0:02 Hello everybody. Good afternoon, good
0:05 morning. And I am delighted to be the
0:07 one chosen to introduce to you this
0:09 really distinguished group of people
0:10 that we've got here sitting around the
0:13 table. Six I think of the most
0:15 brilliant, most consequential people on
0:17 the planet today. And I don't think
0:18 that's an overstatement.
0:21 So these are the winners of the 2025
0:24 Queen Elizabeth Prize for Engineering.
0:26 and it honors the laureates here that we
0:29 see today for their singular impact on
0:30 today's artificial intelligence technology.
0:32 technology.
0:34 Given your pioneering achievements in
0:37 advanced machine learning and AI and how
0:38 the innovations that you've helped build
0:41 are shaping our lives today, I think
0:43 it's clear to everyone why this is a
0:46 really rare and exciting opportunity to
0:48 have you together around the table. For
0:50 me personally, I'm I'm really excited to
0:53 hear you reflect on this present moment
0:55 that we're in, the one that everybody's
0:57 trying to get ahead of and understand
0:59 and your journey, the journey that
1:02 brought you here today. Um, but also to
1:04 understand how your work and you as
1:06 individuals have influenced and impacted
1:08 one another and the companies and the
1:11 technologies that you've built. And
1:13 finally, I'd love to hear from you to
1:16 look ahead um and to help us all see a
1:18 bit more clearly what is to come, which
1:20 you are in the best position to do. So,
1:24 I'm so um pleased to have you all with
1:26 us today and and looking forward to this
1:28 to the discussion. So, I'm going to
1:30 start going from the zooming out to the
1:33 very personal. I want to hear from each
1:36 of you your a personal kind of aha
1:38 moment in your career that you've had
1:41 that you think has sort of impacted the
1:42 work that you've done or was a turning
1:44 point for you that brought you on this
1:47 path to why you're sitting here today
1:48 whether it was kind of early in your
1:50 career in your research or or much more
1:52 recently what you know what was your
1:55 personal moment of awakening um that
1:57 that has impacted the technology do we
2:00 should we start here with you Yeshua
2:04 >> thank you yes uh with pleasure. I would
2:08 um go to two moments. One when I was a
2:10 grad student and I was looking for
2:13 something interesting to research on and
2:16 I read some of Jeff Hinton's early
2:19 papers and I thought wow this is so
2:22 exciting. Maybe there are a few simple
2:24 principles like the laws of physics that
2:26 could help us understand human
2:28 intelligence and help us build
2:30 intelligent machines. And the second
2:32 moment I want to talk about is two and a
2:34 half years ago after chat GPT came out
2:38 and I realized uhoh what are we doing?
2:40 Uh what will happen if we build machines
2:44 that understand language uh have goals
2:46 and we don't control those goals? What
2:48 happens if they are smarter than us? Uh
2:51 what happens if people abuse that power?
2:54 So that's why I decided to completely
2:56 shift my research agenda and my career
2:59 to try to do whatever I could about it.
3:01 >> That's that that's two kind of very you
3:03 know diverging things very interesting
3:05 build what tell us about your moment of
3:06 like kind of building the infrastructure
3:08 that's that's fueling what we have.
3:09 >> I'll give you two moments as well. So
3:11 the first was you know in in the late
3:13 90s I was at Stanford trying to figure
3:15 out how to overcome what was at the time
3:17 called the memory wall. fact that
3:19 accessing data from memory is far more
3:21 costly in energy and time than doing
3:23 arithmetic on it. And it sort of you
3:25 know struck me to organize computations
3:28 into these kernels connected by streams.
3:29 So you could do a lot of arithmetic
3:30 without having to do very much memory
3:33 access. That basically led the way to
3:34 what became called stream processing and
3:37 ultimately GPU computing. Um and we we
3:39 originally built that thinking we could
3:41 apply GPUs not just for graphics but to
3:44 general scientific computations. So the
3:46 second moment was I was having breakfast
3:48 with my colleague Andrew Ing at Stanford
3:50 and at the time he was working at Google
3:52 finding cats on the internet you using
3:55 16,000 CPUs in this technology called
3:56 neural networks
3:58 >> which fay had something to do with those
4:01 >> and uh um he uh he he basically
4:03 convinced me this is a great technology
4:05 so I with Brian Kenzo repeated the
4:08 experiment on 48 GPUs in Nvidia and when
4:09 I saw the results of that I was
4:12 absolutely convinced that this is what
4:13 Nvidia should be doing. we should be
4:14 building our GPUs to do deep learning
4:16 because this has, you know, huge
4:18 applications in all sorts of fields
4:20 beyond finding cats. And that was kind
4:22 of an aha moment to really start working
4:25 very hard on specializing the GPUs for
4:26 deep learning and and to make them more effective.
