YouTube2Text

OpenAI Co-Founder: Start Building With AI Before You Feel Ready | Greg Brockman — Transcript

by Silicon Valley Girl · 7,129 words · 1,010 segments · language en · Watch on YouTube

Full transcript

  1. 0:00We're going to see this massive
  2. 0:01renaissance of entrepreneurship. Like I
  3. 0:02think it is coming. I think it's really
  4. 0:04going to kick in over the next year to
  5. 0:05two years.
  6. 0:06>> This is Greg, co-founder of Open AI, the
  7. 0:09company behind Chat GPT, which more than
  8. 0:11a billion users use every week.
  9. 0:14>> I'm pretty good at writing software. I
  10. 0:15don't write software anymore. I direct
  11. 0:17an AI for it to write software. I
  12. 0:19provide lots of [music] feedback and
  13. 0:20guidance.
  14. 0:20>> When people are building software, how
  15. 0:22do you strategize around building
  16. 0:24something that's not going to be made
  17. 0:25obsolete with the new models?
  18. 0:28I'm just about to show you something
  19. 0:29that I've built recently. Um, am I
  20. 0:32ambitious enough with AI?
  21. 0:33>> Well, I think the answer to being
  22. 0:35ambitious enough will always be no
  23. 0:37because there's always more we can do
  24. 0:39>> for someone who is willing to step up
  25. 0:41their agent game and they have 1 hour.
  26. 0:43Where do they start this weekend?
  27. 0:45>> I'd say first of all,
  28. 0:48>> this video is sponsored by HubSpot.
  29. 0:51>> I want to talk all things practical. I
  30. 0:53was recently talking to Nobel Prize
  31. 0:55laurate in economics and the one theme
  32. 0:58kept coming up that we have such capable
  33. 1:00AI models but we lack applications. What
  34. 1:04do you think that you're launching this
  35. 1:06year will help change that and help
  36. 1:07people realize what's the power of AI?
  37. 1:10>> Well, it's been very interesting to see
  38. 1:12first of all how many people use chat
  39. 1:15GBT. You know, we have 1.2 billion
  40. 1:17weekly active users. We have like 300
  41. 1:19million people who use chat every single
  42. 1:21week for health queries, right? to help
  43. 1:22them in their personal lives. At the
  44. 1:24same time, how many of those use cases
  45. 1:26just scratch the surface? And we see
  46. 1:28this in the metrics that if people use
  47. 1:31chat for three different use cases, so
  48. 1:33three different things that you realize
  49. 1:34chat could do for you, you become a
  50. 1:35power user. You become you retain you
  51. 1:37you keep using chat and you want to use
  52. 1:38it more. But getting to those three use
  53. 1:40cases, it's such a hard challenge.
  54. 1:43>> And in some ways, it kind of makes sense
  55. 1:45because okay, so here's this AI that can
  56. 1:47do anything, but what's one thing it can
  57. 1:48do, right? It's like sometimes
  58. 1:50>> what's the magic thing?
  59. 1:51>> Exactly. Right. And so really finding
  60. 1:53that magic that speaks to you, but also
  61. 1:54at the same time, I think it shows this
  62. 1:56massive opportunity, this massive space
  63. 1:58that's almost a gap where you have
  64. 2:00you're talking to an AI and the AI knows
  65. 2:02what it's capable of and it can suggest
  66. 2:03to you, hey, you asked me to do this
  67. 2:05thing for you. Actually, I could do even
  68. 2:06more, right? Here's an idea. I can make
  69. 2:08a whole presentation a spreadsheet. So
  70. 2:10that's actually been something we've
  71. 2:11been exploring. But I feel like there's
  72. 2:12something here about the fact that what
  73. 2:14you want out of AI is simplicity, right?
  74. 2:17You don't want an interface that has
  75. 2:18buttons and sliders and model pickers.
  76. 2:20Like none of that. That's not the AI we
  77. 2:21were promised.
  78. 2:22>> You want something you can talk to that
  79. 2:24can solve problems, that can help you,
  80. 2:25can help solve goals, but that can
  81. 2:27actually even be proactive, that can say
  82. 2:28to you, "Hey, you got this email that's
  83. 2:30in your inbox that you've been ignoring
  84. 2:32this person for the past couple days and
  85. 2:34that actually I did some research and it
  86. 2:36turns out that here's like the right
  87. 2:37answer to the question and I've drafted
  88. 2:39an answer. Do you approve?"
  89. 2:41>> And by the way, the use case I just
  90. 2:43mentioned, it's not theoretical with
  91. 2:44dots. I actually got that exactly uh
  92. 2:47this this this use case happening to me
  93. 2:49right before we we started this podcast.
  94. 2:51And so I think we're at a point now
  95. 2:53where AI is actually able to be much
  96. 2:57more proactive and to remove so much of
  97. 2:59the burden of figuring out like what is
  98. 3:01this system capable of and to really
  99. 3:04help you and then help you be able to
  100. 3:06achieve whatever it is that you want to
  101. 3:07do.
  102. 3:08>> Are there any other use cases that were
  103. 3:09unimaginable a year ago that are
  104. 3:11possible now and that people should be
  105. 3:13trying out? Well, I think so many,
  106. 3:15right? There's a whole spectrum of them
  107. 3:16from and I think the way that I always
  108. 3:18recommend going to to discovery is you
  109. 3:20got to play with the system, right? You
  110. 3:21got to see what it's capable of. And so,
  111. 3:23when it comes to coding, like these
  112. 3:24models, they're now better at writing
  113. 3:26software than basically anyone I know,
  114. 3:29>> right? I'm pretty good at writing
  115. 3:30software. Like, I don't write software
  116. 3:32anymore. I direct an AI for it to write
  117. 3:34software. I provide lots of feedback and
  118. 3:36guidance. So, it's always fun to like
  119. 3:37make a little game and just kind of get
  120. 3:39this this quick hit on just seeing like,
  121. 3:40okay, how long did it take to deliver on
  122. 3:42that? and then ask for something more
  123. 3:44complex. I I think that we're at a point
  124. 3:46now where Astra is so good at Blender,
  125. 3:48so good at 3D modeling. And so it's kind
  126. 3:51of fun to ask it to produce 3D models of
  127. 3:54a photo and those kinds of things. You
  128. 3:56can see it using software like using
  129. 3:58Microsoft Paint to to sketch out a
  130. 4:00photo. But you can ask for more real
  131. 4:02things, too. Like I know for example a
  132. 4:04friend who was designing or wanted to
  133. 4:07have a like guest house built and used
  134. 4:10Astra to actually produce CAD plans and
  135. 4:12then brought those to a contractor and
  136. 4:14now it's being built.
