But the selections then, were they pre-programmed, so to speak? No,

so they were generated on the spot.

So back then it was really slow.

So this was one of the biggest challenges we had back then was the speed and the latency of the models.

And It was fun because we were working with cloud providers that they suggested to use their automated and managed cloud infrastructure, and we were like, it doesn't work.

Your VRAM is too slow or it's too small.

We have bigger models than that, and now they're so tiny compared to the big models today.

We'll have to test it on complex builds, but it's exciting for sure.