00:04today on no priors we're speaking with
00:07Mustafa Suleiman co-founder of deepmind
00:09the pioneering AI lab acquired by Google
00:11in 2014 for 650 million dollars and now
00:15co-founder and CEO of inflection along
00:17with Reed Hoffman and Corinne simonyan
00:19inflection just launched their first
00:21public product pie last week Mustafa
00:24welcome to no priors thanks so much for
00:25joining us thanks for having me I'm
00:27super excited to be here yeah we're very
00:30excited to have you today I think one
00:31thing that'd be great to maybe start
00:32with is just a little bit of your
00:33personal story because I think you have
00:34a really unique background you're very
00:36well known obviously for um deep mind
00:38and your pioneering work in the AI world
00:40but I think before all that you worked
00:42on a Muslim youth helpline you started a
00:45partnership in consultancy that was
00:47focused on conflict resolution to
00:48navigate social problems just love to
00:50hear a little bit more about the early
00:51days of things that you did before a
00:53deep mind and maybe we can talk a little
00:54about deep mind and in sort of more
00:55recent stuff as well yeah sure I mean
00:58the truth is I'm I was very much a kind
01:00of changed the world kid growing up like
01:03um a big believer in Grand Visions doing
01:07good having a huge impact in the world
01:09and that was always kind of what drove
01:13um so when I I grew up in London and
01:19um but at the end of the second year of
01:21My Philosophy degree I was kind of
01:23getting a bit frustrated with this sort
01:26theoretical you know nature of it all it
01:30was full of hypothetical moral
01:34um so a friend that I met at Oxford was
01:37starting a telephone Counseling Service
01:38a kind of helpline and it really
01:41appealed to me it was a non-judgmental
01:44non-directional secular support service
01:47for young British Muslims and this was
01:50about six months after the 9 11 attacks
01:52and so there was quite a lot of like
01:54rising islamophobia and the government
01:57was talking a lot about anti-terrorism
01:58and you know in general I think that
02:01like sort of Migrant communities were
02:04feeling the pressure and this was a
02:07support service that was staffed
02:09entirely by us by young people I was 19
02:12at the time and yeah I spent uh almost
02:16three years working pretty much
02:17full-time on on that and it was an
02:19incredible experience because it was
02:21basically my first startup and you know
02:24fundraising was the name of the game
02:25except the numbers were much much
02:27smaller than they are these days uh and
02:30you know the surface the service was
02:32staffed by almost a hundred volunteer
02:35young people which was just amazing
02:37because we felt like we can actually do
02:40something you know it was quite
02:42liberating and energizing to actually
02:45give this a shot and you know I was very
02:47much inspired by the kind of Human
02:48Rights principles it was deliberately
02:50not religious even though it used some
02:54of the kind of culturally sensitive
02:56language that helped people feel heard
02:59um so yeah it's had a very formative
03:00impact on my my Outlook yeah I know it's
03:03super interesting and I think we can
03:04talk more about that in the context of
03:05AI in a little bit one other thing that
03:07you did is you also started a
03:09consultancy where you worked as a
03:11negotiator and facilitator and I believe
03:12you worked with clients like the United
03:13Nations the Dutch government and others
03:15could you tell us a little bit more
03:17about that work as well
03:18yeah I mean I was always trying to
03:20figure out how to scale my impact and
03:23you know I quite quickly realized that
03:26delivering a sort of one-to-one service
03:28via a non-profit was not going to scale
03:32a great deal even though it had an
03:36um you know on a kind of human-to-human
03:37level and so I was super interested in
03:41these like meta structures like how does
03:43you know the UN actually influence you
03:46know Behavior at at the country level
03:49um and you know how could we run more
03:52efficient decision-making processes
03:56um where there's tension and
03:57disagreement so we worked all over the
03:59world actually in israel-palestine and
04:03um in Cyprus between the the Greeks and
04:05the Turks my colleagues worked in South
04:08Africa and colomba Colombia Guatemala
04:10and I think it really taught me that
04:14learning to speak other people's social
04:17languages is actually an acquired skill
04:20and you really can do it with a with a
04:22little bit of attention to detail and
04:24some patience and care it's kind of a
04:27superpower being able to deeply hear
04:30other people and make them feel heard
04:32such that they're better able to
04:33empathize with people that they disagree
04:35with and that that's been an important
04:37theme throughout my kind of career
04:39something I've always been interested in
04:41so I think I think I co-founded that and
04:44worked on it for I think three years and
04:47soon realized the limitations of like
04:49large-scale human processes I mean in
04:522009 uh 2009 I worked I facilitated one
04:57part of the climate negotiations in
04:59Copenhagen and uh yeah it was a kind of
05:02a remarkable experience like you know
05:04192 countries literally a thousand ngos
05:09and activists many different academics
05:11everyone proposing a different solution
05:14a different definition of the problem
05:17and you know and one way it was sort of
05:20inspiring to see so many different
05:21cultures and ideas coming together to
05:23try to form consensus around an issue
05:26that was clearly of existential
05:29on the other hand it was just like
05:31deeply depressing that we weren't able
05:33to achieve consensus it took another
05:35decade to even get mild consensus on
05:38this or half a decade 2015.
