00:00hey guys I'm Sarah Wang General partner
00:02on the a16z growth team generative AI
00:05has kicked off a paradigm shift that is
00:07already Transforming Our World in our AI
00:10Revolution series we talk to the people
00:12who are actually building the technology
00:14to understand where we are where we're
00:16going and the big open questions in the
00:18field Our Guest this episode is Adam
00:21D'Angelo Adam has built and grown
00:23companies that have connected billions
00:25of people across the globe he was the
00:27CTO of Facebook until 2008 before
00:312009 after being interested in AI for
00:33decades Adam joined the board of open AI
00:36in 2018 and now he's using his
00:39experience scaling some of the biggest
00:40consumer companies in the world to build
00:42po a platform that brings AI to the
00:45masses in this conversation Adam speaks
00:48with a16z General partner and my
00:50colleague David George about building AI
00:52infrastructure for creators the
00:54multimodel multimodal future of AI how
00:57AI will shape knowledge sharing on the
00:59internet and much more there's a lot to
01:01dig into here so let's start at the
01:03beginning of Adam's journey in AI he's
01:05going to take us back to his college
01:06Years in 2005 I was actually I was very
01:09excited about AI early on in in my
01:12career I I remember trying to build some
01:15AI products in in college actually and
01:20it was just very difficult the
01:22technology just wasn't it wasn't there
01:24it wasn't at the point where it was
01:25you're going to be able to make
01:27something that was that was ready for
01:28consumers and meanwhile I just watched
01:30social networking start to boom and you
01:34can actually look at a lot of social
01:36networking technology as it's almost an
01:38alternative to AI so instead of trying
01:41to get the computer to do everything you
01:45could just connect people with other
01:46people over the Internet who who could
01:48do those things in the same way that
01:50like globalization can be a a substitute
01:53for automation um social networking is
01:57you know you can think of it as as like
01:58letting people access everyone else in
02:00in the world for for entertainment for
02:03fun for communication for whatever
02:05whatever you want to do I think it was
02:06just incredibly powerful technology and
02:10given that AI wasn't quite there yet um
02:13that was just kind of like that that was
02:14the main thing that um that there was to
02:16do to to apply all the technology to so
02:19so I I first got interested in in social
02:21networking and then through my
02:24experience at Kora we started out with a
02:26product that was entirely human driven
02:30people would come and ask questions and
02:32they would put topics on them and other
02:34people would sign up to answer questions
02:37and they would tell us about what they
02:41knew about by by tagging themselves with
02:43these topics and we would we would try
02:44to Route the questions to the people who
02:46knew about the particular topic and it
02:50was all manual but we knew we knew that
02:52at some point we're going to get to the
02:53point where software would be able to
02:56generate answers we we ran some
02:58experiments using GPT three to generate
03:01answers and compare them to the the
03:03answers that that humans had written on
03:05Kora and a lot of the time
03:08gbd3 could not write as good of an
03:10answer as what the best human answer was
03:13that that that had been written but it
03:15could write an answer instantly to any
03:17question and the the constraint for cor
03:19had always been the amount of time that
03:21highquality answer writers had to to
03:24answer questions and so thing that was
03:26really new about llms is was the ability
03:29to at extremely low cost generate an
03:32answer instantly to any question through
03:34that experience we we realized that a
03:37chat kind of experience where you can
03:39write a question and and then get an
03:41answer instantly from from AI was more
03:44likely to be the the best Paradigm for
03:46interacting with AI as opposed to this
03:48kind of like publication Paradigm that
03:51that cor had yeah of course and So based
03:54on all that we we landed on building po
03:56as a new a new chat oriented AI product
03:59I think many people will be familiar
04:01with po but explain just for us how does
04:03the product work uh how do you how do
04:05you find it in the first place how do
04:06you interact with it uh in the same way
04:08that Kora Aggregates Knowledge from many
04:12different people who have knowledge and
04:13want to share their knowledge we want po