  137. 4:15>> And so it's a very empowering thing.
  138. 4:16Normally you'd have to hire someone else
  139. 4:18like I don't know how to make CAD plans,
  140. 4:19but now we have this intelligence in our
  141. 4:21pocket that can can actually do these
  142. 4:23kinds of things for us. To me, one of
  143. 4:25the things I really like to do is to try
  144. 4:26to push the AI in different dimensions
  145. 4:29of of my life. And so I think that voice
  146. 4:31mode is something that makes AI so
  147. 4:33accessible, right? You just go into
  148. 4:34voice mode and you start talking to it.
  149. 4:35Voice mode now is hooked up to tools. So
  150. 4:38you can actually ask it to do anything
  151. 4:40that your chat can do. And if you're
  152. 4:41hooked up to your email, to your
  153. 4:43calendar, now you can ask it about like
  154. 4:46what's my schedule for today or like
  155. 4:48what should I, you know, I want to free
  156. 4:49up some time for some deep thought. How
  157. 4:50can I do that? And it can actually go
  158. 4:52and have context on what's going on in
  159. 4:53each part of your calendar. So, I think
  160. 4:55that in my mind that there's this
  161. 4:57activity of just stepping back and
  162. 4:59thinking about what's a slightly more
  163. 5:00ambitious thing I can ask my AI to do
  164. 5:02than it's done for me to date. Let's see
  165. 5:04if it works.
  166. 5:05>> Yeah. And I'm just about to show you
  167. 5:06something that I've built recently. I
  168. 5:08think we're in this era where everyone
  169. 5:10can build a personal productivity app.
  170. 5:12Multiple. And uh this one that I've
  171. 5:14built, so I basically asked it to create
  172. 5:17a an app that's going to prepare me for
  173. 5:19the podcast. I tell it who I'm
  174. 5:21interviewing for how long and what's my
  175. 5:24success metric and it's basically the
  176. 5:26views on YouTube and then it goes ahead
  177. 5:27and creates a podcast. It does the
  178. 5:29research uh suggests topic to discuss
  179. 5:32this is great but how do I make it learn
  180. 5:35from my experience right and is there
  181. 5:37anywhere I can push this further cuz I'm
  182. 5:40going to record this conversation it's
  183. 5:41going to come here and it's going to
  184. 5:42check now it's checking my previous
  185. 5:43conversations but what do you think is
  186. 5:45missing here and am I ambitious enough
  187. 5:49with AI
  188. 5:50>> well I think the answer to being
  189. 5:52ambitious enough will always be no
  190. 5:54because there's always more we can do
  191. 5:55right and that's the beauty of AI is
  192. 5:56that it's really about solving problems
  193. 5:58helping you achieve what you want to
  194. 6:00achieve. You know, there's no ceiling to
  195. 6:01what you might imagine, right? And so,
  196. 6:03it's really just about sort of tuning
  197. 6:05that you were able to achieve a thing
  198. 6:07means maybe you can dream a little bit
  199. 6:08bigger. And so, in my mind, yeah, I
  200. 6:10think this I think this is super super
  201. 6:11cool app. So, f first of all, I in terms
  202. 6:14of just doing the background research
  203. 6:16and maybe I when you when this was being
  204. 6:18built, they already went and looked at
  205. 6:20all my interviews that I've ever done
  206. 6:22and did.
  207. 6:23>> Ah, there we go. Did all the background
  208. 6:24research.
  209. 6:25>> It did. It did the background. The only
  210. 6:26thing is it still can't connect to
  211. 6:28social media, but I think it's it's a
  212. 6:29plug-in thing. Yep.
  213. 6:30>> But then yes, it went through your
  214. 6:33conversations.
  215. 6:34>> There we go. And then all of Openai's
  216. 6:37past press releases. Oh, there we got a
  217. 6:39questions tab. I like that.
  218. 6:40>> Yeah, we got a questions tab. That's
  219. 6:42right. And that's the title and
  220. 6:43thumbnail already here. And these are
  221. 6:45the questions that I can copy.
  222. 6:46>> Yes. I'd be curious how much it helps
  223. 6:48with the post-processing and the actual
  224. 6:49production of the final of the final
  225. 6:51video once it's done. Right. because
  226. 6:53this thing has all this context on me,
  227. 6:55on open AI, on the kinds of things that
  228. 6:58you've covered in your your past
  229. 6:59episodes, and then how do you want to
  230. 7:01cut it? Like, I've I've heard there's so
  231. 7:03many good use cases of Astra helping
  232. 7:05with post-production for for videos and
  233. 7:08proactively being able to do that. And
  234. 7:09so, I think actually be I'd be very
  235. 7:11curious if you could have an app where
  236. 7:13it's just like the only thing you have
  237. 7:14to do is basically look over the
  238. 7:17questions, make sure you like them, and
  239. 7:18if you don't, provide some feedback, and
  240. 7:20it can do better. You show up, you
  241. 7:21actually do the interview, and then
  242. 7:23everything else is taken care of. How
  243. 7:24much work do you put in right now from
  244. 7:26the interview part to something actually
  245. 7:27being live on YouTube?
