William Bryk and the book that became a search engine

William Bryk wanted to write a history of civilization. This was an ambitious spare-time project for a software engineer, but the difficulty that caught his attention was surprisingly mundane. Before he could explain how people had lived, he had to find out how they had lived. Searching for accounts of medieval diets across different regions meant fighting through a tool that seemed better suited to a quick lookup than a sustained investigation. The footnotes were beginning to demand their own engineering department.

At Cresta, where he worked on AI, Bryk was spending his free hours researching the past. The experience helped sharpen an idea he had carried since school: the quality of the information people receive shapes what they understand and what they do. If the tools for finding that information are poor, the consequences extend well beyond one frustrated writer. He eventually left to start Exa. A book about civilization had helped turn his attention toward one of its working parts.

Today, Bryk is Exa’s co-founder and CEO, based in San Francisco. The company builds web search infrastructure for AI applications. Its customer might be a coding agent looking for documentation, rather than a person typing a question into a browser. Yet the person behind this machinery remains recognizably the reader who wanted a better way into a complicated subject. His recurring question is simple: why should finding what you mean be so difficult?

When the robot stood up

Bryk grew up in New York City and studied computer science and physics at Harvard, attending from 2015 to 2019. His undergraduate interests ranged from machine learning to the practical business of getting robots to work. By 2017, he was co-president of the Harvard Undergraduate Robotics Club. That title came with a useful education in the distance between an attractive idea and an object that actually moves.

The club had gone through a period when membership dwindled and no robots were completed. Then a four-legged machine rose to its feet after a long night of work. Bryk later recalled that moment as the point at which the group could see its project was possible. The BB-shooting robot went on to win at Shepherd University’s ShepRoboFest. By November 2017, the club had more than 100 members and eight project teams.

“We could build this,” he said. It is a modest sentence with a useful verb. The important discovery was shared: a group of students could make the thing they had been discussing. Reading his later descriptions of Exa, it is tempting to see the same preference for people gathered around a difficult problem, learning enough to make it work. The projects changed scale; the attraction of building together remained.

His curiosity also escaped the workshop. In a 2017 essay for Harvard Science Review, he asked readers to reconsider the familiar experience of existing at all. The opening imagined arriving in the universe with a fully capable mind and no accumulated knowledge. It was a student’s invitation to recover surprise. Long before search became his business, he was writing about how easily people stop noticing questions.

“We could build this.”

William Bryk, recalling the robotics club’s turning point, 2017

An old roommate, a new kind of query

Bryk’s engineering background included internships at SpaceX and Zoox. He has offered a rather good anecdote about the latter world of autonomous vehicles: he delayed learning to drive because he expected self-driving cars to arrive sooner. By his January 2025 telling, he had learned roughly two years earlier. Even a founder who believes in technological progress occasionally has to make alternative transport arrangements.

At Harvard, Bryk and his roommate Jeff Wang had already built a small search engine together. Their partnership therefore preceded Exa and the current appetite for AI infrastructure. When they started the company in 2021 and joined Y Combinator’s summer batch, they were returning to a problem they knew. GPT-3 had provided a new reason to think the problem could be approached differently.

Bryk saw a gap between a language model that could respond to a paragraph’s meaning and a search experience that often disappointed him on much simpler requests. He wanted retrieval to benefit from that same sensitivity to language. The initial ambition was a better search engine for people. The company was called Metaphor, a name that suited an attempt to search through relationships in meaning.

The technical idea involved learning from links people shared online. Rather than only matching words in a query against words on a page, the team worked on models that could connect a description with a relevant destination. Bryk’s interest lay in the selection itself. An eloquent answer built from unsuitable material would still leave the reader badly informed. Good prose cannot rescue a bad reading list.

The customer arrived two weeks later

Exa’s first search engine launched in November 2022. Two weeks later, ChatGPT appeared. In Bryk’s account, requests for API access followed: developers building AI applications wanted their software to retrieve information from the web. The team had been trying to improve how humans searched. A new class of users was asking for the same underlying machinery.

That change made the business more specific. A person can open a page, scan it and decide whether to continue. An AI application needs information supplied in a form it can process, often across repeated searches. Bryk’s team moved toward providing search through an API, a connection other software could use. In early 2023, the company announced its web search API for AI.