04:16to be a way for people to access AI from
04:20many different companies and and many
04:21different people who are are building on
04:23top of AI um and so you can come to PO
04:27and use it to talk to a very wide
04:30variety of models that that are
04:32available today and then we have all
04:33these other products that people have
04:35built on top of these models um and
04:37we've got an open API where anyone can
04:39hook in so anyone who's training their
04:42own model so like any of these research
04:44teams anyone who's doing fine tuning
04:47they can take their model and put it on
04:50PO and what we we allow is for them to
04:53reach a big audience quickly um so so
04:56you know we thought about as a company
04:58Kora what what is the role that we're
05:00going to play in this new world with AI
05:03and what are the strengths that we had
05:05and and what what have we learned over
05:06over the past 10 years building an
05:09operating Kora and there's actually a
05:11lot of this kind of like consumer
05:13internet knowhow that's important in
05:15getting a product to to mass market so
05:17this is things like building
05:20applications across IOS and Android and
05:23Windows and Mac localization of the
05:25interface AB testing subscription
05:30all these other kinds of small
05:32optimizations that you need to to make a
05:35good consumer product we want po to be a
05:37way for anyone who's creating AI whether
05:39it's one of the the big Labs or or an
05:42independent researcher we want it to be
05:44a way for them to to get that uh model
05:47and make it available to to mainstream
05:50users all around the world there's a lot
05:52that you just said that I would love to
05:54go deeper on so one one of the things
05:55that you said you sort of listed off all
05:57the models that you make available
05:59there's one Theory which is you know one
06:02one model one company uh is going to
06:05provide everybody you know the solution
06:07that they need for everything right
06:08there's another theory which is there's
06:10going to be tons of different you know
06:12models for different use cases uh the
06:14world's going to be multimodel and
06:16multimodal the theory behind PO is that
06:19the the future is going to be multimodal
06:21and multimodal why do you think that's
06:23the case I think nobody knows how the
06:25future is going to unfold but we think
06:28that there's going to a lot of diversity
06:31in in the kind of products that people
06:33build on top of these models and and and
06:35and in the models themselves I think
06:37there there are a lot of trade-offs
06:39involved in making one of these models
06:41you have to decide what data are you
06:43going to train on it what kind of
06:45fine-tuning are you going to do what
06:47kind of uh instructions do you is the
06:50model going to expect you to give as a
06:52user does what kind of expectations do
06:54you want to set with your users about
06:56what to use the model for and I I think
06:59the same way that the the early internet
07:03had this huge explosion of of different
07:06applications I think we're going to see
07:08the same thing from PR AI so so early on
07:10in in in the internet the web browser
07:13came along and made it so that anyone
07:15who is making an internet product they
07:18didn't need to build a special client
07:21and get distribution to people around
07:24the world they could just build a
07:26website and this one web browser could
07:28visit any website sure and in the same
07:31way we want po to be a single interface
07:33that that can be used so that people can
07:36use that to talk to to any model we're
07:38betting on diversity just because
07:40there's there are so many talented
07:42people around the world who are going to
07:44be capable of of tuning these models you
07:47can tune the open source models today
07:49there's also products from openi and
07:51anthropic and and I think Google's close
07:54to having something where where you'll
07:55be able to fine-tune all these models
07:57and everyone has their own data sets
07:59everyone has their own special
08:01technology that they can add to the
08:03models and and I I think through the
08:06combination of all of this we're just
08:07going to see a very wide diversity of of
08:09things you can do with with AI there's
08:12two things that I'd like to maybe go
08:13deeper on there uh so one is you know
08:16the idea of like what constitutes the
08:18product itself what is it today and then
08:20what is it going to have to become uh
08:22and then secondly you know the the the
08:24idea of like the long tail right like
08:26bet on the long tail incentivize them
08:28give them a platform abstract away a
