# The Incongruent Economics of AI with Bertin Martens | Markus Academy | Ep. 134

https://www.youtube.com/watch?v=g22iT96DG6M

[00:11] thanks to all for coming and joining us again for another webinar organized by Princeton for everyone worldwide.
[00:18] today we have Burton Martens with us.
[00:21] hi Burton.
[00:25] it's great to have you with us.
[00:27] you will talk about the congruent economics of artificial intelligence AI.
[00:33] so we're trying to learn more about the connection between economics and artificial intelligence.
[00:38] so before we start I would like to make a few opening remarks.
[00:40] first I would like to see when we have a big transition like AI will lead to a technological shift not just a shock but a shift and um there's typically a Cher effect.
[00:54] so what you can see here we have a certain and the technological Frontier.
[00:58] you have perhaps before a and then
[01:00] there's a transition period over which we get reach a higher Innovation Frontier level and of course this Frontier level has to be implemented.
[01:09] we have to adapt it we have to take the good elements of the new technology avoid the bad elements.
[01:13] Ben regulation social norms have to change and this takes time and experimentation and typically what emerges whenever we have new technologies is a CH of effect so the adoption curve adaptation curve is is more like this.
[01:29] so first actually is a set back because the old structures there's no investment anymore and then there's some cut back and then it takes a while to turn around until the J curve effect then we reached a new frontier if everything goes well.
[01:39] however if this Innovation occurs not slowly but fast it can actually happen that you know in this case when there's a much faster Innovation jump almost a jump to to the higher level it can happen that actually this downturn is more dramatic and we hit a Tipping Point and then we spiral.
[02:00] the whole system might spiral out of control.
[02:02] so there's a lack of resilience.
[02:05] and this lack of resilience is you know we have to study this transition phase.
[02:08] and of course the question is how quickly will the transition happen and how quickly can all the adoption and the regulation and the social norms adjust.
[02:17] so the speed of two systems and the question what's the relative speed between the two settings and that's important to understand in order to understand you know whether we manage this transition well.
[02:29] so what in general what we have artificial intelligence will create some habit and network effects.
[02:32] so as the artificial intelligence technology is adopted it will become more and more interwoven with the economy.
[02:41] there will be an AI ecosystem inside the economy connected and it also means that the AI becomes much more systemic and there's typically there are positive Network effects if others adopt the technology I would like to adopt it as well and I have some spillovers from the others adoption but also become much more dependent so we have dispersed.
[03:02] societal knowledge at the moment this might be much more integrated or it might also be kept in silos and there might be across the whole economy there might be much more Network effects and essentially the economy becomes addicted to this new technology are very very dependent and then it will become very costly to pull the blug later on so even though we say you know whenever the technology gets out of control we can actually pull the plug we might not be able to because we might be very dependent if you just think of the smartphones you know if you want to take a plane these days it's very hard to fly without a smartphone so we're very dependent on the smartphones if you suddenly were to say we get rid of all the smartphones it would be very very difficult to implement such a change or outlaw smartphones and the same will be through also for other Technologies including artificial intelligence.
[03:50] another point I would like to make is that the EU is to some extent in a dilemma everybody talks the EU is not on the front tier uh but I think there's another dilemma I would like to point out and that's with regard to social.
[04:03] norms uh the llms will form social norms.
[04:07] how we design the large language models.
[04:09] how we align them to certain criteria.
[04:12] will actually affect social norms in our society in the future.
[04:15] we will interact a lot not only with other humans but also with AI agents.
[04:21] and this will shape this this social norms and the European Union always emphasizes the European values.
[04:26] how important it is to have the European values but the gbdr gbdr restricts European data available for training the llms.
[04:35] so if the data is are not used Europe datas are not used for the lmms large language models then European values will not be reflected in the algorithm as well or in the underlying model as well.
[04:47] so the social norms will not reflect to the same extent uh the European values as it reflects the Indian or us Norms or other part of the world.
[04:58] and this way actually it's to some extent it's a a dilemma if you have very strict data protection and you don't.
[05:05] allow this models to be based or train based on European data
[05:08] it will also not reflect the European attitude to certain elements
[05:14] finally I would like to say a few words about intellectual property rights of and the interaction with llms
[05:18] and the patterns of course how these intellectual property rights will be designed will shape the industry and the whole AI ecosystem
[05:30] there's a a data Internet content you can put patterns on the data itself on internet intellectual property on the internet content to the extent it's about the past data
[05:40] it's not clear that you have to incentivize people to create new internet content because it was already created on the past and nobody really expected to be enumerated from that
[05:52] for future data probably want to have some uh enumeration to make sure that there's enough data cont content created for to train future large language models in terms of outputs of
[06:06] course there are prompts which are essentially inputs as well that could be also patented or protected to some extent and then there is uh with regard to the outputs whether there needs to be some pattern as well but the question is here a little bit uh to what extent do we need patterns because we have this low marginal cost High fixed cost so very few firms can actually enter there's already a lot of Market concentration and there's not much free entry possible given the high fix cost and it depends of course on the financial Market structure whether who can actually raise so huge amounts of funding in order to enter the market so if you innovate and if you have initially funding and can uh provide an llm you might already have a large advantage of that and and also there are network externalities if you have positive netive externalities you automatically have a monopoly coming as a not natural monopoly you don't need on on top of it some patent protection so finally let me conclude
[07:08] with the poll questions and uh here are the poll questions and unfortunately something happened today with our poll questions so it didn't work so let me just State you the poll questions perhaps you can come back later to them
[07:21] the first question was adoption of llms is it similar to the adoption of the personal computer in 1980s 1990s
[07:30] yes or is it much is it not similar because it's smoth llms are much much faster adopted compared to uh the personal computers
[07:38] second question was about data walls is this a constraining factor constraining input factor for training LMS or can we substitute them easily with some more compute worse data or synthetic data or some other things
[07:54] and the third question was about who really reaps the big benefits of generative AI is it this large language models this large big like open eye Google and so forth what are these more specialized providers who you know based
[08:08] on fragmented data silos and U are.
[08:12] they're benefiting a lot or is it more the customer the customer who using these uh this products because the products will be provided at fairly low cost and the customers other firms and downstream firms can actually benefit most of that.