On January 25, 2024, Metaphor became Exa. Bryk explained the choice with a numerical joke: exa denotes ten to the eighteenth power, while a googol is ten to the hundredth. His argument was that useful knowledge is a smaller, more organized collection than the surrounding flood of information. A search company was advertising subtraction. The name also gave him a tidy way to express what he wanted to keep.

The rebrand came with Highlights, a way to extract relevant passages from retrieved webpages. In a February 2024 interview, he described users who wanted a handful of results, others who wanted thousands, and still others gathering large datasets. The work ranged from feeding an answer to assembling material for analysis. The common requirement was relevance at a scale a human reader would struggle to manage.

A search box with higher expectations

Bryk’s January 2025 essay, “A Perfect Search Engine,” made his ambition explicit. He recalled his father asking whether Google was already adequate when he proposed a new search engine in 2021. It is the sort of sensible parental question that can make a startup sound like an elaborate refusal to take advice. Bryk’s answer was to describe requests conventional search had trained people to avoid.

He imagined finding people with particular combinations of experience, locating related ideas across different kinds of media, and retrieving every item that met a set of conditions. These were descriptions of a desired future. The word “perfect” was his standard for the work ahead, rather than a settled verdict on the product. A list that misses half its members cannot become complete by acquiring a handsome interface.

His earlier essay on “superknowledge” separated the ability to reason from the ability to retrieve information. A model might be capable of working through a difficult question while lacking the material needed to answer it. Bryk wanted retrieval to become a technical priority in its own right. This explains why his company is comfortable working underneath someone else’s application: the information entering the system is where he wants to make a difference.

He has also acknowledged practical compromises. In a September 2024 conversation, Bryk explained that exact lookups can suit keyword search, while meaning-based methods suit other requests. The aim was to combine approaches intelligently. That detail matters because it gives the philosophy a working texture. Building better search requires deciding which method helps with a particular question, as well as making a persuasive argument about the future.

The figures got bigger. The problem stayed familiar.

Exa’s financing charts trace the increasing scale of that effort. In July 2024, it announced $22 million in combined seed and Series A funding, led by Lightspeed with participation from NVentures and Y Combinator. In September 2025, Bryk announced an $85 million Series B led by Benchmark at a $700 million valuation. Benchmark’s Peter Fenton joined the board.

On May 20, 2026, Exa announced a $250 million Series C led by Andreessen Horowitz at a $2.2 billion valuation. The company reported serving more than 5,000 companies and over 400,000 developers, naming Cursor, Cognition, HubSpot, OpenRouter and Monday.com among its users. It also said its crawlers tracked more than 500 billion URLs. These are company-reported figures, and they describe infrastructure and adoption rather than a complete solution to search.

The work has brought an unexpected business pairing. In April 2026, Exa announced integrations with Google’s Gemini Enterprise offerings, including grounding responses in fresh web results and making the Exa Agent available through the Agent Marketplace. Bryk’s criticism of familiar search tools can coexist with supplying another part of the same industry. The company’s information can travel through somebody else’s interface.

Meanwhile, the day-to-day challenge becomes more exacting. In March 2026, Exa released WebCode evaluations for search used by coding agents. The team described the importance of fresh documentation and clean extraction: an agent may build on retrieved information over many steps. For Bryk’s original concern, this is a concrete test. A source that seems roughly relevant can still send a useful-looking answer in the wrong direction.

The soundtrack, and the classroom

Bryk’s public descriptions of leadership favor independent thought. He wants recruits to work through real problems, and has described one- or two-day work trials that reveal how they collaborate as well as how they code. His cultural references are less conventional management reading: Spider-Man, Batman and Ethan Hunt figure among his inspirations. He has said he wakes to “Firefight,” from the Mission: Impossible soundtrack. The search infrastructure business apparently has room for an opening score.

A June 2026 return to Ramaz, his former school, supplied a smaller measure of achievement. Bryk told students that he had wanted to accomplish something impressive enough to be invited back to speak. He pointed to the classroom where his concern about access to reliable news had begun. He also discussed unresolved questions around AI, including copyright and how systems choose answers, without presenting them as finished problems.

That visit brought the grand language of organizing knowledge back to a recognizable place: a classroom, some questions, an alumnus explaining what he had decided to work on. The history project had helped send him toward search, but the concern was older. Bryk wanted people to be able to understand the world they were living in. He now pursues that wish through machines that retrieve information for other machines. Somewhere at the end of the chain, a person still needs to know.