Transcriptly

Building AI Search Engine for the GPT-5 Era | Will Bryk, Exa

EO08:52Added Aug 31, 2026

“In five years, we can have systems that completely automate all human labor, AI does all the repetitive work and humans do all the novel work.” Will Bryk is the co-founder and CEO of Exa, an AI search engine startup. Unlike traditional keyword-based search engines, Exa analyzes the meaning

Watch on YouTube →
Contributed by 刘嘉琪

Transcript

Transcript format

  1. Intro

  2. 00:00In 5 years, we could have AGI systems
  3. 00:01that completely automate all human
  4. 00:03labor. AIs do all the repetitive work
  5. 00:04and humans do all the novel work. Every
  6. 00:06human is going to become a product
  7. 00:07manager of a team of AIs. How are you
  8. 00:09supposed to plan for that? I think it's
  9. 00:11extremely hard to predict where the
  10. 00:12world will be in 5 years. One-year plans
  11. 00:14make sense right now. Threeear plans are
  12. 00:16really hard and 5ear plans are
  13. 00:18impossible. So, because the AI market is
  14. 00:20changing so fast, like every month, like
  15. 00:22new systems come out that make new
  16. 00:24things possible. The right way of
  17. 00:26navigating that is to think from first
  18. 00:28principles about like what does the
  19. 00:29market need that still will be true in a
  20. 00:31year. Because if you're thinking about
  21. 00:32what does the market need right now, a
  22. 00:33month later something new is going to
  23. 00:34come out and they're not going to need
  24. 00:36it anymore. So you have to think a
  25. 00:37little more longterm. You have to be a
  26. 00:38little more strategic than in the past.
  27. 00:40You know, for example, like when I was

  28. Argument with Greg Brockman

  29. 00:41in college, AI was not at all as
  30. 00:43prominent as it is today. Everyone today
  31. 00:45talks about AI, but there were a few
  32. 00:46people who were aware of what was going
  33. 00:48on. I was lucky enough to go to a
  34. 00:50Westworld watching party at OpenAI and I
  35. 00:52was actually on a bean bag with Greg
  36. 00:54Brockman um just chilling on this bean
  37. 00:56bag and we were arguing about the
  38. 00:58scaling hypothesis and the scaling
  39. 00:59hypothesis is this idea that you keep
  40. 01:01putting more compute into transformers
  41. 01:04and they'll just keep getting better and
  42. 01:05that's how we get to
  43. 01:07AGI. That was a crazy idea when I was in
  44. 01:10college. The vast majority of people
  45. 01:11didn't believe it. I didn't believe it.
  46. 01:12Greg was arguing that if we just keep
  47. 01:14scaling these things, we'll get there.
  48. 01:16We'll get to AGI. And I was arguing that
  49. 01:18we need new types of methods. I think we
  50. 01:20were both right in our own way, but we
  51. 01:22kept seeing progress from GB1 to GB2,
  52. 01:24GB3, eventually GB4. And at some point
  53. 01:27along that trajectory, I was like, "Holy
  54. 01:29cow, like these systems when you scale
  55. 01:31them, they just get really good." I
  56. 01:34think the people who were earlier to
  57. 01:35that conclusion did better. And we kind
  58. 01:37of apply the same scaling logic to Exa.
  59. 01:40We are building transformer-l like
  60. 01:41systems for search. And we also know
  61. 01:44that if you keep packing data and
  62. 01:45compute into the search engine, it will
  63. 01:47get better and better. That's a
  64. 01:48hypothesis we have. It's like our own
  65. 01:50scaling hypothesis for search. It's very
  66. 01:51different from traditional search
  67. 01:52engines like Google or Bing which
  68. 01:54basically have stayed the same for you
  69. 01:56know many years whereas like X is
  70. 01:58getting better like this. And so we're
  71. 02:00thinking a lot about like where is the
  72. 02:01future going? Uh so we see a world where
  73. 02:03there are agents everywhere. GBD5 level
  74. 02:05AI agents navigating the web doing all
  75. 02:07sorts of tasks. This future is coming.
  76. 02:09They're going to need search all of
  77. 02:11them.