05:40and I think that was sort of an
05:43eye-opener for me I was like the world's
05:46governance systems are not going to keep
05:49up with both the exponential challenges
05:53that we Face from globalization and
05:55carbon emitting but also like technology
05:58and and that was the next thing that I
06:00saw on the landscape so how did this
06:02lead into your interest in Ai and you
06:03know I believe that you met Demis when
06:05you were quite young and I think he and
06:07your other co-founder worked together
06:08later in a lab but I'm a little bit
06:10curious like how your background and
06:12interests in these sorts of global
06:14issues then transformed into an interest
06:16in Ai and the founding of deepmind
06:19yeah well around about that time
06:20actually like I guess it was like 2008
06:23or so I was starting to keep an eye on
06:27um Facebook's rise and I was like this
06:29is incredible I mean it's like a two or
06:32three year old platform at that point
06:33and it had hit like 100 million monthly
06:36actives and that was just a mind-blowing
06:38number to me and it was obvious that
06:41this wasn't just the kind of neutral
06:43platform for giving people access to
06:45information or connecting people with
06:46other people the the frame because I had
06:49come from a conflict resolution
06:50background our entire approach was like
06:53what is the frame of a conversation like
06:56how do you organize space how do you
06:59prepare individuals to have a
07:01constructive disagreement how do you
07:03like set up the the environment
07:05basically to facilitate dialogue and so
07:08that was the lens through which I looked
07:10at Facebook I was like well this is a
07:12frame there's a choice architecture here
07:14there are significant design choices
07:16which are going to incentivize certain
07:18behaviors obviously at that point I
07:20wasn't really ranking but even just
07:22having a thumbs up or like the choice of
07:25you know which button you place in what
07:27order and how you arrange information on
07:29the page and what all of that drives
07:33um in one way or another and you know
07:35that was a big realization to me because
07:37I was like well this is actually
07:40the default approach to human connection
07:43at a scale that is like completely
07:46unimaginable I mean perhaps only akin to
07:48you know the default expectations in a
07:51religion for example everyone grows up
07:53with an idea that there is a
07:56you know a patriarchy a male God that
07:59you know that there's a particular what
08:01role for women like that's that you know
08:03until a few decades ago that was just an
08:05an implied sort of undertone to an
08:09entire social structure for thousands of
08:11years uh and that's kind of what I mean
08:13by frame there's this sort of these
08:14implicit design choices which cause
08:16hundreds of millions of people to change
08:18their behavior yeah and I think that
08:20that's super interesting because I
08:21remember working on a bunch of Facebook
08:23apps at the time when the platform
08:25and people were purposefully thinking
08:27about that stuff but on the micro level
08:29right how do we get more users how do we
08:31get people to convert how do we drive
08:34and so everybody I think was very
08:35explicitly thinking about this as a
08:37behavioral change platform but not at
08:39the level of society
08:40you know we were thinking about it in
08:42the context of just like how do you get
08:43more people to use this thing you know
08:44and so I think it's really interesting
08:46that people then later realize the big
08:48ramifications of this in terms of you
08:50know how that actually Cascades in terms
08:51of social behaviors and other things how
08:53did that lead to starting Deep Mind
08:56well it was clear to me from that moment
08:58on like I left Copenhagen in 2009
09:02thinking this is not the path to
09:05significant positive social change it
09:08still needs to continue and I support
09:09those processes obviously but I'm just
09:11saying it it's just not something that I
09:12feel I could continue to work on full
09:15time and so my heart was set on
09:18technology at that point so I reached
09:20out to Demis who uh was the brother of
09:23my best friend from when I was a kid we
09:25got together we had a coffee we went and
09:27actually we played poker
09:30um at one of the casinos in London
09:32because we both love games both super
09:33competitive uh both good at poker and on
09:38that night I think we both got knocked
09:40out pretty early in the tournament so we
09:42sat around drinking diet coke uh talking
09:45about ways to change the world and we
09:48basically were you know having exactly
09:50this conversation like you know is it
09:53going to be I mean obviously at that
09:54point I was most the inspired by
09:57platforms and software and social apps
10:00and connectivity and so on whereas
10:03um you know Demis was way more in the
10:05kind of Robotics land and sci-fi land I
10:07mean he was he was fully thinking that
10:10you know the way to manage the economy
10:13the way to make economic decisions was
10:16to simulate the entire economy right and
10:18and he thought that he was very much
10:21obviously had just come off the back of
10:23his games like Evil Genius and black and
10:25white and so on which were kind of
10:26simulation based games so I think that
10:27was his default frame at that point
10:30um yeah and then we spent many months
10:32talking and spend a lot of time with
10:34Shane Legg as well and Shane was really
10:37the core driver of the ideas and the
10:39language around artificial general
10:40intelligence I mean he had worked on
10:45um uh with Marcus Hooter on definitions
10:48of intelligence I found that super
10:50inspiring I think that was actually the
10:52turning point for me that it was pretty
10:54clear that we at least had a thesis
10:57around how we could distill
10:59the sort of essence of human
11:01intelligence into an algorithmic
11:03construct and it was it was his work in
11:07I think he I think for his PhD thesis he
11:10put together like 80 definitions of
11:14and aggregated those into a
11:17single formulation which was how do we
11:21um you know the intelligence is the
11:22ability to perform well across a wide
11:25range of problems and he basically you
11:28know gave gave us a measurement an
11:30engineering kind of measurement that
11:32allowed us to constantly measure
11:33progress towards you know whether we
11:36were actually producing an algorithm
11:38which was inherently General I.E it
11:40could do many things well at the same
11:41time is that the working definition you
11:44use for intelligence today
11:53um I I think that there's a more nuanced
11:56version of that I think that's a good
11:58definition of intelligence but I think
12:00in a weird way it's over rotated the
12:03entire field on one aspect of of
12:06intelligence which is generality
