Will Bryk

Will Bryk co-founded Exa, formerly Metaphor Systems, with Jeff Wang in 2021. As CEO, he has pursued neural web search built to understand complex requests from people and AI systems. His public concept of superknowledge describes comprehensive retrieval over available information—a capability he argues increasingly capable language models will still need.[7][1][6]

Exa's initial training objective used the text surrounding links on the web: hide a link and train a model to predict the document it refers to. In his January 2025 interview, Bryk clarified that this was document prediction, rather than generation of a memorized URL string. For example, text recommending a “really cool aerospace startup” could teach the system to retrieve the linked company's document. The team later supplemented this initial objective with synthetic data and supervised fine-tuning.[1]

Bryk called the objective “a neural PageRank”: frequently shared documents supply a signal of importance, while the different ways people describe the same document teach semantic associations. His example was a Paul Graham fundraising essay referenced in many different ways. He also emphasized that retrieval quality depends on the training distribution and the team's selection of high-quality sources.[1]

Bryk's account of building Exa's first version includes a consequential implementation detail: he retained the first 1,024 tokens of document content for debugging. When users subsequently asked for content alongside search results, that stored text made the request easier to fulfill. Serving content became useful to customers building AI applications that needed the documents themselves, rather than only lists of links.[1]

He also participated with fellow engineers in an exercise labeling search results for niche queries. Bryk recalled that unfamiliar subjects, such as biological experiments, made the task difficult and tempted them to judge relevance by matching keywords. In January 2025, he reported that the team was using LLMs as search labelers, citing their ability to interpret those documents.[1]

Bryk distinguishes two problems in complex search: deciding what the user means and finding a comprehensive set of matches. For a request such as hardware startups in San Francisco, examples and feedback can help establish whether robotics or wearables belong in the result set. Comprehensiveness, in his account, requires a separate decision about how much computation to spend searching. Exhaustive retrieval is the goal of this approach, rather than a demonstrated guarantee.[1]

He argues that useful searches can justify minutes, ten minutes, or even a day of computation. Previews let users refine their intent before committing to a larger job. Bryk described Exa's search as “a combination of an algorithm and an agent,” with a loop that decides when to investigate more deeply. He also acknowledged that an agent can fail because it lacks the user's context, and agreed that users may need to participate throughout the work.[1]

These ambitions involve practical limits. Bryk explained that running LLMs over billions of documents at query time is infeasible; preprocessing documents and reranking a smaller candidate set help manage costs. In January 2025, he reported Exa's purchase of a $5 million H200 cluster for training and research, alongside a mix of owned infrastructure and AWS inference. Despite interest in searching company documents, he advocated keeping the team focused on web search.[1]

In his July 12, 2024 essay, “We need superknowledge before superintelligence”, Bryk defined intelligence as reasoning over an input and knowledge as retrieval from a repository. He argued that improvements in reasoning would remain bottlenecked by inadequate retrieval. His proposed superknowledge API would handle complex requests over available information, such as an organizational chart of AI startups filtered by their founders' experience and sorted by employee count.[6]

Bryk linked this mission to Exa's usage-based revenue model, arguing that charging for retrieval aligns the company with users' desire for control over the information they receive. In January 2025, he favored an efficient, smaller language model acting as an intelligence module paired with a search tool. He also argued that an LLM should translate user intent into detailed search queries, making support for paragraph-length requests a central design requirement.[6][1]

At AI Engineer World's Fair 2025, Bryk was associated with the launch of Exa Research. The June 13 event recap also named him best speaker in Search.[3]

The earlier history of Exa puts these technical choices in context: Bryk recalled entering Y Combinator in summer 2021 with a better-than-Google pitch, prompted by the contrast he saw between GPT-3's language understanding and conventional search.[1]

Bryk studied computer science and physics at Harvard, researched human–AI interaction, and led the robotics club. He was one of Cresta's first engineers, building real-time AI products. He also worked at SpaceX; recalling that experience in 2025, he said his enthusiasm for Elon Musk had motivated him to join one of Musk's companies. Bryk remembered building a miniature search engine in college.[5][1]

Bryk and Wang were Harvard roommates. Bryk recalled meeting Wang on the first day of college and remaining best friends. Wang brought experience building data and web infrastructure at Plaid. Their company was still known as Metaphor when Bryk appeared as its CEO in Latent Space's December 2023 NeurIPS recap; by his January 2025 interview, it was Exa.[5][1][2]

Bryk described the culture he and Wang built as intense but fun, with considerable laughter. One concrete decision was importing a nap pod from China so people could rest at work; he recalled struggling with the delivery driver to unload the roughly 500-pound pod before a passerby helped. The team also used its own search engine operationally: Bryk reported finding prospective customers through Exa and using it to identify people interested in search for recruiting.[1]