Transcriptly

Investors Don't Make Your Product Better | AssemblyAI, Dylan Fox

EO14:12Added Aug 31, 2026

Dylan Fox is the founder and CEO of AssemblyAI, where he's built one of the most accurate and developer-friendly speech AI platforms in the industry. Under h...

Watch on YouTube →
Contributed by 刘嘉琪

Transcript

Transcript format

  1. Intro

  2. 00:00As a founder, you just have this
  3. 00:02instinct for some market. That's the
  4. 00:04hard thing about being an entrepreneur,
  5. 00:05being a founder is like most people are
  6. 00:08not going to believe in what you're
  7. 00:09trying to do until all of a sudden they
  8. 00:11do. I loved the idea of the product that
  9. 00:14we were building. As long as I feel
  10. 00:17really happy about the customers that we
  11. 00:19have and they're happy, then that's like
  12. 00:21the validation I'm in search of. You
  13. 00:23have to just like stay the course and
  14. 00:25keep going and keep showing up even if
  15. 00:27you know it's hard and even if you have
  16. 00:29doubts and moments, you just have to
  17. 00:30like keep going.
  18. 00:33My name is Dylan Fox. I'm the founder
  19. 00:35and CEO of Assembly AI. We've built the
  20. 00:38industry's most accurate and easiest to
  21. 00:40use developer platform for speech AI.
  22. 00:43We've raised over $130 million in
  23. 00:45funding to date from Excel Insight
  24. 00:48Partners Daniel Gross and Nat Freeman
  25. 00:50and Smith Point. That's led by Keith
  26. 00:53Block from Salesforce.

  27. One Bowl of Pasta, Seven Days of Code

  28. 01:04My older brother, he would order all the
  29. 01:06hardware online and he would, you know,
  30. 01:08be in our basement making computers and
  31. 01:10I would watch him and being around
  32. 01:12computers as a kid playing video games.
  33. 01:15I was addicted to these MMRPG games. I
  34. 01:19love just working on technology. So in
  35. 01:22college, I started this company that
  36. 01:26helped other college organizations fund
  37. 01:28raise online and people that donated to
  38. 01:31this student organization would get like
  39. 01:34local rewards in their community. It was
  40. 01:36a terrible idea and it was, you know,
  41. 01:38didn't go anywhere. But we learned a lot
  42. 01:41about starting company and being a
  43. 01:43founder through that experience. What I
  44. 01:46learned from that experience was that I
  45. 01:48loved being a founder and I loved
  46. 01:50programming. And with programming, it
  47. 01:52wasn't so much the programming that was
  48. 01:55addictive. It was just this open-ended
  49. 01:58world that you could just build stuff.
  50. 02:00Any idea you had with programming, you
  51. 02:02could make it into something real and
  52. 02:04then you could get feedback from users
  53. 02:06and you could then keep building and
  54. 02:09keep building. And so it was this
  55. 02:10process of like building something that
  56. 02:12I found just so addictive. So after
  57. 02:15college, I didn't get a job. Uh we had
  58. 02:18shut the startup down. I just opened up
  59. 02:21a bunch of credit cards and I went like
  60. 02:24$30,000 into credit card debt. Basically
  61. 02:26just like learning how to program and
  62. 02:28like reading programming books and
  63. 02:31building apps and just like all day.
  64. 02:33That's what I would do in my apartment.
  65. 02:35And I would make like one big bowl of
  66. 02:38pasta every Sunday and just like eat
  67. 02:40that all week with Diet Coke. As I was
  68. 02:43just like spending all day just
  69. 02:44programming and building stuff and
  70. 02:46seeing if I could launch anything new.
  71. 02:48After, you know, almost 2 years of doing
  72. 02:50that and going into credit card debt, I
  73. 02:52was like, "Okay, I need to go get a
  74. 02:53job." But I had found that I was really
  75. 02:56most interested in machine learning and
  76. 02:58natural language processing. And so
  77. 03:00there was a team in San Francisco, the
  78. 03:02company called Cisco that was hiring for
  79. 03:05machine learning engineers to focus on
  80. 03:07building natural language processing
  81. 03:10products and interfaces around Cisco's
  82. 03:12collaboration products. Got really into
  83. 03:15neural networks and more advanced
  84. 03:17machine learning and deep learning while
  85. 03:18I was out there at that job. And that
  86. 03:20was like 2015 2016 time