12:08you know and I think um openai and um
12:12then subsequently anthropic and others
12:14have taken up this default sort of
12:17Mantra that like it all that matters is
12:20can a single agent do everything you
12:23know can it be multi-modal can it do
12:25translation and speech generation
12:29I think there's another definition which
12:31is valuable which is the ability to
12:33direct attention or processing power to
12:37the Salient features of of uh an
12:40environment given some context
12:45um actually what you want is to be able
12:47to take your raw processing horsepower
12:49and direct it in the right way at the
12:52right time because it may be that a
12:55certain tone or style is more
12:58appropriate given a context it may be
13:00that a certain expert model is more
13:03suitable or it may be that you actually
13:05need to go and use a tool right and
13:07obviously we're starting to see this
13:10um and in fact I think the key and we
13:13can get into this obviously in a moment
13:15but I think the key element that is
13:17going to really unlock this field is
13:20actually going to be the router in the
13:23middle of a series of different systems
13:25which are specialized some of which
13:27don't even look like AI at all they
13:29might just be traditional pieces of
13:30software databases tools and other sorts
13:32of things but it's the router uh or the
13:35kind of central brain which is going to
13:38need to be the key decision maker and
13:39that doesn't necessarily need to be the
13:41largest language model that we have
13:43it's really interesting because I feel
13:45like a lot of what you described is
13:46actually how the human brain seems to
13:49in terms of you have something a little
13:51bit closer to a mixture of experts or
13:53Moe model where you have the visual
13:55cortex responsible for visual processing
13:57and then you have a other piece of the
13:59brain specifically responsible for
14:00empathy and you have mirror neurons and
14:02you know it feels like the brain is
14:03actually this Ensemble model in some
14:05sense with some some routing depending
14:07on the subsystem you're trying to access
14:09you know the generality approach seems
14:11like a really it almost goes at odds
14:14with some of those pieces of it unless
14:15you're just talking about some part of
14:16the hippocampus or something right well
14:17I think that's long been the inspiration
14:19right I think for everybody these neural
14:21networks are the obvious example but in
14:23many other elements reinforcement
14:27um you know and so on are are all brain
14:29inspired and I think that you know
14:31there's been a lot of talk about you
14:33know sparsity as well which is sort of
14:36what you're describing and you know so
14:38far we've we've had to do you know very
14:40dense all-to-all connections because we
14:43sort of haven't quite learned the
14:44algorithms for sparse activations but I
14:47think that's going to be a very
14:48promising area and you know in many ways
14:51what I'm describing doesn't actually
14:52require sparse activations because you
14:55know you you actually could just train a
14:58decision-making engine at the middle to
15:00know when to use which size model right
15:02so maybe in some contexts you would want
15:05the highest quality super expensive 20
15:08second latency model and in most other
15:10contexts a super fast 3 second Mini
15:12model might work fine I think that's
15:15going to be the the key unlock actually
15:18and and quite sort of remarkably that's
15:20an engineering problem
15:22um perhaps more than it is a an AI
15:24problem which you know is is just a
15:27pretty surreal moment give you know just
15:28if you actually observe that given where
15:30we are in the field and stuff when you
15:32um start a deep mind I think it was
15:34reasonably unpopular to do what you were
15:36doing right and so I think you ended up
15:37getting funded by um Founders fund and
15:40Peter Thiel and Elon Musk but I remember
15:41at the time there was like three or four
15:43parties that funded a lot of AI things
15:45and then nobody else was really doing it
15:46in terms of the types of approaches you
15:48were taking in terms of saying we're
15:49going to build these big AI systems that
15:52can do all sorts of things right yeah I
15:54mean it was wacky like I I can't say
15:57that enough like it it was especially
16:00for the first two years so because we
16:03founded in 2010 and for the most of the
16:06sort of spring and summer of 2010
16:09actually most of that year I was going
16:12to Gatsby computational Neuroscience
16:14unit at UCL sneaking in with Demis and
16:17Shane to just sit in on the lunches that
16:19uh Peter Diane ran and I remember Shane
16:23like sort of saying to me like you know
16:26the language here is machine learning
16:28yeah you can say I don't say AI yeah I I
16:32was like okay okay I'll keep my mouth
16:34shut don't worry like we certainly don't
16:38um you know and and that that was a kind
16:41of that was pretty weird I mean that you
16:42know they weren't you know there weren't
16:44very many funders for us like you know
16:46Peter Thiel you know to his credit uh
16:49did actually have significant Vision
16:51here although he sold pretty early I
16:53think and now doesn't seem to be in the
16:55game so but uh yeah he certainly he
16:57certainly saw it first
16:59um and you know I think that all changed
17:03pretty quickly first with uh you know
17:05Alex net of course in 2012 and then with
17:09dqn uh the Atari paper in 2013
17:13um you know and then a kind of
17:14succession of breakthroughs after
17:16alphago and people got more more sort of
17:18aware of it but it still surprises me
17:21the extent to which the rest of the
17:22world is like suddenly waking up and
17:24obviously we've seen that like crazy in
17:25the last six months so yeah and then I
17:28guess last question on sort of your time
17:29with uh Google and deepmind and because
17:31I think there's a lot of really exciting
17:33things to talk about in the context of
17:34inflection and sort of the broader field
17:36in world what are some of the things you
17:38were most excited to have the team
17:39create a Deep Mind Over the years or
17:41some of the breakthroughs that you're
17:43yeah well I mean in some ways we we
17:47definitely sort of pioneered the Deep
17:49reinforcement learning effort and I
17:52um you know in principle it's a very
17:56promising Direction I mean you clearly
17:58want some mechanism by which you can
18:01learn from raw perceptual data and that
18:04directly feeds into a reinforcement
18:07learning algorithm that can update and
18:11essentially iterate on that in real time
18:13with respect to some reward function
18:15whether that's online or offline like
18:17directly interacting with the real world
18:19in real time or it's you know in in a
18:21kind of batch simulation mode
18:24um you know and and that turned out to