4:27 effective.
4:28 >> And when was that what year?
4:30 >> Um, the breakfast was in 2010 and I
4:31 think we repeated the experiment in 2011.
4:32 2011. >> Okay.
4:32 >> Okay. >> Yeah.
4:33 >> Yeah.
4:36 >> Jeff, tell us tell us about your work.
4:39 One very important moment was when I in
4:42 about 1984 I tried using back
4:46 propagation to learn the next word in a
4:48 sequence of words. So it was a tiny
4:51 language model and discovered it would
4:53 learn interesting features for the
4:55 meanings of words. So just giving it a
4:58 string of symbols it just by trying to
4:59 predict the next word in a string of
5:01 symbols it could learn how to convert
5:03 words into sets of features that
5:05 captured the meaning of the word and
5:06 have interactions between those features
5:08 predict the features of the next word.
5:10 >> So that was actually a tiny language
5:13 model from 1980 late 1984
5:17 um that I think of as as a precursor for
5:19 these big language models. The basic
5:20 principles were the same. It was just
5:23 tiny. We had 100 training examples. It
5:24 took 40 years to get us here though.
5:26 >> And it took 40 years to get here. And
5:28 the reason it took 40 years was we
5:29 didn't have the compute and we didn't
5:32 have the data and we didn't know that at
5:34 the time. We couldn't understand why we
5:35 weren't just solving everything with
5:36 back propagation.
5:39 >> Which takes us cleanly to to Jensen. We
5:41 didn't have the compute for 40 years and
5:44 here now you are building it. Tell tell
5:47 us about your moments that of real kind
5:50 of clarity.
5:54 Well, for my career, um, I was the, uh,
5:56 first generation of chip designers that
5:58 was able to use higher level representations
5:59 representations
6:02 and design tools to design chips.
6:07 and and uh that that discovery
6:11 um uh was helpful when I learned about a
6:14 new way of developing software
6:16 uh around the 2010 time frame
6:19 simultaneously from three different labs
6:22 uh what was going on in uh uh University
6:26 of Toronto researchers uh reached out
6:29 reached out to us at the same time that
6:33 uh researchers at the NYU reached out to
6:36 um as well as uh in Stanford reached out
6:39 to us at the same time and I I I saw the
6:41 early indications of what turned out to
6:42 have been deep learning around the same
6:47 time uh using uh a framework uh and a
6:52 structured design to uh create software
6:54 and that software turned out to have
6:57 been incredibly effective.
7:00 Uh and that second that second observation
7:02 observation
7:06 uh is seen again using frameworks rep
7:08 higher level representations
7:11 structured types of uh structures like
7:14 the deep learning networks. I uh was
7:17 able to develop software uh w was very
7:20 similar to designing chips for me and
7:22 the patterns were very similar and I
7:24 realized at that time maybe we could
7:26 develop software uh and capabilities
7:30 that that scale very nicely as we've
7:33 scaled uh chip design over the years and
7:35 so that was that was a quite a quite a
7:36 moment for me
7:37 >> and when do you think was the moment
7:40 when the chips really started to help
7:43 scale up today's sort of the the LLMs
7:45 that we have today because you you said
7:47 2010 that's still a 15 year.
7:50 >> Yeah. The the thing about about Nvidia's
7:52 architecture is is once you're able to
7:55 get something to run well on a GPU
7:57 because it became parallel, you could
8:00 get it to run well on multiple GPUs.
8:04 that same sensibility of scaling uh the
8:06 algorithm to run on many processors on
8:08 one GPU. This is the same logic and the
8:10 same reasoning that you could do it on
8:12 multiple GPUs and then now multiple
8:14 systems and in fact you know multiple
8:17 data centers and so that once we
8:20 realized we could do that effectively
8:23 then then the rest of it is about about
8:26 uh imagining how far you could
8:28 extrapolate this capability. you know,
8:30 how much data do we have? How large can
8:33 the networks be? How much dimensionality
8:35 can it capture? What kind of problems
8:38 can it solve? Uh the all of all of that
8:42 is is really engineering at that point.
8:44 You know, the the observation that that
8:47 uh the deep learn deep learning models
8:50 are so effective uh is is really quite
8:54 the the the spark. The rest of it is
8:56 really engineering extrapolation.
