Tabular AI is suddenly where some of the most interesting conversations in machine learning are happening. I want to be part of those conversations, in public. So I am starting a video series.
It is called Ground Truth. The premise is simple: I will sit down with the people doing real applied AI work, and have the kind of conversation I would have anyway, just with the cameras on.
The inaugural episode is live today. My guest is Gaël Varoquaux.
Why this series, and why now?
For the past three years, almost all the oxygen in machine learning has gone to language and images. The progress has been astonishing. But it has also created a strange blind spot. Most enterprise data is not text or images — it's tables. Trillions of dollars of value have been locked up in rows and columns. A new class of model — Large Tabular Models ('LTMs') — are emerging that are architected specifically to process huge structured datasets and make critical enterprise predictions: who will churn, who will pay, where the fraud is, what to charge, when to act.
That is the territory Ground Truth wants to map. I want to talk to the people in the field, working on the actual problems, and ask the questions that are usually only asked privately.
What is Ground Truth?
It is conversations from the field. With the people deploying AI inside real companies, doing the unglamorous work of making models meet messy data. The point is to think out loud about what is actually happening, what is working, what is not, and what we do not yet understand.
This is a topic that is really close to my heart. I wanted to make sure this series is not a founder podcast, not a research seminar, and not a thinly disguised marketing channel. There are plenty of all three already, and the world does not need another. What seems to be missing is the practitioner conversation: the kind of exchange you have over coffee with another senior person who has actually shipped these systems, where the question of whether a thing works is more interesting than the question of whether it is impressive.
And honestly, it is also simply a great way for me to talk with extremely bright minds.
Ground Truth sits alongside First Principles, our research interview series hosted by my colleague Marta Garnelo. Where First Principles asks the people building the science, Ground Truth reports from the field. Together, I hope, they give you a picture of how this technology is actually moving.
The inaugural episode
Gaël felt like the obvious first conversation. If you have written a line of data science code in the last decade, you have almost certainly used his work. He co-created scikit-learn, which has been downloaded more than four billion times. He is research director at Inria and chief science officer at probabl. He is also one of the people behind TabPFN and TabICL, the open-source LTMs pushing the frontier of what foundation models can do for structured data.
Sitting down with him was, honestly, a career highlight. The kind of person you spend a career hoping to get an hour with.
We covered a lot of ground. Why scikit-learn won when better-funded alternatives did not. The politics of benchmarking, and his team's finding that some models that look great on standard tests fall apart when you evaluate them the way enterprise data actually breaks. Why he thinks gradient boosting is not going away, even as LTMs improve.
The moment that has stayed with me most is something he said toward the end of our conversation. I asked him what he would tell a PhD student today, choosing between research on LLMs or tabular AI. His answer was direct:
"It is obvious. You want to go the tabular way. It is really becoming trendy, so you are at the right moment. Things are happening. There is a huge potential. If you go LLM, you are in an incredibly crowded space."
It is a striking thing for someone in his position to say. And it is the kind of observation Ground Truth wants to surface and spend more time with.
What's next
A new Ground Truth episode every few weeks. I have a list of people I want to talk to: data science leaders at major enterprises, engineers who have shipped LTMs in production, founders building infrastructure for tabular AI. If there is someone you want to hear from, tell me.
The Gaël conversation is up now. Hope you enjoy it.
Alex Gerbeaux leads applied research at Fundamental, where he works on deploying Large Tabular Models inside enterprises.