18:27be very valuable for a specific type of
18:31um where a game-like environment had a
18:33very structured scalar reward and we
18:37could play that game many millions of
18:41um that's part of the reason why we
18:42started the alpha fold project because
18:45it's actually my group that was looking
18:48around for other applications of dqn
18:51like alphago-like uh tools and uh in a
18:56um that we did one week
18:58um someone stumbled across across this
19:00problem we'd actually looked at it back
19:02in 2013 when it was called fold it which
19:05was a very small scale kind of version
19:08of this and just for a contextard
19:10interrupt um you know Alpha fold was
19:12focused on folding proteins which at the
19:13time was a really hard problem right
19:15people were trying to do this molecular
19:17modeling they couldn't really make any
19:18real Headway and lots of the traditional
19:20approaches and then your group at
19:22deepmind really started pioneering
19:23hearing how to think about protein
19:24folding in a different way so sorry to
19:26interrupt I just want to give context
19:27for people listening so I think the
19:29hackathon was probably 2016 and then as
19:34soon as we saw the hackathon that you
19:36know start to work then we actually you
19:38know scaled up the effort and hired
19:40um you know a bunch of outside
19:41Consultants to help us with the domain
19:43knowledge and then I think the following
19:45year we um entered the Casp competition
19:48so you know these things take a long
19:50time uh and you know sort of longer than
19:52I think people realize there's a lot of
19:54it was that was a very big effort by
19:55deepmind and eventually it became a a
19:57company-wide Strike Team
20:00um so in in hindsight these things do
20:02take a huge amount of effort yeah the
20:05fascinating thing here is that you know
20:06the work started with alphago which was
20:08how to play go better right or how to
20:11beat people to go and then the same
20:13underlying approach could then be
20:15morphed and applied to protein folding
20:17which I think is an amazing sort of leap
20:18or connection to make and you know I
20:20used to work as a biologist and I
20:22remember you'd spend literal years
20:24trying to crystallize proteins in
20:26different solutions you do all these
20:28different salt concentrations in each
20:29well so the protein would crystallize
20:31you could hit it with x-rays and then
20:33you'd interpret those x-rays to look at
20:34the structure right and so you had to do
20:36this really hard sort of chemistry and
20:38physics to get any information about a
20:40protein at all and then you folks with
20:42the machine ran through every protein
20:45sequence literally in in the in every
20:48database for every organism and you're
20:50able to then predict folding which is
20:52it's pretty amazing it's very striking
20:54yeah I mean I I think the the if if I
20:58were to sort of summarize the core
21:00thesis of deepmind it was that it would
21:03be possible to the motivation for
21:05generality was that you would be able to
21:10um you know a rewarding behavior in one
21:13environment and transfer in a more
21:16compressed or efficient representation
21:18the insights that had made you
21:20successful in one environment to the
21:22next environment right that transfer
21:23learning has always been the key goal
21:26and that was one of the one of the very
21:30exciting proof points that it is you
21:32know um increasingly looking likely that
21:34that's possible so you know I definitely
21:36think that's that's pretty cool because
21:38when you think about the sorts of
21:39problems that we're facing in the world
21:41today we don't have obvious answers
21:43lying around there's no like genius
21:45Insight that's just waiting to be
21:47applied we actually have to discover new
21:49knowledge and I think that's the that's
21:51the attraction of artificial
21:53intelligence that's why we want to work
21:55on these you know on these models
21:57because you know we're we're sort of at
22:00the limit of what you know the smartest
22:02humans in the world are are capable of
22:05inventing and we have you know very
22:08pressing urgent Global challenges you
22:11know from food supply to water to
22:14decarbonization to clean energy
22:17Transportation you know with a rising
22:19population that we really want to solve
22:21so there are that you know amidst all of
22:24the stresses and the fears about
22:26everything that's being worked on at the
22:28moment it is important to keep in mind
22:30that there is a an important North Star
22:32that everybody is working towards and we
22:34just got to keep focused on those goals
22:36rather than sort of be too sidetracked
22:38by um some of the fears
22:41let's talk about inflection what was the
22:43motivation for starting another company
22:49I guess back in sort of
22:512018-2019 it wasn't clear that neural
22:56networks were going to have a
22:57significant impact in language because
23:00you think about it intuitively
23:02um for the for the previous sort of five
23:04years cnns had been effective at
23:07learning structure locally right so
23:10pixel in an image in the input so pixels
23:13in an image that were correlated in
23:16space tended to produce you know sub
23:19features which were you know a good
23:22representation of what you were trying
23:23to predict maybe there were lines and
23:25edges and they grew into eyes and faces
23:27and scenes and so on and that kind of
23:30hierarchy just intuitively seemed to
23:31make sense and seemed to apply to audio
23:34and other modalities right whereas if
23:37you kind of think about it a lot of the
23:39structure of predicting the next word or
23:42letter or token in a sentence
23:45seems to exist in a very very very
23:48spread out you know far removed from the
23:50immediate next step of the prediction
23:52right and so it didn't look like that
23:55was working and then to be honest like
23:57when GPT 3 came out that was like a big
24:03um I I had seen the gpt2 work and hadn't
24:08quite clicked for me that this was
24:10significant it was really only when I
24:12started saw the gbt3 paper that my eyes
24:15were wide open to this possibility it's
24:18pretty amazing that you could attend to
24:20you know a very very spot seemingly
24:23sparse representation and use that to
24:27predict something which on the face of
24:29it seemed like there were billions of
24:30possibilities of what might come next in
24:32a sentence where maybe tens of millions
24:34or something but a lot and for me it was
24:37early 2020 that I went uh to work at
24:40Google and uh I got involved in the
24:44large language model efforts I got
24:46involved in the Mina team that was
24:48called at the time I know that you guys
24:49had gnome on the show recently
24:52um no one's super awesome and it was me
24:54and noem Daniel cockley uh and a few
