OpenAI CEO Sam Altman | AI for the Next Era — Transcript

Full transcript

1. 0:09all right let's start a little bit more
2. 0:12pragmatic but then we'll Branch out so
3. 0:14one of the things I think a lot of folks
4. 0:16here are interested in is based off the
5. 0:21apis that very large models will create
6. 0:23what are the real business opportunities
7. 0:25like what are the ways to look forward
8. 0:28and then how given the the apis will be
9. 0:31available
10. 0:33to multiple players how do you create
11. 0:35distinctive businesses on him yeah
12. 0:38um so I I think so far we've been in the
13. 0:40realm where it's you know you can do
14. 0:41like an incredible copywriting business
15. 0:43or you can do like a sort of like
16. 0:45Education Service or whatever
17. 0:47um but we I don't think we've yet seen
18. 0:49the kind of like people go after the
19. 0:51like you know trillion dollar like take
20. 0:53on Google's
21. 0:54um and I think that's about to happen
22. 0:56like maybe it'll be successful maybe
23. 0:57Google will do it themselves but like I
24. 1:00would guess that with the quality of of
25. 1:03language models we'll see in the coming
26. 1:04years
27. 1:06um
28. 1:07you know there will be like a serious
29. 1:08challenge to Google for the first time
30. 1:10for for a search product
31. 1:12um and I think people are really
32. 1:13starting to think about like how did the
33. 1:15fundamental things change
34. 1:17um and that's going to be really
35. 1:19powerful uh I think that a like a human
36. 1:23level
37. 1:24chat bot interface that actually works
38. 1:27this time around like I I think like you
39. 1:29know many of these trends that like we
40. 1:31all made fun of were just too early like
41. 1:32the chatbot thing was good it was just
42. 1:34too early now it can work and I think
43. 1:36you know having like new medical
44. 1:39services that are done through that
45. 1:41where you get great advice or new
46. 1:42Education Services like this these are
47. 1:45going to be very large companies I think
48. 1:47we'll get multimodal models and not that
49. 1:49much longer and that'll open up new
50. 1:51things I think people are doing amazing
51. 1:53work with sort of agents that can use
52. 1:55computers to do things for you use
53. 1:57programs and this idea of like a
54. 2:00language interface
55. 2:02um where you know you say a natural
56. 2:04language what you want in this kind of
57. 2:05like dialogue back and forth you can
58. 2:07iterate and refine it and the computer
59. 2:08just does it for you you see some of
60. 2:10this uh with like Dolly and co-pilot in
61. 2:12very early ways but I think this is
62. 2:15going to be a massive Trend and you know
63. 2:18very large businesses will get built
64. 2:20with this as the interface and more
65. 2:22generally that like these very powerful
66. 2:24models will will be
67. 2:26one of the genuine new technological
68. 2:29platforms which we haven't really had
69. 2:30since mobile
70. 2:32and there's always like an explosion of
71. 2:34new companies right after
72. 2:35so that'll be cool and and what do you
73. 2:38what do you things are
74. 2:39given that the large language model we
75. 2:41provided as an API service what are the
76. 2:44things that you think that folks who are
77. 2:46thinking about these kind of AI
78. 2:47businesses should think about is how do
79. 2:49you create an enduring differentiated
80. 2:51business so
81. 2:53you know they're they're I think there
82. 2:55will be a small handful of like
83. 2:57fundamental large models out there that
84. 2:59other people build on but right now what
85. 3:01happens is you know company makes large
86. 3:05language model API other people build on
87. 3:07top of it and I think there will be a
88. 3:08middle layer that becomes really
89. 3:10important where uh
90. 3:12I'm like skeptical of all of the
91. 3:14startups that are trying to sort of
92. 3:15train their own models I don't think
93. 3:16that's going to keep going but what I
94. 3:17think will happen is there'll be a whole
95. 3:19new set of startups that take an
96. 3:21existing very large model of the future
97. 3:23and tune it uh which is not just
98. 3:27fine-tuning like all the things you can
99. 3:28do I think there will be a lot of access
100. 3:29provided to create the model for
101. 3:32medicine or using a computer or like the
102. 3:36kind of like friend or whatever and then
103. 3:39those those companies will create a lot
104. 3:41of enduring value because they will have
105. 3:43like a special version of they won't
106. 3:45have to have created the base model but
107. 3:47they will have created something they
108. 3:48can use just for themselves or share
109. 3:50with others that has this unique data
110. 3:52flywheel going that sort of improves
111. 3:54over time and all of that so I think
112. 3:55there will be a lot of value created in
113. 3:57that middle layer
114. 3:58and what do you think some of the most
115. 4:00surprising ones will be it's a little
116. 4:02bit like for example you know a surprise
117. 4:05from a couple years ago and we talked a
118. 4:07little bit to Kevin Scott about this
119. 4:09this morning as we opened up which is
120. 4:11train on the internet do code right so
121. 4:14so what do you think some of the the
122. 4:17surprises will be of you didn't realize
123. 4:20it reached that far
124. 4:22I think the biggest like systemic
125. 4:24mistake in thinking people are making
126. 4:25right now is they're like all right you
127. 4:27know maybe I was skeptical but this
128. 4:28language model thing is really going to
129. 4:30work and sure like images video too but