  246. 7:28>> Oh, a lot of work. I have a I have a big
  247. 7:30team, but it's also we've documented a
  248. 7:32lot and I've been doing this for 12
  249. 7:34years, so we have all the files that
  250. 7:35we're referencing, so it's easier to
  251. 7:37build with AI now, but it's still every
  252. 7:39episode is at least 40 hours.
  253. 7:41>> Wow. And if you could get back that
  254. 7:43time, right, so that kind of all of the
  255. 7:44the mechanics are taken care of. So you
  256. 7:46can spend so much time really thinking
  257. 7:47about what the right interview is and
  258. 7:50all of the the other things that go into
  259. 7:51this or really thinking about if you
  260. 7:53could scale you could do way more
  261. 7:54interviews like yeah what would you do
  262. 7:55with that additional time
  263. 7:56>> that's the question to you with small
  264. 7:58businesses who that are applying AI for
  265. 8:00their day-to-day jobs what is the role
  266. 8:02of the business owner what should they
  267. 8:04be thinking more of and where's actually
  268. 8:06the bottleneck for the business now
  269. 8:08>> I think that there's something
  270. 8:09fundamental so first of all we see so
  271. 8:10many small businesses that are being
  272. 8:12started because people feel secure with
  273. 8:15the fact that they have chatbt, right?
  274. 8:16That there's so many pieces of expertise
  275. 8:18that if you want to get a business off
  276. 8:19the ground, it's hard,
  277. 8:21>> right? But you actually have a PhD level
  278. 8:23intelligence, a world expert, business
  279. 8:26consultant. You have all these things in
  280. 8:28your pocket now for free. It's like
  281. 8:30pretty remarkable, right? And if you're
  282. 8:31willing, if you if you're able to pay,
  283. 8:32you can get even so much more depth and
  284. 8:34so much more compute towards any problem
  285. 8:36you want. So small businesses easier to
  286. 8:38start than ever and we're seeing it very
  287. 8:40concretely. And then people are able to
  288. 8:42I when you talk to many of these small
  289. 8:44business owners that there's just so
  290. 8:45many tasks that they outsource to an AI
  291. 8:47that you know certainly there's the ones
  292. 8:49that they knew that they had to do but
  293. 8:50there's so many tasks that they would
  294. 8:51just no one would have done it before
  295. 8:53and so little holes in their business
  296. 8:54strategy are thinking about how do you
  297. 8:56actually produce this document. The
  298. 8:57other thing too is is small business
  299. 8:59owners who don't even realize what they
  300. 9:00could be doing. like one of our our
  301. 9:02employees was telling me that he was um
  302. 9:05shopping for for watches in this like
  303. 9:06antique watch store and that he started
  304. 9:09talking to the owner about AI and that
  305. 9:11the owner was like, "Oh, AI, I can't use
  306. 9:12that for anything." Like, you know, I'm
  307. 9:14a I'm a watch store. And he was like,
  308. 9:16"No, no, AI can help everyone. Let me
  309. 9:17show you." And so he talked about some
  310. 9:19of the problems he has and he's like,
  311. 9:20"Well, yeah, he doesn't really have a
  312. 9:21website." And so you just Exactly.
  313. 9:25Right. And just like chatty
  314. 9:26[clears throat]
  315. 9:27voice being like, "Build me a website.
  316. 9:28Build an awesome website." And then the
  317. 9:30owner said, "Hey, I also have each watch
  318. 9:32has its own story, right?" And so it's
  319. 9:33just like then becomes very easy like,
  320. 9:34"Okay, talk to the AI about here's the
  321. 9:36story behind each watch." And you end up
  322. 9:37with this beautiful visualization of
  323. 9:38everything. And you just realize that
  324. 9:40there's this this way that technology
  325. 9:42becomes alive that normally it's almost
  326. 9:44inaccessible to so many people. And so I
  327. 9:46think that this really pushing on if you
  328. 9:48had like the world's best programmer on
  329. 9:51your payroll, what would you put them to
  330. 9:53work on? And and that now you can do I
  331. 9:56think you mentioned somewhere that it's
  332. 9:57actually ambition that small businesses
  333. 9:59have to start working on, right?
  334. 10:01>> Yes. Yes. And I I think that
  335. 10:03>> I think it's fascinating.
  336. 10:04>> I I agree and I think it is really about
  337. 10:06we can all raise a ceiling of ambition,
  338. 10:08right? We can all accomplish more people
  339. 10:10who really lean into that. I think we're
  340. 10:12already starting to see really
  341. 10:12remarkable things. I'll tell you another
  342. 10:14example by the way is um that that we
  343. 10:16have uh there are these two two brothers
  344. 10:18who are who are blind who use our
  345. 10:20technology and uh that they use it to uh
  346. 10:23both in their daily life and and for for
  347. 10:25their work and even for tasks like being
  348. 10:27able to plug uh the the cable into the
  349. 10:30right port on a computer, right? You can
  350. 10:32use chatbt to to actually basically
  351. 10:35provide that information when normally
  352. 10:36they would have to ask someone for for
  353. 10:38help. And so I think this kind of
  354. 10:39empowerment, you know, being able to
  355. 10:40help people be able to be more
  356. 10:42independent, be able to live their life
  357. 10:44the way that they want, all of that,
  358. 10:45it's happening. It's happening right now
  359. 10:46at massive scale.
  360. 10:48>> When someone is thinking of building
  361. 10:49something in software, when OpenAI is
  362. 10:51building so much around like work and
  363. 10:53productivity, what do you think makes an
  364. 10:55app worth paying for?