24:58others and it was just unbelievable what
25:02was being built there and um when I
25:05joined is pretty small models and um
25:08very quickly we scaled it up it became
25:11the the Lambda group
25:13um and we started seeing how it could
25:16potentially be used in various kinds of
25:18search started looking at retrieval
25:20grounding for improving factuality
25:22started getting a feel for all the
25:24hallucinations and so on and that was
25:27really a mind-blowing few years to me
25:31while I was there sort of in the in the
25:33last year in 2021 I tried pretty hard to
25:36get things launched at Google we were
25:38all kind of banging on the table being
25:39like come on this is the future and uh
25:42you know um obviously David Luan from
25:44Adept was also in and around that group
25:46so the three of us in our own ways were
25:48pushing pretty hard for for launch and
25:53it wasn't meant to be uh just you know
25:55timing is everything and you know Google
25:57just wasn't wasn't the right timing for
25:59Google for various reasons
26:01uh and you know I was just like look
26:04this this has to be out there in the
26:05world this is this is clearly the new
26:07wave of technology and so yeah in
26:10January I left got together with Karen
26:14um who I work with at d mine for seven
26:17years we bought his company back in 2014
26:19at deepmind he led the
26:21um deep learning scaling team at
26:23deepmind for years and worked on all the
26:24big breakthroughs at deepmind uh and
26:26then of course Reed Hoffman who's been
26:28my uh one of my closest friends for like
26:3010 years and we've always talked about
26:32starting something together and um I was
26:34like this is the obvious thing now is
26:36the time for sure and so the rest is
26:38history you know we've we've it's been a
26:41wild ride since then
26:43a little bit better than somebody who's
26:44been such a Pioneer in the field uh and
26:46working on this all the time it's still
26:48constantly surprised as I am also
26:50constantly surprised
26:53um I remember when you were first
26:55starting to get this going I another
26:58thing I was surprised by is the focus
27:01you I mean I came around to it in
27:03writing the investment memo but you know
27:06you have this focus on the idea of
27:07companionship rather than information as
27:10the right initial approach uh you've
27:12talked about worked on thought about
27:14empathy for humans and other populations
27:16for a long time it seems
27:18counter-intuitive like what why
27:20companionship yeah it's a great question
27:24I think to step one step back from that
27:27first I think my core insight about what
27:30was missing for Lambda was interaction
27:36um in a funny way that was exactly what
27:38was motivating Karen too
27:41um having beaten all the the academic
27:43benchmarks and achieved Sota many times
27:46he had come to the same conclusion I had
27:48seen the same thing from Lambda what we
27:51was user feedback and
27:54um actually when you think about it all
27:56of our interfaces today are
27:58fundamentally about interaction you know
28:00you're giving your browser feedback all
28:02the time you're giving uh you know that
28:05web service feedback same with an app or
28:07anything that you interact with it's
28:09actually a dialogue and so the way I'd
28:12position Lambda at Google is that you
28:15is the future interface and Google is
28:19already a conversation it's just an
28:21appallingly painful one right you say
28:23something to Google it gives you an
28:26answer in 10 Blue Links you say
28:28something about those 10 Blue Links by
28:29clicking on it you it it generates that
28:32page you look at that page you say
28:34something to Google by how long you
28:36spend on that page what you click on it
28:38how much you scroll up and down etc etc
28:40and then you come back to the search
28:42login and you update your query and you
28:44say something again to Google about what
28:46you saw that's a dialogue and Google
28:48learns like that and the problem is it's
28:51you know using uh 1980s Yellow Pages to
28:55have that conversation and actually now
28:57we can do that conversation in fluent
28:59natural language and I think
29:01the problem with what Google has sort of
29:04I guess in a way accidentally done to
29:07the internet is that it has basically
29:11shaped content production in a way that
29:14optimizes for ads and everything is now
29:17SEO to within an inch of its life you
29:19know you you go on a web page and all
29:23the text has been broken out into sub
29:25bullets and subheaders and you know
29:29separated by ads and you know you spend
29:32like five to seven or ten seconds just
29:34like scrolling through the page to find
29:36a snippet of the answer that you
29:37actually wanted like most of the time
29:39you're just looking for a quick snippet
29:40and if you are reading it's just in this
29:42awkward format and that's because if you
29:44spend 11 seconds on the page instead of
29:47five seconds that looks like high
29:49quality content to Google and it's quote
29:51unquote engaging so the content creator
29:53is incentivized to keep you on that page
29:56and that's bad for us because what we
29:58want is a we assignments well we as
30:01humans All Humans clearly one a high
30:04quality succinct fluent natural language
30:07answer to the questions that we want and
30:10then crucially we want to be able to
30:12update our response without thinking how
30:15do I change my query and like write this
30:18we've learned to speak Google like it's
30:20a crazy environment we've learned to
30:22Google right that's just a weird lexicon
30:25that we've co-developed with Google over
30:2720 years no like now that has to stop
30:30that's over that moment is done and we
30:32can now talk to computers in fluent
30:36natural language and that is the new
30:39um so that that's what I think is going
30:40on maybe we should back up for a second
30:42and just tell people about what pi is
30:47building on all of that we think that Pi
30:51I think that everyone
30:53in the next few years is going to have
30:55their own personal AI right so there's
30:57going to be many different types of AI
31:00um there will be business AIS government
31:02AIS non-profit AIS political AIS
31:05influencer AIS brand AIS all of those
31:09AIS are going to have their own
31:10objective right aligned to their owner
31:13which is to promote something sell
31:16something persuade you of something
31:18and my belief is that we all as
31:21individuals want our own AIS that are
31:24aligned to our own interests and on our
31:26team and in our corner and that's what a
31:28personal AI is and Ours
31:30is called Pi uh personal intelligence it
31:34is as you said there to be your
31:37um we've we've started off as with with
31:44um empathetic and supportive and we
31:47tried to sort of ask ourselves at the
31:49beginning like what makes for great
31:52conversation when you have a really
