Nathan Labenz: evaluation

Transcript context

…Do you have a point of view? I do. So this is kind of similar to... how up until recently, when you'd ask a question on ChatGPT, you'll say, Hey, my information cutoff is whatever, a year ago, and I can only answer questions up to that point. And then web search as a tool came in, and now all of these platforms would use web search to give you the most up-to-date information so that their world information can be It is more about the intelligence to figure out when to use the tool to retrieve the information and then make sense of it and then give it back to you, versus having been trained, pre-trained with all that information upfront. To me, that pattern is exactly what's playing out, right? So in the enterprise world, you have a lot of information, and then your text to SQL and RAG solutions can bring that information in for the agent, for the platform to reason with and then give you information. The nice thing about that is it is substantially cheaper. The model keeps getting better as the underlying premier model keeps improving. And it's also, relatively easily tunable. You can update it, you can change things and so forth. So that means for me, majority of businesses would continue to benefit from this architecture rather than codifying all of that information in the weights of the model. They'll just use the information and then use tools to retrieve parts of the information that are relevant. The exception to that is what we discussed earlier, which is if there are certain tasks that require either high throughput, low cost, if you have a lot of data and in an area that the model has not seen before, then it might make sense to go create custom models for those specific tasks. So I do believe there is going to be increasingly a need for task-specific small models in large corporations or when we have that need. But still, the majority of the use cases will be more retrieval-oriented. So I think that's a great first-pass answer. If I think, though, even just about my own ability to search through my own stuff, my own Gmail, my own Google Docs, One of the intuitions I have pretty strongly is if I were to give you full access to my Gmail and give you full access to my Google Docs, you couldn't search through it nearly as well as I can. And that's despite the fact that you're clearly smarter than me. So I'm like, there seems to be something about the fact that I have had this like pre-training on this corpus that allows me to even just search through it a lot better. Because if nothing else, like one of the intuitions is I know when I found what I'm looking for. You might not know, you could do 100 searches in my Google Drive and never be quite confident you got the absolute best document for whatever the question is. Whereas if it's my Google Drive and I created all those documents, when I hit that document, that is the one that's yes, this is the one. Now, yes, I remember this now. This is the one. So I have that sort of confidence that... I've got to the answer, if nothing else, I think, which strikes me as really hard. And I've seen this when I try to give my own, just whatever, give Claude access to search my drive. It also struggles in that way. It doesn't know how many times the search or if it's found the right thing or sometimes satisfied too easily, whatever. So I still feel like there's something there where you could expect that a model that really had a sort of more in the weights familiarity with the data could do a better job of navigating it. And then maybe it just comes down to upgrade cycles are terrible for this kind of thing. And yeah, as you said, like you want to keep taking advantage of better and better models. This potentially, Dwarkesh has obviously influenced the discourse recently with kind of focus on continual learning. So maybe you need either a new architecture that's more suited to that or some sort of new training paradigm that would be more suited to it. I guess maybe one way to phrase this is, if that were to flip, if you imagine a world a year from now where it's no longer the case that the best approach is to pick the best models and leave them as they are, but tune them through the search as you described, and it instead becomes one of these things where they actually do have this deeper familiarity with all the enterprise's data, what would have changed to flip us from one paradigm to the other? I was trying to kind of think about how humans do this. We'd go into an environment that we don't know, and then we'll go do a bunch of searches, and then we'll read to create some kind of knowledge. And then as you build out that knowledge, there's intuition that comes with it. So you don't need to keep referring back to it, and then somehow that turns into intuition. Right now, I think what these models are doing is the first part. I'll go pick the information. As the context windows of these models also keep getting better, I can stuff more and more information in these models and then get an answer. What intuition is not really understood, so I don't really know how that changes the dynamic. What would change if a model is trained with your data? You clearly need much less data to steer it to a certain direction. You'll have much more consistency in the responses. I don't think you can ever get away from feeding it information, up-to-date information, and so forth. But what I would imagine happen is First of all, that model that you want to do certain tasks doesn't have to write a poem in French. So you'll benefit from using the weights more efficiently for the tasks that you want to do. And therefore, perhaps, again, you may not need as large a model, so you get benefits from more optimizations to reduce the cost, increase the speed, and so forth.…

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