Confronting Our AI Future · Closing Keynote — transcript（原文存档）

项	值
形态	视频 / 大会闭幕主题演讲 + 炉边对谈 · 自动转录字幕（逐行带 MM:SS 时间码）
活动	Stanford HAI「Confronting Our AI Future: Hope, Fear, and the Choices Ahead」，2026-10-28，Hoover Institution, David & Joan Traitel Building, Stanford
议程位置	4:00–4:15 PM PST「Discovery Loop: Accelerating Science and Engineering」（Jeff Dean）；4:15–4:45 PM PST「Closing Fireside with Jeff Dean and Fei-Fei Li」
主讲 / 对谈	Jeff Dean（Discovery Loop 联合创始人兼 CEO，前 Google 首席科学家）/ 李飞飞 Fei-Fei Li（Stanford HAI 创始主任、Sequoia 讲席教授、World Labs 联合创始人兼 CEO）
主持开场	Stanford HAI 主持人（未具名）
时长	约 48 分钟（本份字幕末帧 48:14）
侃姐发来	2026-11-09 20:54，Downloads 里的 txt（59515 B / 3561 行）
转录性质	自动生成。含口吃、重复词、断句破碎、专名误识（fee Lee / John Levven / Oral Vignyals / Asimoff 类）、漏标的说话人切换；时间码有 05:60、07:60、14:60 等溢出写法，原样保留
来源	活动页 https://hai.stanford.edu/events/confronting-our-ai-future-hope-fear-and-the-choices-ahead；视频 https://www.youtube.com/watch?v=ooGKNwJWdb8（经同期中文报道与第三方摘要页交叉核对，未逐帧校验）
状态	正文一字未动、未校对。专名校正与转录错误清单见 ../notes/2026-11-09-confronting-our-ai-future-jeff-dean-feifei-li-zh.md
定位方式	本份有时间码，译文用 ¶ 段落编号 + tMM:SS 锚点；不补写、不外推时间码