31:53flowing smooth you know generative
31:56interaction with somebody what's the
31:58essence of that and I think there's a
31:59few things like the first is the other
32:02person really does listen to you right
32:04and they demonstrate that they've heard
32:07you by reflecting back some of what
32:10they add something to the conversation
32:12you know so it's not just regurgitation
32:14but they introduce another nugget
32:18um they ask you follow-up questions and
32:20they're curious and interested
32:23um in what you say and you know
32:25sometimes there's a bit of spice right
32:27they throw in something silly or
32:29surprising or random or kind of wrong
32:31and it's endearing and you're like oh
32:33like that that we're connecting and so
32:37we've tried to as in our first version
32:39and this really just is a first version
32:41like this is actually not even our
32:42biggest model at the moment
32:44um so we're just putting out a first
32:46version that is skinned for this kind of
32:48interaction so that we can sort of learn
32:50and improve and you know it really makes
32:53for a good companion
32:55um someone that is thoughtful and kind
32:57and interested in in your world as a
33:00first start you're working on these sort
33:03of personalized intelligence or personal
33:04agents and you mentioned how you think
33:06in the future there'll be all these
33:07different types of agents for
33:08representing different businesses or
33:10causes or political groups or the like
33:12what do you think that means in terms of
33:14how the web exists and how it's
33:16structured so to your point the web is
33:17effectively really based on a lot of SEO
33:19and a lot of sort of Google as the
33:22what happens to web pages or what
33:24happens to the structure of the internet
33:27I think it's going to change
33:29fundamentally I think that most
33:32Computing is going to become a
33:34conversation and a lot of that
33:36conversation is going to be facilitated
33:38by AIS of various kinds so your pie is
33:42going to give you a summary of the news
33:44in the morning right it's going to help
33:47you keep learning about your favorite
33:49hobby whether it's cactuses or you know
33:52like motorcycles right and so you know
33:55every couple days it's going to send you
33:57new updates new information in a summary
34:00snippet that really kind of Suits
34:02exactly your reading style and your
34:05interests and your preference for
34:07consuming information whereas a website
34:09you know the traditional open internet
34:12just assumes that there's a fixed format
34:13and that everybody wants a single format
34:15and generative AI clearly shows us that
34:18we can make this Dynamic and emergent
34:20and entirely personalized so you know if
34:23I was Google I would be pretty worried
34:24because the car this is that that old
34:27school system does not look like it's
34:29going to be where we're at in 10 years
34:31time it's not going to happen overnight
34:33there's going to be a transition but
34:35these kind of succinct Dynamic
34:37personalized interactive moments are are
34:40clearly the future in my opinion
34:43the other group of people that is
34:45clearly worried is anybody with a with a
34:48website where their business is that
34:49website I spent a lot of time talking to
34:51Publishers in April because they were
34:53freaking out and uh what what advice
34:56would you have for people who like
34:57generate content today well I think that
35:01you know an AI is kind of just a website
35:04or an app right so you can still have
35:08like let's say that you have a Blog
35:11about baking and so on you know you're
35:13you can still produce super high quality
35:16content with your AI and your AI will
35:20you know be I think a lot more engaging
35:25um for other people to talk to so to me
35:27any brand is already kind of an AI it's
35:30just using static tools right so so you
35:34know for a couple hundred years the ad
35:36industry has been using color and shape
35:38and texture and text and sound and image
35:42to generate meaning right it's just they
35:45release a new version every six months
35:48or every year right and it's you know
35:52the same thing that applies to everybody
35:53like TV ads used to be right whereas now
35:56that's going to become much more Dynamic
35:58and interact so I I really don't
36:01subscribe to this view that there's
36:02going to be like one or five AIS I think
36:05this is like completely misguided and
36:08fundamentally wrong there are going to
36:10be hundreds of millions of AIS or
36:12billions of AIS and and they'll be
36:14aligned to individuals so what we don't
36:16want is autonomous AIS that can operate
36:18completely independently and wander off
36:20doing their own thing that I I'm really
36:21not into that vision of the world that
36:23doesn't end well right but you know if
36:26your blogger you know has you know their
36:30own AI that represents their content
36:32then I imagine a world where my pie will
36:36go out and talk to that AI and say yeah
36:40like my Mustafa is super interested to
36:43learn about baking he can't crack an egg
36:45so where does he need to start right and
36:48then Pi will have an interaction and be
36:50like oh that was really kind of funny
36:51and interesting and stuff will really
36:53like that and then Pai will come back to
36:55me and be like hey I found this great AI
36:57today maybe we could set up a
36:59conversation you'll find something super
37:01interesting or they recorded this this
37:02little clip of me and the other AI
37:04interacting and here's a three three
37:07minute video or something like that
37:08right that'll be how new content I think
37:10gets produced and I think it will be
37:12your AI your pie your personal AI that
37:15acts as interlocutor accessing the other
37:18world which is basically by the way what
37:19Google does at the moment right Google
37:21crawls other you know essentially AIS
37:24that are statically produced by you know
37:26the existing methods and has a little
37:28interaction with them ranks them and
37:30then presents them to you bacteria
37:32Original Point on Facebook I think um
37:34one thing Facebook has been uh
37:36criticized for is the creation of
37:39where the only information that you see
37:41is information that you know you you
37:43kind of inherently believe or the feed
37:44is kind of tailored to you and if you
37:47think about some of these AI agents one
37:48could argue they're going to be the
37:49extreme form of this right in the
37:51downside case and the upside case
37:53obviously there's other versions of this
37:54but the downside cases it will just
37:56constantly use the feedback from you to
37:59reinforce things you already strongly
38:00believe whether they're correct or not
38:02and so I'm a little bit curious how you
38:03think about this as we go through this
38:05new platform shift and you mentioned
