And can, can that, can that actually be essentially automated or, or pipelined so that you can take a regular off-the-shelf model, certainly not a big one, but, you know, a regular model of, let's say, Hugging Face, and shove it through a thing and it gets optimized to the point that it's runnable on, on hardware that by rights shouldn't be able to run it?

Yeah, I mean, that's kind...

That was the hope, that this would be the deliverable of my PhD, and, and I guess we achieved it, or at least I [laughs] ...

And I think I achieved it for, at least for, for CNNs, so convolution neural networks, that you can kind of like do this.

So, so the idea is that you start, like you said, with a big baseline model you can get from PyTorch or from, from Hugging Face, I don't know, your favorite ResNet or whatever you want to have, and then we can kind of like start from there.

Of course, you also need a dataset, kind of like a problem you want to learn.

And then we have this kind of like multi-step pipeline which first does stuff like neural architecture search, hardware neural architecture search.

[laughs]