertainty or diversity sampling to prioritize what gets labeled next. What I found most useful was less the technique and more the caveat: active learning quietly assumes you already have a stable test set to measure against. If you don’t, and if the data you’re labeling is drifting under you, the whole framework gets shaky fast. His team’s paper on the subject was literally called rebuilding trust in active learning, because the field had no shared benchmarks and everyone was quietly claiming state of the art on incomparable setups. Unglamorous work, benchmarking, but he was blunt that it’s the only scientifically honest starting point.
Why prediction alone doesn’t move the business
The part of the conversation I keep coming back to is his explanation of why his team pushed into causal machine learning. A churn model can be extremely accurate and still do nothing for the business, because knowing who will leave is not the same as knowing who can be retained. Some customers are lost no matter what you do. Others were never leaving anyway. The only group worth spending a marketing budget on is the slice in between, and finding that slice requires shifting the question from prediction to uplift: what is the difference between acting and not acting. He was direct that no model, however large, solves this on its own, because the hard part is an assumption about how the historical data was collected, not a modeling choice.
Tabular data was never the interesting kid at the party, and that’s fine
I asked him about the current wave of large tabular foundation models, and he gave an answer I want more people building in this space to sit with. Tabular data, he argued, has never been the flashy topic, even before language models existed. Enterprises have always run on tables, but nobody put a spreadsheet on a billboard. He sees real promise in these models for low-data use cases where a customer doesn’t have enough history to train a gradient-boosted tree properly, and some hope that they could be more robust to the kind of data drift that quietly breaks models in production, fraud rings changing behavior, customer demographics shifting under a model that only ever learned one pattern. But he was equally clear that he doesn’t expect a single model architecture to resolve assumptions that live in how data was collected in the first place.
The next chapter: trust, formal methods, and AI for math
The back half of the conversation turned toward where Léo is spending his time now. He traced it back to DeepMind’s AlphaProof reaching silver-medal performance at the International Mathematical Olympiad, not because of the medal, but because the system used the formal language Lean to verify its own reasoning was actually correct. That, he said, is what changes how the next generation learns mathematics.
It also shapes how he thinks about trust in AI more broadly, and this is where I think he said something genuinely important. With code, you write it once and can verify it once, then run it forever. With a data analytics query, you write it once, run it once, and if the answer is wrong you may never find out, you just make a decision on the wrong number. He does not think blind trust in an AI system is ever the right posture. He wants an intermediate layer people can actually check, without needing to read the underlying code themselves, which is part of why he has put his own money behind Formel AI, a startup applying formal verification techniques to make AI outputs checkable rather than merely plausible.
The conversation with Léo is up now. Hope you enjoy it.
Alex Gerbeaux leads applied AI at Fundamental, where he works on deploying Large Tabular Models inside enterprises.