  87. Why I Became Interested in Voice AI

  88. 03:24when I when I got the job at Cisco and I
  89. 03:26knew I always wanted to start another
  90. 03:28company. I'd always spend nights and
  91. 03:30weekends just still tinkering with
  92. 03:32random ideas I would have. And I think
  93. 03:34the Alexa launched around the time I was
  94. 03:37in Cisco. So the Alexa product was like
  95. 03:40the first computer you could talk to.
  96. 03:42And I as a machine learning engineer
  97. 03:44that was working in natural language
  98. 03:45processing. I wanted to start
  99. 03:47experimenting with my own ideas for
  100. 03:50voice interfaces or voice products,
  101. 03:52voice driven products. And the leading
  102. 03:55company at the time that was building
  103. 03:57that technology was this big company.
  104. 03:59And I contacted them to try to get
  105. 04:01access to their developer SDK. And they
  106. 04:03mailed me a CDROM. It was like a $10,000
  107. 04:07evaluation agreement that you had to
  108. 04:09sign. I didn't even have a CDROM drive
  109. 04:12to like load their SDKs. It was just
  110. 04:14this complete archaic experience. As a
  111. 04:17developer, I really wanted this super
  112. 04:20accurate, really easy to use developer
  113. 04:23platform because I saw back then that
  114. 04:26the technology was going to just get
  115. 04:28orders of magnitude better and that was
  116. 04:30going to make what was a small market
  117. 04:32huge. That was a really exciting thing
  118. 04:35that I just wanted to build and work on.
  119. 04:38as a founder, you just have this
  120. 04:40instinct for some market and that's what
  121. 04:43gets you excited about working in that
  122. 04:45market or building in that market. And
  123. 04:47so for me, that instinct was voice
  124. 04:50interfaces and speech AI technology.
  125. 04:53That is one of the most important
  126. 04:54modalities for AI. And so we're really
  127. 04:57excited about that potential and
  128. 05:00especially that potential to create
  129. 05:02really accurate and amazing AI for
  130. 05:04speech and voice and then just put it
  131. 05:06into the hands of developers to build
  132. 05:08really creative stuff with and really
  133. 05:10amazing apps with.

  134. No Product, No Customer, Still Got into YC

  135. 05:17I left my job and then a few months
  136. 05:20later got into Y Combinator. I had no
  137. 05:23clue I was going to get into Y
  138. 05:24Combinator. I just wanted to submit the
  139. 05:26application as really a thought exercise
  140. 05:29to like crystallize what I was working
  141. 05:31on, what I was going to do. You know, I
  142. 05:33figured like no chance I'm getting in
  143. 05:34alone. No progress, no traction, no
  144. 05:36product, just an idea. But there was a
  145. 05:38YC partner at the time, Daniel Gross,
  146. 05:41who had worked at Apple, had worked
  147. 05:44around Siri, got an email, it's like,
  148. 05:46"Hey, what's your accuracy rate from
  149. 05:49Daniel?" And then the next day they were
  150. 05:51like, "Hey, we'd love you to come in for
  151. 05:52an interview." And so I bought a ticket
  152. 05:54home. I flew back to San Francisco where
  153. 05:57I was living at the time. The next day,
  154. 05:59drove down to to Y Combinator for the
  155. 06:02interview, got in and I'd submitted the
  156. 06:05application. It was like 30 days late,
  157. 06:07past the deadline. I went in the first
  158. 06:09day and it was like all these other
  159. 06:10companies had so much progress, so much
  160. 06:12traction, and I was just getting started
  161. 06:15and it was a really hard idea to get
  162. 06:16started with like creating AI models for
  163. 06:19speech. That was probably one of the
  164. 06:20most stressful periods of my life. It
  165. 06:22was like those those three months in YC
  166. 06:25where I was just like really trying to
  167. 06:27get things off the ground working mostly
  168. 06:29by myself. A lot of founders think that
  169. 06:32getting into Y Combinator raising
  170. 06:34capital is like just going to make
  171. 06:35things happen. It doesn't. No investors
  172. 06:38are going to hand you customers, make
  173. 06:40your product better, fix things. You
  174. 06:43still have to make everything happen. So
  175. 06:45we just again like showed up every day
  176. 06:48and just tried to make progress and just
  177. 06:50like kept at it. I've really just tried
  178. 06:52to focus on is like as long as I feel
  179. 06:55really happy about the product that
  180. 06:57we're making and the customers that we
  181. 07:00have and they're happy, then that's like
  182. 07:02the validation I'm in search of is do I
  183. 07:05feel happy about our product? Are
  184. 07:07customers happy? Like those are the
  185. 07:09things I try to get validation from, not
  186. 07:11what do other people think about our
  187. 07:14company or what we're doing.