[08:25] I don't have your poll answers so but um I think these are interesting uh questions.
[08:31] and who will provide these large language models will there be a few big large firms or many many smaller firms because training the models will go down as new technology comes along.
[08:44] and finally I think a very interesting question from a European angle is should the European Union mimic the US strategy and you know subsidize it and raise public investments in order to have their own large language model or should the European Union actually rather go a different way given the huge expenditures needed to run such a a program.
[09:03] so let me uh pass on the microphone to Burton Martin who is the
[09:09] expert on artificial intelligence in connection with economics and we're looking forward to his presentation.
[09:16] thanks again for joining us and I should say you will get find an executive summary on YouTube uh downst in the comments line if you want to get the executive summary just click on the link below.
[09:30] thank you Marcus for this introduction and good afternoon everyone uh let me first share my slides is that okay and visible to everyone.
[09:50] perfect so um thank you for having me in this uh Marcus Academy series of um the previous week already uh Anton Corin gave a presentation on some economic aspects of AI and so I will
[10:10] build a little bit on that and talk about the in congruent economics of AI.
[10:18] now what do I mean by in congruences in AI economics this is not about incompatibilities or contradictions.
[10:28] these are are things that are just poorly fitting together and they force us to think a bit more and a bit deeper to find some out of the box Solutions.
[10:36] I'll will talk about four in congruences in my view first is the exponentially growing fixed cost of AI models and the tension that it creates with slow productivity growth and what are possible ways out of that.
[10:49] the second is a number of challenges in competition policy with regard to AI models.
[11:00] third is the challenge of moving from a decentralized learning system which we are used to in our um decentralized
[11:11] market economies to much more centralized learning and the pressure that this puts on individual data rights.
[11:19] um and last but not least I'll talk a bit about in congruences in EU AI policy options.
[11:27] I I'm based in Europe I deal a lot with European policies on AI and so a big question mark has come up here in Europe is should we try or should we aim for the AI technology Frontier or should we try to prosper below that Frontier.
[11:45] and if time permits at the end I will try to step out this narrow economic mindset and and offer a few more broad thoughts on this so let's start with the first one the exploding modeling cost versus slow productivity growth.
[12:02] Marcus already gave this jcurve um technology adoption explanation in his introduction and so this is
[12:11] precisely what I would like to talk about a little bit more.
[12:14] so as Anon Corin already explained last week uh basically AI modeling costs are growing exponentially over the period 2016 to basically 2024.
[12:27] so the last 8 years we started with the first Transformer generative AI model somewhere over here in 2016 and they cost no more than $1,000 to train.
[12:39] and so if we go to the latest models nowadays we are over $100 million and probably some of these models are now reaching $150 to $200 million although these figures are not always disclosed um.
[12:56] so basically six orders of magnitude increase in modeling cost over an 8-year period that is enormous uh.
[13:05] so the composition of these
[13:12] costs is we know from several studies is mostly staff actually but also the AI chips count for 30% infrastructure cost other infrastructure than chips interconnection cost because usually the training of these mod models and the use of these models happens over a number of servers spread out over over the world and then last but not least electricity cost electricity is relatively cheap in the US it's much more expensive in the EU and in other parts of the world sometimes so can see there where the comparative advantage is um Buton do you have an idea are there a lot of people working in this industry or is it just a small number of very very highly compensated people and it's very hard to train new people are they're very special in a sense there is a very small number of uh experience and very hard to train people but there are large numbers of of of of not so highly trained people.
[14:16] who have to do a lot of uh grinding tasks in this so uh filtering data
[14:23] cleaning data testing models and so on
[14:27] so there are thousands of people
[14:29] sometimes involved in
[14:31] this um the infrastructure cost for models can be about 10 times the modeling cost itself and bearing in mind
[14:40] that the Chip's amortization rate is about 100% over an 8mon period less than a year
[14:46] so you buy these expensive Nvidia chips and eight months later they're basically out of the market because Nvidia itself comes up with a more advanced model and you have to throw out all your chips
[15:00] well you can use them for other purposes of course but if you want to be in the frontier you have to move to something else
[15:03] and we have several reports that say in 2024 this year there's more than 200 billion US dollar invested in AI models and infrastructure it's hard to
[15:16] verify that figure um because companies are not always very transparent about this and what counts as investment and what not um but I think this is a good ballark figure.
[15:29] be in mind it is probably higher in reality but that that should give you a good indication.
[15:35] clearly at this growth rate four to five times per year this cost explosion is not sustainable and even for big tech companies um once we get over a trillion dollars a year uh this will become a hard path to follow for their Financial Resources.
[15:57] it may even exceed macroeconomic resources available if we extrapolate this to the mid-30s mid 2030s we we've easily exceed world GDP forecast so something will have to change in that trajectory um what drives this cost explosion is model scaling.
[16:19] loss and this was something that was already established in a paper dating back to 2020.
[16:24] So basically if I reformulate this in economic terms uh data to train the model the parameters the number of variables that you allow internally in the model and the compute capacity that you need to handle all this or complimentary inputs.
[16:42] Uh as an economist I would say these is a lonf production thing or something close to that.
[16:50] There is some substitution elasticity between the data and the parameters.
[16:56] So larger models are a bit more data efficient but what we see happening today is that um most models are overtrained.
[17:03] Uh meaning that you run more iterations than what you actually need and you have more variables than what you actually uh need above the minimum the so-called Chilla rate of variables to parameters of parameters to data.
[17:20] Sorry, um, so, so, uh, there is not much movement in that now.
[17:25] Scale improves cognitive performance.
[17:29] That has been shown over the last couple of years and in many ways in many tasks.
[17:37] These large models exceed human benchmarks, uh, human cognitive benchmarks already.
[17:43] But the question for an economist, of course, is, does this cognitive performance improvement translate into revenue, into productivity gains, uh, for the companies that that use these models?
[17:57] So, and this brings us to the productivity debate, of course.
[18:04] There are a number of uh task-level and firm-level empirical studies on AI models nowadays and gradually more studies become available.
[18:18] They show various levels of productivity growth between five and 20% at firm level or even at task level.