  78. How Is Exa Different from Google?

  79. 02:14Hey, I'm Will. I'm the CEO of Exa. We're
  80. 02:16building the next generation of search.
  81. 02:17One good way of understanding Exa is in
  82. 02:19contrast to traditional search. So,
  83. 02:21traditional search engines use mostly
  84. 02:23keywords. Okay? So, if you're using a
  85. 02:24traditional search engine and you want
  86. 02:26to find startups working on futuristic
  87. 02:28hardware in the Bay Area, traditional
  88. 02:31search engines will use keyword
  89. 02:32matchings. The results they give you
  90. 02:34will be documents that contain the words
  91. 02:35startup and hardware and Bay Area. But
  92. 02:38startups that are working on futures of
  93. 02:39hardware in the Bay Area, they don't
  94. 02:40typically have those terms. Like you
  95. 02:42might have a a rocket company in SF. A
  96. 02:45traditional search engine won't be able
  97. 02:46to find that rocket company. But Exa can
  98. 02:48because uh we understand the meaning of
  99. 02:50documents. We understand the meaning
  100. 02:51that oh, if it's a rocket company in SF,
  101. 02:53then it does match startups working on
  102. 02:55futures of hardware in the Bay Area. We
  103. 02:58believe it's possible to have perfect
  104. 02:59search over the web, meaning whatever
  105. 03:01information you want, you get exactly
  106. 03:03that. We help companies integrate this
  107. 03:05high-quality knowledge into their
  108. 03:07applications. So we recently raised $17
  109. 03:09million from Lightseed and Nvidia. Our
  110. 03:11revenue is doubling every quarter. We're
  111. 03:12building the next generation of
  112. 03:17search. I came into college wanting to

  113. The Movie That Changed Me

  114. 03:20study physics to like I want to
  115. 03:21understand how the universe works and I
  116. 03:23thought physics was the right way to do
  117. 03:24that. Something big that influenced me
  118. 03:25was watching the social network cuz I I
  119. 03:27was studying at Harvard and the social
  120. 03:29network took place in Harvard. It was
  121. 03:30actually a very accurate movie and it
  122. 03:32was very inspiring to see this guy
  123. 03:34change the world just on his laptop. I
  124. 03:36realized that you could have like a
  125. 03:37massive influence just coding on your
  126. 03:39laptop in a way that you couldn't as
  127. 03:41much with physics. It was very clear
  128. 03:43that that that the AI could understand
  129. 03:45the problems of physics. So I went into
  130. 03:47computer science and I think that turned
  131. 03:48out to be right because now like the AI
  132. 03:50is getting so good that it should be
  133. 03:52able to just tell us the answer or
  134. 03:54infuse it into our brains. Before EXA on

  135. Google Really Fails

  136. 03:57the side I was writing a history book. I
  137. 03:59got really excited about world history
  138. 04:00and I decided I'm just gonna write a
  139. 04:02book about it because I'm like no one
  140. 04:04has captured in a book my the level of
  141. 04:07excitement that I had and so I was doing
  142. 04:08a lot of research for the book. I
  143. 04:10quickly realized that Google is actually
  144. 04:12not good enough for that type of
  145. 04:13research. Like Google is great for
  146. 04:15surface level investigations. Once you
  147. 04:17start like trying to go deeper and
  148. 04:19trying to understand any topic deeply on
  149. 04:21the web, Google really fails. For
  150. 04:23example, if I want to find like all the
  151. 04:24research papers on uh poverty in ancient
  152. 04:27Rome, it's actually really hard to find
  153. 04:28that on Google. Not every paper will
  154. 04:30mention the word poverty. And so I was
  155. 04:31doing the research for this book and it
  156. 04:33was really hard to find things. And then
  157. 04:35at the same time, GB3 had recently come
  158. 04:37out and GB3 was this magical creature
  159. 04:39really that I could talk to and it could
  160. 04:42understand like anything I say at a very
  161. 04:44deep complex level. And so the thinking
  162. 04:46was what if we could apply the same
  163. 04:49technology of GB3 to search? What if you
  164. 04:51can make a search engine that actually
  165. 04:53understands you at a deep level? And
  166. 04:55it's been the same goal ever

