0:04 welcome to a another conversation on AI
0:07 I don't think we get enough of this
0:09 conversation going I do want to thank
0:11 Richard AAS and the FI team for really
0:16 increasing the conversation this year on
0:18 AI because I think there is no greater
0:22 uh topic of import on the financial side
0:26 on the leadership side education Side
0:28 Medical side it's transforming
0:30 everything we have uh three incredible
0:33 CEOs here um who are representing uh
0:38 variety of different parts of the AI
0:41 emergence um I'm going to start by
0:43 asking each of them to take just one
0:45 minute introduce themselves and what
0:47 they're doing and then we're going to
0:49 jump into where is this going how fast
0:51 is it going how big is it going to get
0:54 you know we'll ask the question what is
0:55 after chat GPT Prem let's begin with
0:58 yourself awesome thank you you thank you
1:00 Peter um I'm pre maraju I'm the CEO of
1:04 stability AI we are one of the leading
1:07 open- Source image video and 3D um
1:11 models in in the world and past GPT
1:16 pictures are worth a thousand words and
1:18 we're making quite a few of them and in
1:19 fact 80% of all the images that were
1:24 generated by AI last year in 2023 were
1:26 driven by our model stable diffusion
1:29 amazing Richard hi everyone really
1:31 excited to be here my name is Richard
1:33 soer I'm the CEO and founder of u.com
1:36 y.com it's a productivity engine which
1:38 is the next Generation after a search
1:40 and an answer engine so we really make
1:43 people more productive across a whole
1:45 host of different kinds of organizations
1:47 from hedge funds to universities to
1:49 companies insurance companies and so on
1:51 Publishers news agencies and uh almost
1:54 everyone else in between who has sales
1:56 service marketing research analysis uh
1:59 and so on I also run a venture fund
2:01 called aiex Ventures that invests in
2:04 early stage uh preed seed AI companies
2:06 and startups been very fortunate that
2:09 when I was a professor at Stanford I had
2:10 two students who created this cute
2:12 company called hugging face invested in
2:14 that a 5 million valuation worth four
2:16 and a half billion now so Fun's doing
2:18 that's that's bragging that's just
2:20 straight up bragging I wish I could brag
2:23 like that Dr Kaiu
2:25 Lee hi uh I've been working on AI for
2:28 about 43 years I was two at no at the uh
2:33 in college when I started Ai and um I
2:35 think that may have started before my
2:37 colleagues were born but um I actually
2:39 worked on machine learning and um and I
2:43 have a PhD carnegi melon and I have
2:45 worked at uh Apple Microsoft Google uh
2:48 some of you may know me as with my books
2:51 AI superpowers and AI
2:53 2041 uh my part-time job is I I run
2:56 sinovation Ventures which invests
2:58 globally and then my full-time job is I
3:02 run 0 one. it's a uh gener generative AI
3:07 company uh we build a large language
3:09 model we're currently ranked as the
3:11 third company with the highest
3:12 performance only next to the best models
3:15 from open and Google um and you can find
3:18 it online uh we're also building uh
3:20 consumer and Enterprise Products um
3:23 we're based in China uh but our products
3:26 are accessible globally and also we
3:30 extensively uh do open source as well so
3:34 incredible and and first of all Kaiu is
3:36 a legend and one of the greatest leaders
3:38 globally in this field so very honored
3:40 to have him on here um Prem I want to
3:43 start with you um uh you very famously
3:48 were able to recruit James Cameron onto
3:51 your board uh and since stability is
3:54 creating video and is creating sort of
3:57 the future of
3:58 Hollywood um
4:00 I am curious about two things one did uh
4:05 did Jim get it right with the
4:09 Terminator um and and secondly uh you
4:14 know there's been a lot of conversation
4:16 about the disruption of Hollywood um
4:19 that we're going to have AIS creating
4:21 the future of all movies all content and
4:23 so forth so you said beyond you know GPT
4:26 models were you know images worth of
4:29 th000 words talk to us about what this
4:32 what this future is what is going to
4:35 happen in sort of the visualization
4:36 world of of TV and Hollywood love it so
4:39 did Jim get it right with Terminator uh
4:42 let's hope not I guess but the um but
4:45 what a great movie it was and I love
4:47 when he actually he jokes about it he
4:48 says I told you guys like you know this
4:50 is coming and now it absolutely is here