08:30bunch of the infrastructure that they
08:31don't know how to build um and harness
08:33really what they're great at right um so
08:36on the first what is the product like is
08:38the model like is today the AI model
08:40many people probably say like is largely
08:42the product um what are the advances
08:45that you anticipate seeing that are
08:47going to sort of change the way people
08:49interact with these um enable new kinds
08:52of products being built you know one way
08:54of of thinking about that is are the
08:55model providers themselves going to be
08:57the ones that build all the products if
08:59you're a large model Creator and you
09:01have you know you have tens of employees
09:04that you can allocate to to building a a
09:08consumer product and you have the the
09:10culture to do that then then you can you
09:13know you can go direct to Consumer and
09:14and you can build a good product I think
09:17that most of the people who are training
09:20these models are not in in that position
09:24if you want to take this take your model
09:25and bring it to to Consumers all around
09:28the world you got to think about you
09:29need an IOS app you need an Android app
09:31you need desktop apps you need a web
09:34interface you need you need to do
09:37billing in all these different countries
09:39you got to think about taxes and um
09:42there's just there's just a lot of work
09:44and you could either spend you know
09:47you're you you raised some Venture
09:48funding you could either spend some of
09:50that funding on hiring out a whole team
09:52and developing all those competencies or
09:55you can spend that on making your model
09:57even better yeah and
09:59I think different startups will choose
10:01different paths here but I I think for a
10:03lot of them the the right path is going
10:06to be to just set up an API or plug into
10:09the PO API and and use that to to get to
10:14um to a lot of consumers very very
10:15quickly yeah talk about the role that
10:17the sort of long tail of creators then
10:19plays and like how do you want to engage
10:21with them and like what's the incentive
10:23for them to want to build on top of Po
10:25as opposed to other places yeah so so we
10:27have a a revenue sharing program that
10:30allows people to to get paid as a result
10:33of people using their their bots on PO
10:36and and there's a you know it costs a
10:39huge amount of money to to provide
10:42inference for for these models and so
10:45it's almost no other platforms provide
10:48this kind of Revenue share today so if
10:50you have a model that requires a lot of
10:51gpus to to do inference on then this is
10:56this is really your best place to to
10:58come and and you have a real business
11:00you can you know you can cover your
11:01inference costs and make more and we
11:03think a ton of innovation is is going to
11:06come from from these companies there's
11:08other companies that are building things
11:09on top of some of the big models so say
11:12from from open aai and in that case they
11:14have to pay the open a inference cost
11:16which is another sort of source of Need
11:18for um for for money and so in the the
11:22PO Revenue share model works works in
11:24the same way where where it'll let you
11:26afford your your cost that you're then
11:28paying on to to any other inference
11:30provider yeah absolutely what are some
11:32of the really fun and interesting uh
11:35things that creators have have already
11:37built on top of Po a lot of people right
11:38now are excited about image models we
11:41have uh there's uh stable diffusion SC
11:44cxl and and then we let users go and and
11:48do some prompting to customize it to to
11:51provide art of a particular style so
11:53there's there's like these like anime
11:56style sdxl bots on on PO those are
11:59popular there's this company called
12:01playground they're making a product for
12:04people to edit images but in the process
12:09they've created a a pretty powerful
12:11model and and they have that model
12:13available on PO and that's um that
12:16that's that's gotten pretty popular
12:17recently yeah it's so cool that you can
12:19have a long tail of these creators make
12:22their own sort of opinionated style uh
12:25of these these base models but I think
12:27there's something to that where you
12:28provide you know the the sort of
12:30infrastructure and support uh and then
12:32sort of let the let the users or
12:34creators um you know do what they do
12:36best yeah and it's all you know I think
12:39it's it's super early days right now but
12:42I think what we're going to see over
12:43just the next year or two is going to be