[18:20] within a firm
[18:21] extrapolating that to the entire economy of course is much more difficult
[18:27] and so dar mlu is rather pessimistic and he says he expects less than 0.7% per year additional productivity growth
[18:39] on the other hand we have private companies like Goldman Sach of course invest in these sectors
[18:44] they are more optimistic and they say we can reach 1.5% productivity growth additional productivity growth per year
[18:51] uh that's of course on the assumption that AI will not only automate a production process but will automate the automation of production process
[19:02] so becomes a more sharply exponentially rising productivity growth curve
[19:09] I thought was 7% over 10 years not per year in total
[19:18] 7% well I would have to check
[19:21] that much more pessimistic in a sense.
[19:24] um so um.
[19:29] we The The J curve effect of course slows all this down a little bit over time but then may we may experience this pickup and acceleration again at a later stage.
[19:41] um of course there's a lot of productivity gain below the frontier uh people who are not using these large uh llms or generative AI models.
[19:55] they are smaller specialized and derived AI model and they can help firms of course to adapt and and to adopt AI in their firm environment and they deliver important productivity graines.
[20:07] we just don't know how much yet that is going to give in the longer run but we can turn this question around and say how much productivity grave do we need to sustain on the ongoing AI.
[20:23] development and so I did just a quick
[20:26] back of the envelope calculation so
[20:29] assume that firms invest1 trillion Us
[20:31] doar in Ai and that's a figure with the
[20:35] current cost explosion that we could
[20:37] reach sometime in 2026 2027 so in the
[20:41] next couple of years and assume that
[20:45] this investment in the eye and the use
[20:47] of these models triggers a
[20:49] 1.5% overall productivity growth so in
[20:52] line with the Goldman Sach uh optimistic
[20:56] estimate in that case the required GDP
[20:59] scale to reach Break Even would be
[21:02] around 67 trillion
[21:04] us now the total GDP of the advanced
[21:08] economies is around 75 trillion us doar
[21:11] according to the IMF World economic
[21:14] outloop for this year uh so that's more
[21:18] or less in the same order of magnitude
[21:20] so I think1 trillion dollar you AI
[21:25] investment is incredible choke point
[21:27] when you reach even with optimistic
[21:29] productivity growth assumptions somehow
[21:32] this will remain stuck unless of course
[21:35] productivity growth would accelerate
[21:38] significantly uh North house for
[21:41] instance in his 2020 paper explores the
[21:44] conditions for what he calls a
[21:45] singularity in growth so all sudden
[21:48] acceleration but this would require an
[21:51] enormous increase in the share of
[21:53] investment in GDP currently worldwide
[21:56] investment is around what 25% % of GDP a
[21:59] quarter of course in China is is much
[22:02] higher in other countries is a bit
[22:04] lower but if we would go much beyond
[22:07] that and so we would invest all of that
[22:09] in AI some sort of Singularity could be
[22:13] reached but this would require an
[22:15] enormous compression of the consumption
[22:18] component in GDP and whether that is
[22:20] politically acceptable of course is
[22:22] another
[22:23] question so what are the solutions um I
[22:26] can I ask you a quick question it could
[22:28] be that these numbers are overstated as
[22:30] well because what happens to a large
[22:32] extent is that Nvidia is investing in
[22:35] these companies as well and J trust
[22:37] sells at a very high price their chips
[22:39] to let's say open eye and then at the
[22:41] same time they also invest in it so
[22:43] there's a little bit of a secularity
[22:45] there I give you my chips for a billion
[22:47] dollars and I invest these billion
[22:49] dollars in you and boost up your
[22:52] valuation is this too
[22:54] critical yeah um their investment
[22:57] strategies
[22:59] but yeah the company pays for the chips
[23:01] and Ai and Nvidia reinvests part of that
[23:05] return into additional things I presume
[23:07] in the company gives them some Financial
[23:10] Resources to to buy the comp
[23:13] complimentary inputs that they need to
[23:15] make these chips work so still part of a
[23:18] big pot for AI investment I would
[23:21] say so of course AI developers and firms
[23:25] are aware of this and try to reduce this
[23:27] fixed cost
[23:29] um U and there are several ways now of
[23:32] doing this one movement that we see is
[23:36] to move from investing heavily in the
[23:39] pre-training of AI models to move some
[23:42] of these costs to posttraining once the
[23:44] model is ready you you enhance the model
[23:48] you improve the model by adding and
[23:50] uploading proprietary data into the
[23:53] model or realtime data access into the
[23:57] model and so you don't have to do that
[23:59] in the training phase you do that later
[24:02] and you can do that for specific
[24:03] application versions of the model so you
[24:06] don't need to do that for for for the
[24:08] entire model and having access to
[24:10] realtime data is of course very useful
[24:13] because then you can resp reply to uh to
[24:17] current questions about current issues
[24:19] which uh models cannot do if they don't
[24:22] have access to realtime data for most of
[24:25] the big models today the cut off dat for
[24:29] the data sets that are used in
[24:30] pre-training is around 2020
[24:34] 2021 and actually most of the companies
[24:37] are still using all data sets from that
[24:40] period uh for doing most of the training
[24:43] work um there's also a shift away from
[24:47] um web-based training data to uh social
[24:51] media data and of course to the extent
[24:53] that you have realtime access to social
[24:55] media platforms uh you can incorporate
[24:59] more recent information in in your
[25:02] training data the other movement that we
[25:05] see is to move from training to
[25:07] inference to the use of the models and
[25:10] expand the possibility for models to
[25:13] start taking into account a much larger
[25:16] context when they receive a question you
[25:18] can now upload for instance you ask a
[25:21] question on on on chemistry and you can
[25:23] upload an entire chemistry handbook or
[25:26] recent paper and say do it according to
[25:29] this method and so that facilitates or
[25:33] guides the model through the way to work
[25:35] towards a solution but of course this
[25:37] takes more inference time is more
[25:39] expensive on recurrent costs for running
[25:42] the B there are also efforts to make
[25:45] Muscle more in to to make them do
[25:49] reasoning do logical reasoning again
[25:52] this takes more inference time but uh