---

00:00 All right, we've spent the day on hard 00:03

00:03 questions about AI. Who govern it 00:05

00:05 governs it? How fast it's moving and how 00:08

00:08 we design it around people. So for our 00:11

00:11 closing keynote, we'll turn to one of 00:13

00:13 the most hopeful promises, using AI to 00:16

00:16 accelerate scientific discovery itself. 00:19

00:19 Jeff Dean is co-founder and CEO of 00:21

00:21 Discovery Loop, a new company focused on 00:24

00:24 AI for science and engineering. And 00:26

00:26 before that he spent 27 years at Google 00:29

00:29 where he helped build systems like rapid 00:32

00:32 map reduce big table and create 00:34

00:34 tensorflow and served as chief science 00:37

00:37 scientist and co-lead of gemini. His 00:40

00:40 talk is titled discovery loop 00:42

00:42 accelerating science and engineering 00:45

00:45 and afterwards he'll be joined on stage 00:47

00:47 by professor fee Lee for a closing 00:49

00:49 conversation. FFE is the founding 00:52

00:52 director and senior fellow at Stanford 00:54

00:54 HAI, the inaugural Sequoia Professor of 00:57

00:57 Computer Science here at Stanford and 00:59

00:59 co-founder and CEO of World Labs. She 01:03

01:03 recently took on the role as special 01:04

01:04 adviser on AI to President John Levven 01:07

01:07 and as the co-chair of HAI's advisory 01:10

01:10 council. HAI would not be here today if 01:12

01:12 it weren't for her vision and commitment 01:14

01:14 to putting people at the center of AI. 01:17

01:17 And that founding vision remains at the 01:19

01:19 heart of what we do here at the 01:21

01:21 institute. So now please join me in 01:24

01:24 welcoming Jeff Dean. 01:26

01:26 [applause] 01:32

01:32 >> All right. Wonderful. Uh it's delightful 01:35

01:35 to be here uh with all of you today. Um 01:38

01:38 I'm going to talk to you about discovery 01:39

01:39 loop. Our mission is that we are trying 01:42

01:42 to build AI systems that can 01:44

01:44 automatically solve important problems 01:45

01:45 in machine learning, science, and 01:47

01:47 engineering. And we think this will 01:49

01:49 ultimately lead to more discoveries in 01:51

01:51 the world. And that's uh what we're 01:53

01:53 excited about. 01:55

01:55 So there's this common cycle across many 01:58

01:58 many disciplines of kind of the 02:00

02:00 universal loop of science and 02:02

02:02 engineering where you propose 02:03

02:03 experiments, you implement and run those 02:05

02:05 experiments, you look at the results, 02:07

02:07 you examine whatever you measurements 02:09

02:09 you can make about them and you decide 02:11

02:11 what to do next which often means doing 02:13

02:13 another experiment. 02:16

02:16 So our approach is we want to automate 02:18

02:18 these kinds of experimental loops. Um, 02:20

02:20 we think if you can drive down iteration 02:22

02:22 time, the time to run one time time 02:25

02:25 through these loops and you can propose 02:27

02:27 and automatically run many experiments, 02:29

02:29 you can learn from each evaluation so 02:31

02:31 that the subsequent experiments you 02:33

02:33 propose are better and are informed by 02:35

02:35 the previous experiments you've run and 02:37

02:37 scale both the quality and the volume of 02:40

02:40 experiments, you're going to get 02:41

02:41 breakthrough results. So that's 02:44

02:44 in essence what we're out to do. Um, if 02:48

02:48 you squint at this nice list of US 02:51

02:51 National of Academy of Engineering Grand 02:54

02:54 Challenge problems, these were problems 02:57

02:57 uh proposed as really important for us 02:59

02:59 to solve in the 21st century. And we're 03:01

03:01 a quarter of the way through roughly. 03:03

03:03 Um, if you squint at kind of the sub 03:06

03:06 problems that you can imagine need being 03:08

03:08 needed to be solved for some of these 03:10

03:10 areas, in particular the ones in red. We 03:12

03:12 think these kinds of automated discovery 03:14

03:14 loops can actually assist and help with 03:16

03:16 all of these different kinds of areas. 03:18

03:18 [snorts] 03:22

03:22 Now, one of the things that's important 03:23

03:23 to us and partly why we formed this as a 03:25

03:25 separate company is we believe 03:27

03:27 technology can have an amazingly 03:29

03:29 positive impact in the world if we do it 03:31

03:31 right. but that we all have to work to 03:34

03:34 make sure that it has this positive 03:36

03:36 impact. Um, and so we established our 03:39

03:39 company as a public benefit corporation 03:41

03:41 and our public mission is to develop 03:43

03:43 technology that addresses important 03:45

03:45 societal problems and to provide broad 03:47

03:47 access to its benefits to improve 03:50

03:50 people's lives. 03:53

03:53 So a little bit about the four of us, we 03:55

03:55 h I have uh three amazing co-founders 03:57

03:57 Sanjay Gimat Quacle and Oral Vignyals. 04:00

04:00 uh we've worked together for depending 04:02

04:02 on the pair-wise relation uh somewhere 04:04

04:04 between 14 and 30 years. Uh we really 04:07

04:07 enjoy working together on difficult 04:08

04:08 problems and we've got a lot of 04:11

04:11 different kinds of experience. So we've 04:13

04:13 helped build uh Google products. We all 04:15

04:15 came from Google. So these are all 04:17

04:17 Google centric things we've worked on. 04:19

04:19 Um lots of different products. You know 04:22

04:22 we've worked together on many different 04:23

04:23 generations. Google search, the initial 04:25

04:25 ad system, content ads, Gmail news 04:28

04:28 translate, um infrastructure systems in 04:31

04:31 the second column, uh in particular, uh 04:34

04:34 focus initially at the top part of that 04:37

04:37 column are things that help you uh uh 04:40

04:40 organize, manage large amounts of data, 04:42

04:42 run large scale computations over data 04:44

04:44 to extract interesting insights. The 04:47

04:47 bottom four things there are really 04:49

04:49 multiple generations of largecale 04:51

04:51 machine learning infrastructure to 04:53

04:53 enable training more and more capable 04:55

04:55 models and TPUs are machine learning 04:58

04:58 focused hardware that uh we were 04:59

04:59 involved in. Um then we've worked on a 05:02

05:02 bunch of different AI research problems 05:04

05:04 uh including some things that are kind 05:06

05:06 of at the core of lots of different uh 05:08

05:08 sort of uh highly capable models today. 05:11

05:11 Things like model distillation or 05:13

05:13 mixture of experts models. uh neural 05:15

05:15 architecture search is kind of a early 05:17

05:17 version of these kind of automated loops 05:19

05:19 that are trying to identify uh capable 05:22

05:22 um you know uh neural model 05:24

05:24 architectures for solving problems and 05:26

05:26 then applications of AI to different uh 05:29

05:29 domains including some in science. So I 05:32

05:32 was uh Quack and I were both involved in 05:34

05:34 alpha chip which is reinforcement 05:35

05:35 learning for chip design and uh 05:37

05:37 placement and routing and chip design. 05:39

05:39 Uh Oral was involved in alpha fold um 05:42

05:42 and neural weather forecasting. So good 05:45

05:45 mix of experience. Uh we've also managed 05:48

05:48 pretty large teams. So I managed uh the 05:50

05:50 Google rating team up to 600 people and 05:52

05:52 the uh Google research and AI teams up 05:55

05:55 to about 4,400 people before I decided I 05:57

05:57 was spending a lot of time in meetings 05:60

05:60 about Q2 OPEX spending. And I pressed 06:02

06:02 send on an email that took me from 4,400 06:05

06:05 to four reports. Uh and then Jeff Hinton 06:08

06:08 retired the next week. So then I had 06:10

06:10 three. Um and uh and then I was uh Oriel 06:15

06:15 and I were both uh then co-technical 06:17

06:17 leads of the Gemini project and Quac was 06:20

06:20 our post training lead. We're we're also 06:22

06:22 quite proud of many alumni of our teams 06:24

06:24 who've gone off and done lots of amazing 06:26

06:26 things including founders of companies 06:28

06:28 like Anthropic OpenAI uh physical 06:31

06:31 intelligence uh which is probably known 06:32