4:53 um and why did why would someone like
4:55 him get involved in stability yeah great
4:58 question so you I I had the great
5:00 Fortune of of working on Avatar 2 with
5:02 him when I was the CEO of wether digital
5:04 before I joined as CEO of stability and
5:07 that movie took over four years to make
5:10 and that's because it was fully rendered
5:13 and I think if you fast forward to 5 to
5:16 10 years from now the vast majority of
5:20 film and television and visual media as
5:22 we know it today is not going to be
5:23 render is going to be generated and in
5:26 fact in Avatar there were certain shots
5:29 there were certain uh that took 6,000
5:32 7,000 hours of compute time to render
5:35 one single frame thousands of hours that
5:38 literally can be reduced down to minutes
5:40 now so I think Jim just wants a whole
5:43 lot of life back and when you think
5:46 about like the creative process we all
5:48 watch films we watch movies we love them
5:50 from the time we've born to our last
5:52 memory it's some it's a commodity we
5:54 never get sick of um we never we never
5:58 not want to watch it
6:00 um and so there's this insatiable
6:02 appetite out there in the world to
6:04 consume stories and to create stories
6:07 and I think that we should just
6:08 accelerate that the problem with the
6:10 film production process is time and
6:11 money so what he really wanted to do is
6:13 rip those things out so we can move from
6:15 a render to a generated model are we
6:18 going to see a situation where we're
6:20 ever going to have ai generating entire
6:23 movies because it knows my preferences
6:26 what I love and it's like the perfect
6:28 movie for me
6:30 you know personally I kind of hope not
6:33 um I don't think that
6:35 actually uh the creative process I think
6:38 needs to start with a human and I think
6:40 that human needs to dictate these tools
6:42 in separate agents to actually make that
6:44 story and so I'm hoping that you'll
6:48 probably want to hear stories that other
6:49 people want to tell you all right well
6:51 let's take a different direction then
6:52 sure am I going to see uh Marilyn Monroe
6:56 and you know all stars of the Past
6:58 coming back
7:00 is there a need for human actors if you
7:02 can generate absolutely lifelike uh
7:05 actors and actresses perfectly I mean I
7:08 can't see a situation where they're
7:09 still around yeah I think that it's
7:12 actually quite it's faster when you're
7:14 talking about the Film Production it's
7:15 actually easier to just shoot plates on
7:18 an actor just shoot real photography and
7:20 get their performance I think there's
7:22 that's the visible layer of of
7:23 production people gravitate toward it a
7:25 lot I think that AI will enhance those
7:27 prod those um performances I think the
7:30 physicality of a director with a camera
7:33 and an actor in front of it is a very
7:34 important part of the creative process
7:37 and I don't think that that's going to
7:38 go away too soon and in fact I think
7:42 about the things that aren't going to
7:43 change just as much as I think is going
7:44 to change but I do think after they take
7:47 one take the director is going to say I
7:49 got it because they're going to be able
7:50 to do what you're talking about which is
7:51 manipulate that performance may ask one
7:53 more question to you before I move on
7:55 what is the most dramatic change we're
7:57 going to see in film and TV 10 years
7:59 years from now as we see digital super
8:01 intelligence we like what's what's the
8:03 craziest vision of what we're going to
8:05 see in entertainment I think we're going
8:07 to see on the magnitude of 5 to 10 to
8:10 20x more content being created I think
8:12 we're going to see a variation of time
8:16 where it's going to be a two-minute like
8:17 you said you may want to have 20 minutes
8:19 before you go to bed you want to see a
8:20 movie that that's you'll have different
8:22 type of time signatures and I think that
8:24 you're going to have an explosion of
8:26 content creation an explosion of number
8:28 of artists in the world
8:31 I'm going to come back in 10 years and
8:33 see if you're right about that
8:35 okay uh Richard
8:37 um a lot of your work was instrumental
8:41 in the early days of bringing neural
8:43 Nets to natural language
8:45 processing um so what do you see as the
8:48 next Frontier Beyond NLP so just explain
8:52 if you would what NLP is and where is it
8:54 going next yeah natural language
8:57 processing NLP used to be a a sub area
9:00 of AI and it has I think influenced
