Soumith Chintala: belief

Transcript context

…So one of the things with the scaling loss, like Chinchilla is optimal to balance training and inference costs. I think at Meta's scale, you would rather pay a lot more maybe at training and then save on inference. How do you think about that from infrastructure perspective? I think in your tweet, you say you can try and guess on like how we're using these GPUs. Can you just give people a bit of understanding? It's like, because I've already seen a lot of VCs say, Llama 3 has been trained on 600,000 GPUs and that's obviously not true, I'm sure. How do you allocate between the research, FAIR and the Llama training, the inference on Instagram suggestions that get me to scroll, like AI-generated stickers on WhatsApp and all of that? Yeah, we haven't talked about any of this publicly, but as a broad stroke, it's like how we would allocate resources of any other kinds at any company. You run a VC portfolio, how do you allocate your investments between different companies or whatever? You kind of make various trade-offs and you kind of decide, should I invest in this project or this other project, or how much should I invest in this project? It's very much a zero sum of trade-offs. And it also comes into play, how are your clusters configured, like overall, what you can fit of what size and what cluster and so on. So broadly, there's no magic sauce here. I mean, I think the details would add more spice, but also wouldn't add more understanding. It's just gonna be like, oh, okay, I mean, this looks like they just think about this as I would normally do. So even the GPU rich run through the same struggles of having to decide where to allocate things.…

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