[25:55] this seems to work rather well in a
[25:56] number of models open ai's latest 01
[26:00] version claims that it's rather strong
[26:02] on reasoning and some tests have shown
[26:05] that this is indeed a case and models
[26:08] can of course accumulate memory from
[26:11] past experiences and start learning on
[26:13] their own after their training phase
[26:17] stopped uh on the one hand this is risky
[26:20] because models May learn the wrong
[26:21] things and people may use mod and guide
[26:23] them towards the wrong
[26:25] answers and that's what happened when
[26:27] the first models were released that
[26:30] could do memory accumulation so that was
[26:33] quickly stopped by companies but now
[26:36] this seems to be more stable and robust
[26:38] and and and so these memory questions we
[26:41] will see coming up especially with AI
[26:44] assistance that memorize the task you
[26:47] gave the model and it will remember what
[26:50] your preferences are and and and where
[26:53] you want the model to go what sort of
[26:55] decisions you want it to take
[26:58] but of course this raises new questions
[27:00] about access to real time and
[27:02] proprietary data often and it creates
[27:06] Network effects uh that were not present
[27:10] in the first generation of generative AI
[27:12] models but when users can interact and
[27:17] become more models become more path
[27:19] dependent and become also more dependent
[27:21] on what other users have learned or the
[27:24] tasks given by other users to the model
[27:27] that creates Network effects the more
[27:29] people use the model the better it will
[27:31] become at reasoning at giving answers
[27:34] and network effects as we know may
[27:36] reduce uh
[27:38] competition how from search engines
[27:41] where you also have essentially the same
[27:44] effects the network effects everybody
[27:46] uses Google hence Google has a better
[27:47] search engine exactly
[27:50] yeah so onwards to the second in
[27:54] congruence competition versus
[27:57] Monopoly so I identify three competition
[28:00] policy issues in in uh AI models first
[28:04] the one I already mentioned is the
[28:06] exploding overall fixed costs and if
[28:09] that explosion continues even at a
[28:11] slightly slower Pace uh we might reach a
[28:15] state where it will force collaboration
[28:17] even between big Tech firms because the
[28:19] costs become too heavy uh and so this
[28:23] will create a big natural monopoly and
[28:26] the sheer Global challenges in our
[28:28] notion of regulating such a natural
[28:31] monopoly across the globe will be very
[28:34] challenging I think for competition
[28:36] policy how to do
[28:37] that the second one is monopolistic or
[28:41] exclusionary Behavior within the
[28:44] vertical layers of the AI
[28:46] stack and competition authorities are
[28:49] now paying a lot of attention to this uh
[28:52] lot of ongoing
[28:53] investigations and so how can we make it
[28:56] more competitive uh I'll say a little
[28:59] bit on that and third uh collaboration
[29:02] deals between startups and Big Tab
[29:04] companies are they designed to
[29:06] circumvent merger and access rules or
[29:09] are they necessary tools to make the
[29:12] whole thing work and then there an
[29:14] alternative to vertical
[29:16] integration um first on regulated
[29:19] monopolies so exploding fixed cost of
[29:22] course collaboration between big Tech
[29:24] firms uh when they want to distribute
[29:26] their workloads and share their
[29:28] resources would be would look very
[29:30] suspicious from a competition or
[29:32] antitrust point of
[29:34] view uh on the other hand I don't see
[29:37] how it would be possible to regulate
[29:39] price the quantity of models the quality
[29:41] of models these are all totally
[29:44] unexplored uh domains for competition
[29:48] policy uh it would also be hard to
[29:51] combine this with the strong competition
[29:54] that we now observe between AI models we
[29:57] have new AI models major AI models
[30:00] coming out almost once a week uh we have
[30:04] new derived versions of AI models coming
[30:06] out once every couple of hours so there
[30:10] there is a lot of competition and if if
[30:13] this would not be possible any more due
[30:15] to huge fixed costs then we have an
[30:18] innovation problem I think as well but
[30:21] but can I ask you a question do you
[30:23] expect that these big firms will merge
[30:25] or will collaborate or will it be more
[30:28] shake out that some firms will just give
[30:29] up and one monopolist will remain at the
[30:33] end of the
[30:34] day sh out the ShakeOut was first happen
[30:38] of course among the smaller firms but
[30:40] the big firms it's another issue they
[30:42] sit on a lot of cash usually they can
[30:45] hold on for a while but what we see
[30:46] happening for big Tech firms in AI is
[30:48] they're already trying to shift their
[30:51] workloads to let's say second tier um
[30:55] infrastructure and Computing work
[30:58] providers and there is some degree of
[31:00] sharing of resources
[31:02] already uh if they shift workload to
[31:06] second tier companies that's not so
[31:08] suspicious from a competition policy
[31:10] point of view but if they would start to
[31:11] redistribute workload between themselves
[31:14] that would of course be a different game
[31:17] how this is going to work out it's
[31:19] difficult to say and you would say it's
[31:20] very different from the surge engine
[31:23] competition we saw Yahoo Google and alav
[31:26] Vista searching
[31:28] because the fixed costs were not so high
[31:29] do you think it's it's a different
[31:30] structure here compared to oh yes yes
[31:33] for search engines fixed costs were
[31:35] considered to be high but it's still
[31:37] much much lower than what we observe in
[31:39] AI at the moment
[31:42] yeah so then competition bottlenecks
[31:45] within the vertical stack where the AI
[31:48] stack so you have chips you have the
[31:50] Computing infrastructure software layers
[31:53] on top of that and the data you use uh
[31:56] then the models as an output but the
[31:58] models are an input into the business
[32:00] outlets and so this vertical stack
[32:03] what's happening there we know on the AI
[32:04] chips for the time being Nvidia has a
[32:07] near Monopoly for the most performance
[32:09] ships um some others are are coming up
[32:13] AMD is coming up we have seen over the
[32:16] last month the first models coming out
[32:19] with trained on AMD chips but we don't
[32:22] have any data on the efficiency of that
[32:24] training so we have to wait for that uh
[32:28] that mon Monopoly May weaken over the
[32:31] next couple of years I don't expect it
[32:33] to go best very soon
[32:35] though um on the other hand what can you
[32:38] do about it um can you sell Nvidia chips
[32:42] by by open uh um open open uh um price