8:59 Fei, tell us about your your moment.
9:02 >> Yeah, I also have two moments to share.
9:04 So around 2006
9:07 and 2007, I was transitioning from a
9:10 graduate student to an a young assistant
9:14 professor and I was among the first
9:16 generation of machine learning graduate
9:20 students um reading papers from young
9:24 Yoshua uh Jeff and I was really obsessed
9:28 in trying to solve the problem of ob uh
9:30 visual recognition which is the ability
9:34 for machines to see meaning in objects
9:38 in everyday pictures and uh we were
9:40 struggling with this problem in machine
9:42 learning called generalizability
9:45 which is um after learning from certain
9:48 number of examples can we recognize
9:52 something a a new example new sample and
9:54 I've tried every single algorithm under
9:57 the sun from baset support vector
9:59 machines to neuronet network and the
10:02 missing piece that my student and I
10:06 realized is that data is missing that uh
10:09 uh you know if you look at the evolution
10:11 or development of uh intelligent animals
10:14 like humans we were inundated with data
10:17 in the early years of development but
10:21 our machines were starved with data. So
10:24 we um decided to do something crazy at
10:26 that time to create a internet scale
10:30 data set uh over the course of three
10:34 years called imageet that uh uh in uh
10:38 included 15 million images handcurated
10:41 um by by people around the world across
10:45 22,000 categories. So, so for me the aha
10:48 moment at that point is big data drives
10:50 machine learning
10:51 >> and it's now it's it's the limiting
10:54 factor the building block of all of the
10:56 you know algorithms that we're seeing with
10:56 with
10:58 >> yeah it's part of the scaling law of
11:02 today's AI and the second aha moment is
11:05 um 2018
11:08 I was the first chief scientist of uh AI
11:13 at Google cloud uh part of the the work
11:16 we do is serving all vertical industries
11:18 under the sun, right? From healthcare to
11:21 financial services, from entertainment
11:24 to uh manufacturing, from agriculture to energy.
11:26 energy.
11:30 And that was a few years after the the
11:32 what we call the image that Alex moment,
11:36 a couple of years after Alph Go, and I realized
11:37 realized
11:39 >> Alph Go being the algorithm that was
11:40 able to beat humans at playing the
11:44 Chinese board game Go. Uh yes and as the
11:47 chief scientist at Google I realized
11:49 this is a civilizational technology
11:52 that's going to impact every single
11:55 human individual as well as sector of
11:59 business and uh if humanity is going to
12:02 go enter an AI era what is the guiding
12:06 framework so that we not only innovate
12:09 but we also bring benevolence
12:11 to uh through this powerful technology.
12:14 technology to everybody and that's when
12:17 I returned to Stanford as a professor to
12:20 uh co co-found the human center AI
12:23 institute and and uh propose the human-
12:28 center AI framework so that we can keep
12:30 humanity and human values in the center
12:32 of this uh technology.
12:33 >> So developing but also looking at the
12:35 impact and what's next which is where
12:37 the rest of us come in.
12:39 >> Um Yan do you want to round us out here?
12:41 What's what's been your highlight?
12:44 >> Yeah, probably go back a long time. Um,
12:46 I realized when I was in undergrad, I
12:48 was fascinated by the question of AI and
12:50 intelligence more generally and
12:52 discovered that people in the 50s and
12:56 60s that worked on
12:58 training machines instead of programming
13:00 them. I was really fascinated by this
13:02 idea probably because I thought I was
13:04 either too stupid or too lazy to
13:05 actually build an intelligent machine
13:07 from scratch, right? So it's better to
13:11 let itself be um like train itself or
13:12 self-organized and that's the way you
13:16 know intelligence in in in life uh
13:18 builds itself. It's uh it's
13:20 selforganized. So I I thought this
13:22 concept was really fascinating and I
13:24 couldn't find anybody when I graduated
13:25 from engineering. I was doing chip
13:28 design by the way um wanted to go to
13:30 grad school. I couldn't find anybody who
13:32 was uh working on this but connected
13:33 with some people who kind of were
13:36 interested in this and discovered Jeff's
13:39 papers for example uh and uh he was the
13:41 person in the world I wanted to meet
13:45 most in 1983 when I started grad school
13:48 and we eventually met two years later um and
13:49 and
13:51 >> and today you're friends would you say?
