Soumith Chintala: commitment

Transcript context

…Mainly a performance pitch, some amount of cross compiling pitch. Yeah, I think the performance pitch for Mojo was like, we're going to be performant even if you have a lot of custom stuff, you're going to write arbitrary custom things and we will be performant. And that value proposition is not clear to us from the PyTorch side to consider it for PyTorch. So PyTorch, it's actually not 250 operators, it's like a thousand operators. PyTorch exposes about a thousand operators and people kind of write their ideas in the thousand operators of PyTorch. Mojo is like, well, maybe it's okay to completely sidestep those thousand operators of PyTorch and just write it in a more natural form. Just write raw Python, write for loops or whatever, right? So from the consideration of how do we intersect PyTorch with Mojo, I can see one use case where you have custom stuff for some parts of your program, but mostly it's PyTorch. And so we can probably figure out how to make it easier for say Torch.compile to smoothly also consume Mojo subgraphs and like, you know, the interoperability being actually usable, that I think is valuable. But Mojo as a fundamental front end would be replacing PyTorch, not augmenting PyTorch. So in that sense, I don't see a synergy in more deeply integrating Mojo. So call out to Mojo whenever they have written something in Mojo and there's some performance related thing going on. And then since you mentioned Apple, what should people think of PyTorch versus MLX?…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.