06:32 here in Stanford uh various other 06:35

06:35 things. Um, 06:37

06:37 and we think we have a nice blend that 06:40

06:40 mixes both AI research skills and 06:43

06:43 distributed system uh skills. So, uh, 06:46

06:46 these are the most cited AI researchers 06:48

06:48 and distributed systems researchers per 06:49

06:49 Google Scholar. Um, and there's uh us in 06:53

06:53 the red boxes. Also, the people with 06:56

06:56 stars have all worked in our teams, 06:57

06:57 which is nice. 07:01

07:01 We also have fun together. So, we've 07:04

07:04 traveled in many different continents. 07:05

07:05 Uh, there we are in London in the upper 07:08

07:08 right, Patagonia in the upper left, uh, 07:10

07:10 Vietnam in the bottom left, Quac and I 07:12

07:12 are coming home from dinner on Uberlike 07:15

07:15 moped services, and California. 07:19

07:19 Um, so, one of the things we've been 07:21

07:21 focused a lot on is how do we assemble 07:24

07:24 an amazing founding team? And so that's 07:27

07:27 what we've been up to for kind of the 07:29

07:29 first um you know uh period of existence 07:32

07:32 of the company is hiring amazing people. 07:34

07:34 Um and they've come from lots of 07:37

07:37 different places. You see all the logos. 07:39

07:39 Um many of the people on the team have 07:41

07:41 experience founding various companies or 07:43

07:43 various important projects. Uh we have a 07:45

07:45 lot of great team experience in you know 07:48

07:48 leading the Go programming language 07:51

07:51 project. uh you know uh key contributors 07:54

07:54 to lots of different models Gemini Deepc 07:56

07:56 Kimmy K3 uh the co-author of the Kimmy 07:60

07:60 K3 paper and author of the tension 08:02

08:02 residuals work joined us um recently we 08:05

08:05 have a great mix of kind of very 08:07

08:07 experienced people and people who are uh 08:10

08:10 rising amazing uh stars earlier in their 08:13

08:13 career um and roughly 10% of the company 08:16

08:16 has the ACM prize in computing and 10% 08:19

08:19 have uh III olymp medals. So, we're 08:23

08:23 we're pretty pretty excited about the 08:25

08:25 team we've hired. We're having we've 08:26

08:26 also uh really focused on hiring for um 08:30

08:30 kind of amazing uh ability to work 08:33

08:33 together, which was uh great fun. Um so, 08:37

08:37 why now? 08:39

08:39 You know, we now have these really 08:41

08:41 capable frontier models that can do lots 08:43

08:43 of amazing things. Um and partly that 08:46

08:46 came about because there's been 08:48

08:48 innovation over the last say 10 to 15 08:50

08:50 years in the AI space of you know lots 08:54

08:54 of different kinds of things. The 08:55

08:55 hardware layer has improved tremendously 08:57

08:57 in performance both TPUs and GPUs and 08:60

08:60 sort of more specialized things. 09:02

09:02 Software abstractions have gotten better 09:03

09:03 so that people can implement things they 09:05

09:05 care about more e easily and readily and 09:08

09:08 it can be mapped onto the hardware more 09:09

09:09 efficiently. Um model architectures have 09:12

09:12 improved. you know uh word tovec and 09:14

09:14 then sequence to sequence and then 09:16

09:16 transformers and have all been important 09:19

09:19 innovations in what kind of models we 09:21

09:21 should train. Uh training algorithms 09:23

09:23 have gotten a lot better. Uh things like 09:25

09:25 distillation enable very um high quality 09:28

09:28 but smaller models and inference 09:31

09:31 algorithms have actually improved quite 09:33

09:33 a lot. So you can actually apply more 09:34

09:34 compute at inference time and get better 09:37

09:37 results with things like chain of 09:39

09:39 thought or uh you know inference time 09:41

09:41 compute scaling where you try many ways 09:43

09:43 of solving the problem and then kind of 09:44

09:44 examine uh what's happened. 09:48

09:48 Um but a great base model isn't 09:50

09:50 sufficient for solving some of the 09:51

09:51 problems that we want to solve. So in 09:54

09:54 particular, we think you start with 09:55

09:55 great base models, but then you want 09:58

09:58 agents plus models to be able to really 10:01

10:01 effectively use tools to accomplish 10:03

10:03 really complicated tasks. We want the 10:05

10:05 agents and models to be especially good 10:07

10:07 at creativity. Like coming up with 10:09

10:09 amazing scientific or engineering 10:11

10:11 breakthroughs is not just kind of uh 10:14

10:14 doing things that have already been done 10:16

10:16 in a lot of cases. It's actually putting 10:18

10:18 together lots of insights, coming up 10:20

10:20 with something genuinely new and seeing 10:22

10:22 how well that works. And we also think 10:26

10:26 agents and models should learn from the 10:28

10:28 previous experiments and previous 10:29

10:29 experience. Both in trying to solve this 10:32

10:32 particular problem, but also over time 10:34

10:34 you will see systems that maybe have 10:36

10:36 solved 5,000 problems. and the 5,000 10:39

10:39 first one they can learn from the 10:41

10:41 problems that are sort of most closely 10:43

10:43 related to that particular new problem 10:45

10:45 they're trying to solve. 10:48

10:48 So at this traditional experimental loop 10:50

10:50 you often have a lot of manual effort at 10:52

10:52 each stage. Um and when you try to solve 10:55

10:55 a larger problem uh that becomes even 10:57

10:57 more complicated because you have to 10:60

10:60 typically decompose those larger 11:02

11:02 problems into a bunch of subpros or 11:03

11:03 subpieces. you end up with actually a 11:07

11:07 difficult decomposition problem in and 11:09

11:09 of itself. That's the left hand side. 11:11

11:11 And then each one of these subpros or 11:13

11:13 sub areas has lots of people running 11:16

11:16 experiments, figuring out what worked, 11:18

11:18 what didn't work, trying different 11:19

11:19 approaches. Um, you end up with a lot of 11:21

11:21 coordination overhead. Uh, and resource 11:24

11:24 conflicts can occur, large system 11:26

11:26 effects. So this decomposition that you 11:28

11:28 have to do is actually quite hard. Um, 11:30

11:30 and it's hard to combine the results 11:32

11:32 across the sub problems you set out. And 11:35

11:35 perhaps even worse, the way you've 11:37

11:37 decided to decompose things in a very 11:38

11:38 large scale project is actually pretty 11:41

11:41 hard to change once you've kind of got 11:43

11:43 the project up and running and going. 11:46

11:46 And so you can't really revisit those 11:48

11:48 early decisions very effectively when 11:50

11:50 you have a uh, you know, a large scale 11:52

11:52 engineering effort like building an 11:54

11:54 airplane or training the best frontier 11:57

11:57 model or whatever. Um so what we think 12:00

12:00 is possible is you can take a relatively 12:04

12:04 small number of domain experts use 12:07

12:07 automated systems that can ideally 12:10

12:10 automatically decompose into interesting 12:12

12:12 subpros and then automatically solve 12:14

12:14 those sub problems. But uh it may be 12:18

12:18 that human guidance helps the 12:20

12:20 decomposition and then humans sort of 12:23

12:23 direct what are the important subpros to 12:25

12:25 solve and really enable a small number 12:29

12:29 of people with domain expertise to 12:31

12:31 really tackle large impactful problems. 12:34

12:34 And we think this is possible if you 12:36

12:36 build the right infrastructure. You know 12:39

12:39 you want to have tens of thousands of 12:41

12:41 automated experiments. You want a system 12:43

12:43 that can actually handle that, can 12:45

12:45 understand what is the value of running 12:47

12:47 this experiment versus running that one 12:49

12:49 and decide what to do next. Um, 12:52

12:52 automated decision-m with reinforcement 12:54

12:54 learning or evolutionary algorithms. 12:56

12:56 These earlier decisions can then be more 12:58

12:58 revisited more fluidly. Uh, and every 13:01

13:01 earlier experiment should inform every 13:03

13:03 future experiment. 13:06

13:06 So, what have we done so far? first 13:08