130. 4:32but it's not going to be generating net
131. 4:34new knowledge for Humanity it's just
132. 4:36going to like do what other people have
133. 4:38done and you know that's still great
134. 4:39that's still like brings the marginal
135. 4:42cost of intelligence very low but it's
136. 4:43not it's not going to go like create
137. 4:45fundamentally new it's not going to cure
138. 4:47cancer it's not going to add to the sum
139. 4:49total of human scientific knowledge and
140. 4:51that is what I think will turn out to be
141. 4:52wrong that most surprises the current
142. 4:54experts in the field yep so uh let's go
143. 4:57to science then there's the next thing
144. 4:58what are some of the things whether it's
145. 5:00building on the apis you know uh use of
146. 5:03apis by scientists where what are some
147. 5:07of the places where science will get
148. 5:08accelerated and how so I think there's
149. 5:10two things happening now and then a
150. 5:12bigger third one later
151. 5:13um one is there are these science
152. 5:15dedicated products whatever like Alpha
153. 5:18fold and those are adding huge amounts
154. 5:20of value and you're gonna see in this
155. 5:21like like way more and way more I like I
156. 5:25think I if I were like
157. 5:28you know had time to do something else I
158. 5:30would be so excited to like go after a
159. 5:32bio company right now like I think you
160. 5:34can just do amazing things there
161. 5:36um the
162. 5:38anyway but there's like another thing
163. 5:40that's happening which is like tools
164. 5:42that just make us all much more
165. 5:44productive uh that help us think of new
166. 5:46research directions that sort of write a
167. 5:48bunch of our code so you know we can be
168. 5:50twice as productive and that impact on
169. 5:53like the net output of one engineer a
170. 5:55scientist I think will be the surprising
171. 5:57way that AI contributes to science that
172. 5:59is like outside of the obvious models
173. 6:01but even just seeing now like what I
174. 6:03think these tools are capable of doing
175. 6:05copilot is an example there's you know
176. 6:07be much cooler stuff than that
177. 6:09um that will be a significant like
178. 6:12change to the way that technological
179. 6:14development scientific development
180. 6:16happens but then so those are the two
181. 6:19that I think are like huge now and uh
182. 6:21lead to like just an acceleration of
183. 6:24progress but then the big thing
184. 6:26um that I think people are starting to
185. 6:29explore is um
186. 6:32I hesitate to use this word because I
187. 6:34think there's one one way it's used
188. 6:35which is fine and one that is more scary
189. 6:37but uh like AI that can start to be like
190. 6:41an AI scientist and self-improve and so
191. 6:44when like can we automate like can we
192. 6:46automate our own jobs as AI developers
193. 6:49very first the very first thing we do
194. 6:51can that help us like solve the really
195. 6:53hard alignment problems that we don't
196. 6:54know how to solve like that honestly I
197. 6:56think is how it's going to happen
198. 6:58um
199. 6:59the the scary version of
200. 7:00self-improvement like the one from the
201. 7:01science fiction books is like you know
202. 7:04editing your own code and
203. 7:07changing your optimization algorithm and
204. 7:09whatever else
205. 7:10um but there's a less scary version of
206. 7:12self-improvement which is like kind of
207. 7:14what humans do which is if we try to go
208. 7:16off and like
209. 7:18discover new science uh you know that's
210. 7:20like we come up with explanations we
211. 7:22test them we think like we whatever
212. 7:25process we do uh that is like special to
213. 7:27humans
214. 7:28teaching AI to do that I'm very excited
215. 7:32to see what that does for the total like
216. 7:34I'm a big believer that the only real
217. 7:38driver of human progress and economic
218. 7:40growth over the long term is the the
219. 7:43structure the societal structure that
220. 7:45enables scientific progress and then
221. 7:46scientific progress itself and uh like I
222. 7:50think we're gonna make a lot more of
223. 7:51that
224. 7:52well especially science that's deployed
225. 7:54in technology say a little bit about how
226. 7:56what uh I think probably most people
227. 7:59understand what the alignment problem is
228. 8:00but it's probably worth four sentences
229. 8:02on the alignment problem
230. 8:04yeah so the alignment problem is like
231. 8:06we're going to make this incredibly
232. 8:07powerful system and like be really bad
233. 8:11if it doesn't do what we want or or if
234. 8:14it sort of has you know goals that are
235. 8:16uh either in conflict with ours um and
236. 8:19many Sci-Fi movies about what happens
237. 8:20there or goals where it just like
238. 8:22doesn't care about us that much and so
239. 8:24the alignment problem is how do we build
240. 8:26AGI that that does what is in the best
241. 8:30interest of humanity how do we make sure
242. 8:32that Humanity gets to determine the you
243. 8:34know the future of humanity and how do
244. 8:37we avoid both like accidental misuse
245. 8:40um like where something goes wrong we
246. 8:42didn't intend intentional misuse where
247. 8:43like a bad person is like using an AGI
248. 8:46for great harm even if that's what other
249. 8:48person wants and then the kind of like
250. 8:49you know inner alignment problems where
251. 8:51like what if this thing just becomes a
252. 8:52creature that views this as a threat
253. 8:55um
254. 8:56the the way that I think the
255. 8:58self-improving systems help us is not
256. 9:00necessarily by the nature of