  167. Good Products Make Customers Knock

  168. 05:00since. The first year and a half of EXO,
  169. 05:03we did research into uh search models
  170. 05:06into how can we take transformer models
  171. 05:08and apply them to a search engine. No
  172. 05:10one's really done that before. It took a
  173. 05:12long time to figure out how to do it
  174. 05:14well. And that required persistence.
  175. 05:16like we were just banging our head
  176. 05:17against the wall for a year and a half
  177. 05:18trying out different models, trying out
  178. 05:20different data sets and eventually we
  179. 05:22got something that worked really well.
  180. 05:23If we didn't have the persistence, you
  181. 05:24know, 6 months in, we might have given
  182. 05:25up, but we didn't. Early November 2022,
  183. 05:28we launched the first version of Exa to
  184. 05:30the public. We built a search engine
  185. 05:32that was perfect for AI applications.
  186. 05:33Basically, AI systems, they have all
  187. 05:35this intelligence, but they're lacking
  188. 05:36in knowledge. And so, when they need
  189. 05:38knowledge, they go make a call to Exa
  190. 05:40and get exactly that knowledge. Then
  191. 05:42catch came out a few weeks later and
  192. 05:45that was a big moment uh for the
  193. 05:47information ecosystem. But for us it was
  194. 05:48really interesting because we started
  195. 05:49getting requests for API access to our
  196. 05:52search engine. We started getting
  197. 05:53requests for API access uh first from a
  198. 05:55friend who was actually like living
  199. 05:57downstairs. I told him no sorry we don't
  200. 05:59have API access. And I didn't really
  201. 06:01think much of it. Uh but then we kept
  202. 06:02then we got uh a request for API access
  203. 06:05from someone from some company in
  204. 06:06Germany. And we also told them no sorry
  205. 06:08we don't have an API. And then we kept
  206. 06:10getting requests for API access. And we
  207. 06:12realized that because of chatbt, people

  208. How ChatGPT Changes Exa

  209. 06:15were starting to build AI applications
  210. 06:17all over the place for all sorts of
  211. 06:19businesses. And all these AI
  212. 06:20applications needed search. Like the AIS
  213. 06:22themselves needed to search and that's
  214. 06:24when we started to realize, okay, X
  215. 06:26could be really useful for these AI
  216. 06:28applications. Yeah. So our initial
  217. 06:30customer found us. And so one lesson
  218. 06:32there is just like be a really good
  219. 06:34listener to the market. Like what are
  220. 06:35people repeatedly saying they need? and
  221. 06:37you might have some idea of what you
  222. 06:40know what you're going to sell but then
  223. 06:41if people keep requesting something like
  224. 06:43API access maybe you should start
  225. 06:44listening to them like we cared more
  226. 06:46about uh learnings from the customer
  227. 06:48than getting a lot of revenue and so
  228. 06:50yeah we were like opening our ears to
  229. 06:52what do the customers need uh over time
  230. 06:54developed a hypothesis about how we
  231. 06:57should price and what types of customers
  232. 06:59we should

  233. Preparing for the GPT-5 Era

  234. 07:03pursue. I think you can guess where
  235. 07:06companies like OpenAI and Enthropic are
  236. 07:08going to build based on like what are
  237. 07:11like what are big markets that they
  238. 07:12could tackle like agents like automating
  239. 07:14work is a clear huge market and so
  240. 07:17they're clearly going to do that. Okay.
  241. 07:18So now you know what types of things
  242. 07:20they want to do. Can they do it? Well
  243. 07:22then you think about fundamentals of
  244. 07:24LLMs. You think like okay LLM can as
  245. 07:27long as you can make training data for
  246. 07:28some objective they will get better at
  247. 07:30that objective. Can you make training
  248. 07:32data for agentic behavior? Definitely
  249. 07:34you just have a bunch of examples of
  250. 07:37navigating the web in order to buy plane
  251. 07:39tickets and if you have a million
  252. 07:41examples of that then now the LLM will
  253. 07:43know how to buy plane tickets. Yes. So
  254. 07:45any sort of task such that you could
  255. 07:47create data for it LLM will get near
  256. 07:50perfect at that task. So that's like a
  257. 07:51fundamental way of thinking about this.
  258. 07:53So will AI get really good at navigating
  259. 07:55the
  260. 07:58web? Yes. That's pretty easy because
  261. 08:00it's it's easy to generate lots of
  262. 08:02navigating the web data. Will AI get
  263. 08:04really good at
  264. 08:07robotics? Yes, but probably on a longer
  265. 08:09term horizon because gathering lots of
  266. 08:12robotic training data is harder because
  267. 08:14it's hardware and hardware is hard to
  268. 08:16work with. So what do we learn from
  269. 08:17that? We learned that AI's navigating
  270. 08:18the web will come pretty soon. AI's
  271. 08:21navigating robotics will take longer. So
  272. 08:23just from like simple first principles,
  273. 08:25you can like guess where things are
  274. 08:26going. Picture in your mind how crazy
  275. 08:28good technology will be and AI will be
  276. 08:30in a year and then build for that world.