David Singleton: evaluation

Transcript context

…I'm going to go in a different direction now. AI very hot right now. Has AI impacted the way you all build product yet? And if not, where do you think it starts to go in the next few years? Yeah, we're very excited about AI. Now to just take one step back, Stripe has been using machine learning and advanced machine learning techniques at the heart of our products since the very early days. Radar is our solutions for payment fraud and it has used, I mean this is the very core of our products and has used machine learning since we introduced it. We also, if you think about the nature of Stripe, we operate in an environment where we have to do a lot of work to kind of catch bad actors in the system, fraudsters or fraudulent businesses. So we've been employing a lot of machine learning techniques there again for years and we couldn't operate the company without them. So the difference between ML and AI, I think when we think about AI today we're mostly talking about the applications of large language models, but those are really based on this new technology called Transformers. It was a paper that came out of Google a while back about four years ago, five years ago. And we put transformer technology into those systems at some point more than a year ago. And it's proved to have a very profound impact on our ability to solve those problems for our users, which is great. But we are also excited about the applications of large language models. I mean there's really two dimensions to this. Number one, we feel privileged to be able to help serve and power a lot of businesses getting started in this space. One big difference between AI startups and others is it's actually quite expensive to operate an AI startup in terms of the compute resources. Like running this stuff in these GPU machines is quite expensive. So typically these companies need to have a monetization model from day one. And I'm really excited to say most of the AI on these is running on Stripe and we're working very hard to make sure we serve all their needs. Open AI, for instance, using us for monetizing chat GPT plus and all their other products. But they don't just stop there. We're actually helping power all of their subscriptions and revenue tracking and financial operations around the business. And the point about that is we've been putting very large engineering teams on this stuff for years and that means anyone who wants to build a reconciliation product and a subscriptions business model, if you want to do that yourself, you have to put big engineering teams on them. But if you can use Stripe instead, you can actually focus those really precious engineering resources on keeping up with the rate of breakneck innovation in this space. So we're excited about that, but we're also really excited about the applications of these AI techniques on our own ability to serve our users better. And so for instance, at sessions we'll be talking about a few of these. We started working with OpenAI when they had the beta. It was a kind of private beta of GPT four available at the beginning of this year. r instance, at sessions we'll be talking about a few of these. We started working with OpenAI when they had the beta. It was a kind of private beta of GPT four available at the beginning of this year. The first thing we thought of was we've got a lot of documentation on Stripe and we put a lot of care attention into making it good. But if you want to achieve something with your Stripe integration, you're typically going to spend quite a lot of time reading docs and kind of synthesizing them as an end user. We realized we can have GPT four read all of our docs, store them as embeddings and then answer questions for developers. And so we now have that available in kind of early stage release inside of our documentation. It's turning out to be pretty valuable. We've also been able to apply these techniques to other areas of our products. So something that we're going to show an early version of IT sessions is we have a really powerful product as part of our revenue and financial automation suite, which we call Sigma. It allows you to write sequel queries across all of your Stripe data so you can interrogate your Stripe data to understand your business and really fine grain detail. Which country has the fastest growing sales, which segment in which country has particular attributes to it? Very powerful product, but it's one that is potentially challenging for non-developers to adopt. You do have to know how to write sequel queries unless you just use the ones in the menu in order to get the full value out of the product. Well it turns out with large language models, we can apply this technology where you can ask questions in natural language and our engine will actually write the sequel query for you. And we've had to do a lot of work to kind of tune that to make sure that it's reliable and understandable and it's going really well. And we also see great applications in actually applying these technologies to make us more efficient internally, answer users questions faster, help each other out more quickly. And so we're doing those things as well. One concrete thing that we did, which I'm pretty excited about is not long after we saw chat GPT come out, we realized it would be really awesome if we could apply that to many use cases inside of Stripe. But as you can imagine, a lot of the data that we are dealing with on a daily basis is very sensitive to our business, to our users and we care a lot and we have a lot of governance around this, but we wanted to make that technology bill. We couldn't say, Hey Stripes, go and use chat GPT. So we stood up an integration to GPT four and an internal UI to use models like that. But here's the thing that I want to kind of relate to all of your viewers. We find that we built presets into that. So when you work with large language models, you want to write a prompt that then helps get the model into a state where it's going to be able to solve the particular problem at hand. And we find that writing prompts is something that's accessible to folks across lots of different job families, not just engineers. Folks in our marketing team, folks in our user support team have been able to use this as well.…

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