257. 9:01self-improving but like we have some
258. 9:03ideas about how to solve the alignment
259. 9:05problem at small scale
260. 9:07um and we've you know been able to align
261. 9:08open ai's biggest models better than we
262. 9:10thought we'd we would at this point so
263. 9:11that's good
264. 9:13um we have some ideas about what to do
265. 9:15next
266. 9:16um but we cannot honestly like look
267. 9:17anyone in the eye and say we see at 100
268. 9:19years how we're going to solve this
269. 9:20problem
270. 9:21um but once the AI is good enough that
271. 9:23we can ask it to like hey can you help
272. 9:25us do alignment research
273. 9:26um I think that's going to be a new tool
274. 9:28in the toolbox yeah like for example one
275. 9:30of the conversations you and I had is
276. 9:32could we tell the uh the the agent don't
277. 9:38be racist right as opposed to trying to
278. 9:40figure out all the different things
279. 9:41where the weird correlative data that
280. 9:44exists on all the training settings
281. 9:46everything else may lead to racist
282. 9:48outcomes it could actually in fact do a
283. 9:51self-cleansing totally once the model
284. 9:53gets smart enough that you can that it
285. 9:55really understands what racism looks
286. 9:57like and how complex that is you can say
287. 9:58don't be racist yeah exactly
288. 10:01um
289. 10:02what do you think are
290. 10:05the kind of Moon shots
291. 10:08that in terms of evolution of the next
292. 10:11couple years
293. 10:13that people should be looking out for
294. 10:15in terms of like evolution of where AI
295. 10:17we'll go
296. 10:20um
297. 10:22I'll start with like the higher
298. 10:23certainty things I I think language
299. 10:26models are going to go just much much
300. 10:28further than people think
301. 10:30and we're like very excited to see what
302. 10:34happens there
303. 10:35um I think it's like what a lot of
304. 10:37people say about you know running out of
305. 10:38compute running out of data like that's
306. 10:40all true but I think there's so much
307. 10:42algorithmic progress to come that
308. 10:46that we're going to have like a very
309. 10:48exciting time
310. 10:50um another thing is I think we will get
311. 10:51true multimodal models working and so
312. 10:54you know not just text and images but
313. 10:56every modality you'd like in one model
314. 10:58able to easily like uh
315. 11:00you know fluidly move between things
316. 11:03um
317. 11:04I think we will have models that
318. 11:06continuously learn so like right now if
319. 11:08you use GPT whatever it's sort of like
320. 11:10stuck in time that it was trained and
321. 11:12the more you use it it doesn't get any
322. 11:13better and all of that I think we'll get
323. 11:15that changed so very excited about all
324. 11:18of that and if you just think about like
325. 11:21what that alone is going to unlock and
326. 11:23the sort of applications people will be
327. 11:25able to build with that
328. 11:26um that that that that that would be
329. 11:28like a huge victory for all of us and
330. 11:30just like a
331. 11:31like a massive step forward and a
332. 11:34genuine technological Revolution if that
333. 11:35were all that happened
334. 11:37um
335. 11:39but I think we're likely to keep making
336. 11:40research progress into new paradigms as
337. 11:43well um we've been like pleasantly
338. 11:44surprised on the upside about
339. 11:47what seems to be happening and I think
340. 11:49uh you know all these questions about
341. 11:52like new knowledge generation how do we
342. 11:53really Advance Humanity uh
343. 11:56I think there will be systems that can
344. 11:57help us with that
345. 11:59so one thing I think would be useful to
346. 12:01share because
347. 12:04uh folks don't realize that you're
348. 12:06actually
349. 12:07making these strong predictions from a
350. 12:09fairly critical point of view not just a
351. 12:11you know we can take that Hill
352. 12:13say a little bit about some of the areas
353. 12:15that you think are current kind of
354. 12:16illusionally talked about like for
355. 12:18example Ai and fusion oh yeah so I like
356. 12:22one of the unfortunate things that's
357. 12:24happened is uh
358. 12:26you know AI has become like the Mega
359. 12:29buzzword
360. 12:30um which is usually a really bad sign I
361. 12:32hope I hope it doesn't mean like the
362. 12:34field is about to fall apart
363. 12:35um but historically that's like a very
364. 12:37bad sign for you know new startup
365. 12:39creation or whatever if everybody is
366. 12:40like I'm this with AI and that's
367. 12:42definitely happening now
368. 12:43um so like a lot of the you know we were
369. 12:46talking about like are there all these
370. 12:46people saying like I'm doing like these
371. 12:48you know RL models for Fusion or
372. 12:50whatever and as far as we can tell
373. 12:51they're all like much worse than what
374. 12:53like you know smart physicists to figure
375. 12:56it out
376. 12:57um I think it is just an area where
377. 12:58people are going to say uh everything is
378. 13:01now this plus AI many things will be
379. 13:04true I do think this will be like the
380. 13:06biggest technological platform of the
381. 13:07Generation
382. 13:09Um but I think it's like
383. 13:12we like to make predictions where we can
384. 13:15be on the frontier
385. 13:16understand predictably what the scaling
386. 13:19laws look like or already have done the
387. 13:20research where we can say all right this
388. 13:21new thing is going to work and make
389. 13:23predictions out from that way and that's
390. 13:25sort of like how we try to run open AI
391. 13:27um which is you know do the next thing
