Keep in mind, this is the fast one it had NEVER seen. So, handed nothing but realistic initial conditions, Walrus seemed to reproduce a decades-old discrepancy, in addition to several other statistics adding believability.

Now, where do foundation models come into the story? Well, as you may know, foundation models (like chatbots) are taught using vast amounts of diverse data - "pretraining."

This makes them particularly good at generalizing to new problems, since the old and new often have conceptual overlap. Within the @PolymathicAI collaboration, we have developed one such foundation model: "Walrus," which has digested all sorts of fluids-related simulations.

One way to frame this is as purely data-driven evidence supporting the hypothesis that initial conditions play the most significant role in this longstanding sim-experiment gap. And, it's... so incredibly cool to me that this seemed to work. Anyways there are a lot more details in the paper, please do check it out!