  188. EO Partner Highlight

  189. 07:18As a founder, you already know ideas are
  190. 07:20the easy part. It's the execution,
  191. 07:23actually building the product, that
  192. 07:24slows everything down. That's where
  193. 07:26Lovable comes in. It's not just an AI
  194. 07:29tool. It's your ondemand engineering
  195. 07:31team. Simply describe your idea. Lovable
  196. 07:35then builds a full frontend, backend,
  197. 07:36and database so you can launch real
  198. 07:38production ready software without
  199. 07:40writing code. It's already powering over
  200. 07:42a 100,000 new products a day, helping
  201. 07:442.5 million builders turn ideas into
  202. 07:47software just by describing what they
  203. 07:49want. No devs, no delays, no excuses.
  204. 07:54They're launching in weeks, not months.
  205. 07:56And guess what? These teams are still
  206. 07:57tiny. In fact, team EO is also using
  207. 08:00Lovable to build our upcoming EOS school
  208. 08:02platform, and we're loving it. If you're
  209. 08:04a non-technical founder or just want to
  210. 08:06build without bottlenecks, try Lovable
  211. 08:09today for free. Use the promo code EO2YT
  212. 08:13to get 20% off your first purchase of
  213. 08:15the Lovable Pro plan.

  214. How a Startup Wins with Deep Subject Matter Expertise

  215. 08:23Most of my day is spent talking to
  216. 08:25customers, working with our product
  217. 08:27teams, playing with our product, and
  218. 08:29seeing where it's working, where it's
  219. 08:31not. founders and and startups, you
  220. 08:33really need to have this deep deep
  221. 08:36subject matter expertise in your market
  222. 08:38and in your customers in your product. I
  223. 08:41think in a lot of ways it's like
  224. 08:42undervalued and underappreciated versus
  225. 08:45functional expertise. Don't always have
  226. 08:47a lot of confidence in your functional
  227. 08:48expertise. And I think that's where
  228. 08:50startups can win. You know, there's a
  229. 08:52lot of people out there that have
  230. 08:53amazing functional expertise, but don't
  231. 08:55have the deep subject matter expertise
  232. 08:58that you have or that your team has over
  233. 09:01a certain market or a certain customer
  234. 09:03or product. And so, when I talk to
  235. 09:05customers, I don't want the flattery. I
  236. 09:07want like where does our product suck?
  237. 09:09Some of the questions I ask are like,
  238. 09:10hey, what are the top three things that
  239. 09:12you don't like about our product? Or if
  240. 09:15you were in charge of our road map, what
  241. 09:16would you prioritize? Those types of
  242. 09:18questions are really helpful because,
  243. 09:20you know, they give you that feedback
  244. 09:22and that insight over like where do you
  245. 09:24need to continue to push. I'm never
  246. 09:26satisfied with our progress. I really
  247. 09:29excited about what our customers are
  248. 09:30building. Like I see some of the apps
  249. 09:32they launch and they build and it's like
  250. 09:34like I want to go talk about them with
  251. 09:36my friends cuz they're so cool and
  252. 09:37inspiring and exciting. And as all these
  253. 09:39new applications around speech are
  254. 09:41continuing to take off, the amount of
  255. 09:43speech data we're handling it just
  256. 09:44continues to grow rapidly. We'll handle
  257. 09:48this month alone about five pabytes of
  258. 09:51speech data through our API platform.
  259. 09:54That's about 10x the size of the entire
  260. 09:56Spotify library and catalog. And usage
  261. 09:59to our developer platform is growing
  262. 10:01over 250%
  263. 10:03year-over-year. So, the scale is like
  264. 10:05pretty insane.