9:03 pretty much every other area of AI and
9:05 uh there lots of different algorithms
9:07 you could train and 2010 I had this
9:09 crazy idea to train a single neural
9:11 network for all of NLP and 2018 we
9:14 finally really built the first model uh
9:16 that invented prompt engineering where
9:18 you can just ask one model all the
9:20 different questions you have and over
9:22 time of course you can ask questions not
9:23 just over text but also over images and
9:25 so I think next one of the answers to
9:27 the the panel's main topic of what's
9:29 after chat gbt is that we have many more
9:32 multimodal models you'll be able to have
9:34 conversations over images you have
9:36 seamless inputs and outputs in not just
9:39 the modality of text but also
9:40 programming which is a huge unlock uh
9:43 visual videos images voice sound but one
9:47 really interesting modality that not
9:49 many people have quite realized yet is
9:52 that of proteins proteins are
9:54 essentially the basic Lego blocks of all
9:56 of biology everything in our body is
9:58 governed by prot proteins and you can
10:00 create a protein just like you can ask a
10:04 large language model to write AET for
10:06 you or a poem for your wife you can ask
10:09 an llm to create a specific kind of
10:10 protein it will only bind to SARS Cove 2
10:14 or only bind to a specific type of
10:17 cancer in your brain and what that means
10:21 is that we will unlock a lot of
10:22 different aspects in medicine so I'm
10:24 extremely excited about the future of
10:25 LMS going into different modalities and
10:27 we're seeing that with Deep Mind
10:29 products in you in Alpha proteo and and
10:33 such so we had a conversation in back
10:36 but I didn't hear the answer and the
10:39 question is basically is there an upper
10:41 limit to
10:44 intelligence and you know we've talked
10:47 about and we just did a conclave on
10:49 digital
10:50 superintelligence and how fast we're
10:52 going to get there and what does it mean
10:55 um as we think about AI becoming more
10:57 and more intelligent yes I want speak to
10:59 Elon he said okay 2029 2030 equal to
11:02 intelligence to the entire human
11:04 race is it just you know a million times
11:08 more and then a billion times more and
11:10 then a trillion times more is there an
11:12 upper limit to intelligence yeah so
11:14 really interesting question so just to
11:17 talk about Alpha fold and Google for a
11:18 second as you mentioned it like that was
11:20 really interesting uh to understand how
11:22 proteins fold because that will help you
11:25 understand how they are likely to
11:26 function interact in your body what we
11:28 did in 2020 is create the first LM that
11:30 generates a completely new kind of
11:32 protein and it was 40% different uh to
11:35 any naturally occurring protein and it
11:37 actually we synthesized it in the wet
11:39 lab this was at Salesforce research did
11:41 scientist there and it was an
11:42 antibacterial Lio type of protein that
11:44 is basically has antibacterial
11:46 properties and just to put that into
11:47 perspective was really close to covid-19
11:50 so make sure you weren't um got to be
11:54 careful what you say online sometimes um
11:56 but what was interesting is that
11:58 multiple startups have now started from
12:00 this line of research and and I think
12:02 it's hard for people to Fathom like how
12:04 much that can change medicine in terms
12:06 of upper bounds of intelligence it's a
12:08 really interesting question can it just
12:10 keep going and going going I think you
12:12 have to basically look at the different
12:14 dimensions of intelligence right there's
12:17 language intelligence visual perception
12:19 intelligence reasoning knowledge
12:22 extraction uh and a few others physical
12:25 manipulation and just I'll show you just
12:28 one example I don't want to talk talk
12:30 about this for hours but visual
12:32 intelligence right there are you know
12:34 for a long time people have looked at
12:35 just the electromagnetic uh frequency
12:37 spectrum of human vision and there you
12:40 know classifying every object on the
12:42 planet is actually not that hard and the
12:44 upper limit is classifying all the
12:46 objects um on the planet and we're
12:48 probably going to reach that and we're
12:49 not too far away from it but that's just
12:51 human Vision AI could eventually see all
12:55 the way down to gamma frequencies and
12:56 see and try to perceive atoms right and
12:59 there you actually start to hit limits
13:02 uh of physics like Quantum limits of