13:54 >> Yes. Oh, we we we we had lunch together
13:58 in 1985 and we could finish each other's
14:02 sentences. Basically, he had uh
14:04 um I had a a paper written in French at
14:05 a conference where he was a keynote
14:08 speaker and and managed to actually kind
14:10 of decipher the the math. It was kind of
14:12 sort of like back propagation a little
14:14 bit to train multi-layer nets. It was
14:17 known from the 60s that the limitation
14:19 of machine learning was due to the fact
14:22 that we could not train machine with
14:23 multiple layers. So that was really my
14:26 obsession and it was his obsession too
14:29 and um and so I had a paper that kind of
14:31 proposed some some way of doing it and
14:32 he kind of managed to read the math. So
14:35 that's how we hooked up and
14:36 >> and that's what has set you on this path.
14:36 path.
14:39 >> Right. So and and then after that you
14:41 know once you can you can train complex
14:43 systems like this you ask yourself
14:45 questions. So how do I build them so
14:46 they do something useful like
14:48 recognizing images or things of that
14:51 type? And at at the time Jeff and I had
14:53 this debate when I was a postoc with him
14:57 in the late 80s. Um I I I thought um the
14:59 only machine learning paradigm that was
15:01 well formulated was supervised running.
15:03 You you show an image to the machine and
15:05 you tell it what the answer is, right?
15:07 And he said no no no like the only way
15:09 we're going to get to make progress is
15:11 through unsupervised running. And I was
15:15 kind of dismissing this at the time. Um,
15:17 and what happened in you know the mid
15:21 2000 when he Yosha and I sort of start
15:24 getting together and restart the
15:26 interest of the of the community in deep
15:29 learning. We actually kind of uh made
15:33 our bet on unsupervised learning or self
15:35 reinforcement loop. Right?
15:37 >> This is not reinforcement. So this is
15:39 basically discovering the structure in
15:41 data without training the machine to do
15:43 any particular task which is by the way
15:45 the way LLMs are trained. So an LLM is
15:48 trained to predict the next word but
15:49 it's not really a task. It's just a way
15:52 for the system to learn a good kind of
15:54 uh representation or capture the
15:57 >> is there no reward system there that
15:59 sorry to get geeky but is there no
16:00 nothing to say this is correct and
16:02 therefore keep doing it because
16:03 >> well this is correct if you predict the
16:06 next word correctly right
16:07 >> from the rewards in reinforcement
16:08 learning where you say that's good
16:10 >> yeah okay
16:14 >> um and so in fact uh I'm going to blame
16:18 it on you uh it turns out produced this
16:22 big data set called imageet and uh which
16:24 is which was labeled and so we could use
16:26 supervised learning to train the systems
16:28 on and that turned out to work actually
16:31 much better than we expected and so we
16:33 temporarily abandoned the whole program
16:34 of working on self-supervised
16:36 unsupervised learning because supervised
16:38 learning was working so well we figured
16:40 out a few tricks
16:41 >> Joshua stuck with it
16:43 >> I said I didn't
16:46 >> no you didn't I didn't either but uh but
16:48 it it kind of
16:50 refocus the entire industry and and the
16:52 research community if you want on sort
16:54 of deep deep learning supervised
16:55 learning etc. Mhm.
16:58 >> And it it it took another few years
17:00 maybe around 201617
17:04 to uh tell people like this is not going
17:06 to tell take us where we want. We need
17:08 to do self-s supervised learning now and
17:10 that's what LLM really are the best
17:11 example of this. >> Okay.
17:11 >> Okay.
17:14 >> But uh what we're working on now is
17:16 applying this to other types of data
17:18 like like video sensor data which LLM
17:21 are really not very good at at all. Um
17:22 and that's a new challenge for the next
17:24 few years. So that brings us actually to
17:26 the present moment and I think you know
17:29 you'll all have seen this crest of the
17:31 interest from people who had no idea
17:33 what AI was before who had no interest
17:34 in it and now everybody's flocking to
17:36 this and this has become more than a
17:38 technical innovation right that's a huge
17:41 business boom it's become a geopolitical
17:44 strategy issue um and you know
17:46 everybody's trying to get their hands
17:48 around what this is so or their heads
17:50 around it Jensen I'll come to you here
17:52 first to I want you all
17:56 to reflect on this moment now here
17:58 Nvidia in particular has it's basically
18:01 in the news every day hour week you know
18:02 and you have become the most valuable
18:04 company in the world so there's
18:06 something there that people want
18:08 >> you'll be to hear that
18:11 >> yeah you know tell us about do are you
18:13 worried that we are getting to the point
18:16 where people don't quite understand and
18:18 we're all getting ahead of ourselves and
18:19 there's going to be a reckoning that
18:20 there's a bubble that's going to burst
18:22 and then it will write itself self and
18:25 if not what is the kind of biggest
18:27 misconception about demand coming from
18:30 AI that is different to say the dotcom
18:32 era or that people don't understand you
18:34 know if if that's not the case
18:37 >> uh during the dotcom era during the the
18:43 bubble the vast majority of the fiber
18:46 deployed were dark
18:48 meaning the industry deployed a lot more
18:51 fiber than it needed Mhm.
