Google Gives You 33 Million Results. Exa Wants 54.

Will Bryk opens with a distinction that runs through the hour. Google is a search engine: you roughly know what you want. When you do not, you are doing research, and that is where Exa is supposed to shine. He prefers a database metaphor to a search box. Every query should take the world's knowledge and filter it down to just what you asked for. Google's habit of printing "about 33 million results" at the top of the page is, in that frame, a confession. "What am I supposed to do with 33 million results?"

Nathan Labenz has spent months trying the new information products and giving each a nickname. ChatGPT is still his workhorse for ad hoc work; Perplexity is the answer engine for a fast, accurate paragraph; you.com's Research Mode is an autonomous assistant that searches, reads, and synthesizes multi-page reports; Elicit is a structured literature-review tool that lays papers out in a table. Exa still returns links, like a traditional search engine. After using it he decided the gap is quality, controllability, and depth, and he started calling it an example engine — a machine that emits lists you can feed into other AI workflows.

Bryk tells Labenz that Exa is a search engine built for AI systems, which in practice means redesigning the algorithm rather than wrapping Google. People have lived with Google for a long time, he says, while models like GPT-3 and GPT-4 now understand text at near-human level. The bet was to combine that understanding with a search stack that had barely moved in a decade. Exa is a research startup: its own GPU cluster, its own foundation models for search, and an appetite for the complex queries Google is "really bad at." They had been at it for a couple of years when this March 2024 conversation was recorded.

A line on Exa's site had stuck with Labenz: Google has limited people's imaginations about what you can find on the internet. Bryk says people think search is solved, then try to get a list of startups applying AI to law and discover that Google will not give them a list of startups. Add "in the Bay Area" and it still will not. Keep stacking modifiers and you realize the engine is simply not doing that job. Whenever you leave Google — LinkedIn for people, PitchBook for companies, Twitter for high-quality news, a dating app for a date — it is because Google failed to filter the world's information into exactly the knowledge you wanted.

The queries that expanded his own sense of what search could be are the ones you would not even bother typing into Google. Paste a designer's page and ask for similar designers: similarity search with a whole webpage as the query. Find researchers in the Bay Area who have Rust experience. Find articles that really argue why Rust beats Python for multiprocess applications. "You don't think to type that into Google because you know it's just not going to work." The vision is to take the petabytes that are already out there and, in real time, filter them to exactly what you asked for.

Labenz offers "example engine" as a label. Bryk admits that even the people who have been building it are not sure what to call it; research engine might be closer. Google, and most tools sitting on top of it, still start from a keyword engine and add a post-processing step. Search, for him, is "Taylor Swift boyfriend" or your own name: you know the target. Research is explaining, in natural language, the type of thing you want and getting a matching list. He likens Exa to a library. You walk up to the librarian and say you want startups applying AI to law. Some of those companies never use the words "AI" or "law"; they say they apply a machine-learning algorithm to legal documents. A smart librarian can still point you there.

The company is API-first. The current consumer interface is a list of links because they want people to see exactly what the API returns, not because a link list is the right shape for consumer search. Language models, Bryk says, will be a huge part of that interface. They are exploring something more like a researcher's workbench — the kind of product they hope customers will build on Exa.

Labenz and a friend are advisers at Athena, where AI assistants are supposed to help clients hit goals they do not themselves know how to unpack. One example: increase lung oxygen capacity. The assistant has no special expertise, but the expertise is on the web. Exa supplies a list of expert articles; language models turn those into concrete actions, such as ordering a machine or signing up for a course. Almost every company Labenz talks to wants to scale some form of lookalike search — more accounts like the ones that closed, more résumés like the star hire. Models can already personalize outreach at roughly human quality. The unsolved input is the list itself. Garbage in, garbage out, and that is where he thinks Exa actually shines.

Bryk's favorite professional example is a paper-writing tool that uses ChatGPT to unblock authors with ideas for what to write next. ChatGPT is not smart enough alone, so it needs search, and a traditional engine is a narrow pipe. With Exa you can paste a paragraph of your own thinking and ask which paper would naturally follow that idea — a query Google cannot run. If the product is opening the world's knowledge to academics who spent two decades on a limited tool, he says, that is why he started the company. Another tool is a chatbot for venture capitalists that recommends startups they would not have found through ordinary channels. Better search, in his phrase, can enable new marketplaces.

The stranger end is people search. Paste a dating profile, get twenty similar people; Bryk knows someone who has met several people in real life because of Exa. For a while the consumer page said "Rediscover the internet," because the interesting pages were always there and the old algorithm hid them. Someone described the feeling as uncovering a hidden temple of knowledge. His division of labor is clean: Exa is the knowledge engine, ChatGPT is the combining engine. Find a hundred relevant documents, hand them to GPT-4 and eventually GPT-5, and that is when the experience gets magical. Most of the products he is proud of have an AI system doing the searching.