  365. 10:56>> There's definitely parts of what app
  366. 10:59building used to be that's becoming
  367. 11:00commoditized. It's becoming accessible
  368. 11:02to everyone. It's becoming massively
  369. 11:04democratized. And I think that's a
  370. 11:05really important thing, a beautiful
  371. 11:07thing, right? that there can just be
  372. 11:08more software and so then the question
  373. 11:10of yeah what is differentiation becomes
  374. 11:12even more important. I think that there
  375. 11:14is going to be more room than ever for
  376. 11:18software that adds value in unique ways
  377. 11:20and some of this is really about deeply
  378. 11:22understanding the problem. Some of this
  379. 11:23is deeply understanding your users,
  380. 11:25right? You think about something like if
  381. 11:26you want to build something for
  382. 11:28education, there's teachers, there's
  383. 11:30parents, there's students, right? That
  384. 11:31there's so many stakeholders that you
  385. 11:34have to really think about. There's
  386. 11:35administrators. How do you build a
  387. 11:37product that actually gives the right
  388. 11:38information, has the right guard rails,
  389. 11:40that is useful in the right ways and who
  390. 11:42are you selling to? All these questions.
  391. 11:43So, I think that these like human
  392. 11:45factors of what is the core of the
  393. 11:47business? What is it that you have
  394. 11:49discovered? What is it that's your
  395. 11:50insight? What is your unique way of
  396. 11:51looking at the world? And I think that
  397. 11:53the mechanics of how you instantiate
  398. 11:56that, how does you know do you have to
  399. 11:58write in C++ or Swift or something like
  400. 12:00who cares about that? Like that was
  401. 12:01never what the barrier should have been.
  402. 12:03>> Yeah, I've seen that YC startups recent
  403. 12:05batches are 60% B2B and my intuition is
  404. 12:09telling me with these tools available to
  405. 12:11us if I want to become an entrepreneur.
  406. 12:13I've seen so many problems in consumer.
  407. 12:15Do you think we're going to enter this
  408. 12:17era when more and more people are
  409. 12:18building consumer businesses? Well, I
  410. 12:20think that the line between what
  411. 12:22enterprise and consumer are is going to
  412. 12:24blur.
  413. 12:25>> I think that what a company is is going
  414. 12:27to become just less well-defined in some
  415. 12:31ways or it's going to be less like,
  416. 12:32okay, a company is like these like big
  417. 12:34masses of humans that are organized in
  418. 12:37certain ways. I think people are going
  419. 12:38to be so empowered, right? We're going
  420. 12:39to see this massive renaissance of
  421. 12:41entrepreneurship. Like, I think it is
  422. 12:42coming. I think we're seeing the leading
  423. 12:44edges of it, but it's really, I think,
  424. 12:46really going to kick in over the next
  425. 12:47year to two years. And I think that the
  426. 12:50question of are you selling to consumers
  427. 12:52or enterprises, there's some things that
  428. 12:55will remain true in terms of the split.
  429. 12:57Like I think that when I think about
  430. 12:58what enterprise sales is, it's fun
  431. 13:00fundamentally about relationships. But
  432. 13:01if you think about it as relationships,
  433. 13:03it's like well the reason that you can't
  434. 13:05really do enterprise sales to consumers
  435. 13:07is because there's just too many
  436. 13:08consumers, right? It's just too high of
  437. 13:10a scale. It's too expensive. But if you
  438. 13:12have AI that really amplifies your
  439. 13:13ability to have relationships, really
  440. 13:15meaningful ones and and really take care
  441. 13:16of people at massive scale, then
  442. 13:18actually maybe the enterprise sales
  443. 13:19motion becomes something that makes
  444. 13:20sense even for more consumer businesses.
  445. 13:22But the mechanics of what a business is,
  446. 13:24I think, will evolve. I think that the
  447. 13:26how you relate to your users will
  448. 13:29evolve. But the fundamentals of that you
  449. 13:32want to be delivering value and you want
  450. 13:34to build a relationship with the people
  451. 13:36who are receiving that value and have it
  452. 13:37be a birectional exchange. All of that
  453. 13:40is going to become much more front and
  454. 13:41center regardless of which domain.
  455. 13:43>> Greg is talking about how much more you
  456. 13:45can do now even with a tiny team. If
  457. 13:47that makes you think maybe I can start
  458. 13:50that business, I want you to think about
  459. 13:52your first customer. You want to listen,
  460. 13:55understand what they need and do a great
  461. 13:57job. And you also have notes,
  462. 13:58follow-ups, and a spreadsheet to update.
  463. 14:00Our sponsor, HubSpot, helps with that.
  464. 14:02Its new smart CRM is a CRM that updates
  465. 14:05itself, keeping your customer
  466. 14:07information and conversations together.
  467. 14:09Its AI notetaker captures customer
  468. 14:11calls, drafts follow-ups, and suggests
  469. 14:13next steps and updates to customer
  470. 14:15records for you to approve, so more of
  471. 14:17your attention can go to the people
  472. 14:19you're building this business for. Check
  473. 14:21out HubSpot Smart CRM through the link
  474. 14:23in the description. And thanks HubSpot
  475. 14:25for sponsoring this episode. Now, back
  476. 14:27to Greg. when people are building um
  477. 14:30software, how do you strategize around
  478. 14:33building something that's not going to
  479. 14:34be made obsolete with the new models?
  480. 14:36>> There's some activities I think scale
  481. 14:38well as the models get smarter and some
  482. 14:40activities that are kind of working
  483. 14:41around limitations in the model. The
  484. 14:43second one is pretty hard to be durable
  485. 14:45and so I think that there are at various
  486. 14:47points been startups that really focus
  487. 14:48on it. take the model, we add some
  488. 14:50prompts, we have some very handcrafted
  489. 14:52whatever. And it's not to say that
  490. 14:54there's there's a point in time where
  491. 14:55that is valuable. And sometimes it's
  492. 14:57even about the interface, right? That at
  493. 14:59each mo level of model capability,
  494. 15:01there's almost a new interface that
  495. 15:02suggests itself. In 2022, we had an AI
  496. 15:06that was personalized enough that it was
  497. 15:08worthwhile reading the responses GBD4.
