Shawn Wang: preference

Transcript context

…Yeah, so Yangqing and I used to be framework rivals, PyTorch, I mean, we were all a very small close-knit community back then. Caffe, Torch, Theano, Chainer, Keras, various frameworks. I mean, it used to be more like 20 frameworks. I can't remember all the names. CCV by Liu Liu, who is also based out of SF. And I would actually like, you know, one of the ways it was interesting is you went into the framework guts and saw if someone wrote their own convolution kernel or they were just copying someone else's. There were four or five convolution kernels that were unique and interesting. There was one from this guy out of Russia, I forgot the name, but I remembered who was awesome enough to have written their own kernel. And at some point there, I built out these benchmarks called ConNet benchmarks. They're just benchmarking all the convolution kernels that are available at that time. It hilariously became big enough that at that time AI was getting important, but not important enough that industrial strength players came in to do these kinds of benchmarking and standardization. Like we have MLPerf today. So a lot of the startups were using ConNet benchmarks in their pitch decks as like, oh, you know, on ConNet benchmarks, this is how we fare, so you should fund us. I remember Nirvana actually was at the top of the pack because Scott Gray wrote amazingly fast convolution kernels at that time. Very interesting, but separate times. But to answer your question, Alessio, I think mainly Lepton, Fireworks are the two most obvious ones, but I'm sure the fingerprints are a lot wider. They're just people who worked within the PyTorch Cafe2 cohort of things and now end up at various other places. I think as a, both as an investor and a people looking to build on top of their services, it's a uncomfortable slash like, I don't know what I don't know pitch. Because I've met Yang Tsing and I've met Lin Chao. Yeah, I've met these folks and they're like, you know, we are deep in the PyTorch ecosystem and we serve billions of inferences a day or whatever at Facebook and now we can do it for you. And I'm like, okay, that's great. Like, what should I be wary of or cautious of when these things happen? Because I'm like, obviously this experience is extremely powerful and valuable. I just don't know what I don't know. Like, what should people know about like these sort of new inference as a service companies? I think at that point you would be investing in them for their expertise of one kind. So if they've been at a large company, but they've been doing amazing work, you would be thinking about it as what these people bring to the table is that they're really good at like GPU programming or understanding the complexity of serving models once it hits a certain scale. You know, various expertise like from the infra and AI and GPUs point of view. What you would obviously want to figure out is whether their understanding of the external markets is clear, whether they know and understand how to think about running a business, understanding how to be disciplined about making money or, you know, various things like that.…

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