00:05hi no prior listeners time for a host
00:07only episode this week elad and I are
00:10going to talk about what's going on at
00:12open aai of course video qar uh what
00:16might be next in research and some
00:18predictions okay elad we have to start
00:21with the Saga from this past week what
00:24is your take on the outcome and the
00:26second order effects from a second order
00:30um this seems like overall positive news
00:32for everybody involved so it looks like
00:35on the openi side they're back to being
00:37in a really positive stable situation I
00:39think they still have like the leading
00:40model in GPT 4 um they've reworked the
00:43board which seems like a positive thing
00:45so you know imagine if this had happened
00:48two years from now three years from now
00:49Etc so it it seems like it would n
00:51increase the stability of the company in
00:54governance and a few things like that or
00:55the nonprofit and the company in
00:56governance so as an external viewer seem
00:58like a uh pay painful thing to go
01:00through but the flip side of it is it
01:02seems like they're moving forward and
01:04moving ahead in a positive way um and
01:07then in parallel I think it may have
01:08ramifications to other areas um that we
01:10can talk about if useful like what are
01:12the second order aspects of this but be
01:14great to hear what you think yeah I
01:16think the first uh obvious lesson is
01:19that governance matters right and this
01:22isn't an area where I think most
01:24companies are that experimental but I
01:26think a lot of entrepreneurs are likely
01:28to think twice about placing their
01:31Destiny in the hands of groups with
01:33explicitly mixed incentives now I'd say
01:36generally nonprofit governance is not
01:39every organization but as a class known
01:41to be kind of abysmal right because
01:43performance is hard to measure
01:44objectively and so it often ends up more
01:47about politics and specific
01:49relationships and Status games than
01:51outcomes the clarity around how much you
01:54know any board matters was kind of a a
01:57wakeup call for people the second lesson
01:59that a lot of people will take away or I
02:02think should from this whole Saga is
02:05that money matters right the factors of
02:07production are labor and capital and
02:09compute is the AI specific form of
02:11capital Microsoft holds the compute here
02:14and that clearly matter this is
02:15amazingly well managed uh and supported
02:18by saan Kevin and then the class of like
02:22really special labor here the team
02:25rebelled and the board obviously
02:27underestimated the level of support that
02:30Sam and Greg had from them and then I
02:33think one thing that is often unsaid
02:36because it's a little bit less
02:38idealistic is that a lot of open AI
02:39people were very upset this last week
02:41about not just the destruction of the
02:43mission which I think was absolutely
02:45like genuine but also destruction of the
02:47value they' built and been promised a
02:51piece of with the 86 billion uh tender
02:54offer it's just a reminder that labor
02:56and capital are like leverag they're
02:58stakeholders and there's no there's no
02:59free launcher control without skin in
03:01the game and it I think they're um
03:04likely shouldn't be from the view of
03:07many of the people involved in this yeah
03:09I think you're raising an important meta
03:11point which is basically what are the
03:12incentives that different organizations
03:14have in place and ignoring open AI I
03:16think there's a lot of boards which
03:17added board members for reasons that
03:20were politically motivated or motivated
03:22by different regulations getting passed
03:23that for certain board changes Etc and
03:25you also see that in executive teams and
03:28I think it's really important for people
03:29to go back and rethink okay who should
03:31be on my board and what why what are
03:33they representing relative to the board
03:35what expertise do they provide or what
03:37insights are they bringing or what
03:38strategic views are they bringing same
03:40with your executive team and also what
03:43incentives and you know there's this um
03:46this view that's kind of been moving
03:48around Silicon Valley in terms of the
03:49professional managerial class right
03:50people who have Alliance not to the
03:53organizations they work at but external
03:57incentives and those external incentives
03:59could be speaking at Ted or going to
04:01Davos or getting Kudos or an award from
04:05a specific organization versus doing
04:07what's actually their Duty as a as a
04:10representative of the various
04:12shareholders of a company in the context
04:13of a corporation and so there are these
04:16fiduciary duties that may be being
04:17breached by um other incentives for
04:21different actors who've been added over
04:22the last you know 5 10 years to boards
04:24to Executive teams Etc and I think it's
04:26really worth rethinking like who do I