18:55 >> Today almost every GPU you could find is
18:58 lit up and used.
19:01 And so uh the reason why I think it's
19:04 important to take a take a step back and
19:07 understand and understand what AI is,
19:09 you know, for a lot of people AI is Chad
19:12 GBT and it's image generation and and it
19:14 that's all true. That's one of the
19:17 applications of it. Um, and AI has
19:18 advanced tremendously in the last
19:21 several years. The ability to not just
19:24 memorize and generalize, but to reason
19:27 and effectively think and ground itself
19:29 through research. It's able to produce
19:31 answers and do things that are much more
19:35 valuable now. It's much more effective.
19:37 and the number of companies that are
19:39 able to build businesses that are that
19:41 are helpful to other businesses. For
19:44 example, a software programming company,
19:46 an AI software company that that we use
19:49 called Cursor, uh they're very
19:51 profitable and we use their software
19:54 tremendously and it's incredibly useful.
19:57 uh or a bridged or open evidence who are
19:59 uh serving the healthcare industry doing
20:01 very very well producing really good
20:05 results and and so so the AI capability
20:08 has grown so much and as a result we
20:09 were seeing these two exponentials that
20:11 are happening at the same time on the
20:12 one hand the amount of computation
20:15 necessary to produce an answer has grown
20:17 tremendously on the other hand the
20:20 amount of usage of these AI models are
20:23 growing also exponentially these two exponentials
20:24 exponentials
20:28 are causing a lot of demand on compute.
20:30 Now when you take a step back, you ask
20:33 yourself fundamentally what's different
20:35 between AI today and the software
20:38 industry of the past. Well, software in
20:40 the past was pre-ompiled
20:42 and the amount of computation necessary
20:45 for the software is not very high.
20:47 >> But in order for AI to be effective, it
20:49 has to be contextually aware. It has to
20:51 it can only produce the intelligence at
20:53 the moment. You can't produce it in
20:56 advance and retrieve it. That's you know
20:59 that's called content. AI intelligence
21:01 has to be produced and generated in real
21:04 time. And so as a result we now have an
21:08 industry where the computation necessary
21:09 to produce something that's really
21:12 valuable in high demand is quite
21:15 substantial. We have created an an
21:17 industry that requires factories. That's
21:20 why I I remind ourselves that AI needs
21:23 factories to produce these tokens to
21:25 produce the intelligence and this is
21:28 this is a a once you know once in a it's
21:30 never happened before where the computer
21:33 is actually part of a factory and and so
21:35 we need hundreds of billions of dollars
21:37 of these factories in order to serve the
21:39 trillions of dollars of industries that
21:42 sits on top of intelligence. You know,
21:44 you go come back and take a look at at
21:47 software in the past. Software in the
21:49 past is they're software tools. They're
21:53 used by people. For the first time, AI
21:56 is intelligence that augments people.
21:59 And so, it addresses labor. It addresses
22:00 work. It does work.
22:01 >> So, you're saying no, this is not a bubble.
22:02 bubble.
22:03 >> I think this we're we're well in the
22:04 beginning of the buildout of
22:08 intelligence. And and the fact of the
22:10 matter is most people still don't use AI
22:13 today. And someday in the near future,
22:16 almost everything we do, you know, every
22:18 moment of the day, you're going to be
22:20 engaging AI somehow. And so between
22:22 where we are today where the usage is
22:24 quite low to where we will be someday
22:27 where the usage is basically continuous,
22:29 that buildout is is you know what
22:32 >> and if even if the LLM runway runs out,
22:34 you think GPUs and the infrastructure
22:36 you're building can still be of use in a
22:37 different paradigm and then I want to
22:39 open up to others to talk. LLM is a is a
22:42 piece of the AI technology. You know,
22:46 AIS are systems of models, not just LLMs
22:48 and LLM are big part of it, but there
22:51 are systems of models and and uh the the
22:53 technology necessary for for AI to be
22:55 much more productive from where where it
22:57 is today irrespective of what we call
23:00 it. Um we still have a lot of technology
23:02 to develop yet.