13:08 three weeks we did 400 interviews mostly 13:11

13:11 at Cafe Venetia on University Avenue uh 13:14

13:14 to find amazing people which I talked 13:17

13:17 about some of them uh finding office 13:19

13:19 space we're in the old Zibbo restaurant 13:22

13:22 so we have a funky funky uh space for us 13:25

13:25 uh we've built our own agent harness 13:27

13:27 that we're using for all of our own 13:29

13:29 development and for automation uh that 13:31

13:31 harness has some really nice 13:33

13:33 collaboration features so you can think 13:35

13:35 of it as a system that you interact with 13:38

13:38 uh a model uh to have conversations and 13:40

13:40 ask it to solve things, but you can also 13:42

13:42 invite your teammate in to collaborate. 13:45

13:45 And so then the two of you can talk to 13:47

13:47 uh a model and an agent. Uh and you can 13:50

13:50 also uh have other people add comments 13:52

13:52 to the thing you're trying to do or 13:54

13:54 propose a design to the to the agent 13:56

13:56 before you propose that. Um, we've built 13:60

14:00 really agent first infrastructure so 14:03

14:03 that it will scale to having tens, 14:06

14:06 hundreds of thousands of agents doing 14:08

14:08 things rather than sort of hundreds of 14:10

14:10 people doing things. Uh, we think that's 14:12

14:12 really important and we've got probably 14:14

14:14 on the order of 20 automated improvement 14:16

14:16 loops running in the way in which we 14:19

14:19 imagine uh that we want to solve 14:21

14:21 problems. uh and they're we can see that 14:23

14:23 they're running and improving uh 14:25

14:25 evaluating methods for making these 14:27

14:27 experimental iterations more effective 14:29

14:29 and faster. How do you learn from the 14:31

14:31 previous experimental results and the 14:33

14:33 previous uh measurements you're able to 14:35

14:35 take in order to um learn faster? 14:40

14:40 Um, so this is kind of where we are is 14:43

14:43 building this software loop to automate 14:46

14:46 machine learning tasks in particular 14:49

14:49 because we think that unlocks actually a 14:51

14:51 lot of uh autonomous science. Um, but 14:53

14:53 initially we're focused on how do you 14:55

14:55 make automated systems that can solve 14:57

14:57 machine learning research and machine 14:59

14:59 learning engineering problems. Uh, and 15:01

15:01 our goal is really to make domain 15:03

15:03 experts 20x to 50x more productive. You 15:05

15:05 know, we're not trying to take people 15:07

15:07 out of what we're doing. We're just 15:09

15:09 trying to make uh really empower amazing 15:12

15:12 people or or domain experts who want to 15:15

15:15 solve particular problems and enable 15:16

15:16 them to solve those problems with less 15:19

15:19 effort on their part. 15:21

15:21 Um so we think 15 years of breakthroughs 15:24

15:24 in machine learning have led to very 15:26

15:26 capable models setting the stage for 15:29

15:29 automating ML science and engineering to 15:32

15:32 dramatically accelerate discovery. Thank 15:34

15:34 you. Oh, and we're hiring apply at 15:37

15:37 discovery.com. 15:39

15:39 Thank you. [applause] 15:46

15:46 >> Oh, this one. Okay. Sorry, I took your 15:48

15:48 water then. 15:49

15:49 >> No worries. 15:51

15:51 All right, that was a great talk. 15:53

15:53 Thanks, Jeff. And uh I think this is the 15:57

15:57 last session or Okay, so we're we're 15:60

15:60 between the audience and snacks or 16:03

16:03 >> never a good place to be. 16:04

16:04 >> Yeah. So um I was hoping that we can be 16:09

16:09 a little bit nerdier for this session 16:11

16:11 and and get into a little bit of the 16:13

16:13 tech, but uh uh we're supposed to have a 16:16

16:16 mutual discussion. So I'm going to wear 16:18

16:18 the dual hat of asking you questions but 16:21

16:21 participate uh in the discussion. Um 16:24

16:24 thanks for for that talk because I was 16:27

16:27 going to ask you why as my first 16:29

16:29 question for for discovery loop. 16:31

16:31 Congratulations. 16:32

16:32 >> Thank you. 16:33

16:33 >> And pretty much all four of you are more 16:36

16:36 or less Google lifers 16:39

16:39 especially you joined Google 1999 right? 16:42

16:42 Yeah. So it is very clear what the macro 16:46

16:46 incredible technological um progress we 16:49

16:49 have seen the past few years but I'm 16:51

16:51 still curious was there a moment in the 16:55

16:55 past I would have guessed a year or two 16:59

16:59 years that the technology has moved into 17:03

17:03 a state that you and Qua and San and 17:06

17:06 Oreo are like we've got to do something 17:10

17:10 different and we cannot do it at Google. 17:13

17:13 >> Yeah. Yeah, I mean I think we all have 17:15

17:15 been excited about seeing automated 17:18

17:18 systems that can actually sort of help 17:21

17:21 people develop software systems in a 17:23

17:23 more automated way or um you know Quark 17:26

17:26 did a lot of work on neural architecture 17:28

17:28 search even a decade ago which you can 17:30

17:30 think of as an early version of that 17:32

17:32 kind of recursive uh sort of 17:34

17:34 self-improving system where you say 17:37

17:37 please find me great model architectures 17:38

17:38 for computer vision problems or um you 17:42

17:42 know language problems 17:44

17:44 And so we knew that kind of approach 17:46

17:46 worked. Uh but in that day you were sort 17:49

17:49 of specifying the search space as like a 17:51

17:51 junk janky kind of pile of Python code 17:54

17:54 rather than natural language as you can 17:56

17:56 today. You can sort of describe to a you 17:59

17:59 know a capable model plus agent system. 18:02

18:02 You know please go off and explore model 18:04

18:04 architectures that might help interpret 18:06

18:06 MRI images or whatever and it will go 18:08

18:08 off and do that. And you know, I think 18:11

18:11 that's just a really profound shift in 18:15

18:15 what we imagine computer systems and AI 18:19

18:19 systems to be able to do, right? And if 18:21

18:21 you can get those things working 18:23

18:23 reliably, that enables a lot of stuff. 18:25

18:25 And in particular for science and 18:27

18:27 engineering, we think that's a really 18:29

18:29 ripe set of domains for how can you 18:31

18:31 really advance positive impact in the 18:34

18:34 world. 18:35

18:35 >> I see. So it [clears throat] it was 18:36

18:36 really the the loops the self 18:39

18:39 improvement um that that got you guys 18:42

18:42 feel that you want to you want to 18:44

18:44 >> some might say we're a bit loopy but yes 18:48

18:48 >> it is and also I would also say that it 18:51

18:51 was important to us to you know we've 18:54

18:54 enjoyed many times starting very small 18:57

18:57 projects 18:58

18:58 >> and that that feeling of magic when you 19:00

19:00 have five people working on something 19:02

19:02 and you add a sixth and then you add a 19:05

19:05 seventh and 10th and so on. That that 19:07

19:07 really is fun. Um, and you can do that 19:09

19:09 in in large companies, but we also 19:11

19:11 wanted to do this as a public benefit 19:13

19:13 corporation. And we felt like that was 19:14

19:14 an important reflection of our own 19:16

19:16 values and um, you know, taking a bigger 19:19

19:19 plunge of now you have to figure out 19:22

19:22 real estate and all the other things 19:24

19:24 that involve uh, setting up a company. 19:27

19:27 Yeah. 19:27

19:27 >> Is kind of fun. 19:28

19:28 >> Yeah. Cool. Yeah. Even for myself when 19:31

19:31 we started world labs it was a moment of 19:33

19:33 seeing the technology it's different 19:35

19:35 from you guys seeing computer v vision 19:38

19:38 and uh generative models have converged 19:41

19:41 at the moment where we can create 19:44

19:44 generative models beyond just words was 19:46

19:46 the moment for for us to start world 19:49

19:49 labs. 19:49

19:49 >> Yeah I mean like the generative models 19:51

19:51 you you all produced and that are now 19:53

19:53 cap possible are pretty amazing. 19:55

19:55 >> Yeah. Okay. So let's talk about this 19:58

19:58 self uh improvement and recursive 20:01

20:01 self-improvement. This let's nerd out a 20:04