[32:49] bidding for these chips uh I think that
[32:53] that's not a feasible
[32:55] solution um the same for compute
[32:57] infrastr structure it's not a
[32:58] monopolistic but an oligopolistic Market
[33:01] with mainly Google Amazon and Microsoft
[33:04] being the big players there but there
[33:05] are many small players that are trying
[33:08] to scale up uh but they don't have the
[33:11] resources to buy t or hundreds of
[33:14] thousands of Nvidia chips so there will
[33:17] have to be merges between these smaller
[33:19] scale uh compute infrastructure
[33:22] providers to do that to achieve that and
[33:26] I don't see this happening so
[33:28] quickly uh then there's a bottleneck on
[33:31] the software layers on top of the chips
[33:34] Nvidia has the Cuda software and the
[33:36] Nvidia chips run on that software and
[33:39] most AI models code is written for that
[33:42] Cuda software there are some competing
[33:45] softwares but still uh Cuda plays a
[33:50] dominant role in that and remains to be
[33:52] seen to what extent that will also
[33:55] solidify nvidia's Market position
[33:58] the same is true from Microsoft this
[34:00] software so now with large language
[34:02] models we can translate software easily
[34:04] from one language to another language so
[34:06] this network effects are much less
[34:07] powerful because of these large language
[34:10] models can also translate the Cuda code
[34:13] into some other language e more easily
[34:15] nowadays or is this not part of the all
[34:19] large language model um some people
[34:21] claim this should be fairly
[34:22] straightforward others have more doubts
[34:25] about this I'm not a computer scientists
[34:28] I cannot give an authorative answer to
[34:30] that
[34:31] question um same happens in Microsoft
[34:35] Ser software we saw recently that Google
[34:37] launched the complaints against
[34:39] Microsoft for the way they use licensing
[34:43] deals for their server software it's
[34:45] very difficult even for companies like
[34:47] Google at least it claims uh that is
[34:50] very difficult to enter into that market
[34:52] and keep its position in that market and
[34:54] of course there are network effects
[34:56] there as well
[34:58] then there is the data wall um we are
[35:01] slowly moving away from chomon craw web
[35:04] page text because uh the total amount
[35:07] available is simply not enough to train
[35:10] the most advanced models anymore and so
[35:12] we are moving to social media data we're
[35:15] also moving to synthetic data uh to
[35:18] provide sufficient supply of data for
[35:22] model training but there are only a few
[35:25] large social media companies in the
[35:26] world and of course they have privileged
[35:29] and exclusive access to these social
[35:31] media data it seems that there are deals
[35:34] in the making on sharing these social
[35:37] media
[35:38] data uh I'm not sure how they work
[35:42] though at the moment uh and so companies
[35:44] like uh that's why XI formerly Twitter
[35:50] is now taking up so rapidly because they
[35:53] have access to the huge Twitter x uh
[35:56] social media database meta is very
[35:58] strong because it has Facebook uh
[36:01] Instagram and uh and S social media
[36:06] data
[36:07] um there's strong competition in the
[36:10] model output layer as I said so there uh
[36:13] many models per month now coming out and
[36:16] many models per hour sorry not per day
[36:18] per hour uh so there's enormous amount
[36:21] of competition there uh so there's no
[36:24] reason to worry about this but many of
[36:28] these models are done by smaller AI
[36:31] startups and they need to plug their
[36:33] models into a business model in order to
[36:36] uh to to generate Revenue with that and
[36:39] the revenue has to pay for producing the
[36:41] next model and for running their models
[36:44] and the inference part of the model so
[36:46] there are a number of global established
[36:48] very large scale business model in
[36:50] advertising online for example Google
[36:52] and meta are very strong in that and
[36:54] productivity software both for business
[36:57] is for consumers and for servers
[37:00] Microsoft is very strong in that if you
[37:02] can plug your models in there of course
[37:04] then uh you're running on on Rails and
[37:07] this goes fast if you want to start your
[37:10] own business model this is very hard to
[37:12] do even companies that are successful in
[37:15] this like open AI they open their model
[37:18] App Store with chap GPT it's very
[37:21] successful App Store there millions of
[37:24] derived models now available there but
[37:26] still how much revenue does that
[37:28] generate uh 5 billion per year maybe a
[37:31] bit more maybe eight uh that's not
[37:34] enough to keep open AI running to keep
[37:36] the servers and the models running and
[37:39] so they have to look for more and they
[37:41] have to bring in new
[37:43] capital uh they may ask Microsoft to get
[37:47] even more access to the servers and so
[37:51] there is this negotiation model going on
[37:53] between AI startups and big Tech and
[37:56] I'll come back to that in a
[37:57] so can I just summarize what you're
[37:59] saying you say even if open AI gets all
[38:01] the ad revenue from Google in meta
[38:04] combined they will not be able to keep
[38:07] running or is this too strong a
[38:09] statement all the ad revenue from Google
[38:11] and meta combined no
[38:13] no uh that would be a hell of a lot of
[38:16] money now what I say open AI if it gets
[38:20] the the revenue that it currently earns
[38:23] through it its open AI App Store and
[38:26] chat GPT revenue and the revenue may get
[38:30] from Microsoft because tgpt is plugged
[38:33] into several Microsoft products and
[38:35] other all that Revenue combined I think
[38:39] is at the moment not enough to keep open
[38:41] AI running but if they would get a big
[38:43] chunk of the ad revenue from Google and
[38:46] meta or capture that then it would be
[38:49] sufficient but meta is running its own
[38:51] models doesn't use CH GPT yes of course
[38:54] I mean there might be but you could
[38:56] argue now we in this War attrition a
[38:59] game and then there will be one winner
[39:00] at the end and the winner gets most of
[39:02] the ad Revenue down the road
[39:06] yeah whoops see yes no I'm just uh yes
[39:11] yeah
[39:13] so let me go to the third part of it
[39:17] this collaboration deals between
[39:20] startups and big Tech firms I call these
[39:22] coopertition deals combination of
[39:25] collaboration and competition
[39:27] why do I say so well I've tried in this
[39:31] graph to summarize this a little bit so
[39:33] you have uh the AI startups here they
[39:38] they need model inputs uh hyperscale
[39:40] Computing infrastructure Computing ships
[39:43] and data and then they can do model
[39:46] development and they can deploy the