12:46incredible this will this will go from
12:47being sort of useful to some people
12:51right now to being something that's just
12:53critical to to many different tasks that
12:56that anyone is going to try to
12:57accomplish yeah there's an there's a a
12:59really good you know analogous company
13:02that you and I both know very well um
13:04which is Roblox right early days um you
13:07know creators were on there building
13:09games and they were pretty basic the
13:11early days and it was a lot of kids uh
13:14you know kind of learning how to build
13:15games and and then it sort of graduated
13:17eventually to people who were able to
13:19earn a living so I think the ideal for
13:21you would be you build enough scale um
13:24they can build large enough audiences to
13:25to actually be sort of Professionals of
13:27what they're doing yeah and we're we're
13:29we're spending millions of dollars
13:32already on on inference it's it's mostly
13:34going to the the large model providers
13:36right now but we want to let as much of
13:38that as possible go off to these
13:40independent creators cool I want to
13:42shift topics um and get maybe a little
13:45bit more like conceptual AI you CTO at
13:48Facebook at the time where social was
13:49emerging and then right when the
13:51platform shift to mobile was taking
13:53place right so I'd love your thoughts on
13:57what are the similarities to the shift
13:59of mobile and this AI wave and what are
14:01some of the big differences yeah you
14:03know I I I think it's it's very hard to
14:05say I I think with Kora I think we were
14:08a little bit slow to adopt to mobile you
14:11know mobile was one of the things on our
14:13list of many priorities and it needed to
14:15be the number one priority and we needed
14:17to make tough tougher tradeoffs to to
14:20prioritize it you know we needed to do
14:22things like hire a set of different
14:25people who were who were going to focus
14:27on it and and and really take some you
14:30know have a period where we release no
14:32new features and we were just
14:34simplifying things because the the
14:35mobile UI called for a different
14:37experience when you have such a critical
14:40change in the platform structure you
14:43need to rethink so much that it's only
14:45going to happen if you have this very
14:48strong kind of top- down leadership and
14:50so you've done it differently this time
14:52around yeah yeah right so yeah talk
14:54about some of the organizational changes
14:55and you know what you've done to
14:57actually refocus yourselves on the big
14:59thing that's right in front of us here
15:01yeah so I mean I think the first thing
15:02is you know just identifying this trend
15:03and then starting off doing some
15:05experimentation early on um just to
15:08learn and and that didn't require any
15:10kind of strong decisive leadership as
15:13much as it just required paying
15:14attention to the the market and and then
15:17but then from that experimentation that
15:20got us enough conviction that so we in
15:22our case we said hey too much of the cor
15:25product has been built up around this
15:27publication model is sort of
15:30fundamentally premised on the idea that
15:32expert time is going to be scarce and
15:36and the AI the llm time is not scarce in
15:39the same way and so we need to rethink
15:41that this was in uh I think August of
15:452022 we got to this conclusion that chat
15:48is the right Paradigm for this and we
15:50need a new product we don't want to try
15:52to just trying to retrofit everything
15:54into Kora is we thought was just going
15:56to we're going to move too slowly uh so
15:58we so we had a small team just start
16:00working on PO based on based on that
16:02talk about maybe the relationship
16:04between Kora and Po and how do you
16:07actually Envision that changing in the
16:09future and then maybe there's even an
16:10extrapolation of like okay cor and Po
16:14and like human experts and you know AI
16:16experts answering questions do they do
16:19it in the same place is it a different
16:20way of interacting yeah yeah know we
16:22we'd love to have all of this as
16:24integrated as possible you know I think
16:26if you think about maybe the
16:27relationship between Facebook and
16:29Facebook Messenger these are two
16:32products built by the same company but
16:34they they share a lot I I think that PO
16:36and kaora might evolve to a a similar
16:38kind of relationship we'd love to get
16:41more of the human aspects of Kora into
16:44po we'd also love to get the whole core
16:46data set into the the PO Bots and we and
16:50we're also working we've we've launched