20:04 little bit. We're contemporaries in 20:06

20:06 terms of our education through the field 20:09

20:09 of machine learning. I was just thinking 20:11

20:11 in my head how many names this concept 20:16

20:16 has had in its lifetime 20:19

20:19 from me being a grad student. We had 20:21

20:21 incremental learning then we had 20:23

20:23 continuous learning. Then we had 20:25

20:25 metalarning. Then we had um um lifelong 20:30

20:30 learning and um and neuro architecture 20:34

20:34 search. You mentioned that now we have 20:37

20:37 >> self-improvement and then RSI. So do you 20:41

20:41 want to just treat the audience 20:44

20:44 especially the technical part of the 20:45

20:45 audience with some more rigorous 20:49

20:49 definition of what is self-improvement 20:52

20:52 versus what is recursive 20:54

20:54 self-improvement and my goal here is 20:57

20:57 actually to tease apart the fiction from 20:59

20:59 the the facts. 21:01

21:01 >> Sure. Yeah. I mean, I think if you think 21:03

21:03 about those loops that I talked about, 21:06

21:06 those loops can be applied to many kinds 21:08

21:08 of problems. And I I think of like 21:11

21:11 self-improving systems as one thing 21:14

21:14 where you've given the system a goal to 21:16

21:16 go off and improve you typically some 21:19

21:19 measurable objective like accuracy of an 21:22

21:22 ML model and you know it's trying 21:25

21:25 different kinds of model architectures 21:26

21:26 to improve that accuracy. Um 21:28

21:28 >> so pause there. A lot of machine 21:31

21:31 learning algorithm is capable of doing 21:33

21:33 that. 21:33

21:33 >> Yes, 21:34

21:34 >> that itself is not recursive. Correct. 21:36

21:36 >> That is a 21:37

21:37 >> so like self-improving system might be 21:40

21:40 >> we point our system at uh a bunch of MRI 21:43

21:43 data and MRI outcomes and we say I want 21:46

21:46 to develop a model to help interpret MRI 21:49

21:49 images. 21:49

21:49 >> Right? That would be a self-improving 21:51

21:51 system where you can develop that model 21:53

21:53 automatically and can then you know see 21:56

21:56 the accuracy of the different approaches 21:57

21:57 that are being tried hopefully increase 21:60

21:60 as the system runs experiments but 22:01

22:01 that's not improving the base system 22:04

22:04 that is proposing what to try next right 22:07

22:07 >> y 22:08

22:08 >> so recursive self-improvement is how do 22:10

22:10 you actually get that kind of 22:13

22:13 self-improving approach to improve the 22:16

22:16 base system that you're trying to get to 22:19

22:19 propose improve 22:21

22:21 if that makes sense. 22:22

22:22 >> Yes, it makes sense to me. But I think 22:24

22:24 it's it's a very confusing concept for 22:28

22:28 people and now 22:30

22:30 >> especially we're in the Silicon Valley 22:33

22:33 people are just using this in a very 22:34

22:34 loose way and as a scientist actually 22:37

22:37 bothers me. So I I just want to be on 22:40

22:40 this stage to make sure we define uh RSI 22:43

22:43 really well. So let's talk about the 22:45

22:45 physical limit of RSI though because 22:48

22:48 >> RSI is being hyped in a way it's almost 22:52

22:52 like 22:54

22:54 it's almost a system it has no end we 22:58

22:58 can just accelerate and and and uh to 23:02

23:02 the infinity 23:04

23:04 >> what is the hype here what is the 23:06

23:06 fiction from 23:06

23:06 >> well I mean I think 23:09

23:09 if you think about how humans improve 23:11

23:11 the capabilities of models which is what 23:13

23:13 has been happening over the past 10 or 23:14

23:14 15 years, you know, you get really 23:17

23:17 capable researchers and ML engineers 23:19

23:19 together. They run lots of experiments 23:21

23:21 at small scale. They say, "Oh, these 23:23

23:23 these ideas didn't work. These ideas did 23:26

23:26 work." Then we're going to incorporate 23:28

23:28 those into the next generation of the 23:30

23:30 the model. That's how Gemini 2 comes to 23:32

23:32 be after we had Gemini 1 is we ran lots 23:36

23:36 of experiments and and improve things. 23:39

23:39 So that is kind of a normal process of 23:44

23:44 research and engineering. And I think 23:46

23:46 when you start to apply automation to 23:48

23:48 that, yes, you can get some uh 23:50

23:50 improvements in how fast that can 23:52

23:52 happen. You know, people will start off 23:55

23:55 an experiment, the experiment will 23:56

23:56 finish at 2 a.m. They're not into the 23:58

23:58 desk until 8 a.m. to look at the 23:60

24:00 results. So if you have an automated 24:02

24:02 system, it can look at the results, 24:04

24:04 figure out what to do next in a much 24:05

24:05 more automated way. can maybe make a lot 24:08

24:08 better decisions across um an entire 24:11

24:11 large project than each individual 24:13

24:13 making decisions about what to do next. 24:15

24:15 Um so but I think if you think about 24:19

24:19 what needs to be involved in creating 24:23

24:23 the base model and making it better, 24:25

24:25 there's lots of actual training and 24:28

24:28 experimentation that needs to be done 24:30

24:30 and that's not going to happen 24:31

24:31 infinitely fast, right? Like there's 24:33

24:33 just actual physics involved. you need 24:35

24:35 to run some computations. You need to 24:37

24:37 look at the results. The computers need 24:39

24:39 to like do their thing and you need to 24:41

24:41 save the data, look at the have another 24:43

24:43 model look at the results in order to 24:45

24:45 decide what to do next. So I think some 24:47

24:47 people have this illusion that you're 24:49

24:49 going to be able to run this loop 24:51

24:51 infinitely fast. And I just find that 24:54

24:54 kind of uh implausible. 24:56

24:56 >> Yeah. 24:57

24:57 >> And it's good to be grounded in reality. 24:59

24:59 Having automated systems that can try to 25:02

25:02 improve things is definitely going to 25:04

25:04 improve the rate at which we can improve 25:06

25:06 those systems, but it's not going to 25:08

25:08 make them go infinitely fast. 25:10

25:10 >> I think this is a really important 25:11

25:11 point. Actually, I was I I was not 25:14

25:14 applying for a job at Discovery Loop, 25:16

25:16 but I was at the cafe Valencia with 25:19

25:19 Jeff. 25:20

25:20 >> We we have coffee for many reasons cuz 25:23

25:23 we like coffee 25:24

25:24 >> and we have been talking about this RSI, 25:26

25:26 you know, hype and facts. We were 25:29

25:29 actually prepping for this panel 25:30

25:30 >> apparently. But I was thinking, okay, 25:33

25:33 this is something I haven't run by you, 25:35

25:35 but tell me if you agree or you can 25:37

25:37 disagree with me, that's totally fine. I 25:39

25:39 was thinking about humans. Are we a 25:42

25:42 self-improving system or are we a RSI 25:45

25:45 system? And I think we're definitely a 25:49

25:49 self-improving system. I suck at 25:51

25:51 cooking. And if I were to try to cook 25:55

25:55 omelette, the first time will really I 25:58

25:58 will burn everything and sec I would 25:60

25:60 have learned and the second time I would 26:02

26:02 do it a little better. Maybe after 10 26:04

26:04 times I'll have a reasonable omelette 26:06

26:06 that I will eat. My kids will not eat. 26:08

26:08 And that is self-improving. 26:11

26:11 >> But the RSI part has to do with my 26:14

26:14 brain. In this process of learning how 26:17

26:17 to cook omelette, first of all, my brain 26:20

26:20 has to be somehow altered. Some neurono 26:25

26:25 connectivity needs to be altered enough 26:27

26:27 that I have a better system to to cook 26:32

26:32 omelette. Maybe I can grow the six 26:34

26:34 finger 26:36

26:36 if if if I'm we're talking about 26:38

26:38 hardware improvement. And then by 26:40

26:40 cooking more, 26:41

26:41 >> maybe you're AI generated. 26:43

26:43 >> Exactly. [laughter] 26:44

26:44 And then the by cooking more omelette my 26:47

26:47 brain will be further improved. And if 26:50

26:50 that loop happens humans is a RSI 26:53

26:53 system. Is that the right analogy? 26:55

26:55 >> Yeah. I mean I I I think the answer is 26:57

26:57 humans are both right. 26:58

