Machine learning powers AI through two stages: training, where algorithms learn what good looks like, and inference, where trained models make real-time decisions. ChatGPT's breakthrough came from massive internet-scale training data combined with fine-tuning that let it generate sophisticated content like screenplays in Tarantino's style. The conversation explores how these advances apply to marketplaces and commerce.
could you actually just kind of walk us through like maybe the basics of, you know, first off, like what AI is from a technology level and then how that might impact kind of, you know, or maybe even to the point where we're at today in time with AI and, you know, ChatGPT and then thinking about how it kind of impacts the future and just kind of a walkthrough.
AI to me is a capability and I'll talk about machine learning instead. AI is the what, machine learning is the how. And machine learning is, you know, 2 parts. It's part training where you basically teach an algorithm what good looks like. And then there's inference where the algorithm which is trained, then it has to make real-life decisions. So I give the example of Grammarly where the training part is where human linguists put together a corpus of what good English looks like, no pun intended. And then when it infers, that's when you're writing something and it make suggestions like, oh, that's not as crisp as it could be, or that's not grammatically correct and this is what it should look like. And then translating, taking that example and bringing it within the remit of LLMs, large language models, these are models that have actually existed for a while. So I wouldn't, just to be a little bit more precise, the innovation isn't LLMs because if you look at, if everyone's experienced autocomplete from a few years ago, it's actually, that's a large language model that's enabling autocomplete. The big innovation is this process called fine-tuning that OpenAI did with ChatGPT, which really, along with training it across a lot more data, like the entire internet, that one-two punch of the entire internet plus the fine-tuning is what created this like really crazy capability to write, you know, screenplays in the style of Quentin Tarantino.
No, that's super helpful. So it gives us a little bit of a context here. So I guess, you know, since we are, you know, in commerce and kind of marketplaces here in the community, how do we see like that being applied maybe today and then how that can maybe shape the future of marketplaces and kind of commerce too?
For marketplace companies, the first question I have is, you know, what is the core value of your product? And starting with that, like there's 2 levers, there's operational efficiency. Most marketplace companies that I know have complex and abstract processes. That are, that are, that no one really sees, that are below the surface, but that enable the marketplace, whether it's recruiting people, finding, you know, filtering for, for basically doing the things you need to do just to ensure that you create a good match between a buyer and a seller. Right. So LLMs can really resolve a lot of the operational— well, let me be careful what I'm saying. LLMs tend to create multiples of efficiency from an operational standpoint. And so that's one lever. And I can deep dive there since I kind of just said something a little bit loaded. And then the second part is just creating a better customer experience. I give the example of Airbnb and travel. You know, there's another company called Curated, which does— which serves a marketplace of experts with shopping. And they can, you know, they're focused on categories, but with LLMs, they can scale up their category domination and be able to offer just a really great experience where instead of going through a user, a survey or a user flow, I can just kind of say like, hey, I want to get Snowboard, what do I do? And it's just a much more natural front-end experience for a customer to get started, much lower friction than it currently is. So to recap, I think that there's 2 levers, there's operational efficiency, and then there's just like a much better customer experience. And then within operational efficiency, I think it's how many, you know, what are your business processes like and What are the ways to automate them? And then how much arbitrage is there on human capital or human labor? You know, so, you know, there's a couple of companies that come to mind here. You and I both know Robbie's Meta. You know, he has a coaching startup. So, you know, we're experimenting on using LLMs to make the coaches a lot more efficient. You know, of course, it's still the coaches doing the work, but you can use LLMs to just make them a lot better or faster than they would be by an order of magnitude. Which is a tremendous boon for any marketplace that is dependent on humans in the loop to provide value back to the user.
