Transcriptly

How Prompt Engineering Inventor Built $1.5B in 3 Years | You.com, Richard Socher

EO12:04Added Aug 31, 2026

We met Richard Socher, the founder of You.com. Richard spent 17 years proving that AI could understand human language back when the world dismissed it as a "crazy idea." After serving as the Chi

Watch on YouTube →
Contributed by 刘嘉琪

Transcript

Transcript format

  1. Intro

  2. 00:00had the crazy idea in 2010 to use neural
  3. 00:02networks for natural language processing
  4. 00:04which was a very controversial idea at
  5. 00:06the time. A lot of people in the field
  6. 00:07said, "Oh, neural networks, they never
  7. 00:09work. They won't ever work." Most of my
  8. 00:11papers got rejected cuz people hated
  9. 00:13neural nets for NLP. MIT and especially
  10. 00:16Berkeley here in the Bay Area, they
  11. 00:17hated neural networks for natural
  12. 00:19language processing too. We had invented
  13. 00:21prompt engineering and so we thought
  14. 00:22people should get access to this. We
  15. 00:24felt like well someone's got to do it.
  16. 00:26In 2020 when we started.com a lot of
  17. 00:28people said search [music] is dead.
  18. 00:29nothing you can do. But I don't really
  19. 00:31generally care about what's popular. I
  20. 00:33just care about what's meaningful. When
  21. 00:34[music] you once you really love an idea
  22. 00:36and you feel like that idea makes sense
  23. 00:38from first principles, you have to have
  24. 00:40a little bit of that belief inside of
  25. 00:42you that [music] you can make it prevail
  26. 00:44through a lot of rejection and still
  27. 00:45keep on going. Hi everyone, I'm Richard
  28. 00:48Ser. I'm the founder and CEO of.com and
  29. 00:51the founder and partner at AI Accentur.
  30. 00:54U.com is AI search infrastructure. In
  31. 00:57order to make an LM not hallucinate, you
  32. 01:00actually need to have a good search
  33. 01:02infrastructure to inform that LM. People
  34. 01:05searched on Google AI and LMS and agents
  35. 01:08search on you.com. Companies like
  36. 01:10OpenAI, Amazon, Alibaba, Telegraph,
  37. 01:14Windsurf, Harvey use our API
  38. 01:18infrastructure to make their LMS
  39. 01:20upto-date, accurate, and have citations.
  40. 01:22So you can actually verify that the
  41. 01:25facts are correct. It's amazing to
  42. 01:26become a unicorn. And in many ways, it
  43. 01:29also just feels like okay, it's what's
  44. 01:31next? What's the next step?

  45. How the Inventor of Prompt Engineering Built a $1.5B Unicorn

  46. 01:43I originally from Germany. I thought a
  47. 01:45lot about sort of the meaning of life
  48. 01:47when I was younger. In high school
  49. 01:49already, I love natural languages. You
  50. 01:52know, I studied English. I'm originally
  51. 01:53from Germany and [music] French and
  52. 01:55Chinese and but I also love math. And so
  53. 01:58math and languages don't intersect
  54. 02:01often, right? It's very different fields
  55. 02:02of study, but they do intersect [music]
  56. 02:04in a computer where you try to use math
  57. 02:07to understand language. And so I ended
  58. 02:10up deciding study linguistic computer
  59. 02:12science in 2003. [music] It was
  60. 02:14definitely not famous or or popular in
  61. 02:17Germany. Linguistic computer science as
  62. 02:19it was called then was very much a niche
  63. 02:22orchid type of subject that very few
  64. 02:24people were studying. Still remember my
  65. 02:26my dad thinkingh what what will become
  66. 02:28of my son cuz this linguistic computer
  67. 02:30science thing doesn't sound like
  68. 02:31anything useful for ever like uh for a
  69. 02:34long time and but I don't really
  70. 02:35generally care about what's popular. I
  71. 02:37just care about what's meaningful. And