26:58 >> Right. I mean I think through your life 27:02

27:02 you learn a lot of things which is 27:04

27:04 self-improving like you learn how to 27:05

27:05 cook an omelette, you learn how to do 27:07

27:07 calculus problems but you also learn 27:11

27:11 how you learn. 27:12

27:12 >> Yeah. And that learning how to learn 27:14

27:14 actually is what makes you 27:16

27:16 self-improving uh and recursively 27:19

27:19 self-improving because now you can learn 27:21

27:21 to do more things more quickly based on 27:24

27:24 you understanding what kinds of 27:26

27:26 knowledge and the ways to get 27:28

27:28 information presented to you actually 27:30

27:30 makes you most effectively pick it up. 27:32

27:32 >> Yeah. and possibly recursive self-learn 27:37

27:37 recursive self-improvement is po 27:40

27:40 possibly a evolutionary algorithm right 27:43

27:43 so anyway I don't want to go down the 27:45

27:45 rabbit hole but I think it's just a rare 27:47

27:47 opportunity to have this discussion in 27:50

27:50 public because we have so much hype 27:51

27:51 right now about 27:52

27:52 >> concept I mean there's a lot of hype 27:54

27:54 about like oh it's going to go 27:55

27:55 infinitely fast and it's scary I mean I 27:57

27:57 think if you look at it as a pragmatic 27:59

27:59 engineering system doing things it 28:02

28:02 becomes 28:03

28:03 less scary and more understandable. 28:05

28:05 >> Yeah. So, you talked a lot about 28:07

28:07 discovery loop, but you were still a 28:10

28:10 little vague about what scientific 28:13

28:13 problem you want to work on. So, I want 28:15

28:15 to just push you a little bit to the 28:17

28:17 extent you're willing to review. 28:19

28:19 >> Yeah. 28:20

28:20 >> There are so many scientific problems. 28:22

28:22 >> Yes. 28:23

28:23 >> That benefits the society. Which one are 28:26

28:26 you passionate about right now? 28:28

28:28 >> Yeah. So uh 28:31

28:31 I I I for a long time I've been very 28:33

28:33 interested in AI for healthcare. I think 28:35

28:35 that's an area that we will do work in 28:37

28:37 not right now but down the road because 28:39

28:39 I think it's a super important area and 28:42

28:42 there's that would be very aligned with 28:44

28:44 our sort of uh public benefit 28:46

28:46 corporation mission. Um right now we are 28:50

28:50 focused on the kinds of science and 28:52

28:52 engineering problems where you can do 28:54

28:54 the entire loop in digital form because 28:56

28:56 we think that's the one where you it's 28:58

28:58 very amenable to speeding up that entire 28:60

28:60 loop. If you have like actual physical 29:02

29:02 processes involved that be that becomes 29:04

29:04 harder to speed up the loop. There are 29:06

29:06 obviously many important disciplines 29:08

29:08 where you do need physical processes. Um 29:11

29:11 and we have some hybrid you know 29:13

29:13 approaches that we can speed things up 29:15

29:15 in those domains. 29:17

29:17 So let's look at digital domains of 29:20

29:20 science and engineering. We are 29:22

29:22 initially focused on machine learning 29:24

29:24 research and machine learning 29:25

29:25 engineering itself and uh also hardware 29:29

29:29 design because we think those three 29:31

29:31 areas are you know it's good to have a 29:32

29:32 few different areas you focus on not try 29:34

29:34 to focus on everything but um those are 29:38

29:38 ones where we have domain expertise 29:40

29:40 ourselves and we think it's a really 29:42

29:42 important domain because it unlocks 29:44

29:44 other things. So for example in 29:46

29:46 healthcare if you have a system that can 29:48

29:48 automatically solve machine learning 29:51

29:51 engineering tasks 29:53

29:53 then you can point it at lots of 29:54

29:54 problems in in health or you know 29:56

29:56 material science or or whatever and say 29:59

29:59 I would like an amazing model 30:01

30:01 specialized to this particular task 30:03

30:03 which is say self-improvement but not 30:05

30:05 recursive self-improvement. Um so that 30:09

30:09 uh I think is how I would describe us. 30:11

30:11 We're focused on those three domains 30:14

30:14 right now, but we have ambitions to go 30:17

30:17 beyond. 30:18

30:18 >> Okay. So, I was trained in physics when 30:22

30:22 I was undergrad and um in in physics we 30:28

30:28 have this division between 30:30

30:30 experimentalists and theorists, right? I 30:33

30:33 was just thinking about your loop the 30:35

30:35 the way you draw it. you call it propose 30:37

30:37 uh experiment but in physics we tend to 30:40

30:40 call it hypothesis 30:42

30:42 form hypothesis and then you go through 30:44

30:44 um experiments and evaluation and all 30:47

30:47 that I was just thinking the way how 30:50

30:50 powerful AI agents or these uh 30:52

30:52 self-improvement loops whether it's 30:54

30:54 recursive or not doesn't matter is it 30:58

30:58 philosophically taking a stance that it 31:02

31:02 is more a experimental experimentalist 31:05

31:05 approach approach to a scientific 31:07

31:07 question because it's very much about 31:10

31:10 proposing hypothesis and parallel um 31:13

31:13 parallelized experiments 31:16

31:16 um as opposed to the more mysterious 31:20

31:20 theoretical way of doing scientific 31:22

31:22 discovery. You know to this day you 31:25

31:25 would wonder 31:27

31:27 great theorists like Newton like 31:31

31:31 Einstein they don't do that much 31:33

31:33 experiments they read about them but 31:35

31:35 they don't do it that way so do you feel 31:38

31:38 this is 31:40

31:40 skewing towards experimentalism 31:42

31:42 >> I don't think necessarily but I it 31:46

31:46 probably there's a bit more leaning in 31:47

31:47 that direction I do think you could do 31:49

31:49 great theoretical work with a system 31:51

31:51 like this where you can propose 31:53

31:53 Can you see if you can come up with a 31:55

31:55 theory that explains all of this data? 31:58

31:58 >> And I think it could iterate and try to 31:60

32:00 find kind of the minimal description 32:03

32:03 theory that helps explain experimental 32:05

32:05 data. You know, I think a lot of 32:06

32:06 experimental or theoretical physicists 32:09

32:09 are seeing what's happening in the 32:11

32:11 physics community in the experimental 32:13

32:13 domain and then trying to come up with 32:16

32:16 theories that explain, you know, some of 32:18

32:18 the things that people are observing and 32:19

32:19 don't understand yet. So I I think that 32:21

32:21 that's certainly possible and I think 32:23

32:23 actually it's not good to think of these 32:26

32:26 as separate disciplines. 32:28

32:28 >> Right. 32:28

32:28 >> Right. So theoretical physics, 32:30

32:30 experimental physics, you know, 32:32

32:32 chemistry. 32:33

32:33 If you think about a system that 32:35

32:35 actually has say performed 32:38

32:38 a mil solved a million problems with a 32:40

32:40 100 million experiments across a 100 32:43

32:43 domains of science. you know, it's very 32:45

32:45 likely going to have a lot of insights 32:47

32:47 that derive from the fact that it has 32:49

32:49 seen lots of different kinds of, you 32:51

32:51 know, experimental results and seen like 32:53

32:53 what worked and what didn't work in sort 32:56

32:56 of the neighborhood of of interesting 32:58

32:58 problems you're going to ask it to solve 32:59

32:59 in the future. And that I think doesn't 33:02

33:02 have hard domain boundaries the way we 33:05

33:05 as people sometimes tend to instill 33:08

33:08 with, you know, academic department 33:10

33:10 boundaries or whatever. One of the 33:11

33:11 things I like about HAI is it's like a a 33:14

33:14 pan institutional thing that gets 33:16

33:16 everyone talking together. So, 33:18

33:18 >> okay. Well, we have to talk about math 33:19

33:19 then. Um, in fact, when we were 33:22

33:22 proposing to have to invite you to this 33:25

33:25 panel, I I actually was going to did 33:28

33:28 invite Terrence to so I called him but 33:31

33:31 he couldn't come. So, let's talk about 33:33

33:33 math a little bit. Um I have to say 33:38

33:38 I had mixed emotions when we see a 33:42