  498. 15:11So it's like, of course, you want to
  499. 15:12turn make a chat GBT unit at 3.5. and
  500. 15:14GB4, you know, end of of 2025, I'd say,
  501. 15:17is when the coding agent era really
  502. 15:19started. And that was like you had AI
  503. 15:21that kind of shifted from being 20% of
  504. 15:23your software production to 80% of your
  505. 15:24software production. And so, you just
  506. 15:26want a very different product that
  507. 15:28captures that thing. This is the thing
  508. 15:30everyone's going to utilize. Now, I
  509. 15:32think we're starting to enter this era
  510. 15:34of the personalized agent. So, you're
  511. 15:36going to have a cloud-based AI and this
  512. 15:37is being raified number of different
  513. 15:39form factors. Openi, we just released
  514. 15:41Dots. I'm super excited about I think
  515. 15:43it's got the smartest model backing up
  516. 15:45this this kind of of technology and it's
  517. 15:48so useful. It's useful to me. I think
  518. 15:49it's be useful for lots of people. And
  519. 15:51so I think that the core thing is
  520. 15:52thinking about where will you scale as
  521. 15:56the models get smarter. And again, some
  522. 15:58of this is about understanding the
  523. 15:59industry. Some of this about building
  524. 16:00relationships with back to if you think
  525. 16:02about hospitals, right? There's so many
  526. 16:04different stakeholders there, right?
  527. 16:05That there's the patient, the doctor,
  528. 16:07the hospital admin, there's the uh
  529. 16:08insurance provider. There's so many
  530. 16:10different parties. And so how do you
  531. 16:12build trust across each of those and
  532. 16:15have a product that's useful to them and
  533. 16:17then actually be able to even if you're
  534. 16:19yeah when you think about if you have a
  535. 16:20model that's just even better if you are
  536. 16:22already in that industry you have these
  537. 16:24customers that actually your product
  538. 16:26gets better and more defensible as a
  539. 16:28result.
  540. 16:28>> Is there anything you particularly
  541. 16:30looking forward to um seeing startups uh
  542. 16:33use models to build something? I think
  543. 16:34the domain expertise is also really key
  544. 16:36because one thing that's easy to miss is
  545. 16:38that as something like open AI we have
  546. 16:40this technology that's so broad right it
  547. 16:42kind of touches every single aspect of
  548. 16:45economic growth and development people
  549. 16:47in their personal lives work lives
  550. 16:49there's something for everyone something
  551. 16:50that can really help them something to
  552. 16:52help them achieve their goals but often
  553. 16:54in any specific vertical we're so far
  554. 16:56from optimal right that if you you
  555. 16:58almost think of the economy as this
  556. 16:59fractal thing right where it's like you
  557. 17:01zoom in you're like okay here's a sector
  558. 17:03of like healthcare and then you zoom and
  559. 17:04you like really look at the number of
  560. 17:05different companies and the number of
  561. 17:07different ways of doing things and no
  562. 17:08matter what sector you pick, there's
  563. 17:10just so much low hanging fruit. That's
  564. 17:12actually been one of the most
  565. 17:13interesting and surprising things for me
  566. 17:15in building OpenAI is to realize that
  567. 17:17there's just so much suboptimality in
  568. 17:19the economy in terms of how you can
  569. 17:22actually have much more efficient
  570. 17:23process that gives people their time
  571. 17:24back that actually delivers the value
  572. 17:26that being able to connect across
  573. 17:28different fields is not something we do
  574. 17:29very much, right? It's like people
  575. 17:31specialize
  576. 17:31>> and now it makes economic sense to build
  577. 17:33them out because 5 years ago they would
  578. 17:35be like a feature of a big company now
  579. 17:37you can build a small company around
  580. 17:38that feature.
  581. 17:39>> Exactly. And I think that even these
  582. 17:41like classic sectors and domains are
  583. 17:43really because of the limitation of our
  584. 17:46tools. Again you think about how do you
  585. 17:48become good like an expert in any field
  586. 17:50is like you go to school you like
  587. 17:51specialize and as humanity has learn
  588. 17:53more things the what it means to
  589. 17:55specialize becomes narrower and
  590. 17:57narrower. Like if you're a biologist
  591. 17:58now, you get a PhD, there's like one
  592. 18:00specific reaction that you got your PhD
  593. 18:02on, right? And it's like it's not
  594. 18:03connected to the whole system. And I
  595. 18:05think that when it comes to to building
  596. 18:07a business, a question of can you have
  597. 18:09these much more hybrid things. So you
  598. 18:11think about things like edtech as an
  599. 18:12example of like you take people who
  600. 18:15understand about education, they
  601. 18:16understand technology, they mash them up
  602. 18:17and that they're able to innovate and do
  603. 18:19new things. I wonder if we're going to
  604. 18:21find that there's going to be other
  605. 18:22domains that we've never thought about
  606. 18:24that actually mesh together super well.
  607. 18:25>> Totally. That's the future of
  608. 18:26entrepreneurship, matching those domains
  609. 18:28and finding the right solution. Can you
  610. 18:29talk to me about a setup that you have
  611. 18:31that is sophisticated enough for me to
  612. 18:33compare this because sometimes people
  613. 18:36bash me in the comments for asking
  614. 18:38people like you about morning briefing
  615. 18:39agents. So, let's not do morning
  616. 18:41briefing agent. Let's do something that
  617. 18:42helps you run a company and be
  618. 18:43efficient. It's something that people
  619. 18:45see, they will think it's an aha moment
  620. 18:47for their productivity.
  621. 18:48>> I would not poo poo the morning briefing
  622. 18:50agent too much [laughter] because I
  623. 18:52think the morning briefing agent has so
  624. 18:53much depth to it, right? There's the
  625. 18:55like, okay, yeah, it's like it kind of
  626. 18:56gave me some info where it's like,
  627. 18:58honestly, I could have just scrolled the
  628. 18:59notification myself. There's the version
  629. 19:01of it where it says like, here are the
  630. 19:04things that are pending, right? Like
  631. 19:06here are these slacks that you have not
  632. 19:07answered. Um, and you really should get
  633. 19:09back to these people with this
  634. 19:10prioritization order. There's the next
  635. 19:12level of I have like done the research.