  72. Why Linguistics is the Operating System of Intelligence

  73. 02:39then it felt to me like if we could
  74. 02:41really get that to work uh on the
  75. 02:43research side, it would have an amazing
  76. 02:45impact. Ultimately, language is the most
  77. 02:48interesting manifestation of human
  78. 02:49intelligence. Several civilizations were
  79. 02:52able [music] to create written language
  80. 02:53to not all of them, right? And the ones
  81. 02:55that didn't were falling behind and
  82. 02:57we're saying similar things.
  83. 02:58Civilizations that don't use AI now are
  84. 03:00falling behind. And so ultimately I
  85. 03:02think it's very meaningful to help
  86. 03:04[music] us understand language because
  87. 03:06it helps us understand who we are as
  88. 03:08humans. And then it uh also uh [music]
  89. 03:10actually helps in on the journey to
  90. 03:13understanding intelligence. It helps to
  91. 03:15create it because anything we can create
  92. 03:17we can engineer we understand a lot
  93. 03:19better afterward.
  94. 03:20Your brain and mine are jam-packed full
  95. 03:22of neurons that are tightly connected to
  96. 03:24and they talk to each other. in a
  97. 03:26computer. We can therefore build what's
  98. 03:28called an artificial neuronet network or
  99. 03:30the the other technical term is a sponse
  100. 03:32learning algorithm. But we can build a
  101. 03:34neuronet network that simulates all of
  102. 03:36these neurons being connected to and
  103. 03:38talking to each other. And then I was
  104. 03:40very fortunate uh at Stanford to hear
  105. 03:43Andrew talk about deep learning and
  106. 03:45neural nets for computer vision. It made
  107. 03:48sense to me from first principles that
  108. 03:50neural networks would be right because a
  109. 03:52lot of the research actually was about
  110. 03:54feature engineering that were people
  111. 03:56were doing at the time like in sentiment
  112. 03:57analysis you might say oh these are
  113. 03:59positive words and this is how negation
  114. 04:01works and here's like all these like
  115. 04:03linguists would come up with features
  116. 04:04and that clearly wouldn't scale to more
  117. 04:07complex things like translation [music]
  118. 04:09and I wanted to unify also uh the field
  119. 04:12and that eventually led us to prompt
  120. 04:13engineering and inventing that but then

  121. The Moment We Had to Scale Up

  122. 04:15after [music] the PhD I was like now we
  123. 04:17have the main ingredients. We know how
  124. 04:18to make it work. We need large neural
  125. 04:20networks. We need a lot of data. And I
  126. 04:22showed that in all my research papers.
  127. 04:24And now we [music] need to actually
  128. 04:25scale it up. We need to like take those
  129. 04:27ideas and apply them into real
  130. 04:28applications for real people. And I
  131. 04:31think a lot of the papers that came out
  132. 04:32in the last couple of years, they made
  133. 04:34everything a little bit better. But
  134. 04:35those main ideas of endtoend trainable
  135. 04:38neural networks on large data sets, that
  136. 04:41is the main idea. That's [music]
  137. 04:42those those ingredients are the main
  138. 04:44ideas that push the field forward. And I
  139. 04:47felt like it made more [music] sense now
  140. 04:48to majorly scale that. And then in
  141. 04:51academia, you just don't have the
  142. 04:52resources to really scale. And then
  143. 04:54[music] while I was excited about
  144. 04:55scaling it, I should have scaled it even
  145. 04:57more. You know, I thought, oh, I raised
  146. 04:58like 1020 million, but I should have
  147. 05:01raised $200 million or a billion [music]
  148. 05:02dollars to really scale it more. So it
  149. 05:05was clear to me that to really scale it,
  150. 05:07you had to do it in industry. And that
  151. 05:08the technology was ready now to move out
  152. 05:11of academic research into the real
  153. 05:13world. And then when you maximize impact
  154. 05:15and you realize well the main ideas have
  155. 05:17we've been researched now it's time to
  156. 05:18really apply them in the real world and
  157. 05:20so startup uh felt like it made sense.