16:52some of this already to get some of the
16:53PO AI to generate answers that are
16:57available on Kora as these models
17:00continue to scale up the the quality is
17:03going to go higher and higher to the
17:04point where it actually will be as good
17:06as human quality in in a lot of cases
17:09and and so the core of Paradigm actually
17:11I think is is actually becomes more
17:13appropriate for for AI as as the cost of
17:16inference gets gets higher get lower
17:18yeah and model quality gets better yeah
17:19yeah yeah so we'll see what the exact
17:22right relationship is but we think of
17:24this as we're we're building a a network
17:26for people and AI to all share knowledge
17:31together and sometimes the the people
17:33will be getting Knowledge from the AI
17:35and sometimes the AI will need to get
17:37Knowledge from from humans and we'd love
17:39to be as much of a a conduit for that as
17:42as possible yeah and cor or po depending
17:45on how they interact like is a place you
17:47get answers and sometimes your answer is
17:50going to come from an expert and
17:51sometimes it's going to come from AI
17:52yeah right yeah yeah what what do you
17:54think about just the internet like you
17:56extrapolate that out are people going to
17:58be engag engaging with this collection
18:01of bots that have different
18:02personalities and different expertise
18:05and will those be interspersed with real
18:07humans you know will like real humans be
18:10interspersing in the AI like what what
18:12do you think actually happens I I
18:13personally I think that humans are
18:16always going to play some role there's
18:18knowledge that people have in their
18:20heads that is not in is not on the
18:23internet and is not in any book and so
18:26no llm is going to have that knowledge
18:29so yeah Andre Kathy called the llms uh a
18:32lossy compression algorithm of the
18:35internet and it's like it's just of the
18:37internet like there there's experts that
18:38know a lot of stuff that's not that
18:40right right right so you know I I think
18:42there's there's a lot of potential in in
18:44the kind of interplay between between
18:46humans and the the llms going forward I
18:51people um they LM have this problem with
18:55hallucinations right now and and I think
18:57hallucinations are gonna the rate is
18:59going to go down as the models get
19:00better but it's never going to get to
19:03the point where it's 100% perfect and so
19:07there's I I think a there will be a lot
19:09of value placed on the idea that you
19:13know the source of your information you
19:15know which human said it or which which
19:18publication originally printed it and I
19:22expect that that is going to lead to
19:25some kind of some kind of product or
19:27some kind of user experience where where
19:29the llm is helping you sort through your
19:32sources and quoting exact experts or
19:36exact sources as opposed to just
19:38synthesizing it all and giving you
19:40something where you can't exactly trust
19:42where where it came from yeah and is
19:44that a new technology that gets built
19:46outside of the models themselves or do
19:48you think that that's Incorporated
19:49inside of the model I I could see it
19:51going either way I mean it's it's if if
19:53you just look at a a model the raw model
19:57doesn't have access to these other
19:59databases where it can get exact and so
20:03it'll have to be some augmentation of
20:05the model but how tightly integrated
20:08into the model it'll be I I think I
20:10think we don't know yet yeah I agree I
20:11think that's going to be uh going to be
20:13critical you know like it's it's it's
20:15one thing like we've started out with
20:16these use cases of like companionship
20:19and creativity and like hallucinations
20:21are a feature of that right like that
20:24like makes it more fun and exciting uh
20:26especially when you get into you know
20:27business use cases is or more utility
20:29type stuff um you know it's it's it's
20:32obviously needed um what are the other
20:34big advances that you're excited about
20:36just broadly in the AI in the AI space
20:39for for language models I personally the
20:41most excited just about scale just
20:44continuing on the current Paradigm if
20:47you just play this forward there's so
20:49much further that it can go and and you
20:51think the scaling laws will hold are
20:54holding so far they have held my
20:57prediction would be there there are some
20:59issues that need to be overcome but
21:01there's just this incredible industry so
21:04many talented people right now who are
21:08to make this technology advance and
21:11there's so much money behind it the