04:28want on board and what
04:30and it also comes back to some of the
04:31companies that have been resetting
04:33things relative to politics I think
04:34Shopify did a great job for example of
04:37saying we're performance-based culture
04:39we're focused on um you know a very
04:42specific Mission we don't want that
04:43mission to creep we're about we're not a
04:46family you know like if your uncle shows
04:48up drunk and does something bad you
04:50forgive him if a board member shows up
04:52and does that then you know you don't
04:54want them on your board right they're
04:55being irresponsible there's also that
04:57broader context of like how do you want
04:59to think about alignment incentives
05:03motivations and you know is this a good
05:05moment in time to sort of pause and
05:06rethink some of those things relative to
05:08to your own company yeah one friend at
05:10open aai who uh I guess publicly
05:13declared that this reignited their
05:16belief in clear incentives and some um
05:20and good intent and capitalist
05:22structures that has actually seemed
05:23somewhat radical in uh in many Silicon
05:27Valley companies over the last few years
05:29but but I think that is going to get
05:30rethought when you see what happens when
05:32there are unclear or misaligned
05:34incentives yeah there there's two
05:36actually related quotes to that there's
05:37one which is something which I'm going
05:38to get wrong which is something along
05:39the lines of like capitalism is the best
05:41way to take care of people that you
05:42don't know you know it's the means of
05:44actually growing the pie in many cases
05:46and providing for for others through the
05:48sort of incentive of markets but the
05:50other one is a Charlie merism and
05:51unfortunately Charlie Munger passed away
05:53earlier today and obviously he was sort
05:55of a giant of industry and he had this
05:57great saying which was anytime I think I
06:00understand the importance of incentives
06:01I realize that I'm underestimating the
06:03importance of incentives if we just
06:05think about what the other second order
06:07like more commercial effects are I do
06:09think that uh there is an interest in
06:15owning models more in open source models
06:18and in um at least understanding like
06:21Reliance on a single vendor what do you
06:24think here yeah I think um there's a
06:26couple people who built Solutions during
06:28the last week that for examp example
06:29Brain Trust now has a AI proxy where you
06:32can use the open AI SDK to effectively
06:35query multiple different uh models
06:37including mistol um and llama through
06:40perplexity as well as a variety of other
06:42things GPT 4 GPT 3.5 um I think
06:46potentially anthropic and so it just
06:48allows you to be able to both load
06:50balance your uh queries or prompts but
06:53also interchange models more easily so
06:55you can actually look at performance
06:56across them I think Chima similarly has
07:00over the last week that they've released
07:01that helps with some of the proxying and
07:03other things and so I think there's
07:04Solutions like that that have started to
07:06be accelerated into a market that would
07:07have happened inevitably right I think
07:10everybody the journey that I see people
07:12often take is they'll prototype on gp4
07:15they'll look at how good it is and then
07:16they'll either keep it on gbd4
07:18particularly if they need like Advanced
07:19CH logic or other things or if they need
07:21very high throughput in performance and
07:23low cost and sometimes they'll either
07:24switch to gbt 3.5 or they'll see if they
07:27can fine-tune something right and that's
07:28the only people with enormous scale like
07:30I don't see very many fine tunes
07:31happening in general unless you know
07:33somebody has enormous scale and or
07:35proprietary data that they just don't
07:36want to get out right so they'll F tune
07:38mistol or llama or something so already
07:40I think people were thinking about that
07:42and that means you need to build an
07:43orchestration layer you may need the
07:44proxy you may need a VAR of things so
07:46the dimensions that people were
07:50evaluating an llm provider on or whether
07:53not they wanted to um control or uh host
07:58or find tuned themselves just became
08:00more clear right um where like
08:04reliability um became more important but
08:07the reliability latency um cost control
08:11capability questions were sort of
08:13naturally there and to be clear like
08:15open AI leads on capability in many
08:19areas in unique capability in in some
08:22right like cod generation GPT 4V right
08:25you can do amazing things with that and
08:27people uh should go build on those tools
08:29I think the ecosystem will mature and
08:31opening eyes a great partner but uh I
08:33think the questions are just much more
08:35obvious for anybody relying on these
08:37models now yeah and I think honestly a
08:39lot of the bigger Enterprises I knew uh
