Shawn Wang: preference

Transcript context

…Yeah, we're listening. So that's the four words discussion. I think I want to leave some discussion time open for miscellaneous trends that are happening in the industry that don't exactly fit in the four words or are a layer above the four words. So the first one to me is just this trend of open source. Obviously, this overlaps a lot with the GPU poor thing, but I want to really call out this depreciation thing that I've been working on. Like, I do think it's probably one of the bigger thesis that I've had in the past month, which is that we now have a rough idea of the deprecation schedule of this sort of model spend. And, yeah, I basically drew a chart. I'll link it in the show notes, but I drew a chart of the price efficiency frontier of, as of March, April 2024. And then I listed all the models that sit within that frontier. Haiku was the best cost per intelligence at that point in time. And then I did the same chart in July, two days ago, and the whole thing has moved. And Mistral is like deprecating their old models that used to be in the old frontier. It is so shocking how predictive and tight this band is. Very, very tight band and the whole industry is moving the same way. And it's roughly one order of magnitude drop in cost for the same level of intelligence every four months. My previous number for this was one order of magnitude drop in cost every 12 months. But the timeline accelerated because GPT-3 took about a year to drop order of magnitude. But now GPT-4, it's really crazy. I don't know what to say about that. Do you think GPT-Next and Cloud 4 push it back down because they're coming out with higher intelligence, higher cost? Or is it maybe like the timeline is going down because new frontier models are not really coming out at the same rate?…

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