13:05 like what can actually be observable and
13:06 you can go all the way into like seeing
13:10 uh massively larger scale things at the
13:13 universe level and how many uh different
13:16 sensors do you have in that then you can
13:18 process all of that information and AI
13:21 could have billions of uh sensors that
13:24 go out and then you get into really
13:26 interesting limits of like the speed of
13:27 light cone of like so I can talk about
13:30 for hours it's a really tough subject
13:32 but in some cases we are astronomically
13:34 far away from those upper bounds and in
13:36 some cases we already got pretty close
13:38 fasc you talk about work productivity as
13:42 U.C com's objective what does that mean
13:45 and uh I guess the question is the same
13:48 is there any limitation on work
13:51 productivity that we're going to
13:52 see given the fact that I can command AI
13:56 agents and robots to just do anything
13:58 and everything and just and self-improve
14:00 along the way it seems like we're going
14:02 to hit sort of an infinite GDP at some
14:05 point yeah there there's some areas of
14:07 AI where AI can actually get into a
14:11 self-training loop if there's a
14:13 simulation of something that and
14:15 anything that can be simulated AI can
14:17 solve everything in that areas for
14:20 instance chess the game of Go you can
14:23 perfectly simulate it hence the I can
14:25 train and play with itself billions and
14:27 billions of times cre almost infinite
14:29 amounts of training data and hence solve
14:32 every problem in that domain what are
14:34 other domains that we can perfectly
14:35 simulate is programming if you can
14:37 programming languages can be run and
14:40 then you can simulate the outputs
14:41 obviously in the computer and then the
14:43 AI can get better and better and
14:44 eventually get super human uh in terms
14:47 of programming but where I can't
14:49 simulate things uh infinitely many times
14:51 is in like customer service right you
14:53 can have billions and billions of
14:54 customers kind of ask about all the
14:56 different things that uh can go wrong
14:58 with a product that you're sending and
15:00 so in those kinds of areas the limits
15:02 are going to be on data collection can
15:05 you actually fully digitize a process I
15:07 often joke like plumbers are probably
15:09 the safest from AI disruption because no
15:11 one's even collecting data on how to do
15:13 plumbing right you like crawl somewhere
15:16 get different pipes no one's having
15:17 GoPro and 3D sensors and robotic arms
15:19 and so on collecting data for that so
15:21 that will take much much longer um I
15:24 think in terms of work productivity a
15:26 lot of us are going to become managers a
15:29 lot of current employees that are
15:31 individual contributors are going to
15:33 have to learn to manage an AI to do the
15:36 kinds of work that they do and it turns
15:38 out managing is also a skill not
15:40 everyone is a good manager from day one
15:42 you have to really explain to the AI how
15:44 you do a certain kind of job and what
15:47 we've seen with for instance uh a really
15:49 large cyber security company called
15:50 minecast is we've they've had 200 seat
15:53 licenses using their product and then we
15:55 did a workshop with them and actually
15:56 explained to all the different groups
15:58 like this is what you can do and someone
16:00 from marketing can say well I usually
16:01 get this long product description and
16:04 then I have to describe it for these
16:06 different Industries and an email
16:07 campaign and I have to write three
16:09 tweets and three LinkedIn messages all
16:11 this stuff and we're like well just say
16:13 that to this agent and then the I agent
16:15 does it for them they're like wow now
16:17 it's like six to 20 hours of work every
16:19 other week just got automated by
16:22 describing this workflow that I used to
16:24 do manually to a agent and I think that
16:26 will change pretty much all work and
16:28 pretty much every industry Kaiu um I can
16:31 go in a thousand different directions uh
16:34 here uh first of all uh your Venture
16:38 fund Innovations which is how many
16:40 billions of capital a uh we manage about
16:43 $3 billion about3 billion and you've
16:45 been one of the most prolific AI
16:47 investors I've had the pleasure to visit
16:49 you multiple times in China and thank
16:51 you for your amazing Hospitality you've
16:53 now become an
16:54 entrepreneur um and you're running both
