Building AI Search Engine for the GPT-5 Era

Will Bryk is the co-founder and CEO of Exa, an AI search engine startup. Unlike traditional keyword-based search engines, Exa analyzes the meaning behind your queries to deliver deeper and more accurate results-enabling true research, not just simple lookup.

Will believes that to be ready for the coming GPT-5 era, startups have to build for the future from day one. So, what advice does Will Bryk have for startups in the GPT era?

Hey, I'm Will. I'm the CEO of Exa. We're building the next generation of search.

One good way of understanding Exa is in contrast to traditional search. Traditional search engines use mostly keywords. If you're using a traditional search engine and you want to find startups working on futuristic hardware in the Bay Area, traditional search engines will use keyword matching. The results they give you will be documents that contain the words startup, hardware, and Bay Area. But startups working on the future of hardware in the Bay Area don't typically use those terms. You might have a rocket company in SF, and a traditional search engine won't be able to find it. Exa can, because we understand the meaning of documents. We understand that if it's a rocket company in SF, then it does match startups working on the future of hardware in the Bay Area.

We believe it's possible to have perfect search over the web, meaning whatever information you want, you get exactly that. We help companies integrate this high-quality knowledge into their applications. We recently raised $17 million from Lightspeed and Nvidia. Our revenue is doubling every quarter. We're building the next generation of search.

In five years, we could have AGI systems that completely automate all human labor. AIs do all the repetitive work, and humans do all the novel work. Every human is going to become a product manager of a team of AIs. How are you supposed to plan for that?

I think it's extremely hard to predict where the world will be in five years. One-year plans make sense right now. Three-year plans are really hard, and five-year plans are impossible. Because the AI market is changing so fast, every month new systems come out that make new things possible, the right way of navigating that is to think from first principles about what the market needs that will still be true in a year. If you're thinking about what the market needs right now, a month later something new is going to come out and they're not going to need it anymore. So you have to think a little more long-term. You have to be a little more strategic than in the past.

For example, when I was in college, AI was not at all as prominent as it is today. Everyone today talks about AI, but back then there were only a few people who were aware of what was going on. I was lucky enough to go to a Westworld watching party at OpenAI, and I ended up on a bean bag with Greg Brockman, chilling, arguing about the scaling hypothesis. The scaling hypothesis is this idea that you keep putting more compute into transformers and they'll just keep getting better, and that's how we get to AGI.

That was a crazy idea when I was in college. The vast majority of people didn't believe it. I didn't believe it. Greg was arguing that if we just keep scaling these things, we'll get there, we'll get to AGI. I was arguing that we need new types of methods. I think we were both right in our own way, but we kept seeing progress from GPT-1 to GPT-2, GPT-3, eventually GPT-4. At some point along that trajectory, I thought, "Holy cow, these systems, when you scale them, they get really good." I think the people who reached that conclusion earlier did better.

We apply the same scaling logic to Exa. We're building transformer-like systems for search, and we know that if you keep packing data and compute into the search engine, it will get better and better. That's our own scaling hypothesis for search. We're thinking a lot about where the future is going. We see a world where there are agents everywhere: GPT-5 level AI agents navigating the web, doing all sorts of tasks. This future is coming, and they're all going to need search.

I came into college wanting to study physics. I wanted to understand how the universe works, and I thought physics was the right way to do that. Something big that influenced me was watching The Social Network, because I was studying at Harvard and the movie took place there. It was actually a very accurate movie, and it was inspiring to see this guy change the world just on his laptop. I realized you could have a massive influence just coding on your laptop, in a way you couldn't as much with physics.

It was also clear that AI could understand the problems of physics. So I went into computer science, and I think that turned out to be right, because now AI is getting so good that it should be able to just tell us the answer, or infuse it into our brains.

Before Exa, on the side, I was writing a history book. I got really excited about world history and decided I was going to write a book about it, because no one had captured the level of excitement I had. So I was doing a lot of research for the book, and I quickly realized that Google is actually not good enough for that type of research. Google is great for surface-level investigations, but once you start trying to go deeper, trying to understand any topic deeply on the web, Google really fails. For example, if I want to find all the research papers on poverty in ancient Rome, it's actually really hard to find that on Google, because not every paper will mention the word poverty. I was doing the research for this book, and it was really hard to find things.

At the same time, GPT-3 had recently come out, and GPT-3 was this magical creature that I could talk to, and it could understand anything I said at a very deep, complex level. So the thinking was: what if we could apply the same technology behind GPT-3 to search? What if you could make a search engine that actually understands you at a deep level? It's been the same goal ever since.

The first year and a half of Exa, we did research into search models, into how we could take transformer models and apply them to a search engine. No one had really done that before, and it took a long time to figure out how to do it well. That required persistence. We were banging our heads against the wall for a year and a half, trying out different models, trying out different data sets, and eventually we got something that worked really well. If we didn't have the persistence, six months in we might have given up, but we didn't.

Early November 2022, we launched the first version of Exa to the public. We built a search engine that was perfect for AI applications. Basically, AI systems have all this intelligence, but they're lacking in knowledge. So when they need knowledge, they make a call to Exa and get exactly that knowledge.

Then ChatGPT came out a few weeks later, and that was a big moment for the information ecosystem. For us, it was really interesting, because we started getting requests for API access to our search engine, first from a friend who was actually living downstairs. I told him no, sorry, we don't have API access, and I didn't really think much of it. But then we got a request for API access from someone at a company in Germany. We also told them no, sorry, we don't have an API. Then we kept getting requests for API access, and we realized that because of ChatGPT, people were starting to build AI applications all over the place, for all sorts of businesses, and all these AI applications needed search. The AIs themselves needed to search, and that's when we started to realize Exa could be really useful for these AI applications.

So our initial customer found us. One lesson there is just be a really good listener to the market. What are people repeatedly saying they need? You might have some idea of what you're going to sell, but if people keep requesting something like API access, maybe you should start listening to them. We cared more about learnings from the customer than about getting a lot of revenue, so we kept opening our ears to what the customers needed, and over time developed a hypothesis about how we should price and what types of customers we should pursue.

I think you can guess where companies like OpenAI and Anthropic are going to build, based on what big markets they could tackle. Agents, automating work, is a clear huge market, so they're clearly going to do that. Now you know what types of things they want to do. Can they do it? Then you think about the fundamentals of LLMs: as long as you can make training data for some objective, the model will get better at that objective. Can you make training data for agentic behavior? Definitely. You just need a bunch of examples of navigating the web to buy plane tickets, and if you have a million examples of that, the LLM will know how to buy plane tickets. Any task you can create data for, an LLM will get near-perfect at. That's a fundamental way of thinking about this.

Will AI get really good at navigating the web? Yes, that's pretty easy, because it's easy to generate lots of web-navigation data. Will AI get really good at robotics? Yes, but probably on a longer time horizon, because gathering lots of robotic training data is harder. It's hardware, and hardware is hard to work with.

So what do we learn from that? AI navigating the web will come pretty soon. AI navigating robotics will take longer. Just from simple first principles, you can guess where things are going. Picture how good the technology and AI will be in a year, and build for that world.