Shawn Wang: evaluation

Transcript context

…scaling log Which is like, okay, there was this T. R. Texas, I don't know who this person is Yeah, he keeps talking about me. It's fucking amazing Oh, okay. Yeah, he does have some obsessions, but like, he's good. I don't know who he is, but he's good. So he says, if 2024 papers are to be trusted, you don't need most attention, you don't need high precision, you don't need most KV cache, you don't need most feedforward network layers, you don't need a reward model, blah blah. Like, it's like, a lot of efficiency papers are just like, hey, on this small example We cut this thing out. Works fine, or works great, works better, whatever. And then it doesn't scale. Right? Like, or So it's a very interesting observation where like, most efficiency work is just busy work, or like, it's work at a small scale that doesn't, that just ignores the fact that like, this thing doesn't scale, because you haven't scaled it. It's just fine for a grad student, but as for someone who's trying to figure out what to pay attention to, it's very difficult. to figure out what is a worthwhile direction in efficiency. Yeah, that's, that's, that's a good point. I think there's a couple, I agree with you fundamentally that like, it's actually quite easy to tell, like when you see a paper, okay, this one doesn't work, this one works, this one doesn't work. I guess the hippo account will just tell you that, sometimes it's just a diary about this thing doesn't work, this thing works, everything. Right, sometimes it's not like, you know, you can always find a dataset where your efficiency method gets neutral results, right? You can always find one, I have comparable complexity. And you know what's the most, the cutest thing ever? Every time some people propose like this, they run like some zero shot score on like some LME Valhannes or something like that. And at 1B scale, all the numbers are random, basically. Like all your boolkill, they're all like, Random chance performance, right? And they will be like, okay, I get like 50 versus 54, I'm better. But like, dude, that's all random chance, right? Like, you know, sometimes I see people that run experiments that like, And then it's like…

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