08:42always wanted to make sure that if they
08:44really needed to that they could Second
08:45Source something so I don't think this
08:46is a new thing in other words you know
08:49one could argue that no matter what open
08:52does there'll always be at least one or
08:54two other suppliers or vendors or
08:57partners for Advanced llm simply because
09:00the market always wants an alternative
09:02even just for negotiation leverage and
09:03so if you look at other markets for
09:05example in the router world one of the
09:07main reasons Juniper exists is because
09:08everybody wants to make sure that they
09:09can push on Cisco for pricing and so
09:12they always want to have a second Source
09:13that's why Juniper is always 10 to 20%
09:16size of Cisco right it's just second
09:19sourcing or AMD versus Intel for a very
09:23time so I think often markets will end
09:26up with other players just because bner
09:29prizes always want to have that option
09:30if they need it even if it isn't as good
09:33and if anything I think open AI kind of
09:35emerges more stable through this in ways
09:37that people didn't expect simply because
09:39there's going to be more stability at
09:40the board level in a way that people
09:42didn't understand perhaps that there
09:43could have been instability right this
09:45is a strengthening event and a focusing
09:47event for the company at least you know
09:49from what I can tell
09:50exter do you want to talk about Pika
09:56video yeah there are a couple really
09:59amazing launches happening in the video
10:03space what's the cause for this like we
10:06we suddenly have text to video
10:07generation and Avatar cloning in
10:10different ways what do you think is
10:11going to happen in this space
10:12interesting shift has been happening
10:14because basically if you go back a year
10:15and a half mid Journey launched staple
10:17diffusion came out Dolly 2 came out and
10:20there's a whole wave of people saying
10:22that they were going to go build on
10:22diffusion models and image gen was like
10:24the thing that everybody was going to go
10:27do for like two months and then GPT came
10:29out and then everybody was like oh my
10:30God I need to go work on llms and
10:32language and natural language and NLP
10:33and all this stuff and so the
10:35entrepreneurial ecosystem went through
10:38this sort of zigzag where everybody was
10:40going to do image gen and a bunch of
10:42companies started going down that
10:43direction and then that the llm stuff
10:45really kind of was substantiated through
10:46chat TPT and then most people went that
10:48way and a handful of Founders stuck
10:50around on the diffusion model side and
10:53diffusion models you know are really
10:55popping up and obviously there's like
10:56image Transformer and a bunch of other
10:57stuff but they're mainly being used for
10:59image gen for video and for audio
11:02actually and so there's a wave of people
11:04who've continued to work and crank on
11:06this um and they're starting to come out
11:08with really interesting products for
11:10example p is a great example where it
11:11was two um Stanford PhD students who've
11:14been working on diffusion models for
11:15some time and they made this you know
11:17really amazing creative um text to video
11:20engine there are other companies like a
11:22haen or CIA or others that are doing you
11:25know let me use these diffusion models
11:27to clone an avatar to generate an avatar
11:29person so that they can either go into
11:31the metaverse alak OR alternatively they
11:34can use it for marketing purposes they
11:36can use it for internal training they
11:38can use it for all sorts of applications
11:40and then there's some really cool like
11:41audio based things coming out too which
11:43I think are starting off more sort of
11:44tools to create music or just simulate
11:47voice in the context of um a soundtrack
11:50or you like make EDM and you want to add
11:52voice to it right and you can just now
11:54kind of do some really interesting
11:56things there so it seems like there's
11:58this really interesting Renaissance um
12:00that's happening in part due to
12:01diffusion model work and in part due due
12:04to a handful of Founders not getting
12:06distracted by llms which are Super
12:07exciting obviously um but wanting to do
12:10things in video it's a really exciting
12:12Trend and I'm I'm guessing the success
12:13of some of these companies and their
12:16their traction and growth is going to
12:17pull more people over to to work in this
12:20area again I think it was just an area
12:21of less emphasis for the last year
12:22relative to language one of my favorite
12:25dynamics that happens in sort of
12:27Technology e systems is that once people
12:32show that something is possible like a
12:35lot of talent floods in right and you
12:37kind of get oh you get a lot of
12:39competition but you also get um
12:41Innovation coming in waves that could be
12:45with mraw developing open source models