  636. 19:15you know, there's this question about
  637. 19:16like what our strategy should be in some
  638. 19:18particular market and like I've actually
  639. 19:19gone and I even like, you know, sort of
  640. 19:21looked at all the presentations and like
  641. 19:23here's like the things you need to like
  642. 19:24be paying attention to to make this
  643. 19:26decision. That third thing is like where
  644. 19:27I think we're at, right, is like the AI
  645. 19:29that really can do very sophisticated
  646. 19:31research across a bunch of different
  647. 19:32things and provide a very compressed
  648. 19:33version of of what I should be paying
  649. 19:35attention to. thing I'm actually very
  650. 19:36excited about as a next step that we're
  651. 19:37not quite there yet is the morning
  652. 19:40briefing agent so to [clears throat]
  653. 19:41speak that is like hey I actually called
  654. 19:43like you know I called up these people
  655. 19:45like I talked to you know these these
  656. 19:47co-workers of yours I talked to this
  657. 19:48external partner and like I got
  658. 19:50everything into perfect order like I
  659. 19:52just need you to approve this purchase
  660. 19:53order and like we're ready to go on the
  661. 19:55new whatever AI that really is proactive
  662. 19:57and able to do things on your behalf
  663. 19:59like that's coming and I think with DOTS
  664. 20:01we actually have the technology for it
  665. 20:03so I think we're going to start seeing
  666. 20:04people really start to utilize that. I
  667. 20:06think we have [clears throat]
  668. 20:06improvements to make and things like
  669. 20:07that. But I'd say in my mind, any of
  670. 20:09these use cases, there's so much depth
  671. 20:11to them. And and that is to me the key
  672. 20:13thing about AI. If you just look at the
  673. 20:15label of an AI that writes software,
  674. 20:18okay, sure, we've had an AI that writes
  675. 20:19software for 5 years now, but AI that
  676. 20:22writes software today, totally different
  677. 20:24from the AI that wrote software 5 years
  678. 20:25ago. So I have a lot of appreciation for
  679. 20:28the depth of application that's possible
  680. 20:31even if the like label on the tin is
  681. 20:33like the same thing and kind of
  682. 20:34everyone's heard it 100 times.
  683. 20:35>> When you look at these use cases inside
  684. 20:36OpenAI, do you have like an internal
  685. 20:38metric how advances in models is
  686. 20:40actually are actually translating into
  687. 20:42people's productivity?
  688. 20:43>> We published blog posts on this as well.
  689. 20:44So we've done a decent amount of of work
  690. 20:46looking at token usage and looking how
  691. 20:48it correlates with like lines of code or
  692. 20:50commits and PRs and things like that.
  693. 20:52And again, all these are proxy metrics.
  694. 20:53They're not perfect, but I think that
  695. 20:56another way that you can tell is AI
  696. 20:58really changing productivity and how
  697. 20:59much is if you ever have an internal
  698. 21:01outage, like if internally people can't
  699. 21:03get access to a model, um how much they
  700. 21:05feel upset. Everyone's like, "I can't do
  701. 21:07my work." And you're like, "Okay, 6
  702. 21:08months ago you had nothing like this and
  703. 21:10now you're saying you can't do your work
  704. 21:11when you don't have it for two hours."
  705. 21:12Like, how is that possible? But we see
  706. 21:14this all the time. It's so so cool. And
  707. 21:16I feel for myself like the kinds of
  708. 21:18things on the morning briefing agent
  709. 21:20access that I do a lot of is I like
  710. 21:22anytime I have a question about open AI
  711. 21:24I just ask my agent like I don't know
  712. 21:27why I would do anything else to start
  713. 21:28because it's just so efficient. It's
  714. 21:30like often and sometimes I even feel
  715. 21:31like I'm like being a little bit lazy of
  716. 21:33like there's like some message that
  717. 21:35someone has sent me with a doc in it.
  718. 21:36I'm just like trying to vaguely remember
  719. 21:38where was it and
  720. 21:39>> so just connect it to everything. It can
  721. 21:40it's going to do a mail slack whatever
  722. 21:42you're using and it has all the
  723. 21:43information. So whenever you need
  724. 21:44something, you just ask it and it has
  725. 21:46all this.
  726. 21:46>> Yes. And with with computer use, it
  727. 21:48really closes the last mile of being
  728. 21:49able to connect to to all my enterprise
  729. 21:51context. And so then that makes it so
  730. 21:53useful because I really have this chief
  731. 21:54of staff and I actually have a human
  732. 21:56chief of staff is also incredible,
  733. 21:57right? And that I kind of am able to
  734. 21:59like give her so much more leverage by
  735. 22:01not having to ask these basic questions
  736. 22:03that now my AI can answer. And some of
  737. 22:05these, you know, call call basic
  738. 22:07questions are also quite sophisticated
  739. 22:08like for example asking questions about
  740. 22:10different business metrics, right? So it
  741. 22:12actually has to then go and issue
  742. 22:13queries into our data warehouse and come
  743. 22:15up with like you know you're looking at
  744. 22:16like the growth in some market or the
  745. 22:18you know the retention curves or
  746. 22:19something like that. And I remember in
  747. 22:22you know my my last job when I was at
  748. 22:24Stripe that we had a data team that you
  749. 22:26would send like an email to and then a
  750. 22:28week later they would come back with an
  751. 22:29answer to that that query and now I just
  752. 22:31ask the AI and it just does it. It's
  753. 22:33it's incredible.