392. 13:29in front of us when we have high
393. 13:30confidence and Kate take 10 of the
394. 13:32company to just totally go off and
395. 13:34explore
396. 13:35which has led to huge wins and there
397. 13:38will be wait like
398. 13:40oh I feel bad to say this like I I doubt
399. 13:43we'll still be using the Transformers in
400. 13:44five years I hope we're not I hope we
401. 13:45find something way better but the
402. 13:47transform has obviously been remarkable
403. 13:49so I think it's important to always look
404. 13:50for like you know where am I going to
405. 13:52find the next the sort of the next
406. 13:54totally new paradigm
407. 13:56um
408. 13:58and
409. 14:00but but I I think like that's the way to
410. 14:03make predictions don't don't pay
411. 14:04attention to the like AI for everything
412. 14:06like you know can I see something
413. 14:08working and can I see how it predictably
414. 14:09gets better and then of course leave
415. 14:11room open for like the you can't plan
416. 14:13the greatness but sometimes it had the
417. 14:15research breakthrough happens yep so I'm
418. 14:17going to uh ask two more questions and
419. 14:19then open it up because I want to make
420. 14:21sure that people have a chance to do
421. 14:23this uh the broader discussion although
422. 14:26I'm trying to
423. 14:27paint the broad picture so you can get
424. 14:29the crazy aspirations as part of this
425. 14:33what do you think
426. 14:35uh what do you think is going to happen
427. 14:37vis-a-vis the application of AI to like
428. 14:40these very important systems like for
429. 14:41example financial markets
430. 14:43um you know because the very natural
431. 14:45thing would be is saying well let's
432. 14:46let's do a high frequency Quant trading
433. 14:49system on top of this and other kinds of
434. 14:52things what what is it is it just kind
435. 14:54of being a neutral arms race is it is it
436. 14:57what how do how what's your thought in
437. 14:59like it's almost like the life 3.0 yeah
438. 15:02omega's point of view yeah um
439. 15:04I mean I think it is going to just seep
440. 15:06in everywhere
441. 15:08my basic model of the next decade is
442. 15:11that uh the cost of intelligence
443. 15:14the marginal cost of intelligence and
444. 15:15the marginal cost of energy are going to
445. 15:17Trend rapidly towards zero like
446. 15:19surprisingly far and and those I think
447. 15:22are two of the major inputs into the
448. 15:25cost of everything else except the cost
449. 15:26of things we want to be expensive the
450. 15:28status Goods whatever and and I think
451. 15:30you have to assume that's going to touch
452. 15:32almost everything
453. 15:34um because these like seismic shifts
454. 15:36that happen when like the whole cost
455. 15:39structure of society change which
456. 15:41happened many times before
457. 15:42um like the Temptation is always to
458. 15:44underestimate those uh
459. 15:46so I wouldn't like make a high
460. 15:48confidence prediction about anything
461. 15:50that doesn't change a lot or that where
462. 15:53that doesn't get to be applied
463. 15:55um but one of the things that is
464. 15:56important is it's not like the thing
465. 15:58Trends either Trends all the way to zero
466. 16:01they just Trend towards there and so
467. 16:02it's like
468. 16:04someone will still be willing to spend a
469. 16:06huge amount of money on compute and
470. 16:07energy they will just get like
471. 16:08unimaginable amount of intelligence
472. 16:10energy they'll just get unimaginable
473. 16:12amounts about that and so like who's
474. 16:13going to do that and where is it going
475. 16:15to get the weirdest not because the cost
476. 16:16comes way down but the amount spent
477. 16:17actually goes way up yes the
478. 16:19intersection of the two curves yeah you
479. 16:22know the thing got 10 or 100 thing got
480. 16:23100 times cheaper in the cost of energy
481. 16:25you know 100 million times cheaper in
482. 16:27the cost of intelligence and I was still
483. 16:28willing to spend a thousand times more
484. 16:30in today's dollars like
485. 16:33what happens then yep
486. 16:35and then uh last of the buzzword Bingo
487. 16:39part of the the future questions
488. 16:42metaverse and AI what do you what do you
489. 16:45see coming in this you know
490. 16:48I think they're like both independently
491. 16:49cool things it's not like totally clear
492. 16:51to me yeah other than like how AI will
493. 16:54impact all Computing yeah well obviously
494. 16:56Computing simulation environments Asians
495. 16:58possibly possibly entertainment
496. 17:01certainly education right
497. 17:04um you know like an AI tutor and so
498. 17:05forth those those would be Baseline but
499. 17:08the question is is there anything that's
500. 17:10occurred to you that's
501. 17:12I I would bet that
502. 17:15the metaverse turns out in the upside
503. 17:18case then which I think has a reasonable
504. 17:20chance of happening the upside case the
505. 17:21metaverse turns out to be more like
506. 17:23something on the order of the iPhone
507. 17:24like a new
508. 17:27a new container for software and you
509. 17:29know a new way a new computer
510. 17:31interaction thing and AI turns out to be
511. 17:33something on the order of like a
512. 17:34legitimate technological Revolution
513. 17:37um and so I think it's more like how the
514. 17:40metaverse is going to fit into this like
515. 17:42new world of AI then AI fit into the
516. 17:44metaverse but
517. 17:46low confidence the TBD all right
518. 17:49questions
519. 17:54hey there how do you see uh Technologies
