Exa: Organizing the World’s Knowledge
Will Bryk, co-founder and CEO of Exa, discusses the journey and technology behind building a search engine from scratch, specifically designed for AI systems. He explains why traditional search engines fail in an AI-first world and how Exa's full-stack, compute-heavy approach aims to solve the world's "information blocker" problems.
Exa is building a search engine from first principles, designed not for humans, but for AI. Traditional search engines like Google were optimized for human behavior: short keyword queries, a list of links, and a user interface. AI systems, however, operate differently. They are not “lazy” and can process complex, paragraph-long queries, scan thousands of results, and demand only the highest quality knowledge, free from SEO and ads. This fundamental difference requires a completely new set of design choices.
Exa’s approach is to provide infrastructure that works “under the hood” of AI applications, serving companies rather than consumers. This B2B model allows them to build features that incumbents won’t, focusing on the specific needs of AI systems.
A Contrarian, Full-Stack Approach
A core, contrarian decision was to build the entire search stack from the ground up, including their own index and models, rather than wrapping an existing service. This provides critical advantages:
- Deep Customization: Customers can make specific requests, such as searching only over a curated list of a thousand domains or requesting a thousand results instead of the typical hundred. This level of control is impossible when relying on a third-party API.
- Enterprise-Grade Features: Owning the infrastructure enables essential features for clients in sensitive industries, like true zero-data retention for financial customers.
- Destiny Control: It allows Exa to control its own technological path and not be beholden to the limitations of other platforms.
This strategy extends to model development. The team found that off-the-shelf embedding models, while useful for searching millions of documents, fail at the scale and chaos of the hundred-billion-document web. To overcome this, Exa made another contrarian bet early on: investing a significant portion of its seed funding into its own GPU cluster. This “pour compute into the problem” philosophy continues today with a multi-million dollar cluster of H100s, enabling the development of proprietary models trained on the world’s knowledge.
An LLM is trained on the world’s data, but then it compresses it into a set of weights… the problem is it can’t memorize the whole web… The web is a huge amount of petabytes and then the GPT-4 size model is a few terabytes. You literally can’t fit the whole web into the weights of this model.
Evaluating the quality of a search engine at this scale is a major challenge, as no standardized public benchmarks exist. Exa is developing its evaluation frameworks in-house and aims to become a thought leader by publishing its methodologies to help standardize the field.
The Pivot to AI Infrastructure
Exa began four years ago, before the release of ChatGPT, with the goal of building a better search engine for “nerdy humans”—users who, much like AI, crave high-quality knowledge without commercial fluff. They had already built a product that resonated with a technical audience.
Two weeks after their initial launch, ChatGPT was released. Soon after, Exa began receiving requests for API access. The team initially dismissed them, but the persistent demand led to a crucial realization.
“There was this big moment when me and Jeff were chilling in his room and like we were like, ‘Wait a second. What if we’re a search engine for AIs?’ And the second we said the expression, it all like clicked into place and was like, ‘Oh my god, that is the world that is coming.’”
The product they had built for knowledge-hungry humans was perfectly suited for AI systems. The company pivoted almost immediately, embracing its new identity as a search engine for AI.
The Future of Search: Agents and Perfect Information
Looking ahead, the world is moving from single LLM calls to more complex, autonomous AI agents. These agents will perform numerous actions in the background, including potentially hundreds of searches for a single user request. In this paradigm, speed is critical. Reducing search latency from one second to 100 milliseconds represents a 10x improvement in the agent’s overall performance.
Exa’s ultimate vision is to “organize the world’s knowledge for real,” completing the mission Google started. This means solving the “coordination problem” in the world by eliminating information blockers.
- Recruiting: A company should be able to instantly find every engineer in the world who perfectly matches their requirements.
- Sales: A business should know about every potential customer that fits its profile, including new ones the moment they emerge.
The goal is to create a future of “perfect information,” where access to knowledge is no longer a bottleneck, making people wonder how society ever functioned without it.