  754. 22:34>> Do you ever see areas where people are
  755. 22:36underutilizing tokens? Do you want to
  756. 22:38tell people like use more tokens for
  757. 22:40this? Oh, absolutely. And look, I think
  758. 22:42there's two dimensions to it. So, one is
  759. 22:44that there was this era of token maxing,
  760. 22:46right? Where it's just like people were
  761. 22:47just blindly saying throw more tokens at
  762. 22:48it, which is clearly not a primary
  763. 22:49objective, but it is a good proxy
  764. 22:51metric. If you're not trying to optimize
  765. 22:53tokens, you're probably going to then
  766. 22:55see improve productivity correlated with
  767. 22:58increased number of tokens. And we
  768. 22:59really do see it. We have leaderboards.
  769. 23:01We kind of look at that. And I think
  770. 23:02that we see like the people who tend to
  771. 23:04be the most sophisticated in their usage
  772. 23:06are just able to do really incredible
  773. 23:07things. Again, it all depends on the
  774. 23:09domain and things like that, but I love
  775. 23:11seeing, for example, in our finance team
  776. 23:13like that they've really gone full AI AI
  777. 23:16first. And I think that there's just
  778. 23:18like so many things where like workflows
  779. 23:20that used to take a long time for
  780. 23:21closing the books that are now just so
  781. 23:23much faster, right? And none of those
  782. 23:24are like everyone just feels that
  783. 23:26increased productivity and how how
  784. 23:27helpful it is. Uh we have an early some
  785. 23:30early adopters from our comm's team. I
  786. 23:31think that they're just like being able
  787. 23:33to manage like you have an event with a
  788. 23:35bunch of journalists coming and you're
  789. 23:36putting together a seating chart and
  790. 23:37trying to figure out everyone's dietary
  791. 23:39preferences and being able to manage all
  792. 23:40of that as one person because you have
  793. 23:42codeex that's actually taking care of so
  794. 23:44much of the back end for you. Like
  795. 23:45that's all been happening since early
  796. 23:47this year. So we see these kinds of use
  797. 23:49cases for knowledge work productivity
  798. 23:51everywhere. And then of course the thing
  799. 23:53that I think is also very easy to to
  800. 23:56miss and I think is actually coming soon
  801. 23:57for everyone is the scientific discovery
  802. 24:00side of things, right? that we put a ton
  803. 24:02of compute into solving Navier Stokes,
  804. 24:04you know, this this big millennium
  805. 24:06problem and that was like 10,000 agents
  806. 24:09for some number of days like it's a lot
  807. 24:11of compute and we're able to create new
  808. 24:13knowledge for humanity and you think
  809. 24:16about actually at much smaller scale I
  810. 24:18think it still is the case that we're
  811. 24:19seeing lots of breakthroughs happening
  812. 24:21and so I just have this feeling of like
  813. 24:23I think that people who really point
  814. 24:25compute at the right problem right have
  815. 24:28some judgment and some real thought on
  816. 24:29like what really matters where is the
  817. 24:31compute going to be utilized and then
  818. 24:33are willing to push that that token
  819. 24:34budget high. We're really seeing this
  820. 24:36like sort of outsiz impact.
  821. 24:38>> I think it's a big question that a lot
  822. 24:39of people have who are starting using
  823. 24:41agents and automating for the first
  824. 24:43time. What do you start with? How do you
  825. 24:45think about that?
  826. 24:46>> This white page problem is like the
  827. 24:48hardest problem in AI in some ways and
  828. 24:49utilizing AI is this question of where
  829. 24:51do you get started? What do you use? I
  830. 24:52think my answer is always start small.
  831. 24:54Like just try something, type something
  832. 24:56random, push enter, see what happens,
  833. 24:57right? And like I think that then as
  834. 25:00soon as you get that feedback loop then
  835. 25:02you can iterate, right? You can improve.
  836. 25:03And so I think that find like a little
  837. 25:06pain point in your life, right? Whether
  838. 25:07it's maybe it's an email in your inbox
  839. 25:09that you you know that you have been
  840. 25:11treading replying to or you know that
  841. 25:13you know it's going to take a lot of
  842. 25:14work to do to to get there. Um and so
  843. 25:17then the first question is well how do I
  844. 25:18get the email to the AI? Right? And so
  845. 25:20there's one answer which you copy paste
  846. 25:21it. There's another answer where you
  847. 25:23have your Gmail connector connected. And
  848. 25:25so I think that you can start very
  849. 25:27mechanically, very much going through
  850. 25:29the the effort yourself and then
  851. 25:31realize, okay, I've done this three
  852. 25:33times, I've done this five times, and
  853. 25:34now I understand how this process works.
  854. 25:37Let's automate it. And that, by the way,
  855. 25:38is like a general software engineering
  856. 25:40maxim. Like that's very much how I've
  857. 25:41always approached when you build
  858. 25:42software versus when you just kind of do
  859. 25:44it mechanically is after you've done it
  860. 25:45five times, it's probably a good time to
  861. 25:47to start automating. And so I think
  862. 25:48>> I like that the five times metric.
  863. 25:51>> It's a pretty good one.
  864. 25:52>> Exactly. Yeah. And I think I think that
  865. 25:53doing that for AI because that way you
  866. 25:55kind of keep your your finger on the
  867. 25:57pulse of the problem uh which I think is
  868. 25:58very important like at the end of the
  869. 26:00day you know back to the question of
  870. 26:02what are small business owners doing
  871. 26:03with AI accountability is something that
  872. 26:06remains with them right it's still the
  873. 26:07business is operating a certain way and
  874. 26:09the customer either signs up or doesn't
  875. 26:11and you know either pays or doesn't and
  876. 26:13so they should really care about that
  877. 26:14outcome and be deep on the details
  878. 26:16>> another side of this problem is how do
  879. 26:17you decide what not to automate and I
  880. 26:19wanted to ask you about with you know
  881. 26:22your mission in mind. How do you at
  882. 26:23OpenAI decide what not to build next?
  883. 26:26>> I think it's very important that humans
  884. 26:28remain in control and in charge, right?