12:48that are uh actually interesting from a
12:51reasoning perspective at relatively
12:53small parameter size or it could be Demi
12:56and chenling and the Pika team creating
12:59text to video generation models that are
13:02really interesting quite efficiently
13:06perspective and I know we're both
13:08investors here but I've seen a huge wave
13:11of people interested as you said in
13:14media diffusion of different kinds now
13:17that they know it's possible and it has
13:18real benefits right because it's um it's
13:22do and from a data set perspective the
13:25data is reasonably straightforward to
13:26get I mean it's hard to get but it's not
13:28as hard as you know the entire internet
13:30and transcribing voice from videos and
13:32all the rest of it the original um
13:33stable diffusion model supposedly was
13:35trained on like 600k of
13:37GPU yeah these models cost in the
13:39millions to train not tens of millions
13:41at least initially right and so that's
13:43another big difference relative to the
13:45really big foundation models and
13:46language models and all the rest right
13:48and so you can actually imagine that in
13:49the language world you're going to have
13:51a lot more platforms that people build
13:52on in the diffusion model world uh image
13:55video Etc you're going to have more
13:57people kind of grow their own right and
13:58people should still potentially try
14:00things on stable diffusion first just to
14:01test it out it's back to the you know no
14:03GPU before product Market fit but they
14:06can still train their own model in a
14:07very economic way relative to like the
14:09amount of money a startup would raise so
14:12I think it is it is a more accessible
14:13thing in some sense unless you just get
14:14and build on somebody else's llm which
14:16people should do for most things
14:19initially as well yeah one of the things
14:21I I think is really interesting about
14:22this space as well is we've actually had
14:24leading researchers like say you know
14:27we're still very very early in video
14:29video generation is so hard it's data
14:31intensive the data's like as you said
14:33it's not the whole internet but it's
14:34problematic in that a lot of the
14:35training has happened on short clips
14:37people aren't sure how to caption it's
14:39expensive to generate you have like
14:41complex like sliding window approaches
14:43and others to um try to deal with the
14:47like temporal coherence problem there
14:48are many more unknowns about how to uh
14:52progress this type of product
14:54technically do you mean video
14:55specifically or for video specifically
14:57yeah cuz the teams for all these things
14:59are actually quite small right the Pika
15:00team is reasonably small the mid-journey
15:02team for a long time was Tiny and so I
15:04actually think this is a good example
15:05where you can do a lot with very small
15:06teams and to your point there's all
15:07sorts of technical challenges but the
15:09reality is you can get to The Cutting
15:11Edge with like a handful of people in
15:13these fields which isn't necessarily
15:15true as much for you know other types of
15:17models or certain types of models at
15:19least so I do think it's striking um how
15:22few people you actually need to do
15:23something really interesting here and to
15:24your point there's other challenges
15:26coming 3 four years maybe it becomes
15:27harder I grew with you you don't need
15:29more than a handful of people like
15:31there's empirical evidence now maybe a
15:33slightly different point which is
15:36instead of like having to have a certain
15:38size of team to go deliver an llm at
15:42scale there are more degrees of freedom
15:44in how you would innovate technically in
15:45this area um and there's more
15:47disagreement on like how to progress and
15:49I think that's actually just interesting
15:51for startups yeah I think they should
15:53qar so I think that's that's the main
15:56solution to most problems I feel in uh
16:00in AI today okay well do you want to
16:02give me some investment advice given qar
16:04should I go home yeah you should do
16:06mainly qar Centric companies and so you
16:10know if they're doing qar you do it if
16:12they're not doing qar you don't do it so
16:13that's one big piece of advice I think
16:15one of the things I would like to go
16:16back to and um talk about is your point
16:19of view on like hey a bunch of people
16:21moved away from diffusion models to LMS
16:23yeah one of the reasons that people
16:25moved away from diffusion models to to
16:29llms is because there's a lot more uh
16:33sort of text and code in Enterprises
16:37that is obvious right working with
16:40images and video and audio felt more
16:42verticalized like where the B2B use
16:45cases and I I think what we're seeing
16:48increasingly is the creative fields are
16:52uh are actually pretty commercial right
16:55so one of the things I'm most inspired
16:57by and I think there's a lot of of money
