Aravind Srinivas : How a Simple AI Search Engine Disrupted Social Media and Went Viral Overnight!

So you know, I got together with my co-founders Dennis and Johnny, and all we wanted to do was build cool products with LLMs. It was a time when it wasn't clear where the value would be created — is it in the model or is it in the product? But one thing was clear: these generative models have transcended from just being research projects to actual user-facing applications. GitHub Copilot was being used by a lot of people, and I was using it myself and I saw a lot of people around me using it. Andrej Karpathy was using it, people were paying for it. So this was a moment unlike any other moment before, where people were having AI companies where they would just keep collecting a lot of data but then it would be a small part of something bigger. But for the first time, AI itself was the thing.

Yeah. So GitHub Copilot, for people who don't know, it assists you in programming — it generates code for you. I mean, you can just call it a fancy autocomplete, it's fine, except it actually worked at a deeper level than before. And one property I wanted for a company was it has to be AI-complete. This was something I took from Larry Page, which is: you want to identify a problem where if you worked on it, you would benefit from the advances made in AI. The product would get better, and because the product gets better, more people use it, and therefore that helps you to create more data for the AI to get better, and that makes the product better — that creates the flywheel. It's not easy to have this property. Most companies don't have this property; that's why they're all struggling to identify where they can use AI. It should be obvious where you should be able to use AI. And there are two products that I feel truly nail this: one is Google Search, where any improvement in AI — semantic understanding, natural language processing — improves the product, and more data makes the embeddings better. Or self-driving cars, where more and more people drive, it's better, more data for you, and that makes the models better, the vision systems better, the behavior cloning better.

Voo, Tesla doesn't matter. Anything that's doing the explicit collection of data. Yeah, and I always wanted my start also to be of this nature. But you know, it wasn't designed to work on consumer search itself. We started off with searching over — the first idea I pitched to the first investor who decided to fund this, Elad Gil, was: 'Hey, you know, would love to disrupt Google, but I don't know how. But one thing I've been thinking is if people stop typing into the search bar and instead just ask about whatever they see visually through a glass.' I always liked the Google Glass version, it was pretty cool. And he said, 'Hey look, focus. You're not going to be able to do this without a lot of money. Identify a wedge right now and create something, and then you can work towards the grand vision.' Which was very good advice. And that's when we decided, okay, how would it look like if we disrupted or created search experiences over things you couldn't search before? And we said, okay, tables, relational databases — you couldn't search over them before, but now you can, because you can have a model that looks at your question, translates it to some SQL query, runs it against the database — you keep scraping it so that the database is up to date — and you execute the query, pull up the records, and give you the answer.

Correct. So you can't ask natural language questions of a table; you have to come up with complicated SQL. Like, 'most recent tweets that were liked by both Elon Musk and Jeff Bezos.' You couldn't ask these questions before because you needed an AI to understand this at a semantic level, convert that into a structured query language, execute it against a database, pull up the records, and render it. But it was suddenly possible with advances — Copilot, code language models were good. And so we decided we would identify this insight and go again — scrape a lot of data, put it into tables, and ask questions by generating SQL queries. The reason we picked SQL was because we felt like the output entropy is lower, it's templatized, there's only a few set of SELECT statements, COUNT, all these things. And that way you don't have as much entropy as in generic Python code. But that insight turned out to be wrong.