[39:48] models themselves or they can give them
[39:50] to Big Tech firms to plug them into
[39:52] their existing uh business models of
[39:56] course big Tech firm have access to the
[39:59] hyperscale uh Computing and ships
[40:02] infrastructure they do their own model
[40:04] development inhouse as well and so
[40:06] there's competition at the level of
[40:08] models and what we see is that most big
[40:11] Tech firms are now losing using a
[40:13] combination of several AI models some
[40:16] coming from uh AI startups some being
[40:20] in-house
[40:21] developed and so at this level there's
[40:24] competition and at these two levels
[40:26] there is
[40:27] uh collaboration because AI startups
[40:30] have a hard time getting directly to end
[40:32] users here it's easier to go through
[40:35] established business models of big Tech
[40:38] firms um and so this Co competition
[40:42] model is under scrutiny by uh
[40:46] competition authorities at the moment
[40:47] there are many investigations that have
[40:49] been launched and what they are looking
[40:51] for basically is exclusionary Clauses in
[40:55] this uh in this collaboration agreement
[40:57] now the big Tech firm says yes you can
[40:59] get access to my compute but in return I
[41:02] get exclusive access to your model so
[41:05] far no smoking guns have been found as
[41:08] far as I know so explicit examples the
[41:12] startups are not the open AI or an
[41:15] Tropic these are the big tchs in this
[41:17] example or the big are the Amazon and
[41:20] like an Tropic and Amazon collaborate
[41:23] like what was announced in the past few
[41:24] days yeah would this fit this chart or
[41:28] not yes
[41:30] yes startup Amazon is also very
[41:34] interested in model devel deployment in
[41:36] its own business models take for
[41:37] instance the uh Voice
[41:40] assistance and they are very uh very
[41:43] active in in uh home AI smart home AI uh
[41:48] also using uh voice assistance in in in
[41:51] shopping on Amazon so all this is very
[41:55] interesting uh to to
[41:57] [Music]
[41:59] yeah so um these cooption deals are I
[42:04] think on the one hand Pro competitive
[42:07] because it makes Market entry feasible
[42:09] for AI startups they don't have all the
[42:12] complimentary inputs required to train
[42:15] and run and and and generate Revenue
[42:17] with this models if they collaborate
[42:19] with a big Tech the combination of that
[42:23] can work well but it may also be
[42:25] anti-competitive in the sense that it
[42:27] strengthens the market power of course
[42:29] of these big Tech
[42:30] firms and sometimes competition
[42:33] authorities look at this as ways of
[42:35] circumventing merger and uh uh
[42:39] rules um merger and acquisition
[42:42] rules but are there an efficient
[42:44] solution for market failures in in
[42:47] complementary markets is my question and
[42:51] so that is a question that um um
[42:56] competition authorities have have not
[42:58] really looked at very much so far in my
[43:01] view uh the French competition Authority
[43:04] for example if before you get into that
[43:06] can you draw a contrast to the
[43:08] pharmaceutical Industries you also have
[43:10] a lot of startups in pharmaceutical
[43:11] Industries and the big pharmaceutical
[43:13] companies have the distribution networks
[43:15] so the startups invent new drugs and
[43:18] then will be taken over by the big
[43:21] farmers would you say it's very
[43:23] different here in a sense that there's
[43:25] more overlap and than it is in the farma
[43:28] industry or can we learn something from
[43:30] the farma industry or not I'm not so
[43:34] familiar with the farma industry in this
[43:36] regard but uh from what I hear people
[43:38] saying the model is very similar in the
[43:41] sense that startups sometimes count on
[43:45] either being taken over by a bigger
[43:47] company if it fits well into that model
[43:49] so vertical integration is
[43:52] useful uh on the other hand many
[43:54] startups want to pursue their own route
[43:56] in Germany for for example you have
[43:57] bonch was very successful in the co uh
[44:02] period but um they're no longer
[44:04] collaborating so much and going their
[44:06] own way and developing new products and
[44:09] they've lost a lot of market value as a
[44:11] result of that but still they are doing
[44:13] quite well I think
[44:15] yeah
[44:17] um I will come back to that at the end
[44:19] when we talk about non-economic motives
[44:22] for further AI development as
[44:24] well so but a little more on these
[44:28] competition deals
[44:30] so compute data chips and business
[44:33] Outlets as I said before or
[44:35] complimentary inputs in the AI
[44:37] production chain and so going back to uh
[44:41] T's uh paper in 1982 already on
[44:45] economies of scope and which he repeated
[44:47] in
[44:48] 2020 he says basically that if
[44:51] complimentary inputs markets fail or do
[44:54] not work properly then vertical
[44:56] integration is better solution more
[44:58] efficient
[44:59] solution and so what this means in AI
[45:02] terms here in this competition model is
[45:05] uh big Tech firms are afraid of
[45:08] competition authorities over the last
[45:09] couple of years they have already been
[45:11] put under the microscope in terms of
[45:13] scrutiny and so they don't want to go
[45:16] for big fights over merges and
[45:18] Acquisitions and so what they do is find
[45:20] all kinds of way to collaborate with
[45:22] these startups and allow these startups
[45:24] to generate their models without talking
[45:27] about vertical
[45:29] integration and and and so we are
[45:32] somewhere in in in in an in between
[45:34] situation I think uh uh and a lot of
[45:39] that will depend on how competition
[45:41] authorities are going to react so my my
[45:44] question here is is this traditional
[45:46] competition policy view going to hold
[45:48] back AI development or will competition
[45:52] authorities come around and say yes we
[45:55] recognize this need for complimentary
[45:57] inputs and we recognize that U markets
[46:00] are not very ideal in this respect so we
[46:03] allow at least some degree of vertical
[46:06] integration um note here you argue that
[46:08] open Ai and Microsoft are vertically
[46:11] integrated or you would say they are not
[46:12] or something in between they will ask
[46:15] argue certainly not okay most people
[46:19] will say that from a legal point of you
[46:21] there are two separate legal entities
[46:24] and there is nothing so far in their
[46:27] collaboration agreement that points to
[46:29] exclusive dealings so it's not vertical
[46:32] integration um but I wanted to point out
[46:35] here that the EU digital markets act
[46:37] article 6 paragraph 7 actually forces
[46:40] Gatekeepers to open access to all their
[46:43] hardware and software technology