  885. 26:29That the goals come from people. Point
  886. 26:31of this technology is to help people,
  887. 26:34right? To help empower everyone to help
  888. 26:36people achieve more of what they want to
  889. 26:38do. And so I think that that human first
  890. 26:41team humanity kind of of vibe is is very
  891. 26:43important, very core to our mission. And
  892. 26:45that's something that really bleeds
  893. 26:46through to how we think about things.
  894. 26:48And so when we think about what to
  895. 26:51build, we always think about how do you
  896. 26:52put in the right guard rails? How do you
  897. 26:54build trust? How do you ensure that
  898. 26:56there's good oversight, moniability, and
  899. 26:58that that is something that has been a
  900. 27:00very core focus of building all that
  901. 27:01technology for us for really since
  902. 27:04inception?
  903. 27:04>> For someone who is willing to step up
  904. 27:07their agent game and they have 1 hour,
  905. 27:09where do they start this weekend?
  906. 27:10>> I'd say first of all, sign up for chat
  907. 27:12to these dots. It's I I think just like
  908. 27:14a really cool new form factor. Again,
  909. 27:16you can talk, you know, chat a little
  910. 27:17bit, get some sense of what it does. Um,
  911. 27:19but connecting it to the appropriate
  912. 27:21context can really do work, right? It
  913. 27:22has its own cloud computer and it can
  914. 27:24optionally hook to your local computer.
  915. 27:26And I think that this picture that we've
  916. 27:28all had of like the way the coding
  917. 27:29agents work is this like task based
  918. 27:31system that you kind of micromanage that
  919. 27:32lives on your laptop. You close your
  920. 27:34laptop, it stops working. Like clearly
  921. 27:36that was never the AI that was promised,
  922. 27:37right? That was never the AI we pictured
  923. 27:395 years ago. And so I think we now have
  924. 27:40something that really looks like the AI
  925. 27:42that we were always picturing. And so
  926. 27:44again, start with a little task like ask
  927. 27:46it to go buy something for you or try to
  928. 27:50like I I really like texting mine,
  929. 27:51right? Just like go and and and really
  930. 27:53set up some connector so that it feels
  931. 27:55like okay, I'm I've got a communication
  932. 27:56channel to this now and ask it to do
  933. 27:59some recurring task, right? Or ask it
  934. 28:01for to do some research. And the thing
  935. 28:03that's very interesting is that DOTS can
  936. 28:06really, it's quite sophisticated, right?
  937. 28:07It's powered by GP6 Astra. So, it's
  938. 28:09actually got basically this incredible
  939. 28:10brain, like this smart, incredible, most
  940. 28:13intelligent model for any of these kinds
  941. 28:14of assistants powering it. And that
  942. 28:18means that it makes some connections
  943. 28:19between things like you ask it a
  944. 28:20question about, hey, like what's going
  945. 28:22on in this particular geography or
  946. 28:24something? And it will realize that
  947. 28:26that's something you care about. And
  948. 28:27then it might proactive within surface
  949. 28:29if there's a new news report about that
  950. 28:30that region, then maybe it will say,
  951. 28:32"Hey, you're doing
  952. 28:33>> some extra thinking for you, which I
  953. 28:34really appreciate." Do you think that's
  954. 28:36the future of work or do you think in 5
  955. 28:38years it's going to be something we
  956. 28:39can't even imagine now?
  957. 28:40>> Well, I think AI is always surprising.
  958. 28:42So, I think that there's going to be
  959. 28:43somehow that the fundamentals will be
  960. 28:45we'll have smarter AI that's like more
  961. 28:47accessible and more empowering and all
  962. 28:48those things. Um, but that the way in
  963. 28:50which you'll utilize it will be
  964. 28:52surprising relative to what we expect
  965. 28:54today. So, I do think that the form
  966. 28:55factors will continue to change. And I
  967. 28:57think that the like if you think about
  968. 28:59really we're gonna have this this AI
  969. 29:01that's able to solve these unsolved math
  970. 29:04problems in everyone's hands.
  971. 29:06>> And so what are people going to do with
  972. 29:08that? Like what are the kinds of
  973. 29:09breakthroughs that are going to become
  974. 29:10possible? And I think about that for
  975. 29:12medicine, right? For drug discovery is
  976. 29:14going to be incredible. Material
  977. 29:15science, we're just going to live in a
  978. 29:17world where people are going to be able
  979. 29:19to solve problems that before required a
  980. 29:21very small set of specialized experts to
  981. 29:23be mobilized behind them. Like you will
  982. 29:25be able to do that. you will be able to
  983. 29:26solve these challenges that no one else
  984. 29:29has ever thought about and there going
  985. 29:30to be some challenges that no one's ever
  986. 29:32faced before that you're going to be
  987. 29:33able to just like be the first person to
  988. 29:36to solve and I think that we're going to
  989. 29:38do this as this broad community and so
  990. 29:40the tools for that I think there's going
  991. 29:41to be so much more collaboration and
  992. 29:43these AIs are going to uh be something
  993. 29:45that will represent you um that you are
  994. 29:48going to have an AI that that is kind of
  995. 29:50there working for your interest
  996. 29:51companies will have their own AI so it
  997. 29:53will be I think there's like a lot of
  998. 29:55infrastructure to be built and building
  999. 29:56this in a way that's trustworthy and
  1000. 29:58that is observable and that really
  1001. 30:00uplifts humanity like that is I think
  1002. 30:01very core to what we should collectively
  1003. 30:03be doing.
  1004. 30:04>> I think that's an amazing answer and I
  1005. 30:05think we'll have homework to work on our
  1006. 30:08ambition this weekend not just agents.
  1007. 30:10>> All right.
  1008. 30:10>> Thank you so much Greg. That was
  1009. 30:11amazing.
  1010. 30:12>> Thank you. Thank you.

About this transcript

This page contains the full transcript of OpenAI Co-Founder: Start Building With AI Before You Feel Ready | Greg Brockman by Silicon Valley Girl, generated from the public captions YouTube serves with the video. The transcript has 7,129 words across 1,010 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.