16:58 uh company in China and a company in the
17:00 United States uh why did you do
17:03 that well because this this time is for
17:07 real right imagine you know this was my
17:09 dream practice before well this was my
17:12 dream in when I went to college that AI
17:15 was nothing no one knew what it was but
17:17 I felt this was the thing I needed to do
17:19 and then we went through multiple
17:21 winters of AI where uh there's
17:23 disillusionment and I had to do other
17:25 things and about uh you know seven eight
17:28 years ago we saw with um you know deep
17:31 learning it was became clear it would
17:33 create a lot of value so but at the time
17:36 I didn't really see it becoming AGI so I
17:39 was an investor we actually created 12
17:42 AI unicorns in sinovation Ventures but
17:45 this time with generative AI uh the
17:47 speed at which is growing um is just
17:50 phenomenal you could help yourself you
17:52 yeah I felt if I just invested I'd be
17:54 missing out I I would be in the back
17:55 seat I want to be in the in the driver's
17:57 seat by the way everybody I hope you
17:59 feel the same right I I'm very clear
18:02 about saying there are two kinds of
18:04 companies at the end of this decade
18:05 companies that are fully utilizing Ai
18:08 and everyone else is out of business and
18:09 I I fundamentally believe that is it is
18:12 true um you've written a number of books
18:16 uh AI superpowers I commend to all of
18:19 you so since that was
18:22 published what's the biggest changes in
18:25 the global AI race and it is an AI arms
18:27 race going on well it isn't isn't
18:30 because the companies in China are
18:32 largely competing against each other for
18:34 the China market and they're generally
18:36 not I don't mean Nation to National but
18:38 it is between companies around the world
18:41 yeah so you mean Chinese companies what
18:42 are their characteristics so in my book
18:45 a superpowers I described uh the
18:48 American companies are generally
18:50 speaking more breakthrough Innovative
18:52 they come up with new things um and then
18:55 the Chinese companies are better at
18:57 engineering execution attention to
19:00 detail doing the grunt work user
19:02 interfaces user interfaces building apps
19:05 so um in the case of mobile or deep
19:08 learning we saw that Americans invented
19:11 pretty much everything but China created
19:14 a lot of value arguably more uh given
19:18 technologies that were largely invented
19:20 in the US so now we're in this
19:22 generative AI again invented by
19:24 Americans and we're in a in a in a
19:27 unique position where where the
19:29 technology is disrupting itself very
19:31 quickly in the US and elsewhere um so it
19:34 arguably is still the age of Discovery
19:36 and US ought to win but then the Chinese
19:39 companies are able to watch the
19:41 Innovations make some themselves and
19:43 then do better engineering and deliver
19:46 Solutions so the company I'm building
19:49 01 is doing exactly that we don't claim
19:52 to have invented everything or even most
19:55 things we learned a lot from the Giants
19:57 and silicon valy open Ai and others but
20:00 we think we build a more solidly faster
20:04 execute better so an example was I
20:07 talked about how 01 now has is the third
20:10 best model modeling company in the world
20:13 ranking number six in models measured by
20:16 lmis and UC Berkeley but the most
20:19 amazing thing I think the thing that
20:21 shocks my friends in the solic valley is
20:24 not just our performance but that we
20:26 train the model with $3 million and GPT
20:30 4 was trained by 80 to 100 million and
20:34 um GPT 5 is rumored to be trained by
20:37 about a billion dollars so it is not the
20:40 case we believe in scaling law but when
20:42 you do excellent detailed engineering it
20:44 is not the case you have to spend a
20:47 billion dollars to train a great so this
20:48 is really important for the audience
20:50 here because there's a lot of parts of
20:52 the world that don't have access to you
20:56 know 100,000 H you know h100 clusters
21:00 right and the question is oh my God can
21:02 I really build a business or a product
21:06 in pick your favorite country with a
21:08 small number of gpus yeah and I think
21:11 the constraint on gpus forced you to
21:15 innovate right can you speak to that I
21:16 think it's really important we talked
21:17 about that on our last podcast together
21:20 yeah I think you know as a company in
21:22 China first we have limited access to
21:24 gpus due to the US regulations and