520. 17:58uh foundational Technologies like tpg3
521. 18:01affecting
522. 18:02um the pace of life science research
523. 18:04specifically you can group in medical
524. 18:07research there and and sort of just
525. 18:09quickening the iteration cycles and then
526. 18:11what do you see as the rate limiter in
527. 18:14life science research and sort of where
528. 18:16we won't be able to get past because
529. 18:18they're just like laws of nature yeah
530. 18:20something like that
531. 18:22um
532. 18:23so I think the currently available
533. 18:26models are kind of not good enough to
534. 18:29have like made a big impact on the field
535. 18:30at least that's what like most like life
536. 18:32sciences researchers have told me
537. 18:34they've all looked at it and they're
538. 18:35like it's a little helpful in some cases
539. 18:37um there's been some promising work in
540. 18:39genomics but like stuff on a bench top
541. 18:41hasn't really impacted it I think that's
542. 18:43going to change and I think uh this is
543. 18:46one of these areas where there will be
544. 18:47these like you know new 100 billion to
545. 18:49trillion dollar companies started those
546. 18:51those areas are rare but like when you
547. 18:53can really change the way that if you
548. 18:55can really make like a you know future
549. 18:58Pharma company that is just hundreds of
550. 19:00times better than what's out there today
551. 19:01that's going to be really different
552. 19:04um as you mentioned there still will be
553. 19:05like the rate limit of like bio has to
554. 19:07run at its own thing and human Trials
555. 19:09take however long they take and that's
556. 19:10so I think an interesting cut of this is
557. 19:12like where can you avoid that like where
558. 19:14are the the synthetic bio companies that
559. 19:16I've seen that have been most
560. 19:17interesting are the ones that find a way
561. 19:19to like make the cycle time super fast
562. 19:22um and that that benefits like an AI
563. 19:24That's giving you a lot of good ideas
564. 19:26but you've still got to test them which
565. 19:27is where things are right now
566. 19:29um I'm a huge believer for startups that
567. 19:31like the thing you want is low costs and
568. 19:33fast cycle times and if you have those
569. 19:35you can then compete as a startup
570. 19:36against the big incumbents uh and so
571. 19:39like I wouldn't go pick like cardiac
572. 19:41disease is my first thing to go after
573. 19:42right now with like this kind of new
574. 19:44kind of company
575. 19:45um but you know using bio to manufacture
576. 19:47something that sounds great uh I think
577. 19:50the other thing is the simulators are
578. 19:52still so bad
579. 19:54and if I were an a if I were a bio means
580. 19:57AI startup I would certainly try to work
581. 19:59on that somehow
582. 20:00when do you think the AI Tech will help
583. 20:04create itself oh it's almost like a
584. 20:06self-improvement will help make the
585. 20:08simulators significantly better
586. 20:11um
587. 20:11people are working on that now uh I I
588. 20:14don't know quite how it's going but you
589. 20:16know there's very smart people are very
590. 20:18optimistic about that yeah
591. 20:20other questions
592. 20:23and I can keep going on questions I just
593. 20:24want to make sure you guys had a chance
594. 20:26this
595. 20:28uh here yes great Mike is coming
596. 20:32awesome thank you
597. 20:34um I was curious what what aspects of
598. 20:37Life do you think won't be changed by AI
599. 20:43um
600. 20:44sort of did all of the deep biological
601. 20:46things like I think we will still really
602. 20:49care about interaction with other people
603. 20:50like we'll still have fun and like the
604. 20:53reward you know systems of our brain are
605. 20:55still going to work the same way like
606. 20:57we're still going to have the same like
607. 20:59drives to kind of create new things and
608. 21:01you know compete for silly status and
609. 21:04like you know form families and whatever
610. 21:07um so I think
611. 21:09the the stuff that people cared about
612. 21:1350 000 years ago is more likely to be
613. 21:15the stuff that people care about you
614. 21:17know 100 years from now than 100 years
615. 21:18ago
616. 21:20as an amplifier on that before we get to
617. 21:23the next whatever the next question is
618. 21:24what do you think are the best utopian
619. 21:27science fiction universes so far
620. 21:31good question
621. 21:32um
622. 21:35Star Trek is pretty good honestly
623. 21:38uh like I do like all of the ones that
624. 21:41are sort of like
625. 21:42you know we turn our Focus to like
626. 21:44exploring
627. 21:46and understanding the universe as much
628. 21:48as we can
629. 21:49um
630. 21:52it's not this is not a utopian one well
631. 21:54maybe I think the last question is like
632. 21:56an incredible short story uh-huh yeah
633. 21:58that was what that came to mind yep uh I
634. 22:01was expecting you to say Ian Banks on
635. 22:03the culture those are great uh
636. 22:08I think science fiction is like there's
637. 22:11not like one
638. 22:13there's not like one sci-fi universe
639. 22:15that I could point to and say I think
640. 22:16all of this is great but like the
641. 22:18collective optimistic corner of sci-fi
642. 22:21which is like a smallish Corner
643. 22:25um I'm excited about actually uh I took
644. 22:28a few days off to write a Sci-Fi story
645. 22:30and I had so much fun doing it just
646. 22:32about sort of like the optimistic case
647. 22:33of AGI
648. 22:35um
649. 22:35that it made me want to go like read a
650. 22:37bunch more so I'm looking for
651. 22:39recommendations of more to read now