38:06that you identified some of these issues
38:07quite early on with some of the Facebook
38:09or other social platforms how do you
38:11think about that in the context of AI
38:13agents I think that is the default
38:15trajectory without intervention right so
38:19that might be a controversial view but
38:21you know I I think that the platforms
38:27that was the big lie
38:29and I think that was frankly to me very
38:33obvious from the very beginning the
38:35choice architecture is
38:38a ranking it's not a clean feed clearly
38:41there's billions of bits of content so
38:42you have to select what to show and what
38:44to show you know is is a huge uh you
38:48know sort of political cultural
38:49influence on on how we end up and so of
38:53AI is an accelerated version of that
38:57um my take is that all of us AI
39:00companies as well as the old social
39:02media platforms have to embrace the
39:04platform responsibility of curation and
39:08try to be as transparent as possible
39:10about what that curation actually looks
39:13like what what is excluded
39:17um and here I think that
39:20you know the valley probably needs to
39:23be a bit more open to the European
39:27um the reality is that you know we have
39:30to figure out as a society which bodies
39:34we trust to make decisions which
39:37influence recommendation algorithms or
39:43um and if that's a requirement for
39:44transparency of training or if it's a
39:47requirement for transparency with
39:49respect to content that has been
39:50excluded or what has been upvoted or
39:54um fundamentally we have to make these
39:57things accountable to democratic
39:58structures and that means the Democratic
40:00structures need to sort themselves out
40:03pretty sharpish and like actually have
40:06some functioning bodies that can provide
40:08real oversight without everybody like
40:11fainting over the accusations that this
40:14is censorship and being super chillish
40:17about that because you know now really
40:20is the time to like actually get that a
40:22bit more straight and out and and have
40:24some kind of responsible interactions
40:25with these companies because you're
40:26right these are going to be very very
40:30this is my bias command but that seems
40:32like a harder Hill to climb than the AGI
40:38I think I do agree but I hope not yeah
40:42yeah well we can all work on it um so
40:45you you describe Pie as like the first
40:47foray that you guys can get out into the
40:49world and um learn from and improve with
40:51what does Improvement mean like how do
40:53you are you measuring emotional
40:54intelligence what is better
40:56yeah yeah we're certainly mentioning
40:58emotional intelligence we're measuring
41:00the fluidity of the conversation we're
41:02measuring you know how respectful it is
41:04we're measuring how even-handed it is
41:08um you know we've already had a couple
41:09of Errors where it's made some
41:12um politically biased remarks and we've
41:15tried super hard to make sure that it's
41:17even-handed no matter how you know sort
41:21of racist homophobic or misogynist in
41:23any way it's it should never be
41:26dismissive disrespectful or judgmental
41:29um it's there to talk through issues and
41:32make you feel heard and
41:33um take feedback like it tries very hard
41:36to take feedback so yeah that's that's
41:38we're measuring all of those kinds of
41:39things but but the next phase of
41:42obviously where we're headed is that
41:44um we really think that this is going to
41:46be your ultimate personal digital
41:50um it is going to as I said interact
41:52with other AIS to make decisions buy
41:54your groceries and you know know manage
41:57your sort of domestic life and help you
41:59book vacations and you know find you
42:03know fun information and that kind of
42:04stuff so it's going to get you know
42:06increasingly more uh you know down that
42:10um you know the other thing is that
42:14um have the ability to access real-time
42:16content in the web so it'll be able to
42:19you know sort of look up uh the weather
42:21and news and other kind of fresh content
42:24like sports results or provide citations
42:28um and you know increasingly add a lot
42:30more of those sort of practical utility
42:32features that you would expect from you
42:34know your personal intelligence
42:36so in my early conversations with my pie
42:40um uh I I guess maybe I shouldn't be so
42:44surprised it's very human and people
42:45like to talk about themselves but I
42:47immediately invested a reasonable amount
42:49of effort in personalizing it right I'm
42:52like okay I hear a bunch of things about
42:53me that you should know what I'm like
42:55and my interests and how you can be
42:57useful to me what surprised you in usage
43:00or maybe you expect it but what would
43:02surprise our listeners
43:03yeah it's a that's a great question I
43:06mean a lot of people proactively share a
43:09huge amount of personal information and
43:13um our memory is is not that long it's
43:15about a hundred messages which is
43:17actually you know it's still quite a lot
43:22what we would really like is to be able
43:25um grab that knowledge and store it in
43:27your own sort of personal brain and have
43:30Pi be your kind of second mind
43:33um able to remember you know all of your
43:36kind of subtle preferences likes habits
43:39relations and so on to be super useful
43:40to you I think in time some people will
43:42want to connect other data sources like
43:44email and documents and drive I think
43:47some people I'm already starting to see
43:48doing that um and so on it's very
43:51interesting to see what people ask Pai
43:54to ask us to do so they're like can you
43:56tell your developers that I really love
43:58this voice I'm really enjoying talking
44:00to uh you know I think it was P2 one
44:03we've just called them P1 P2 P3 P4 our
44:08um and of course some people are like
44:10can you tell your developers that it
44:12should really know that like I wrote you
44:15know the following stories for Forbes
44:17but like I didn't write this story on
44:19this other topic and I was just like
44:22that was a journalist yesterday or the
44:26um you know so yeah it's seeing what
44:28people give us feedback on is really
44:30really helpful okay inflection today
44:32still a relatively small team what's it
44:35like as a company culturally like and
44:37you guys are recruiting what are you
44:41um we're a pretty small team we're about
44:4430 people and we've hand selected a very
44:49very talented uh Team of AI scientists
44:53and and Engineers everybody uh on the
44:56technical side goes by MTS
44:59super important to us that we don't draw
45:01a strong distinction between researchers
45:04scientists engineers data scientists and
45:08anything else to ask that equality and
45:12respect is really important and we've
45:14seen that go wrong at our you know other
45:16labs previously and I think it's an