  158. Solving for Impact When Everyone Said No

  159. 05:29So I graduated in 2014 then started
  160. 05:31[music] Metammind. Metammind uh was
  161. 05:33basically an AI platform that made it
  162. 05:35very easy to train neural networks. we
  163. 05:37had [music] uh started selling it to but
  164. 05:39for selling you had to be very very
  165. 05:41focused on one small niche but we had
  166. 05:43built this very powerful platform and so
  167. 05:45felt like in the hands of a no pun
  168. 05:47intended sales [music] force that is
  169. 05:48very large we could actually have much
  170. 05:50much more impact and the impact within
  171. 05:52Salesforce was much much larger for the
  172. 05:54way we had built that company and then I
  173. 05:57thought for a long time like I'll just
  174. 05:58be very happy cuz [music] uh Salesforce
  175. 06:00and do amazing research and improve a
  176. 06:02lot of the products we did not [music]
  177. 06:04just prompt engineering but we also
  178. 06:06built the are just language model for
  179. 06:07proteins for instance and biology. And
  180. 06:09then we had invented prompt engineering.
  181. 06:11And so [music] we trained this one
  182. 06:13neural network that can give you all
  183. 06:15different kinds of answers. And so we
  184. 06:17thought clearly people should get access
  185. 06:19to [music] this. and we published a
  186. 06:21paper and you know the paper got cited
  187. 06:22by other people at OpenAI and Alec
  188. 06:24Ratford and Ilia and they said oh this
  189. 06:26is an interesting idea and they extended
  190. 06:28it and and so on but we felt like well
  191. 06:30they're also a research lab so we needed
  192. 06:32to bring this to real people and Google
  193. 06:34was just not doing anything cuz they're
  194. 06:35a monopoly they're making money and
  195. 06:37[music] more and more money just selling
  196. 06:39advertisement and so they didn't see a
  197. 06:42need reason and weren't making any
  198. 06:44interesting [music] sort of
  199. 06:45modifications to fundamentally how we
  200. 06:47search and so we felt like well
  201. 06:49someone's has got to do it. And so we we

  202. The Beginning of You.com: Challenging Google

  203. 06:51started [music] you.com and we felt like
  204. 06:53it had to be a new company to have the
  205. 06:55impact to really become a better way of
  206. 06:58finding information online. And
  207. 07:00eventually we became the first to put an
  208. 07:02LM into a search engine. Imagine you
  209. 07:05know Google Gemini like people also ask
  210. 07:07and you get like answers from AI. We did
  211. 07:11those kinds of things but in 2021. So it
  212. 07:13is different because it was no no one
  213. 07:16that didn't exist you know a research
  214. 07:18background where you the whole idea of
  215. 07:20being a PhD [music]
  216. 07:22is to do things that don't exist right
  217. 07:24to create new ideas and and new models
  218. 07:26uh and and then you can often think
  219. 07:29about what kinds of new ideas and models
  220. 07:31should you build the kinds of [music]
  221. 07:32things that are useful for people and
  222. 07:34when you ask like oh how do I write a
  223. 07:37Fibonacci function or how do I write an
  224. 07:39HTML page that does something it's just
  225. 07:41obvious that it's better to just get an
  226. 07:43answer from an LM than to guess get a
  227. 07:45list of blue links where you then have
  228. 07:46to click on 10 open tabs and open them
  229. 07:49up and then kind of uh find the answer
  230. 07:52somewhere else. Uh and so that just
  231. 07:55seems like from first principles it's
  232. 07:56better to get an answer than lists of
  233. 07:58links that may have the answer or not.
  234. 08:00So that's how we invented that one.

  235. Build What People Will Actually Pay For

  236. 08:08I think the biggest thing for us was the
  237. 08:10pivot into enterprise. That was a really
  238. 08:12good focus. A lot of folks now realize
  239. 08:15they need AI, but only the experts
  240. 08:18realize that in order to make AI
  241. 08:20accurate, in order to make an LM not
  242. 08:22hallucinate, you actually need to
  243. 08:25[music] have a good search
  244. 08:26infrastructure to inform that LM. So, we
  245. 08:29built that infrastructure layer cuz
  246. 08:31we've been at [music] it since 2022.
  247. 08:33What we found is more and more companies
  248. 08:35actually want to use the underlying
  249. 08:39infrastructure for their own solutions
  250. 08:41inside [music] their own products. I
  251. 08:42guess you know there's sort of different
  252. 08:44pivots in the world right you can say oh
  253. 08:46we're make cameras and now we sell soft
  254. 08:49[music] drinks right that's a big pivot
  255. 08:50but we gave people answers and now we
  256. 08:53[music] give people answers but how
  257. 08:55we're selling those answers is different
  258. 08:57and it's good to follow the revenue here
  259. 09:00are a bunch of people who want to use
  260. 09:02the product for free and then here a
  261. 09:03bunch of people who want to use and get
  262. 09:05really good answers over their own
  263. 09:07custom data sets and they're willing to
  264. 09:09pay you follow the people that pay
  265. 09:11follow real revenue, not like, okay,
  266. 09:14hype. Some people say, oh, I want to use
  267. 09:16this product for free, and you're like,
  268. 09:17okay, that's great. But if you build
  269. 09:19something that uh companies are willing
  270. 09:21to pay for, you know, you've built
  271. 09:23something of value.