21:28 secondly the Chinese companies are not
21:30 valued what American companies are I
21:32 mean we're F we're we're valued at a
21:34 fraction of the equivalent American
21:36 company so when when we have less money
21:39 and difficulty to get gpus I truly
21:41 believe that necessity is the mother of
21:44 in Innovation so when we only have 2,000
21:47 gpus well the team has to figure out how
21:50 to use it I as the CEO have to figure
21:52 out how to prioritize it and then not
21:56 only do we have to make training fast we
21:58 have make inference fast so our
22:00 inference is designed by figuring out
22:04 the bottlenecks in the entire process by
22:07 trying to turn a computational problem
22:09 to a memory problem by building a
22:11 multi-layer cache by building a specific
22:14 inference engine and so on but the
22:16 bottom line is our inference cost is 10
22:19 cents per million tokens and that's uh
22:21 130th of what the typical comparable
22:24 model charges and where's it going
22:26 where's the 10 cents going yeah it's
22:30 well the 10 cents would lead to building
22:32 apps for much lower cost so if you
22:35 wanted to build a u.com or perplexity or
22:38 some other app you can either pay open
22:40 AI $440 per million tokens or if you
22:44 have our model it costs you just 10
22:46 cents and if you buy our API it just
22:49 costs you 14 cents we're very
22:51 transparent with our pricing yes Richard
22:54 there's there's a really interesting uh
22:55 Paradox called jevans Paradox from the
22:57 previous industrial re ution a lot of
22:59 smart people back then were working on
23:00 making more efficient steam engines and
23:03 using that use less coal they thought oh
23:05 if we make the steam engines more
23:07 efficient we're going to need less coal
23:09 but instead we needed more steam engines
23:10 everywhere and I think that's exactly
23:12 what's going to happen we're currently
23:13 in the jeevan's Paradox of intelligence
23:16 we're just going to use intelligence in
23:17 many more places everyone is going to
23:19 have their own assistant their own
23:21 medical team that like understands
23:23 everything about them instead of being
23:26 restricted by Intelligence being very
23:27 very expensive yeah I totally agree I
23:29 want to clarify I'm not saying there's a
23:31 fixed workload we're making it cheaper
23:33 I'm saying we're enabling a workload
23:35 much much larger corre I want to ask one
23:37 closing question to all of you we have
23:39 people here who have daughters and sons
23:42 or nephews or brothers and
23:45 sisters what's your advice to someone
23:48 who is 20 years old listening to this um
23:52 or through this what's your advice to
23:53 someone at the beginning of their uh
23:56 sort of academic and professional career
23:59 given what you know is going on in AI
24:01 right now
24:02 Prem I think it's don't waste your time
24:05 learning how to code because I think the
24:08 new language is going to be I think the
24:10 new code language is going to be English
24:12 and I think that uh absolutely learn as
24:15 fast as you humanly possibly can on all
24:18 AI in all Ai modalities and I think if
24:22 you and then once you find your passion
24:23 I think you're going to then find a very
24:25 narrow AI to empower you to do what
24:28 you're what you're really set out to do
24:29 thank you PR Richard I I will disagree I
24:32 think you should still learn how to
24:33 program uh I think that is how you get
24:36 to really understand how this technology
24:39 works at the foundational level uh and
24:41 how it becomes less magic and more
24:43 something that you can yourself modify
24:46 and construct with but you need to
24:47 combine computer science and programming
24:50 with another passion that you can
24:52 actually apply all of that intelligence
24:54 to and ideally the younger you are the
24:56 more you learn the foundations Ma ma
24:58 physics The Sciences I think I'm going
25:00 to cut you off because I'm being yanked
25:02 I want I want to have kyu's final word
25:04 here okay I actually agree and this
25:06 agree with both of you I think people
25:09 should follow their hearts right if you
25:11 dream of becoming a fantastic programmer
25:13 and you can do it you should do what
25:14 Richard says if you think programming is
25:16 the way that make the most money no then
25:19 you should follow what prime says ladies
25:21 and Gentlemen please give it up to these
25:23 three amazing
25:24 CEOs thank you thank you thank you thank
25:28 you
Chrome Extension