45:18important modification because everybody
45:19makes a really big contribution
45:22we're very much an applied AI company so
45:25you know we don't publish and we're not
45:27really focused on Research even though
45:29fundamentally what we do do is applied
45:32research and production I mean we we run
45:35some of the largest language models in
45:38um we have state-of-the-art performance
45:41across many of the main benchmarks uh
45:44with the exception of coding because we
45:46don't have Pi generate code and it's not
45:48a priority for us so it's a yeah it's a
45:51it's a very energetic very high
45:54standards environment
45:55um we're very focused on ICS
45:58so everybody is an exceptional
46:01individual contributor and mostly
46:04self-directed so we don't do managers
46:06just yet it's just two of us doing
46:08management which unbelievably has worked
46:12um because we have such senior
46:14experienced people and they're very
46:16driven they know what to do my
46:18experience of building teams like this
46:19over the last you know decade and a half
46:21is that the best people really just want
46:23to work with really high quality people
46:26be given outstanding amounts of
46:29resources and freedom and focus on a
46:32shared goal so we have a very sort of
46:35explicit company goal every six weeks we
46:38we ship and in our seventh week we come
46:41together in person to do a hackathon and
46:45really push super hard as a team because
46:47that forms great bonds and you know it's
46:49it's really fun you know we have drinks
46:51and dinner and hang out and stuff like
46:52that and it's a week of intensity which
46:54closes our launch and then we plan again
46:57for the next six weeks so it's actually
46:58a really nice Rhythm and I found that
47:00most people make up the second half of
47:02their okrs anyway and a 12-week cycle is
47:05just too long and Bs so like six weeks
47:07is actually perfect and it creates a lot
47:09of accountability and a lot of fun so
47:11you know one thing that a lot of people
47:12talk about is how do these models
47:15what is the basis for the next
47:17generation of these types of models
47:19their performance where does the
47:20asymptote how do you think about
47:21scalability how do you think about the
47:24underlying silicon that drives it is it
47:26a data issue is it a compute issue like
47:28I'm just really interested in how you
47:29think about more broadly these really
47:31large scale models since you folks are
47:32building many of them now
47:35the incredible thing about where we've
47:37got to at this point is that all of the
47:40progress in my opinion is a function of
47:43compounding exponentials right so over
47:46the last decade the amount of compute
47:49that we've used uh to train the largest
47:52models in the world has increased by an
47:56order of magnitude every single year so
47:59I went back and and had a look at the
48:01Atari dqn paper that we published in
48:042013 and that used just two petaflops
48:08right and some of the biggest models
48:11that we're training today uh at
48:13inflection uh use 10 billion petaflops
48:17so like nine orders of magnitude in nine
48:19years this is like just insane so I feel
48:22like it's super important to stay humble
48:25and acknowledge that there is this epic
48:28wave of exponentials which is unfolding
48:32around us which is actually shaping the
48:34industry and so when it comes comes the
48:36predictions you have to just like look
48:38at the exponential it's pretty clear
48:40what's going on that's just on the
48:42amount of compute side the data side I
48:44think everyone's super familiar with
48:45we're using vast amounts of data and
48:47that's continuing but I think the other
48:49thing that people don't always
48:50appreciate is that the models are also
48:52getting much more efficient
48:55um you know one of the big breakthroughs
48:57of last year which got some attention
48:59but probably didn't quite get as much
49:01given how many breakthroughs there were
49:02was the chinchilla paper which I'm sure
49:04you know a bunch of you will be familiar
49:06with but you know those are very very
49:07significant result showing that you know
49:10we can actually train uh much smaller
49:12models with much more data for longer
49:15and that was actually compute optimal
49:17and Achieve essentially comparable
49:19performance to the models that were
49:20previously being trained and so that
49:22gives us an indication that it's very
49:24early in the space for architectures and
49:28these models are highly under optimized
49:30and there's a lot of low-hanging fruit
49:32and so that's what we found uh you know
49:35over the last year and a half so
49:36actually the lead author of chinchilla
49:38Jordan Hoffman is on my team here at
49:41inflection and we have a bunch of really
49:43outstanding people who have produced a
49:45number of really awesome proprietary uh
49:47Innovations building on work like that
49:49and so I think both trajectories are
49:53going to play out scale building larger
49:55models is definitely going to deliver
49:57returns we're obviously pursuing that we
49:59have one of the largest supercomputers
50:01in the world uh you know and at the same
50:04time we're going to see much more
50:06efficient architectures which are going
50:07to mean that many many people can access
50:09these models and it's it's in that sense
50:12it's the coming wave of contradictions
50:13in AI That's uh that's what's happening
50:17I have one last question for you so if
50:19you are working on a book I know you
50:21can't say much about it yet but why
50:24you're a pretty busy guy
50:27I love reading I love writing and I love
50:30thinking about stuff and what I've
50:33realized over the years is that the best
50:36way to sharpen your thoughts is to
50:39create hard deadlines so that was like
50:41one of the main things and I'll be
50:44honest like did I regret multiple times
50:46over the last year and a half agreeing
50:50to a book deal with penguin Random House
50:51at the same time as doing a startup
50:54yes like multiple times I was tearing my
50:57already quite gray hair out but uh it's
51:01nearly finished and it has been
51:02absolutely phenomenal and yeah I've
51:04super enjoyed it the book's called
51:07this book's called the coming wave and
51:09it's about the uh consequences of the AI
51:13Revolution and and the synthetic biology
51:15Revolution over the next decade for the
51:17future of the nation state and try to
51:20um intersect the political ramifications
51:22with with the technology trajectories um
51:25at the same time so it's it's been a lot
51:28my hobbies are also this trivial Mustafa
51:33thank you so much for joining us
51:35congratulations on the launch uh and for
51:37our listeners you can try it at
51:40inflection.ai and find pie in the App
51:42Store thanks so much it's really fun
51:44talking to you both see you soon thank