23:05 >> Can who wants to jump in on on this?
23:05 >> Um I don't think
23:08 >> especially if you disagree. I don't
23:11 think we should call them LLMs anymore.
23:13 Um they're not language models anymore.
23:15 They they >> right
23:15 >> right
23:17 >> start as language models at least that's
23:20 the pre-training but but more recently
23:22 there's been a lot of advances in making
23:25 them agents. In other words, uh go
23:26 through a sequence of steps in order to
23:28 achieve something interactively with an
23:30 environment with people right now
23:32 through a dialogue but more and more
23:34 with a computing infrastructure.
23:37 And the technology is changing. It's not
23:39 at all the same thing as what it was
23:42 three years ago. I don't think we can
23:46 predict where the technology will be in
23:49 two years, 5 years, 10 years. U but we
23:52 can see the trend. So one of the things
23:55 I'm doing is trying to uh bring together
23:58 a group of international experts to keep
24:01 track of what's happening with AI where
24:03 it is going um what are the risks how
24:06 are they being mitigated and and and and
24:08 the trends are very clear across so many
24:11 benchmarks now you know because we've
24:13 had so much success in improving the technology
24:15 technology
24:16 uh in the past it doesn't mean that's
24:18 going to be the same in the future. So
24:21 then then there would be financial uh
24:24 consequences uh if the expectations are
24:27 not met but in the long run I completely
24:28 agree. Um
24:30 >> but currently what about the rest of
24:32 you? Do you think that the valuations
24:34 are justified in terms of what you know
24:37 about the technology the applications?
24:38 >> So I think there are three trends that
24:40 sort of explain what's going on. The
24:42 first is the models are getting more
24:44 efficient. If you look just at attention
24:45 for example, going from straight
24:48 attention to GQA to MLA, you get the
24:50 same or better results with far less
24:52 computation. And so that then drives
24:55 demand in ways where things that may
24:56 have been too expensive before become
24:58 inexpensive of now. You can do more with
25:00 AI. At the same time, the models are
25:02 getting better and you know, maybe
25:03 they'll continue to get better with
25:05 transformers or maybe a new architecture
25:07 will come along, but we will we won't go
25:08 backwards. We're going to continue to
25:10 have better models that also
25:12 >> they still need GPUs even if
25:13 >> absolutely transformer based
25:15 >> um in fact it makes it makes them much
25:16 more valuable compared to more
25:18 specialized things because they're more
25:20 flexible and they can evolve with the
25:22 models better but then the final thing
25:24 is I think we've just begun to scratch
25:26 the surface on applications so almost
25:28 every aspect of human life can be made
25:31 better by having AI you know assist
25:33 somebody in their profession help them
25:35 in their daily lives and you know I
25:37 think we've you know started to reach
25:39 maybe 1% of the ultimate demand for
25:42 this. So as that expands, you know, the,
25:43 you know, number of uses of this are
25:45 going to go up. So I don't think there's
25:46 any bubble here. I think we're, like
25:48 Jensen said, we're riding a multiple
25:50 exponential and we're at the very
25:51 beginning of it and it's going to just
25:51 keep going.
25:53 >> And in some ways, Nvidia is in to that
25:56 because even if this paradigm changes
25:58 and there's other types of AI and other
25:59 architectures, you're still going to
26:02 need the the atoms underneath. So that
26:03 makes sense for you. Did you want to
26:07 jump in Fay? Uh yeah, I do think that um
26:09 of course from a market point of view,
26:12 it will have its own um dynamics and
26:14 sometimes it does adjust itself, but if
26:17 you look at the long-term trend, let's
26:20 not forget AI by and large is still a
26:22 very young field, right? We walk into
26:24 this room and on the wall there were
26:27 equations of physics. Physics has been a
26:29 more than 400 year old uh discipline.
26:32 Even if we look at uh modern physics and
26:36 AI is less than 70 years old if we go
26:40 back to Alan Turing you that's about 75
26:43 years so there is a lot more new
26:46 frontiers that is to come uh you know
26:48 Jensen and Yoshua talk about LLMs and
26:51 agents those are more languagebased but
26:54 even if you do uh self uh introspection
26:58 of human intelligence there's more
27:00 intelligent capabilities is beyond
27:02 language. I have been working on spatial
27:04 intelligence which is really the
27:07 combination or the lynchpin between
27:11 perception and action where um where uh
27:13 you know humans and animals have
27:18 incredible ability to perceive reason
27:22 interact with and uh and create uh
27:25 worlds that goes far beyond language.