16:58in is if you look at Mid Journey one of
17:01the bigest like biggest knocks on them
17:04uh from investors or naysayers early on
17:08was well like how many people want to
17:11make images that's not a hobby that's
17:14not a social network like what percent
17:16of the population are artists and this
17:19was clearly wrong just in terms of the
17:22scale that mid Journey has already
17:24reached there's probably two pieces here
17:27right like one these tools like Pika and
17:30haen and audio generation and mid
17:33Journey they make the pie bigger for
17:36Creative Fields especially since if you
17:39look at Pika or hent they're really
17:40focused on all creators rather than just
17:42like the film industry professional um
17:45and like if you go look at the uh mid
17:48Journey use cases and then I suspect the
17:50Pika and Haun use cases over times
17:53they're very commercial right like a lot
17:55of the things that you named or that
17:57people are experimenting with are really
17:59about communication marketing and
18:02advertising yeah I think if you just
18:03look at it as market cap of incumbent
18:06right adop is a 28030 billion do company
18:09like that's huge right and so I don't
18:11think the creative world is small right
18:15I think a lot of Creator economy
18:16companies have failed in the past which
18:18is a different thing depending on how
18:20you think of the Creator economy right
18:21in terms of you know celebrity based
18:23marketing or whatever but if you
18:25actually look at Enterprises obviously
18:26they use enormous amounts of imagery and
18:27video and other things things to reach
18:29with and interact and brand and you know
18:31associate with their customers and you
18:33look internally you you need to create
18:35imagery for slides or for other things
18:36and communication and you know and so to
18:39some extent you're kind of looking at
18:41different proxies and you say okay well
18:42what are some of the proxies on the um
18:45on the text Spas side and you can say oh
18:46you add up Microsoft and you know a few
18:50other companies like that and you're
18:51kind of getting a rough proxy for some
18:53form of text not really but you know I'm
18:55just simplifying things dramatically and
18:57then you add up adobe and a few other
18:58companies and that's your proxy for for
19:01image gen right and so I think um both
19:04are big and then the question I think
19:06always with the diffusion model based
19:07companies was where are the biggest
19:09application areas and the application
19:12areas also were a little bit different
19:14by what where will Encompass play a role
19:16where will you get blocked by other
19:19companies in the ecosystem versus you
19:21know it's a natural new Greenfield thing
19:22most things aren't truly new
19:24capabilities most things are just like
19:25making certain things dramatically
19:27easier there's the The Duality of that
19:29there's the Market expansion and more
19:31people can do this thing and then
19:33there's a value contraction hey you can
19:36do this at a tenth or 100th of
19:38cost and so often in markets like this
19:41you see both of those things happen at
19:42the same time you're simultaneously
19:45growing the market and shrinking it yeah
19:47I mean you have the um the um famous uh
19:51strategy like your margin is my
19:53opportunity right I was going to say I
19:56mean well these are actually higher
19:56margin things I think really what you're
19:58doing if you look at the sort of um my
20:03understanding I need to look up these
20:04numbers again but I think it's something
20:05like software spend is like I don't know
20:06I'm making it up half a trillion dollars
20:07$500 billion a year and then service to
20:10spend is like three to five trillion a
20:12year right and so really what you're
20:14doing is you're taking Services Revenue
20:17which is very people intensive and low
20:18margin and you're converting it into
20:21higher margin software Revenue but less
20:24of it so maybe you take that 5 trillion
20:26and you turn it into 2 trillion
20:29but is 80% margin 2 trillion versus 30%
20:32margin right the margin dollars actually
20:34expand and you see that in other
20:36Industries that's kind of what andreil
20:37is doing in defense right they're taking
20:39a Cost Plus model right you buy a drone
20:41from Lo Martin for a million dollars and
20:44you you get paid by the government 5%
20:46Cost Plus which means you get 5% as your
20:49margin so you make 50k off of it and
20:51andero will sell the same drone or you
20:53know better drone for $100,000 with 50%
20:56margin I'm making up the margin right
20:57but that's 50k and so you're making the
21:01same margin on a tenth of the price and
21:03so I think one of the ways I think about
21:04Andel as a company is they're taking
21:07very bad low margin revenue from other
21:10defense companies and turning it into
21:13margin healthier Revenue right yeah uh
21:16I'm going to edit the quote and just say