[46:45] segments and so if we translate that
[46:48] into um um Google search into Microsoft
[46:54] uh Microsoft's office soft
[46:57] software uh into metas uh systems social
[47:03] media platforms uh these are all
[47:07] Gatekeepers in their own domains and so
[47:10] they fall under the
[47:11] dma and so European commission is
[47:15] pushing to open access to technology
[47:17] segments so that AI companies could plug
[47:21] in their AI models straight into uh
[47:25] let's say an a meta uh social media app
[47:30] or into a Google Android app or into a
[47:35] Microsoft uh uh browser or Word and
[47:38] Excel software so that there can be
[47:42] plugins at several levels in this
[47:44] vertical chain now this becomes
[47:47] technically very complex how do we we
[47:50] ensure this pluggability plugin option
[47:54] at different levels or different
[47:55] segments in the IAL stack or in the
[47:58] horizontal spread even of these hardware
[48:01] and software segments uh
[48:04] without uh encumbering the the
[48:08] functioning of that hardware and
[48:10] software um this this is a big
[48:13] engineering problem and many companies
[48:15] many companies under the digital markets
[48:17] are actually struggling with and working
[48:19] hard on this uh on on how to make this
[48:23] possible
[48:25] um what's a connection to all the
[48:27] platform economics and you know there
[48:29] must be if you think of platforms yeah
[48:33] is there some connection we can draw
[48:35] from platform economics in this
[48:37] Dimension or even the law has special
[48:40] Provisions for platforms and are these
[48:43] big tech companies considered as
[48:45] platforms or not yeah so the digital
[48:48] markets act uh these big Tech firms are
[48:51] platforms multi-sided markets and they
[48:55] have been designated as Gater because
[48:57] they meet certain quantitative
[48:59] criteria uh in Market
[49:01] size market dominance
[49:05] um so this is at the platform level but
[49:08] in my view what we see happening today
[49:11] is that there is a new ecosystem forming
[49:14] on top of those platforms that will
[49:17] integrate several platforms and these
[49:20] are AI
[49:21] ecosystems you go to an AI model you ask
[49:24] questions that model can send signals to
[49:28] several of these platforms to get
[49:30] answers back from those platforms and
[49:32] then give you that reply and so these AI
[49:36] models become an apex structure on top
[49:38] of those platforms and if these AI
[49:41] models start using a personal assistance
[49:45] with memory and learning over time if
[49:47] they learn also across users then you
[49:50] will get network effects there as well
[49:53] and and so that is where the strongest
[49:55] next Network may come in future but it
[49:59] could be could be made an argument that
[50:01] actually this AI allow to reduce
[50:03] switching costs across platforms yeah
[50:06] this makes actually these other
[50:07] platforms more competitive to each other
[50:09] because the lockin effects are less
[50:11] powerful yeah but the switching
[50:14] possibilities will depend on to what
[50:16] extent platform owners allow this
[50:19] switching or to what extent they allow
[50:21] different AI models to access their
[50:24] platform what the conditions for access
[50:26] who those platforms will be uh and that
[50:29] is something that we don't know
[50:31] yet and uh that actually brings me to my
[50:35] next Point uh in congruence number four
[50:39] is growing tension between the what I
[50:41] call the private and the social value of
[50:45] data uh I'm talking here about copyright
[50:48] issues and personal data
[50:50] protection uh and data that are being
[50:53] used for model
[50:55] training so
[50:58] uh a lot of the data that are being used
[51:00] for model pre-training are actually
[51:02] protected by copyright uh it's music
[51:04] data movie data picture data but also
[51:07] text Data uh harvested on internet web
[51:10] pages and so on and what we see from
[51:14] recent Empirical research is that more
[51:17] and more of these copyright owners opt
[51:20] out of um um use of their data for AI
[51:26] model training uh long pre at Ali have
[51:30] established that today we are already at
[51:32] 20 to 25% opt out and they expect in the
[51:36] next year or so this can easily go to 30
[51:38] or 40% that's an enormous reduction in
[51:41] the availability of training
[51:45] data uh of course you can say well
[51:48] people want to use this data can go to
[51:50] the company and ask for a license and
[51:52] then pay for that license and then have
[51:54] access to the training data but this is
[51:57] not affordable for many startups also
[52:00] it's not clear how this licensing system
[52:02] would work if you have 10 or hundreds of
[52:05] thousands of Rights owners how as a big
[52:08] company do you deal with all those there
[52:10] is no easy automatic system to do that
[52:14] at the moment so what we see that some
[52:16] big tech companies go to some big media
[52:19] companies and say okay let's have a deal
[52:21] with you you're big you can give me a
[52:23] lot of data but they leave aside many of
[52:26] the smaller companies because the
[52:28] transaction costs of trying to negotiate
[52:30] all these these these deals are far too
[52:33] high and the impact of that is that it
[52:36] reduces model competition and model
[52:39] quality it reduces the number of models
[52:41] it reduces the quality it increases the
[52:44] bias in the AI algorithms because they
[52:46] work on a narrower set of data a smaller
[52:50] set of data and so this generates
[52:52] negative
[52:54] externalities from the media industry
[52:56] that are mainly the owners of the
[52:58] copyright to the rest of the
[53:01] economy now in the EU of course um there
[53:05] there's copyright legislation there is
[53:07] the AI act and they try to find a
[53:11] solution for this but this is going to
[53:14] take a long time before that solution
[53:18] can be found so this leads to a period
[53:21] of prolonged regulatory uncertainty for
[53:23] AI investors it's even more difficult in
[53:27] the US where there are more than two
[53:29] dozen court cases pending copyright
[53:32] related court cases against AI
[53:35] companies uh we are slowly seeing some
[53:38] of these C cases trying to unwind there
[53:43] was a
[53:44] recent court case in New York City for
[53:48] open Ai and that court case that
[53:51] judgment seemed to go in the direction
[53:53] of transformative use so a allowing uh
[53:57] use of these data for training purposes
[54:00] for AI because that falls under the fair
[54:03] use copyright exception in the US
[54:05] copyright
[54:06] law whether that will be upheld in other
[54:09] courts and in higher courts remains to
[54:12] be seen of
[54:13] course um a similar issue is developing
[54:17] in the EU with regard to personal data I