27:27 And even today's most powerful
27:30 language-based uh or LLM based models uh
27:33 fail at rudimentary spatial intelligence
27:36 uh tests. So from that point of view as
27:40 a as a discipline as a science there's
27:44 far more frontiers to conquer and to uh
27:47 open up and that brings the applications
27:49 uh you know opens up more applications.
27:51 >> Yeah. and you work at a company and so
27:53 you have the kind of dual perspective of
27:54 being a researcher and working in a
27:56 commercial space. Do you agree? Do you
27:58 do you believe that this is all
28:01 justified and you can see where this is
28:02 all coming from or do you think we're
28:04 reaching an end here and we need to find
28:05 a new path?
28:06 >> So I think there are several point of
28:10 views for which uh we're not in a bubble
28:13 and at least one point of view
28:14 suggesting that we we are in a bubble
28:16 but there is but it's a different thing.
28:18 So we're not in a bubble in the sense
28:21 that um there are a lot of applications
28:23 to develop based on LLMs. LLM is the
28:25 current dominant paradigm and there's a
28:28 lot to uh milk there. This is you know
28:31 what Bill was was saying to kind of help
28:33 people in the daily lives with current
28:35 technology that technology needs to be
28:38 pushed and that justifies all the
28:41 investment that is done on the software
28:42 side and also on the infrastructure
28:45 side. uh once we have you know smart
28:48 wearable devices um in everybody's hands
28:50 assisting them in their daily lives the
28:52 amount of computation that would be
28:55 required as as Jensen was saying to uh
28:57 to serve all those all those people is
28:59 going to be enormous so in that sense
29:01 the investment is not is not wasted but
29:02 there is a sense in which there is a
29:04 bubble and it's the idea somehow that
29:06 the current paradigm of LLM would be
29:09 pushed to the point of having human
29:11 level intelligence which I personally
29:13 don't believe in and you don't either And
29:15 And
29:17 we we need kind of a few breakthroughs
29:19 before we get to machines that really
29:21 have the kind of intelligence we observe
29:23 not just in humans but also animals. We
29:25 don't have robots that are nearly as
29:27 smart as a cat, right? Um and so we're
29:30 missing something big still. Which is
29:33 why AI progress is not just a question
29:36 of more infrastructure, more data, uh
29:38 more investment and more development of
29:39 the current paradigm. It's actually a
29:41 scientific question of how do we make
29:44 progress towards the next generation of AI
29:44 AI
29:46 >> which is why all of you are here right
29:48 because you actually sparked the entire
29:49 thing off and I feel like you know we're
29:51 moving much towards the engineering
29:52 application side but what you're saying
29:54 is we need to come back to what brought
29:57 you here originally um on that question
29:58 of human level intelligence we don't
30:00 have long left so I just want to do a
30:02 quick fire I'm curious can each of you
30:04 say how long you think it will take
30:06 until we do reach that point where you
30:08 believe we're you know equivalent
30:10 machine intelligence to a human or even
30:12 a clever animal like an octopus or
30:16 whatever. How far away are we just just
30:17 the years?
30:19 >> It's not going to be an event. >> Okay.
30:20 >> Okay.
30:22 >> Okay. Because the capabilities are going
30:24 to expand progressively in various domains.
30:25 domains.
30:26 >> Over what time periods?
30:28 >> Over, you know, maybe we'll make some
30:30 significant progress over the next five
30:32 to 10 years to come up with a new paradigm.
30:32 paradigm.
30:35 >> F and then maybe, you know, progress
30:36 will come. But it'll it'll take longer
30:39 than we think. Okay. Parts of machines
30:42 will supersede human intelligence and
30:44 part of the machine intelligence will
30:47 never be similar um or the same as human
30:51 intelligence. They are they are they're
30:53 built for different purposes and they will
30:53 will
30:56 >> when do we get to superseding?
30:58 >> Part of it is already here. How many of
31:02 us can recognize 22,000 objects in the
31:03 world? So part of
31:05 >> do you not think an adult human can
31:07 recognize 22,000 objects?
31:10 >> Um the kind of granularity and fidelity.
31:13 No. How many adult humans can translate
31:14 a 100 languages?
31:15 >> That's harder. Yeah.
31:16 >> So yeah.
31:19 >> So I think we should be nuanced and
31:22 grounded in scientific facts that uh
31:26 just like airplanes fly but they don't
31:29 fly like birds. and u machine-based
31:31 intelligence will do a lot of powerful
31:34 things but there is a profound
31:38 um place for human intelligence to to
31:41 always be critical in our human society.