33:42 serious it was first Navier Stokes and 33:45

33:45 there 720 33:47

33:47 you know papers and then there's more 33:49

33:49 that was just released uh by open AI 33:52

33:52 yesterday that uh AI solving math it's 33:55

33:55 not it's obviously it's not that I don't 33:59

33:59 like what's being discovered it's just 34:01

34:01 that math and the century old, 34:05

34:05 millennial old math problem is such a 34:07

34:07 pinnacle of human intelligence and and 34:10

34:10 mathematicians see not only technical 34:15

34:15 solutions in them, they also see beauty. 34:17

34:17 They they and and now we have a 34:20

34:20 completely different machinery that goes 34:23

34:23 about these problems. Right. Yeah. 34:25

34:25 >> So Terrace and I actually talked quite a 34:29

34:29 bit about that and he recently also just 34:31

34:31 gave a talk about math 2.0 O and talk 34:34

34:34 about how AI can help solve certain type 34:37

34:37 of math problems, but he also does 34:40

34:40 insist there are other math problems 34:43

34:43 that are out of the realm of AI, 34:45

34:45 especially if we don't have solutions, 34:47

34:47 there's no data. What's your take of 34:51

34:51 this pretty 34:54

34:54 big tsunami change in math in the past 34:57

34:57 few literally weeks? Yeah, I mean 34:60

34:60 clearly the models and the sort of 35:03

35:03 agent-based systems that can use those 35:05

35:05 models in highly parallel, you know, 35:08

35:08 agent uh activity are capable of solving 35:11

35:11 really hard unopen, you know, unsolved 35:14

35:14 problems in mathematics and that is a a 35:17

35:17 revelation of what is possible now. I 35:20

35:20 think people are find it finding it 35:22

35:22 dissatisfying and a little uh uh 35:26

35:26 discomforting because often those proofs 35:29

35:29 involve like 500 pages of lean or 35:32

35:32 something. So it's machine checked but 35:34

35:34 no one quite understands the elegant 35:38

35:38 simple way to get to that result. And so 35:40

35:40 I think um it is good to know that the 35:44

35:44 problem is proven or disproven. But 35:48

35:48 understanding how to develop really 35:50

35:50 elegant techniques to come to that 35:52

35:52 result is often what happens as 35:56

35:56 mathematician human mathematicians are 35:58

35:58 wrestling with problems and those often 35:60

35:60 generate new kinds of mathematics that 36:01

36:01 can be used to tackle other kinds of 36:03

36:03 problems. So, 36:05

36:05 >> so you're saying new math pro solutions 36:07

36:07 only come from elegant solutions. 36:10

36:10 >> No, but I think the like the elegant an 36:13

36:13 elegant solution often comes up with a 36:15

36:15 cool technique that can then be applied 36:17

36:17 to other problems. Um, and I don't think 36:20

36:20 people are necessarily seeing the crisp 36:24

36:24 elegant technique in some of these very 36:25

36:25 long machine generated proofs. But 36:28

36:28 knowing that something is solvable is 36:31

36:31 actually uh in lots of different 36:33

36:33 domains. If you know something is 36:36

36:36 possible, it actually makes you strive 36:38

36:38 harder and understand that you can do it 36:41

36:41 and then you actually can do it perhaps 36:43

36:43 a different way or in a more elegant 36:45

36:45 way. And so I think actually yes, it's 36:48

36:48 kind of upsetting about 36:52

36:52 you know the way in which so many of 36:54

36:54 these problems are being solved 36:55

36:55 simultaneously. 36:56

36:56 But I do think you know that is now 36:59

36:59 possible. So people have to figure out 37:01

37:01 how does how do we take advantage of 37:03

37:03 this capability to do even more math. 37:07

37:07 >> We can we can talk so much more about 37:09

37:09 this. But I did promise Jeff I'm going 37:12

37:12 to give him a pop quiz on stage 37:14

37:14 [laughter] and I did not tell him what 37:16

37:16 the problem is. Um but it's okay if you 37:18

37:18 don't know how to answer it because I 37:20

37:20 did not remember myself. Um 37:23

37:23 >> great. No pressure. 37:24

37:24 >> Yeah. [laughter] 37:26

37:26 Do you remember the significance of this 37:28

37:28 date? June 12th, 2017. 37:32

37:32 >> June 12th, 2017. 37:34

37:34 >> Don't cheat with AI. 37:35

37:35 >> Oh, well, is that the Transformer paper 37:37

37:37 on archive? 37:39

37:39 >> Amazing. A+. Okay. We have to uh Yes, 37:43

37:43 that's uh I I had to check which date it 37:47

37:47 was. Did you read the transformer paper 37:49

37:49 on June 12th? 37:51

37:51 >> No, probably June 10th. 37:53

37:53 >> Oh, yes, that's true. you had to approve 37:55

37:55 that probably. So um you know your 37:59

37:59 career I I want to reflect a little bit 38:02

38:02 um you know we can talk so much you you 38:06

38:06 one of your slides showed the four of 38:08

38:08 you collectively have done so much at 38:09

38:09 Google and for our field and uh but 38:13

38:13 let's just use that day as kind of a 38:16

38:16 historical date some kind of beginning 38:18

38:18 of you know the generative model era 38:22

38:22 just share with us in this journey since 38:26

38:26 almost 10 years, nine years now. 38:29

38:29 What in your mind, maybe we don't have 38:32

38:32 time to say all of them, but the top 38:34

38:34 most significant 38:36

38:36 technical moment um that for yourself 38:40

38:40 personally and I also want to follow up 38:43

38:43 by asking you what and intellectually 38:46

38:46 honestly what did you predict correctly 38:50

38:50 and what did you predict incorrectly? 38:54

38:54 >> Yeah. Uh let's see. Um I mean I think a 38:59

38:59 handful of things stand out among a 39:02

39:02 bunch of major improvements to all the 39:05

39:05 things we did like that slide that had 39:06

39:06 all those different advances. You know I 39:09

39:09 think obviously transformers are a much 39:12

39:12 more efficient model architecture for 39:14

39:14 tackling these problems. The the idea 39:16

39:16 that you save all the states you went 39:18

39:18 through and then you attend to them 39:20

39:20 rather than trying to update a small 39:22

39:22 vector with every state and and forget 39:24

39:24 the previous ones was a really great 39:26

39:26 insight and has led to um you know many 39:32

39:32 many many advances because the models 39:34

39:34 can now learn so much more with so much 39:38

39:38 less computation than earlier say 39:40

39:40 LSTMbased models. Um, you know, I think 39:44

39:44 uh uh the distillation work that Oral uh 39:48

39:48 I and Jeff Hinton did uh which was 39:50

39:50 actually rejected from Nurips in 2016 39:53

39:53 saying it's unlikely to have any impact. 39:55

39:55 >> We should have a website that collects 39:57

39:57 all the rejected papers that 39:59

39:59 >> I mean that one's pretty good. We threw 39:60

39:60 it on archive. It has 40,000 citations 40:02

40:02 or something. Um, so I mean that I think 40:05

40:05 has been key to making really capable 40:10

40:10 inference efficient models because you 40:12

40:12 can take a very large model you've 40:13

40:13 trained that's really unwieldy and 40:15

40:15 expensive and you can use use it as the 40:17

40:17 teacher to teach your now nimble 40:20

40:20 lightweight uh um student model is a 40:23

40:23 good thing to say in a university. 40:26

40:26 >> Um and so that that's been good. Thee 40:28

40:28 architecture I think has proven to be 40:31

40:31 good. I think one one thing I kind of 40:33

40:33 thought was that was going to be the 40:34

40:34 beginning of a much more exotic set of 40:37

40:37 model architectures. I think you know 40:39

40:39 the MOE architecture is you you branch 40:41

40:41 out to a bunch of experts and then you 40:43

40:43 come back and then you branch out again 40:45

40:45 and you come back and you learn which 40:47

40:47 experts are good at which things which 40:49

40:49 is super useful. Gives your model a lot 40:51

40:51 of capacity but you then activate only a 40:53

40:53 tiny portion of it for any given token 40:56

40:56 or task or whatever. But it's still very 40:59

40:59 regular like in most all the experts are 41:02

41:02 the same size. Uh they come back 41:04