"Search built for AI" is a provocative phrase, Labenz says, but it is not obvious how it cashes out. Bryk's answer is that an AI searcher is almost the inverse of a human one. Humans are lazy and type a few keywords; an AI can instantly emit a paragraph stating exactly what it wants, so the engine has to handle those complex queries. Google was optimized for things like "Walmart homepage." Humans click two or three links and move on; an AI can gobble a hundred results, so you need to return a hundred, not a handful. Humans searching "Taylor Swift" often do not know what they want and prefer a diverse mix. An AI that asked for a hundred companies does not want a company, then a blog post, then a LinkedIn note, then a Reddit thread.

The last mismatch is format. AIs do not want links and titles. They want the full page, because they do not want to sit in a browser and read. You can wire an AI to Google, Bryk says, but then you are connecting a new kind of creature to something designed for organic humans, and it is a bad pairing. Finding information is such a central part of human experience that he does not want to give this "super new thing" an old technology. Labenz recaps four demands: deeper queries, more results, actual content rather than URLs, and control over the type of thing that comes back.

Labenz asks how this compares with Google's move from PageRank toward BERT-style semantic ranking. Paste a paper abstract into Google and ask for similar papers, Bryk says, and you get copies of the same paper rather than neighbors. Nobody outside Google knows the full algorithm, but the BERT piece looks to him like a rerank at the end: keyword-filter the web to some candidate set, then apply a neural method. That is very different from going neural from the first step. If you filter to 10,000 and the relevant pages never made the cut, reranking will not help. Google is more than keywords — a knowledge graph will inject facts about Obama if you mention him — but Bryk still calls the core a keyword-based algorithm.

Exa's training starts from places on the web where people talk about links. Imagine a Reddit comment: "check out this cool aerospace startup," then a URL. They hide the link and ask a transformer-like model to predict which document followed. Do that hundreds of millions of times and you get a search engine. At query time, "aerospace startup" predicts the most likely links to follow. If people citing "a great fundraising essay" keep pointing at Paul Graham, the model learns to return Paul Graham. Bryk calls it learned PageRank, and says it is only v1.

Labenz's obvious objection is that ChatGPT can also complete "here is a great essay on fundraising https://", token by token, and will invent URLs that go nowhere. Bryk agrees they are not doing that. The model predicts a document ID from an index, not characters of a URL. You cannot just fine-tune a transformer on next-token URL prediction, because that forces the model to memorize addresses, which does not work. They had to invent a new objective. It is embedding-based: the transformer outputs an embedding, and retrieval happens in that space. Getting that precise at scale was about a year of research before they shipped product on top of it — something, he notes, very few startups actually do.

Labenz has a standing hypothesis that systems will increasingly talk to each other in embeddings rather than natural-language strings. If both ends already work in an embedding space, translating into tokens and back is inefficient and lossy — and also harder for humans to interpret. Exa looks to him like a glimpse of that future. Bryk's reply is immediate: embeddings feel like the past. He would be disappointed if they were how we passed knowledge around later. Compress text into one vector, take a dot product against document vectors, and you have already thrown information away. Exa is embedding-based today and they have squeezed a lot from it. It is still fundamentally limited.

"Startups applying AI to law in the Bay Area where the founders have Rust experience" is the running example. You cannot expect one embedding and one dot product to handle stacked modifiers. He describes a spectrum. At one end, query embedding dot document embedding: cheap enough to run over the whole web, and weak. At the other, drop the query text and the document text into a transformer's context window, run all the weights, and emit a score: far more powerful, and impossible to do over the whole web. The middle of that spectrum, he says, is almost unexplored. That is what v2 and v3 are about: moving beyond embeddings. Today's best systems will probably mix keywords and vectors. The ceiling is higher.

Because the model was trained to predict the document that followed a stretch of text, the best way to query Exa is to write a prompt that a link would naturally follow. If you want startups, end with `(startup:`, because on the web that marker is usually followed by a company homepage. Bryk compares the moment to GPT-3 before ChatGPT: a completion model you had to prompt in awkward ways. RLHF made the chat product easy enough for his mom and also less controllable; Exa has not poured resources into that kind of smoothing. Most customers are businesses, and they can ask ChatGPT to write the strange prompt.