652. 22:41um like the sort of less known stuff if
653. 22:43you have anything I will I will get you
654. 22:45some great some recommendations
655. 22:49so in a similar vein one of my favorite
656. 22:51sci-fi books is called childhood's End
657. 22:53by Arthur Clark from like the 60s I
658. 22:56think and the I guess the one sentence
659. 23:01summary is aliens come to the Earth try
660. 23:04to save us and they just take our kids
661. 23:06and leave everything else
662. 23:07so you know there's a slightly more
663. 23:10optimistic than that but yes I mean
664. 23:13there's Ascension into the over mind is
665. 23:15is is meant to be more utopian but yes
666. 23:18okay uh you may not read it that way but
667. 23:20yes well also in our current Universe
668. 23:23yes our current situation
669. 23:26um you know a lot of people think about
670. 23:27family building and fertility and like
671. 23:30some of us have
672. 23:33different people have different ways of
673. 23:35approaching this but from where you
674. 23:36stand what do you see as like the most
675. 23:40promising Solutions it might not be a
676. 23:42technological solution but I'm curious
677. 23:43what you think
678. 23:45other than everyone having 10 kids you
679. 23:47know like how do we of everyone having
680. 23:4910 kids yeah how do you populate how do
681. 23:52you like how do you see family building
682. 23:54coexisting with you know AGI
683. 23:58high tech it's this is like a question
684. 24:00that comes up at open AI a lot like how
685. 24:02do I think about you know how should one
686. 24:04think about having kids there's I think
687. 24:06no consensus answer to this
688. 24:08um there are people who say yeah I'm not
689. 24:11I was gonna I thought I always thought I
690. 24:12was gonna have kids and now I'm not
691. 24:13going to because of AGI like there's
692. 24:15just for all the obvious reasons and I
693. 24:18think some less obvious ones there's
694. 24:19people who say like well it's going to
695. 24:21be the only thing for me to do in you
696. 24:22know 15 20 years so of course I'm going
697. 24:24to have a big family like that's what
698. 24:25I'm going to spend my time doing you
699. 24:27know I'll just like raise great kids and
700. 24:29then I think that's what'll bring me
701. 24:30fulfillment I think like as always it is
702. 24:32a a personal decision I get very
703. 24:34depressed when people are like I'm not
704. 24:35having kids because of AGI
705. 24:37uh the EA Community is like I'm not
706. 24:39doing that because they're all going to
707. 24:40die they're kind of like a techno
708. 24:42optimists are like well it's just like
709. 24:44you know I want to like merge into the
710. 24:46AGI and go off exploring the universe
711. 24:47and it's going to be so wonderful and
712. 24:49you know just I want total freedom but I
713. 24:51think like all of those I find quite
714. 24:52depressing
715. 24:53um
716. 24:54I think having a lot of kids is great
717. 24:56I you know want to do that now more than
718. 24:59I did even more than I did when I was
719. 25:01younger and I I'm excited for it
720. 25:07what do you think will be the way that
721. 25:09most users interact with Foundation
722. 25:11models in five years do you think
723. 25:12there'll be a number of verticalized AI
724. 25:14startups that essentially have adapted
725. 25:15and fine-tuned Foundation models to an
726. 25:18industry or do you think prompt
727. 25:19engineering will be something many
728. 25:20organizations have as an in-house
729. 25:22function I don't think we'll still be
730. 25:24doing prompt Engineering in five years I
731. 25:25think it'll just be like you and this
732. 25:28will be integrated everywhere but you
733. 25:30will just like you know either with text
734. 25:32or voice depending on the context you
735. 25:35you will just like interface in language
736. 25:37and get the computer to do whatever you
737. 25:39want and uh that will you know apply to
738. 25:43like generate an image where maybe we
739. 25:44still do a little bit of prompt
740. 25:45engineering but you know it's kind of
741. 25:47just going to get it to like go off and
742. 25:49do this research for me and do this
743. 25:51complicated thing or just like you know
744. 25:52be my therapist and help me figure out
745. 25:54how to make my life better or like you
746. 25:56know go use my computer for me and do
747. 25:58this thing or or any number of other
748. 26:00things but I think the fundamental
749. 26:01interface will be natural language
750. 26:04let me actually push on that a little
751. 26:06bit before we get to the next question
752. 26:07which is
753. 26:08I mean to some degree just like we have
754. 26:10a wide range of human talents right now
755. 26:12uh and taking a look for example a dolly
756. 26:15when you have like a a great visual
757. 26:18thinker they can get a lot more out of
758. 26:20Dolly because they know how to think
759. 26:21more they know how to iterate the loop
760. 26:23through the the test don't you think
761. 26:25that will be a general truth about most
762. 26:27of these things so it isn't that why
763. 26:30would be natural language is the way
764. 26:31you're doing it it will be there will be
765. 26:34like almost an evolving set of human
766. 26:36talents about about going that extra
767. 26:39mile 100 I just hope it's not like
768. 26:42figuring out to like hack the prompt by
769. 26:43adding one magic word to the end that
770. 26:45like changes everything else I I like
771. 26:47what will matter is like the quality of
772. 26:49ideas and the understanding of what you
773. 26:51want so the artist will still do the