21:13force that's there to help overcome any
21:17road bumps that we hit is so massive so
21:20I expect that it's just going to
21:22continue I think there will be Road
21:24bumps and and there'll be there'll be
21:25issues that need to be worked around and
21:26there will be breakthroughs that
21:27probably need in creativity but we have
21:32many of the smartest people in the world
21:34the most determined people in the world
21:35the most talented people in the world
21:37all focused on this problem and I think
21:41we're going to just continue to see the
21:42kind of exponential growth progress that
21:44that we've that we've had so far for I I
21:47think they'll go on for many years we
21:48talked about the last shift right like
21:50the mobile shift that you live through
21:51and some of the lessons that you had
21:53from it what do you think ultimate
21:55Market structure looks like in in the
21:58Gen space in order to train these
22:00Frontier models you need billions of
22:04dollars of capital and you need many
22:08years of investment in infrastructure by
22:12there's a very small set of people who
22:14can who can do that and so that's
22:17leading to this world where there's only
22:20a small number of players that can be on
22:23the frontier and so it's you know right
22:25now it's open AI Google maybe
22:29anthropic maybe maybe meta can be there
22:32those who can get there I think it's
22:34going to be good business you'll be able
22:35to make a lot of money um you can have
22:37good profit margins um you'll have to
22:39you'll have to work very hard to stay on
22:41the frontier to keep up but it's not a
22:44it's not a commodity mhm I think when
22:46you go six months behind the frontier or
22:49even definitely one year it's brutal
22:53there's just way too many people that
22:55are able to get the the capital and the
22:58the resources to train those models and
23:00so it's going to be either fully open-
23:04sourced or there will be to many
23:06different competitors for anyone to make
23:09a good business at that at that point on
23:11the pure technology I do think there
23:13will be very good businesses at that
23:16level um that you know not using
23:18Frontier models but are combining some
23:21other kind of unique thing with the
23:23model so it might be that you're
23:26providing some tool that the model can
23:27use or you have some unique data that
23:30you're using for fine-tuning or there
23:33might be some other some unique product
23:35you build around the model and then that
23:37ends up being the the source of
23:39competitive strength so I I think
23:41there's going to be this kind of choice
23:42where you're either competing
23:44on scale by being on the frontier or
23:48you're competing on some kind of like
23:50feature differentiation and in that case
23:53you don't need a a Frontier Model and
23:55and some in some cases you'll have both
23:56so you know you might be able to use the
23:59open ey API and combine that with some
24:00unique tool that you're providing and
24:02that that could be a good business as
24:03well yeah you sort of get once you get
24:05beyond the foundation models you get to
24:07kind of like more traditional forms of
24:10business differentiation competitive
24:12differentiation like competitive
24:13Advantage you know sources of modes and
24:15things like that which which I think
24:17totally makes sense yeah and and I think
24:18what's interesting about this is that
24:20it's it's evolving so things are moving
24:22so quickly and so every you know every 6
24:26months the frontier moves forward and
24:28and so the the frontier players are
24:31they're you know they have to invest
24:32more Capital but they but then they have
24:33much more powerful models that open up
24:35even bigger markets um but but then you
24:37know the the open source kind of one
24:40year back Frontier that that that's also
24:43moving forward so that that's the
24:45markets that that can address are are
24:47getting bigger and bigger and and so
24:49it's you know I I think every year that
24:51goes by we're going to have this much
24:55larger market that can be addressed by
24:57the technology and and all the products
24:59that are that are built on top of it so
25:01yeah that that that sort of brings me to
25:02another topic which is related to Market
25:04structure which is you know incumbents
25:06versus startups uh and and we uh in the
25:10seat that we're in we hope that the
25:11startups always win um but in the last
25:14cycle and maybe just from a B2B lens
25:16here uh like in the last cycle um you
25:18know SAS and Cloud um there were a bunch
25:22of things that made it really difficult