41:04 together immediately. I thought we were 41:06

41:06 going to start to have like a 41:07

41:07 flourishing of like really wild exotic 41:10

41:10 more organic looking neural network 41:12

41:12 architectures with like columns of 41:15

41:15 neural activity and they come back 41:17

41:17 together and some like very uh 41:20

41:20 infrequent but uh large cross 41:23

41:23 connections that would let you connect 41:25

41:25 distal parts of your your neural net. We 41:27

41:27 haven't done that yet. 41:29

41:29 >> I mean, we thought about that for neuro 41:31

41:31 architecture search and that didn't pan 41:33

41:33 out either, right? 41:34

41:34 >> Yeah. I mean, the the models that come 41:36

41:36 out of neural architecture search do 41:38

41:38 look much more organic than the kind of 41:40

41:40 human design ones where you have like a 41:42

41:42 bunch of layers. Occasionally, you have 41:44

41:44 a skip connection every layer. The 41:46

41:46 neural architecture search ones have 41:48

41:48 like these crazy like 41:50

41:50 >> connect this one up to 17 layers above 41:52

41:52 and do do something funky. 41:54

41:54 >> Yep. So, so that was the one thing you 41:56

41:56 think you predicted wrong? 41:58

41:58 >> Well, I mean, I predicted many things 42:00

42:00 wrong. Uh, that's one that that kind of 42:02

42:02 stands out as like I thought thee push 42:05

42:05 would go more in that direction. 42:08

42:08 >> Um, 42:10

42:10 you know, I think uh 42:14

42:14 coding capability has proven to be 42:16

42:16 really important. I didn't necessarily 42:17

42:17 foresee that was such a key unlocker for 42:21

42:21 making these models really capable in 42:23

42:23 solving lots of different kinds of 42:24

42:24 problems. 42:25

42:25 >> Um, so 42:26

42:26 >> that that one I I'm with you. We all 42:30

42:30 kind of know but did not 42:33

42:33 predict the level of profoundness. 42:36

42:36 >> So like in the Gemini effort, it was one 42:38

42:38 of the things we focused on but it was 42:39

42:39 not like the thing we focused on. 42:43

42:43 >> Okay. Well, there's one thing that 42:45

42:45 obviously we've been nerding out a 42:47

42:47 little bit there. We're in the HAI 42:50

42:50 conference, so we have to talk about the 42:51

42:51 letter H. The one thing I wonder if you 42:56

42:56 predicted and how do you feel about that 42:59

42:59 is all the societal impact of AI and 43:02

43:02 it's not just impact part of the impact 43:06

43:06 is there's jobs there's the 43:09

43:09 psychological anxiety there is the the 43:12

43:12 the the the right now the almost extreme 43:15

43:15 talk about human extinction and 43:19

43:19 alignment and all that. Well, first of 43:21

43:21 all, which one which one worries you the 43:25

43:25 most? 43:26

43:26 >> Yeah. I mean, I think there this is a 43:30

43:30 incredibly powerful and general 43:32

43:32 technology and it's influencing nearly 43:34

43:34 every aspect of society, right? Like all 43:37

43:37 the things that it can do in mathematics 43:38

43:38 or in coding or in generating images or 43:43

43:43 um you know helping autonomous vehicles 43:45

43:45 drive. So it's unlike previous 43:49

43:49 technological revolutions which were a 43:51

43:51 little bit more narrowly contained and 43:52

43:52 affected deeply five or 10 fields. This 43:56

43:56 is simultaneously affecting lots and 43:58

43:58 lots of things in the world. I think 43:60

43:60 that's a 44:01

44:01 you know it's a source of excitement 44:05

44:05 apprehension you know fear all the 44:08

44:08 things. I actually was a co-author with 44:10

44:10 an amazing set of co-authors on a paper 44:13

44:13 called shaping AI's impact on billions 44:15

44:15 of people. Uh it has my only paper with 44:18

44:18 its own website because we had an 44:19

44:19 inspired co-author shapingai.com. 44:22

44:22 Um but really we looked at seven 44:24

44:24 different areas including uh employment 44:27

44:27 and e and economy um AI for science uh 44:32

44:32 you know dis misinformation 44:34

44:34 disinformation 44:36

44:36 uh media 44:37

44:37 uh and um a few other uh domains and we 44:42

44:42 talked to lots of domain experts in each 44:43

44:43 of those domains and I think that was a 44:45

44:45 a great experience because we got their 44:47

44:47 perspective on how AI was already in 44:50

44:50 impacting their fields and how it was 44:52

44:52 likely to impact fields in the future. I 44:54

44:54 mean, the one I worry about the most is 44:56

44:56 just 44:58

44:58 how do we capture 44:60

44:60 as much of the upsides of what AI can do 45:02

45:02 across all these different domains while 45:05

45:05 minimizing the downsides in many many 45:07

45:07 areas that it uh can affect negatively. 45:10

45:10 And I think that is not purely a 45:13

45:13 technical problem. That is a technical 45:15

45:15 problem, a regulatory and policymaker 45:18

45:18 problem, a citizen problem. and we all 45:22

45:22 need to kind of figure out how do we 45:24

45:24 navigate this uh new technology to make 45:27

45:27 sure that it serves us all extremely 45:29

45:29 well and doesn't uh harm us in ways that 45:33

45:33 we think it could in you know so that's 45:35

45:35 things like making sure when people's 45:38

45:38 jobs are affected they have good ways to 45:40

45:40 adapt to what is now needed in the world 45:43

45:43 and and can you know be um you know 45:47

45:47 gainfully employed and and have a 45:50

45:50 prosperous existence and I think that is 45:52

45:52 a you know a a multiaceted thing that 45:56

45:56 needs to be tackled by lots of people in 45:57

45:57 the world. 45:58

45:58 >> So one thing this is my last question 46:01

46:01 one thing I know about you Jeff that 46:03

46:03 most of the public doesn't know is how 46:05

46:05 much you care about the young generation 46:08

46:08 we don't get to be on stage too often 46:12

46:12 but in my memory the last time we're in 46:15

46:15 a conference together was Grace Hopper 46:18

46:18 conference. Oh, 46:19

46:19 >> you know, you you were going there for 46:20

46:21 [clears throat] multiple years. 46:22

46:22 >> So, right now we're on a university 46:24

46:24 campus 46:25

46:25 >> and frankly this generation of college 46:29

46:29 students are anxious. 46:31

46:31 >> They they're even Stanford students are 46:34

46:34 anxious about jobs, anxious about the 46:37

46:37 future of economy. And as someone who 46:41

46:41 has, you know, such a successful career, 46:45

46:45 been an an and a pioneer in AI and also, 46:49

46:49 you know, just just 46:52

46:52 seeing the future in ways that the 46:54

46:54 public doesn't see yet. What is your 46:57

46:57 parting words for for the students in 46:60

46:60 this audience who are feeling that 47:02

47:02 anxiety and possibly also hope? Yeah, I 47:05

47:05 mean I think it I understand the anxiety 47:09

47:09 for sure because the world is changing 47:11

47:11 quickly and I think if you're planning 47:13

47:13 for a future, it's very hard to plan for 47:16

47:16 a future that is changing quickly. 47:18

47:18 >> Yeah. Um, and so, you know, I think the 47:22

47:22 general advice I have is try to 47:24

47:24 understand what new technology is 47:27

47:27 starting to exist or will exist 47:29

47:29 relatively soon and make sure that you 47:32

47:32 understand how to use that technology 47:34

47:34 for whatever it is you are most 47:36

47:36 passionate about, right? That could be, 47:39

47:39 you know, um, being a doctor, that could 47:42

47:42 be, you know, uh, helping write 47:45

47:45 literature. I could be, you know, 47:47

47:47 writing software more effectively. All 47:49

47:49 of these things, I think, can be aided 47:52

47:52 and empowered by AI, but there's still 47:54

47:54 that human element of like humans figure 47:56

47:56 out what we want to collectively do. 47:59

47:59 >> And you should figure out what it is you 48:02

48:02 want to do and be the best you can at 48:05

48:05 it. And that means using all the tools 48:07

48:07 that are now available. 48:09

48:09 >> So, human passion is still timeless in 48:12

48:12 the age of AI. Thank you, Jeff. Yeah. 48:14

48:14 >> Thank you. [applause]