774. 26:52best with image generation but not
775. 26:54because they figured out to like add
776. 26:56this one magic word at the end of it
777. 26:58because they were just able to like
778. 26:59articulate it with a creative eye that
779. 27:01you know I don't have certainly what
780. 27:03they have is a vision and kind of how
781. 27:05their visual thinking and iterating
782. 27:06through it yeah yeah well obviously
783. 27:08it'll be that word or prompt now but
784. 27:10it'll iterate to to better
785. 27:12all right uh at least we have a question
786. 27:15here
787. 27:19hey thanks so much um
788. 27:22uh I think the term AGI is used uh
789. 27:25thrown around a lot and um
790. 27:28sometimes I've noticed my own
791. 27:29discussions like the sources of
792. 27:31confusion has just come from people
793. 27:32having different definitions of AGI and
794. 27:34so it can kind of be the magic box where
795. 27:36everyone just kind of projects their
796. 27:38their ideas onto it and I just want to
797. 27:39get a sense from you what like how do
798. 27:41you think you know how would you define
799. 27:43AGI and how do you think you'll know
800. 27:45yeah
801. 27:46it's a great point I think there's like
802. 27:48a lot of valid definitions to this but
803. 27:50uh for me
804. 27:52um AGI is basically the equivalent of a
805. 27:56median human that you could like you
806. 27:59know hire as a co-worker
807. 28:01um so and then they could like say do
808. 28:03anything that you'd be happy with a
809. 28:04remote co-worker doing like just behind
810. 28:06a computer which includes like you know
811. 28:10learning how to go be a doctor learn how
812. 28:12to go be a very competent coder like
813. 28:14there's a lot of stuff that a media
814. 28:15human is capable of getting good at and
815. 28:17I think one of the skills of an AGI is
816. 28:19not any particular Milestone but the The
817. 28:21Meta skill of learning to figure things
818. 28:22out and that it can go decide to get
819. 28:24good at whatever you need
820. 28:26um so for me like that's that's kind of
821. 28:30like AGI and then Super intelligence is
822. 28:32when it's like smarter than all of
823. 28:34humanity put together
824. 28:36thanks
825. 28:38um just uh what would you say or in the
826. 28:41next 20 30 years are some of the main
827. 28:45societal issues that will arise as AI
828. 28:48continues to grow and what can we do
829. 28:49today to mitigate those issues
830. 28:53obviously the economic impacts are huge
831. 28:56and I think it's just like if it if it
832. 28:58is as
833. 28:59Divergent as I think it could be for
834. 29:01like some people doing incredibly well
835. 29:04and others not I think Society just
836. 29:06won't tolerate at this time and so
837. 29:08figuring out when we're gonna like
838. 29:10disrupt so much of economic activity and
839. 29:13even if it's not all disrupted by 20 or
840. 29:1530 years from now I think it'll be clear
841. 29:17that it's all going to be
842. 29:19um what like what is the new social
843. 29:22contract like how do my guess is that
844. 29:25the things that we'll have to figure out
845. 29:27are how we think about fairly
846. 29:29Distributing wealth
847. 29:30um access to AGI systems which will be
848. 29:33like kind of the commodity of the realm
849. 29:35and governance like how we collectively
850. 29:37decide what they can do what they don't
851. 29:39do things like that
852. 29:41um
853. 29:43and I think figuring out the answer to
854. 29:45those questions is is gonna just be huge
855. 29:47I I'm optimistic that people will figure
856. 29:50out how to spend their time and be very
857. 29:51fulfilled I think people worry about
858. 29:54that in a little bit of a silly way I'm
859. 29:56sure what people do will be very
860. 29:57different but we always solve this
861. 29:58problem
862. 29:59um
863. 30:01but I do think like the concept of
864. 30:03wealth and access and governance
865. 30:05those are all going to change and how
866. 30:08how we address those will will be huge
867. 30:11actually one thing I don't know what
868. 30:13level of
869. 30:15devs you can share that but one of the
870. 30:16things I love about what openai and you
871. 30:19guys are doing is when you they think
872. 30:21about these questions a lot themselves
873. 30:24and they initiate some research so
874. 30:26you've initiated some research on this
875. 30:28stuff yeah so we run the largest uh Ubi
876. 30:30experiment in the world I don't think
877. 30:32that is uh we have a year and a half a
878. 30:36year and a quarter left in a five-year
879. 30:37project I don't think that's like the
880. 30:39only solution but I think it's a great
881. 30:40thing to to be doing
882. 30:43um and
883. 30:45you know I think like we should have
884. 30:47like 10 more things like that that we
885. 30:48try
886. 30:49um we also try with different ways to
887. 30:51get sort of input from a lot of the
888. 30:52groups that we think will be most
889. 30:53affected and see how we can do that
890. 30:55early in the cycle
891. 30:57um we've explored more recently like how
892. 30:59this technology can be used used for
893. 31:01reskilling people that are going to be
894. 31:02impacted early
895. 31:04um we'll try to do a lot more stuff like
896. 31:06that too yeah so they are the the
897. 31:08organization is actually in fact
898. 31:10uh these are great questions
899. 31:13addressing them and actually doing a
900. 31:14bunch of interesting research on it so
901. 31:15next question
902. 31:17hi so um creativity came up today in
903. 31:20several of the panels you know and um it
904. 31:23seems to me that the way it's being used