  272. Are You Moving Fast Enough to Lead the AI Era?

  273. 09:29Some people think we should like slow
  274. 09:31down. I think we should accelerate more.
  275. 09:33I think we should accelerate everything
  276. 09:35a lot more. It's kind of interesting.
  277. 09:36It's hard to navigate AI because on the
  278. 09:39one hand there's real impact, right? Our
  279. 09:41customers have built over 100,000 agents
  280. 09:44that are automating real tasks for their
  281. 09:46work, right? And [music] they're telling
  282. 09:47us and they're paying us for it because
  283. 09:49it's valuable and it's it works. At the
  284. 09:51same time, there's a lot of hype and
  285. 09:54around AI like how quickly and how far
  286. 09:57are we on super intelligence? Uh are we
  287. 09:59on the right track for that? How much
  288. 10:02could uh a browser automate complete
  289. 10:04tasks without knowing enough about me
  290. 10:06and things like [music] that? And
  291. 10:08sometimes the timelines are a little bit
  292. 10:09off. You know, maybe it will take a
  293. 10:11little bit longer, but the field moves
  294. 10:13so quickly, you have to mostly think of
  295. 10:15like 2 to 4 week cycles to try to move
  296. 10:19quickly. Okay. And so it's important to
  297. 10:21think about what are the right
  298. 10:23applications where you can create
  299. 10:24ideally virtuous data cycles where you
  300. 10:27do something manually. You collect data
  301. 10:30and then you make that decision process
  302. 10:32a little bit better and then at some
  303. 10:34point you've made it good enough that
  304. 10:36you can automate it. And so I for
  305. 10:38instance didn't [music] want to invest
  306. 10:39in a bunch of self-driving car companies
  307. 10:41that said we don't even need a steering
  308. 10:42wheel. [music] We just need to like have
  309. 10:45a full self-driving car. And I was like,
  310. 10:47"Oh man, you can't sell that car until
  311. 10:49you're perfect." And so that's not
  312. 10:50generally good. AI [music] is not right
  313. 10:52away perfect. Humans aren't perfect,
  314. 10:54right? Humans still make driving
  315. 10:55mistakes and so on. AI will make some
  316. 10:57driving mistakes, too. And so it's good
  317. 10:59to have a steering [music] wheel. And so
  318. 11:01the companies that were able to
  319. 11:02eventually get to full self-driving were
  320. 11:04either the really clever ones like
  321. 11:06Tesla. You buy the car, you pay for the
  322. 11:08product, you use the product, you're now
  323. 11:11creating training data by using the
  324. 11:13product. And then the eye can use that
  325. 11:15training data and eventually [music]
  326. 11:16automate uh the process.
  327. 11:20When you [music] see small but
  328. 11:23continuous improvements, that's when you
  329. 11:25you can, you know, be very motivated
  330. 11:27too. So that's one of my mottos is
  331. 11:29better, better, never done, right? You
  332. 11:31can always improve uh yourself, your
  333. 11:33company, your processes. Overall, I
  334. 11:36would summarize it as excitement. I love
  335. 11:38AI. I love AI in all of its facets from
  336. 11:40foundational research and thinking about
  337. 11:42the upper bounds of super intelligence
  338. 11:45uh all the way down to like how do we
  339. 11:46make it really work now and get it into
  340. 11:48the hands of more companies and people
  341. 11:50to like to make their lives better.