← All speakers

Will Bryk

Conference affiliation: CEO · Exa · 2026

Will Bryk is the co-founder and chief executive of Exa, which builds search infrastructure for AI agents. He wants software to query the web with the specificity of a database: finding companies, people, or arguments that meet detailed conditions, rather than returning pages that merely mention the right words. His work combines a technical problem—making neural retrieval affordable at web scale—with a conviction that better access to information is essential to informed decisions.

On this page

Will grew up in New York City and studied computer science and physics at Harvard, where he researched human-AI interaction and led the undergraduate robotics club. His early engineering work included internships at SpaceX and Zoox; at Zoox in 2018, he built teleoperation capabilities for a self-driving vehicle. From August 2019 to June 2021, he worked at Cresta, building AI recommendations for human support agents and leading its voice team.

Alongside his work at Cresta, Will was writing a book about the history of civilization. He connects his decision to start Exa to the conviction that civilization’s future depends on the quality of the information people consume. This is his stated motivation, and it remains central to how he describes the company: finding information should help people understand their options and the world around them.

He started the company originally called Metaphor in 2021, and it joined Y Combinator’s Summer 2021 batch. His co-founder, Jeff Wang, had been his Harvard roommate. GPT-3 supplied the technical opening. Its ability to interpret a paragraph describing a nuanced request suggested that neural networks could bring similar sensitivity to retrieval. Will repeatedly illustrates the gap with a search for shirts without stripes: matching words can miss the exclusion that determines whether a result is useful.

The team spent roughly eighteen months researching models, training data, and infrastructure before launching its first search engine in November 2022. ChatGPT arrived shortly afterward and unsettled their assumptions about the product. They had been building a better search engine for people, without initially planning an API business. Developers then began asking for API access. Will’s account of the transition includes an initial refusal—the API did not exist—followed by the realization that language models needed retrieval for information they could not supply from memory. The web was both too large and too changeable to be contained in model weights.

Serving AI applications gave that original search ambition a business model. Agents could issue detailed queries and process far more results than a person browsing links, while developers would pay for the information their applications needed. In January 2024, Will announced the Exa name alongside customizable highlights for extracting relevant passages from retrieved webpages.

Making a difficult query affordable

Will’s founding thought experiment was to give a language model a query and a document, ask whether they matched, and repeat across the web. The obstacle was the cost of evaluating every document for every search. Embeddings move some computation ahead of the query: neural models turn documents into numerical representations, which retrieval can compare with an embedded query without running a language model over every page again.

His architecture account explains why this requires more than an embedding model. The Exa team discovers and crawls URLs, parses documents, selects material for the index, trains specialized retrieval models, and serves queries through a custom vector database designed for high query volume and variable computation per search. Will applies Richard Sutton’s Bitter Lesson to retrieval, favoring learned methods that can improve with additional training data and computation.

His practical choices show what that approach is meant to deliver:

  • Retrieve the requested entities: In his 2024 announcement, a search for climate-change startups returns company websites rather than articles listing companies. Relevant results need not contain the exact words “climate change” or “startup.” For an agent assembling a dataset, finding the entities themselves avoids another layer of extraction from commentary about them.
  • Match compute to the question: In his March 2025 architecture article, Will distinguishes Exa Search, which returns quickly, from Exa Websets, which spends longer on difficult investigations such as finding companies that satisfy several conditions. A simple homepage lookup and a comprehensive market search justify different computation budgets. His aim is to let applications choose the waiting time appropriate to their task.
  • Combine retrieval methods: Will’s search-agent demonstration uses neural search to discover engineers interested in information retrieval, then keyword searches on their names to find GitHub accounts. Semantic discovery handles the descriptive request; keyword lookup handles a known person and destination. Despite his enthusiasm for embeddings, he gives agents both methods rather than asking one to solve every step.
  • Return the material the application needs: Highlights extract relevant passages instead of requiring a downstream model to process whole pages. In his 2026 talk, Will also describes structured results and Exa Connect, which provides access to participating data providers alongside public-web retrieval. His example combines company research with Similarweb traffic data. A voice agent may prioritize a fast response, while a research application may need specific fields or a more exhaustive result.

Information quality at scale

In an August 3, 2026 update, Will reported that Exa served 80 billion pages while tracking 1.4 trillion URLs. The distinction matters to his approach: the system must crawl more than it serves and train models to exclude low-quality material that could weaken AI outputs. He also described the capacity demands of agent workloads, where a single investigation can fan out into dozens of searches and model-training workloads can produce spikes in traffic. Later that month, he announced Jeff Pinner’s arrival as CTO to help scale Exa’s engineering and research.

Will’s argument for an API business extends beyond demand. In his 2024 writing, he describes customers paying for useful retrieval as an incentive to improve information quality without depending on advertising revenue from attention. His 2026 talk develops the societal stakes: more powerful AI cannot ensure better decisions if people lack relevant information or cannot trust what they find. Near-perfect search is his ambition, rather than an established capability. The concrete work toward it is selective as well as expansive—finding more relevant material, rejecting poor results, and spending additional computation where a question warrants it.

Read the topics behind these talks

2 conference talks

Loading more talks and highlights…

References