No Priors Ep. 72 | With Sarah Guo and Elad Gil

(instrumental music plays) Okay. Hi, listeners. Uh, today you just have me and Elade shooting the ... What's the appropriate term here? Shooting the breeze.

Shooting the breeze.

Um, yeah. But I want to start this shooting the breeze session, uh, by talking about this Goldman Sachs report that everyone's reading which essentially says ... I- I'm just gonna get on my soapbox for a second here. That it, um ... The title is something like Calling Th- The Top on AI. Um, and so for obvious reasons, I don't like it, but I do think it's worth decomposing for a second. I do encourage everybody to go skim this thing. So, there's a bunch of interviews in it, and two of the core ones are from this guy Daron Acemoglu, um, and Jim Cavello. They're respectively, like, MIT professor and the GS Head of Global Equity Research. And Daron is arguing essentially that AI is going to impact less than 5% of all tasks, and the, like, trillion dollars of CapEx that people are spending on training models, um, is, is a waste because AI will be unable to solve the complex problems. It's, it's not built to do that. A- and Jim argues, you know, he argues th- that in contrast with the internet where you are, um, disrupting something expensive from the beginning even early on versus having a very expensive solution that then becomes democratized, uh, you know, AI is very expensive from the very beginning. And then the other argument he makes is that any efficiency gains ch- from AI will be competed away anyway, and so, um, like, you know, none of the companies are gonna gain from this. Um, and so if, if we just, like, talk about Daron first, Daron's arguing about something he doesn't understand. Like, he ... His claim is, you know, how do we know scale works? More data won't make customer support reps better. I, I think, like, that's just a fundamental, like, misunderstanding of the technology and also objectively of what has happened over the last, um, several years of scale and data, uh, improving capability and quality of model outputs.

I think a lot of these folks do, by the way, are just kind of stuck in the old AI world. Like, I haven't read the report, so I'm not talking specifically about these authors, but, um, a lot of people are treating this like old school ML, and they don't seem to realize that there's been sort of a breakthrough in terms of these, um, transformer-based models or other architectures that effectively are, um, both highly, uh, scale-dependent but also provide different types of functionality and features than, you know, y- you're sitting there, and you're, you're munging some data and effectively doing fancy regressions in some sense. So, um, I think that's the other issue here in terms of ... A- a lot of what I hear... This, this happens a lot in healthcare. You know, in healthcare, they always talk about how data is the new oil, and you're like, "Data is not the new oil." (laughs) You know, sometimes data is useful, and well-labeled data can be extremely useful, but, you know, a lot of it is also about the model and the application and everything else. And so I think, um, I, I think there's just this broader misconception in terms of how this stuff works and what it means and, and all the rest of it.