Labenz's first idea for stacked modifiers is to decompose the query, run several nearest-neighbor searches, and combine the sets. Bryk says that intuition is right, and they are hiring for it; it is also how humans handle a hard question. It still fails when facts live on different pages: a legal-AI site rarely mentions that the founder wrote Rust. The way through is knowledge-graph-like structure that never becomes discrete metadata, so you can run fuzzy, arbitrary queries over the graph. Improving search, he argues, is oddly neglected next to generating text, images, and video — you no longer see many PhD students publishing on web search, and Labenz jokes that the best minds of the generation are still elaborating chain of thought.

Publicly, Labenz had seen about $5 million raised, and he wondered how a web-scale index and a GPU cluster fit that budget. Bryk says they try to be frugal, and they are not crawling the entire web. They crawl a high-quality subset their customers care about, which lets a lower-resource startup run more complex algorithms. Continuously crawling hundreds of billions of pages, Google-style, is expensive; crawling fewer pages on a slower cadence is something a startup can do. Other search startups, he thinks, tried to be comprehensive too early and ended up okay at everything instead of great at a few things. The usual advice holds: make a few people love you.

Labenz mentions Brave's index, built from where people actually browse. Bryk's analog is shares rather than dwell time. The model is trained to predict the links people pass around. Nobody shares the SEO blog post about a recipe with ads all over it; they share the good recipe. So the model is trained not to predict SEO. It learns raw, recommendable content. That is the anti-spam of this architecture: the training data is human endorsement, which is why a small index can still feel like a better internet.

On pricing, Labenz's first look said expensive: about $10 per thousand searches for up to 25 results, more for 100, with enterprise deals on top, versus the $1 to $2 per thousand that API products often charge. Talking through lookalike search for sales and recruiting, he reversed himself. Quality at the retrieval step can save iteration time and downstream language-model cost. Bryk says customers rarely complain about price because they arrived after Bing's API or Google wrappers failed to return only people, only companies, or only news. Exa positions itself as premium search you cannot get elsewhere. The usage patterns he will name are retrieval-augmented generation, automated analysis, and the extreme case of pulling hundreds of thousands or even millions of results to fine-tune models.

Labenz's advice to builders is not to optimize cost first. Costs are falling; the first question is whether you can make something that works, using GPT-4 or Exa or whatever is best, and only later take the enterprise plan or wait for prices to drop. Evaluations are a separate headache. Score Exa on Google-shaped queries and Google wins; score Google on Exa-shaped queries and Exa wins; public web-retrieval datasets were built for the old way of searching. Their fix is to break a subjective "is this good" into objective abilities: if the query is essays by Paul Graham, did you return Paul Graham; if the topic is space exploration, is the page about that. Subjective quality becomes a set of internal, checkable skills. Nothing like it sits in open-source benchmarks.

They built their own vector database because web-scale search with the filters they want is not a nearest-neighbor lookup. Open-source vector stores, compared in-house, were not designed for those filters — including searching with one embedding and using another as a filter — and generality made them slow. A 2x or 5x win on a specific workload has massive returns. Bryk likes the database analogy because this is not really search. Query "startups applying AI to law" and you should get 367 results, all of them. Add "Bay Area" as a filter and it drops to 54. Google's "33 million results" leaves you overwhelmed and still not holding the thing you asked for. They are trying to combine neural and non-neural methods in a new kind of database, which is why they wrote their own.

Labenz guesses at multiple embeddings per document, projected onto authorship or geography. Bryk says that fails: there are too many possible directions to precompute, so you need a more general method. Would they sell the database? He refuses. The team is focused on one mission: filter the world's information to the knowledge you need, as fast as possible. Crawling, parsing, training, and serving all had to be excellent, and any of those layers could be a product. A company this size cannot sell vector databases and search and a crawler. Focus.

They are a small team that works intensely and does not experience it as work. A fifty-person team that treats the problem as a job, Bryk says, will move slower than a seven-person team that treats it as play. Labenz notes they are young, some living together in a house that is also the office, bucking Sam Altman's remark that founders are getting older. Two values jumped out from the careers page. "Perform with an optimal speed-accuracy tradeoff" means not every query should take the same compute. Google assumes a few hundred milliseconds no matter what you type; some queries are complex enough to deserve seconds or minutes of scouring. The same rule applies inside the company: some problems deserve a day of thought, some a year. Optimize the exact amount of effort rather than always moving fast and breaking things.

"Wield great power responsibly" follows from being a doorway to the world's knowledge. People who control knowledge control power. Google already shaped what content gets created and what people even know exists. Exa's philosophy is to give users full power to get what they asked for, rather than bias results toward what the company thinks is right. Bryk admits there are moral questions — you could warn people the way Twitter adds context — but with transformer-like technology you can finally give people exactly what they ask, however complex. He expects Exa to be the dominant player. The information-retrieval market is going to get much bigger, so there is more room for multiple companies than before, and still benefits of scale. They are hiring.