31:42 Jensen, do you have
31:46 >> we have enough general intelligence to
31:49 uh translate the technology to an
31:54 enormous amount of uh society useful applications
31:56 applications
31:59 uh in the next coming years and with
32:00 respect to >> Yeah.
32:00 >> Yeah.
32:02 >> Yeah. Yeah. We're doing it today.
32:04 >> Yeah. And so I think I think uh one
32:05 we're already there
32:07 >> and two the the other part of the answer
32:09 is it doesn't matter
32:12 >> because at this point it's a bit of an
32:14 academic question. We're going to apply
32:16 the technology to and the technology is
32:17 going to keep on getting better and
32:19 we're going to apply the technology to
32:21 solve a lot of very important things
32:23 from this point forward. And so okay
32:24 >> I I think the answer is it doesn't matter
32:25 matter
32:27 >> and and it's now as well.
32:30 >> Yeah you decide. Right. If you refine
32:31 the question a bit to say how long
32:34 before if you have a debate with this
32:36 machine it'll always win.
32:39 >> I think that's definitely coming within
32:41 20 years. We're not there yet but I
32:43 think fairly definitely within 20 years
32:45 we'll have that. So if you define that as
32:46 as
32:48 >> AGI it'll always win a debate with you.
32:50 >> We're going to get we're going to get
32:52 there in less than 20 years probably.
32:53 >> Okay. Bill, do you have
32:55 >> Yeah. Well, I'm sort of with Jensen that
32:56 it's the wrong question, right? Because
32:58 our goal is not to build AI to replace
33:00 humans or to be better than humans.
33:02 >> But it's a scientific question. It's not
33:04 that we'll replace humans. The question
33:07 is could we as as a society build something?
33:08 something?
33:09 >> But our goal is to build AI to augment
33:12 humans. And so what we want to do is
33:14 complement what what humans are good at.
33:16 Humans can't recognize 22,000 categories
33:18 or most of us can't solve these math
33:20 olympiad problems. Um so we build AI to
33:22 do that. So humans can do what is
33:24 uniquely human, which is be creative and
33:27 be empathetic and and understand how to
33:29 interact with other people in our world.
33:31 And I think that it's not clear to me
33:34 that AI will ever do that, but AI can be
33:36 huge assistance to humans.
33:39 >> So I'll beg to differ on this. Uh I
33:41 don't see any reason why at some point
33:42 we wouldn't be able to build machines
33:44 that can do pretty much everything we
33:47 can do. Um, of course, for now on the
33:49 spatial and you know, robotic side, it's
33:51 lagging, but there's no like uh
33:53 conceptual reason why we couldn't. So on
33:56 on the timeline, I think there's a lot
33:58 of uncertainty and that we should plan
34:01 accordingly. Um, but there is some data
34:05 that I find interesting where we see um
34:07 the capability of AI to plan over
34:10 different horizons to grow exponentially
34:13 fast in the last six years. And if we
34:16 continue that that trend, it would place
34:19 roughly the level that an employee has
34:22 in their job to uh AI being able to do
34:24 it within about five years. Now this is
34:27 only one category of engineering tasks
34:28 and there are many other things that
34:31 matter. For example, uh one thing that
34:33 could change the game that is that many
34:36 companies are aiming to just to focus on
34:38 the ability of AI to do AI research. In
34:40 other words, to do engineering, to do
34:43 computer science, and to design the next
34:44 generation of AI, including maybe
34:46 improving robotics and spatial
34:48 understanding. So, I'm not saying it
34:50 will happen, but the area of ability of
34:52 AI to do better and better programming
34:54 and understanding of algorithms that is
34:56 going very very fast and that could
34:58 unlock many other things. We don't know
35:00 and we should we should be really
35:02 agnostic and not make big claims because
35:04 there's a lot of possible futures there.
35:08 M so so our consensus is in some ways we
35:10 think that future is here today but
35:13 there's never going to be one moment and
35:15 the job of you all here today has helped
35:17 to guide us along this route um until we
35:18 get to a point where we're working
35:21 alongside these systems. Very excited
35:22 personally to see where we're going to
35:24 go with this. If we do this again in a
35:26 year it'll be a different world. But
35:27 thank you so much for joining us for
35:30 sharing your stories and for talking us
35:31 through this this huge kind of
35:34 revolutionary moment. Thank you. Thank you.
35:34 you.