25:23for the incumbents to actually innovate
25:25there was a business model Innovation
25:27and sort of new talent and Technology
25:29required um which opened the door pretty
25:32widely to startups there's a a take out
25:35there now on AI which is this time is
25:37different and the incumbents are the
25:39real winners right um because the
25:41technology is available by simple API
25:43you can plug it right in and they have
25:45distribution so they should be uh they
25:47should be the winners and you know if
25:49you just sum up Microsoft and and
25:51Google's business apps and all these
25:53things it's probably somewhere between
25:5510 and 20 billion dollars of Revenue
25:56over the next one to two years um I'm
25:59curious if you have a take on that if
26:00that's if that's consistent with how you
26:02see it or if you if you see it
26:03differently yeah I I think I think it's
26:05going to vary so definitely the
26:07incumbents they have they're going to
26:09have access to the technology and
26:11they're going to have distribution and
26:12so that's that's a big advantage that
26:14they have I think the opportunities for
26:18new players in this wave are more in the
26:22cases where the kind of product you want
26:25to build around this technology is
26:27somehow fundamentally different than
26:30what was built before and so as an
26:33example the the hallucination problem
26:36that's in some ways a good thing for
26:38startups because a lot of the existing
26:41products out there have zero tolerance
26:44for anything that's going to um to have
26:47a risk of of producing something wrong
26:50and so so you can see this with uh I
26:53think with perplexity getting share from
26:55Google right now yeah Google can't just
26:58go and put something on all their search
27:00results where it has a you know a few
27:04per chance of being wrong yeah that
27:06would be a huge problem for them
27:08perplexity that can just be the
27:09expectation when you're using that
27:10product that it's you know it's it's
27:12almost always right even though there's
27:14a small chance that that it's wrong I
27:16think that same thing is actually going
27:17to play out in a lot of other cases
27:19where the products you build around this
27:21they need to they need some kind of
27:23fault tolerance um and there needs to be
27:25a user expectation that everything's not
27:27perfect and the cost Advantage can be so
27:30great for this right like if you take a
27:32highly paid person like a lawyer and you
27:35know you you run it through an llm which
27:37it it cost cents versus like a, bucks an
27:39hour like maybe you just should have a
27:41really high fault tolerance and you just
27:43have to double check a lot of the work
27:45and that's just a different workflow
27:46that's different way of engaging right
27:48yeah yeah no and and so you know you
27:49have these these entrenched companies
27:52that maybe have a a very strong brand of
27:55never making a mistake or never messing
27:58up always being reliable and a startup
28:01can just come in and say like okay well
28:03this is going to cost a tenth or 100th
28:06the price but it's going to have a small
28:08chance of getting things wrong and
28:10that's that's actually a lot of people
28:11would prefer that um but it's it's a
28:15real problem for the incumbent because
28:17they they can't compromise their brand
28:18yeah that's a great point I guess just
28:20to close it out uh I'm sure a big part
28:22of the audience here is Founders who are
28:25building in probably earlier stage than
28:26you what what advice uh do you have for
28:29people building in AI I think what I
28:31would do if I was starting a new company
28:34right now is just spend a ton of time
28:35playing with the models and playing with
28:38integrating them with different things
28:40you know there's there's so many
28:41different inputs you can give to the
28:43models you can you can make scrapers
28:46that ingest data from anywhere you can
28:48get data from the user's local screen
28:51you can get data from voice and there's
28:58needs people have and such a huge space
29:01of different like inputs you can combine
29:03to to try to address those needs I think
29:05it's very hard to just kind of like
29:07think top down about where there's
29:10demand in the market I think I think
29:12experimentation is really the way to go
29:14to to generate ideas and and to to set
29:17up a you know a startup that's going to
29:19be able to to build something really
29:21valuable yeah and have a place in the
29:22world for sure awesome well thanks for
29:25being here I appreciate this is fun
29:27thanks for sharing the time thank