905. 31:24like you you have tools for human
906. 31:26creators to go and expand human
907. 31:28creativity so where do you think the
908. 31:29line is between these tools to to allow
909. 31:33a Creator to be more productive in
910. 31:35artificial creativity itself so
911. 31:39um I I think and I think we're seeing
912. 31:41this now that tools for creatives that
913. 31:43that is going to be like the great
914. 31:45application of AI in the short term
915. 31:48um people love it it's really helpful uh
916. 31:51and I think it is at least in what we're
917. 31:53seeing so far
918. 31:54um not replacing it is mostly enhancing
919. 31:57it's replacing in some cases uh but for
920. 32:00the majority of like the kind of work
921. 32:02that people in these fields want to be
922. 32:03doing it's enhancing and I think we'll
923. 32:05see that Trend continue for a long time
924. 32:09um eventually yeah it probably is just
925. 32:11like you know we look at 100 years okay
926. 32:12it can do the whole creative job
927. 32:15um I think it's interesting that if you
928. 32:17asked people 10 years ago uh about Holly
929. 32:20I was going to have an impact with a lot
930. 32:22of confidence from almost most people
931. 32:24you would have heard you know first it's
932. 32:26going to come for the blue collar jobs
933. 32:27working in the factories truck drivers
934. 32:29whatever then it will come for the kind
935. 32:31of like the low skill White Collar jobs
936. 32:33then the very high skill like really
937. 32:36high IQ uh white-collar jobs like a
938. 32:39programmer or whatever and then very
939. 32:40last of all and maybe never it's gonna
940. 32:42take the creative jobs and it's really
941. 32:44gone exactly and it's going exactly the
942. 32:47other direction and I think this like
943. 32:49isn't there's an interesting reminder in
944. 32:51here generally about how hard
945. 32:52predictions are but more specifically
946. 32:54about
947. 32:55you know we're not always very aware
948. 32:57maybe even ourselves of like what skills
949. 32:59are hard and easy like what uses most of
950. 33:02our brain and what doesn't or how like
951. 33:03difficult bodies are to control or make
952. 33:05or whatever
953. 33:07we have one more question over here
954. 33:10hey thanks for being here so you
955. 33:12mentioned that um you will be skeptical
956. 33:14of any startup trying to train their own
957. 33:16language model and it would love to
958. 33:18understand more so what I have heard and
959. 33:20which might be wrong is that large
960. 33:22language models depend on data and
961. 33:24compute and any startup can access to
962. 33:26the same amount of data because it's
963. 33:27just like internet data and compute like
964. 33:29different companies might have different
965. 33:31compute but I guess I see a big players
966. 33:32can sell more compute so how good a
967. 33:35large language model startup
968. 33:37differentiate from another
969. 33:39how would the startup differentiate from
970. 33:41another how would one large language
971. 33:43model startup differentiate I think
972. 33:45it'll be this middle layer um I think in
973. 33:48some sense the startups will train their
974. 33:49own models just not from the beginning
975. 33:51uh they will take like
976. 33:53you know base models that are are like
977. 33:56hugely trained with a gigantic amount of
978. 33:59compute and data and then they will
979. 34:01train on top of those to create you know
980. 34:03the model for each vertical
981. 34:05and and that those startups so in some
982. 34:08sense they are training their own models
983. 34:09just not not from scratch but they're
984. 34:11doing the one percent of training that
985. 34:13really matters for for whatever this use
986. 34:15case is going to be those startups I
987. 34:17think they will be hugely successful and
988. 34:18very differentiated startups there but
989. 34:20that'll be about the kind of like data
990. 34:22flywheel that the startup is able to do
991. 34:24the kind of like all of the pieces on
992. 34:26top of and Below uh like this could
993. 34:29include prompt engineering for a while
994. 34:30or whatever the sort of the kind of like
995. 34:33core
996. 34:34base model
997. 34:35I think that's just going to get too too
998. 34:38complex and too expensive and the world
999. 34:41also just doesn't make enough chips
1000. 34:43so Sam has a work thing he needs to get
1001. 34:46to so and as you probably can tell with
1002. 34:50a very far far ranging thing Sam always
1003. 34:53expands my uh boundaries and a little
1004. 34:56bit unlike the that when you're feeling
1005. 34:58depressed whether it's kids in a house
1006. 35:00you're the person I always turn to
1007. 35:01probably I appreciate that yes so anyway
1008. 35:04I think I think like no one knows
1009. 35:06like we're sitting on this like
1010. 35:08precipice of AI and like people like
1011. 35:10it's either gonna be like really great
1012. 35:11or really terrible
1013. 35:13um
1014. 35:15you may as well like you gotta you gotta
1015. 35:18like plan for the worst you certainly
1016. 35:19like it's not a strategy to say it's all
1017. 35:20going to be okay but you may as well
1018. 35:22like emotionally feel like we're going
1019. 35:24to get to the Great future and we'll
1020. 35:25play as hard as you can to get there and
1021. 35:26play for it yes rather than like act
1022. 35:28from this place of like fear and despair
1023. 35:30all the time because if we acted from a
1024. 35:33place of fear and paranoia we would not
1025. 35:35be where we are today so let's thank Sam
1026. 35:38for spending dinner with us thank you