A pre-trained relational foundation model can make accurate predictions on any database and any predictive task without any model training.

28:21reason over structured relational data. Um, and it's crazy what this model can do. So what this model can do, it can make accurate predictions on any database and any predictive task without any model training. >> Yeah. And I I find that proposition to be almost outlandish. Like they're just numbers with some unknown relationship and you're going to say that you're going to train a model on just the

28:51relationship between random business numbers and it's going to work in in some unknown use case. How make that make sense to me? >> Thank you. Thank you. I I think it's great. I think as I say this people who listen should should be like what is this guy talking? So So thank you. Right? So I agree, right? Because it's easy to say, "Oh, it's a foundation model. Yukhoo, right? Great." But then,

29:14okay, what does it really do? So here's maybe how to think about this. So the key here is to do in context learning, right? The the same way as a language model does in context learning where I give it a prompt, I give it the information, I give it a task, and then it gives me the answer. So what we do here is the system has several several

29:37components. So there is the database uh and then there needs to be a way for me to instruct the pre-trained foundation model what kind of prediction I want right I want to say predict me the sum of purchase prices over the next one month for this particular customer and that maybe is like how much I'm predicting how much the customer is going to spend or I'm saying predict me

30:03you know uh transaction is fraud equals true for transaction ID this much. Okay, so this would be like predict me whether the transaction is fraudulent for this particular transaction ID. Right? So I have a way to specify my predictive task. And now what the system does, the system now goes into the database. It extracts a set of labeled in context examples that then get passed through a