21:18like your ASP is my opportunity right
21:21there's a like a democratization that
21:23happened in the like latter half of the
21:26SAS Revolution or really most of the SAS
21:28Revolution which is instead of there's a
21:32hundred very large Enterprises that have
21:34some sort of CRM because it costs X
21:37dollars to um deploy and Implement seble
21:41then you have you know tens of thousands
21:44of companies who can buy Salesforce and
21:47then companies that figure out how to
21:49efficiently distribute um S&B SAS on the
21:52internet even though that that is still
21:54hard I I think one of the things that is
21:57um interesting here is let's just take
22:00video generation as the example the ASP
22:03of you know hundreds to thousands of
22:07hour make it single digigit dollars to
22:11generate per minute and expand the
22:15audience to many more people right
22:17that's the democratization that is
22:19happening somebody told me a joke the
22:21other day like somebody really negative
22:23on AI investing there only five
22:26businesses in AI that have Breakout
22:28traction right now Foundation models
22:30wuss mid Journey co-pilot and inference
22:34platforms I think there's both truth in
22:35it and like also it's not that funny of
22:38a joke because you're like it's true
22:41like there is a set of things that
22:44people are figuring out that are really
22:47early many of which feel pretty
22:50different than like yes it's it's useful
22:53to look at the incumbent vendors like
22:55Adobe but you're not fighting
22:58really the video editing software spend
23:02you're eating into the production spend
23:04the early internet version of this by
23:05the way is there's only five things you
23:07do on the internet you go to Yahoo to
23:09look for links you buy Pez dispensers
23:13eBay you buy some books on
23:16Amazon and then there's probably like
23:18some like two bullshitty companies that
23:19you would have quoted as like hyra
23:21things right so if you were looking at
23:22the internet Circa 96 997 whatever you
23:26probably would have had a pretty short
23:27list of real use cases and then a bunch
23:28of stuff you just thought was kind of
23:30dumb right and you'd be like look you're
23:31not changing anything like you're still
23:33using Microsoft Office and you're still
23:36using whatever shrink WP software you're
23:38still watching TV right and so I feel
23:41like we're kind of in that era of AI the
23:43stuff is going to happen right now is
23:44the really easy loow hanging fruit and
23:46then a bunch of dumb things are going to
23:48get built that aren't going to work and
23:49dumb is not meant in a pejorative way it
23:51just means like it's very hard to tell
23:53what's actually a good idea in a new
23:54market like this and um that was true
23:57the internet and that was true of mobile
23:59and that was true of cloud there's a lot
24:00of these like waves where there's a
24:02bunch of stuff that gets built right um
24:04so I think it's the same thing right
24:06it's a very positive sign this has been
24:10Traction in such a short period of time
24:13for so many companies if you think about
24:15it it's actually kind of amazing so I'd
24:17actually take the other side of that but
24:19I I totally get the point yeah well with
24:21the AI Focus fund I agree with you
24:24you're going to change the name of your
24:25fund you should call it like conviction
24:27star or something conviction star yeah
24:30but spell with a Q like conviction
24:32conviction star okay LP you heard it
24:34here first conviction star very exciting
24:37I should send you a t-shirt thank you
24:39please do that'll actually be the swag
24:41for no prior season one if you're a
24:43guest you're going to get the no prior
24:45tequila and then a conviction star
24:47t-shirt yeah it's very I'm very excited
24:49about the tequila anybody with the
24:51podcast has have a tequila uh we could
24:53actually call the tequila conviction
24:58that'd be pretty amazing I'm serious
25:01okay it could be like a q- shaped bottle
25:02you know how they have like the really
25:03cool bottles for different things the l
25:06gil guys are are U no prior brand
25:09marketer yeah I think I mov into LA and
25:11starting the brand if you haven't yet
25:14been a guest please write into the show
25:17and you know for the low low price of
25:19one GPU we'll ship you a bottle yeah
25:21we're looking for brand marketer to join
25:24the team too or not actually if you work
25:26for Mr Beast just call me A lot's going
25:29to do it okay aad thank you for joining
25:33me on no priors thank you for joining me
25:35and I look forward to getting my
25:38conviction star t-shirt and Tequila
25:41exciting find us on Twitter at no prior
25:44pod subscribe to our YouTube channel if
25:46you want to see our faces follow the
25:48show on Apple podcast Spotify or
25:50wherever you listen that way you get a
25:52new episode every week and sign up for
25:54emails or find transcripts for every