  265. Just Start with a Website

  266. 10:07Speed is probably more important now
  267. 10:08than ever. There's so many use cases
  268. 10:10that we're seeing developers want to
  269. 10:13build apps around that they need really
  270. 10:15good speech AI for. They need new
  271. 10:18capabilities. They need better tech.
  272. 10:20They need better models. They need, you
  273. 10:22know, all this stuff. They're hungry for
  274. 10:24it because the opportunities are
  275. 10:27enormous. An example of where a lot of
  276. 10:29speech AI models will struggle today is
  277. 10:31with hallucinations. So having an
  278. 10:33in-person meeting with 10 people or
  279. 10:36you're having a phone call and it's
  280. 10:38windy and the quality is bad, that's
  281. 10:41where the AI models still do struggle
  282. 10:43today. And that's an amazing opportunity
  283. 10:46because there's so many applications
  284. 10:47that are limited by those things. And so
  285. 10:50we're really excited to keep making the
  286. 10:53models that we're creating better and
  287. 10:54better. And as a startup company, you
  288. 10:56want to try to optimize everything you
  289. 10:59can for speed. You can start by just
  290. 11:01putting up a website and advertising the
  291. 11:04product that you want to build and just
  292. 11:06putting like a, you know, contact us
  293. 11:08button on there and see what are people
  294. 11:11reaching out about, are people reaching
  295. 11:12out, what do they want from your product
  296. 11:14and that can be really helpful to get
  297. 11:16validation early on are you building the
  298. 11:19right thing uh before you go spend a ton
  299. 11:21of time building. So I think that really
  300. 11:22fast iteration loop is important for
  301. 11:25startups to have with customers and the
  302. 11:28markets that you're you're working in.

  303. Focus on Ours, Not Theirs

  304. 11:34When you're making an AI model like at
  305. 11:36so many points you have to decide
  306. 11:37between trade-offs, right? What type of
  307. 11:39data do you use? What type of thing do
  308. 11:41you optimize for? When you know who
  309. 11:43you're building this AI model for, then
  310. 11:45you can make all those trade-offs a lot
  311. 11:47more intelligently in a way that makes
  312. 11:49your AI model have more product market
  313. 11:51fit for who you're building it for. You
  314. 11:53can always change your focus as you
  315. 11:55learn, right? But I think it's important
  316. 11:56to like have a focus and then you learn
  317. 11:59and then you can focus other areas. like
  318. 12:01we're super laser focused on, okay,
  319. 12:04people building voice agents, people
  320. 12:06building notetakers, people building
  321. 12:08sales intelligence apps, these are the
  322. 12:1010 things they really, really care
  323. 12:12about. And so, let's make sure our
  324. 12:13models are hyper optimized for those
  325. 12:15things. And let's build the right
  326. 12:17training data and let's build the right
  327. 12:19model architectures and let's do all of
  328. 12:21the things we can to absolutely max out
  329. 12:23in those dimensions. And that's how
  330. 12:25we're constantly keeping our models the
  331. 12:28most accurate and the easiest to use for
  332. 12:31the developers that we're building for.
  333. 12:33You might come to assembly and maybe our
  334. 12:35models aren't as good for you as another
  335. 12:39model. And usually if that happens, it's
  336. 12:41because that's not an application or a
  337. 12:43use case that we're really focusing on.
  338. 12:45At least not right now. So for us, the
  339. 12:47way we really maintain our competitive
  340. 12:49advantage is by just really clearly
  341. 12:51focusing on who we're building for and
  342. 12:53not building general purpose tech, but
  343. 12:55building tech that's optimized for a
  344. 12:58specific use case and market.
  345. 13:00One of the biggest things I've learned
  346. 13:02is that every startup, you have your own
  347. 13:04journey. It's easy to compare yourself
  348. 13:06to other startups. you know, you have
  349. 13:08friends that are founders and maybe
  350. 13:10they're a stage or two ahead. But
  351. 13:12startups are not franchise businesses, I
  352. 13:14think, is one of the biggest things that
  353. 13:16I've learned. And what I mean by that is
  354. 13:19like you have to really figure out what
  355. 13:21your journey is. And every journey is
  356. 13:24slightly different. And there's a lot of
  357. 13:25startup dogma that you don't have to
  358. 13:28subscribe to. and in a lot of ways can
  359. 13:31actually make it harder for you as a
  360. 13:33founder because you feel like there's
  361. 13:35all this stuff you have to do when in
  362. 13:36reality you just have to really build a
  363. 13:38great product, make your customers happy
  364. 13:41and that's what you want to focus on.
  365. 13:42That realization has helped me a lot as
  366. 13:45a founder realizing, okay, we're on our
  367. 13:47our own journey. Startups are all look
  368. 13:50different. All their journeys are
  369. 13:51different. And just be really focused on
  370. 13:53making a great product, on making
  371. 13:55customers super happy. And that's that's
  372. 13:57the north star.