[54:20] said before that increasingly model
[54:23] trainers are shifting to social media
[54:26] data
[54:27] to get over the data wall but social
[54:30] media data is of course personal data
[54:32] that you and I and all of us have put in
[54:34] our social media platforms and can they
[54:37] be used for model training without our
[54:40] consent if these companies have to go to
[54:42] each and every one of us to get our
[54:44] personal consent that's not a workable
[54:47] method so what they try to do is build
[54:49] this consent in their general terms of
[54:52] use for their apps and their websites um
[54:56] but is that an acceptable way of doing
[54:58] it so the question is is it legitimate
[55:02] to use these data for model training and
[55:05] that's a question under gdpr data
[55:08] protection regulation article six the
[55:11] problem is that in the EU Nobody Knows
[55:15] the answer to that question even the
[55:16] data Regulators are uncertain what the
[55:19] answer to that question is and then of
[55:21] course companies ask when can you give
[55:23] us an answer next week next month
[55:26] they said well we don't know yet May
[55:29] might be next year might be in two years
[55:32] I was at the seminar a couple of days
[55:35] ago on this question and I can tell you
[55:39] there's a lot of uncertainty about how
[55:41] this can be settled how this can be uh
[55:44] resolved this issue this is penalizing
[55:47] for understand if you know the New York
[55:50] Times is suing open eye that have used
[55:52] the New York Times article for training
[55:54] open eye models
[55:56] if they win or if New York Times wins
[55:58] then the jgpt has to be retrained based
[56:01] on new data they have to withdraw their
[56:04] parameter
[56:05] estimates first of all open ey would
[56:08] have to pay statutory damages to the New
[56:11] York Times and they can run into many
[56:14] many millions of dollars so that would
[56:17] be a severe cost for open AI secondly
[56:19] what you do uh you have to retrain your
[56:22] models and very costly on a reduced data
[56:25] set or pay
[56:27] the New York Times um license fee uh
[56:31] open AI is uh arranging license fees
[56:33] with a number of companies because they
[56:35] want to secure at least access to a
[56:37] number of data especially newspaper
[56:39] companies because they give real-time
[56:42] data on the latest events in the world
[56:45] and so that's very valuable for uh AI
[56:48] models because then they can respond not
[56:50] only to things that happened up to the
[56:53] cut off date for their model training
[56:55] data but up to today and and so that's
[56:58] why they're willing to pay for
[57:01] that but back to uh personal data so
[57:04] this is very penalizing for EU
[57:07] users um because what happens is that
[57:10] the most advanced AI models are being
[57:13] withheld from the EU Market at the
[57:16] moment and and so users in the EU
[57:20] businesses in the EU have no access to
[57:22] that um this is especially penalizing
[57:26] for small language communities where
[57:28] there are
[57:29] insufficient data text Data available
[57:32] from common crawl on the internet and
[57:34] they have to use um social media data if
[57:37] they want to give a sufficient number or
[57:40] volume of data for model training and in
[57:43] the EU we have many of these small
[57:45] language communities so I don't know how
[57:48] this is going to be settled the EU says
[57:50] it wants to accelerate Innovation and
[57:52] productivity growth but regulation is
[57:56] slowing it down at this
[57:58] moment and so this brings me
[58:02] to a more G take on
[58:05] this um so what happens until recently
[58:08] we we lived in decentralized market
[58:10] economies with private agents are
[58:13] collecting Market information they're
[58:15] learning from that and making their
[58:17] decisions withal Technologies and
[58:19] platforms coming up they centralize a
[58:22] lot of Market information and as a user
[58:24] I can go to that Central Market data
[58:27] pool extract some information from them
[58:29] through search and then uh um use that
[58:34] for to my benefit and so there are
[58:37] economies of scale and scope in their
[58:39] reuse and aggregation of data and these
[58:43] benefits generate more social value for
[58:46] data that exceeds the private value but
[58:48] still private users hold the key to
[58:51] access those data so how can we overcome
[58:54] this data market Val this gap between
[58:56] the social and the private value so
[58:59] exclusive private rights are an obstacle
[59:01] to uh realize these externalities and we
[59:05] have to somehow find ways in our laws
[59:08] and regulations to overcome that to some
[59:11] extent while still preserving privacy
[59:14] preserving copyright as an incentive for
[59:17] investment in
[59:19] media and AI models they take another
[59:22] Quantum Leap in De centralization what
[59:24] we said before this AI ecosystems on top
[59:27] of platforms they centralize Global
[59:30] Knowledge but they are subject to
[59:32] proprietary rights to data and knowledge
[59:34] so having real time access to the New
[59:36] York Times or other newspapers other
[59:38] propriety data is a problem so should AI
[59:42] machines have a right to access
[59:44] propriety data at least for their
[59:45] training or even for post training uh
[59:49] this is a hot subject of debate at the
[59:52] moment how's it done for search engines
[59:55] sorry
[59:57] yeah for search engines there was a
[59:58] compromise now that you can point to
[01:00:01] some newspaper articles but you can
[01:00:04] search across this the articles in the
[01:00:06] New York Times yeah for for search
[01:00:09] engines the compromise was that they
[01:00:11] could extract a small string of text and
[01:00:14] the length of that string was determined
[01:00:16] by
[01:00:17] law um and how much they could exract
[01:00:20] and show in the search engine this is
[01:00:22] not going to work for AI training
[01:00:25] because you need to hold text of an
[01:00:27] article or paper or text Page or
[01:00:29] something and then there's a question by
[01:00:32] Jeffrey how does it apply to Medical
[01:00:34] Data it's the
[01:00:35] same
[01:00:36] um yeah let's say for medical data U at
[01:00:42] least in the European Union we have a
[01:00:45] separate regulatory regime that will
[01:00:48] kick into action very soon which is
[01:00:51] called the European health data
[01:00:53] space and that regulation provides for
[01:00:57] pooling of all Health Data in EU member
[01:01:01] states and across the EU even because
[01:01:04] the value of that pool is obvious for
[01:01:07] researchers and especially for
[01:01:09] researchers into rare diseases having
[01:01:12] access as country level may not be
[01:01:13] enough you need may need the whole EU
[01:01:16] data Po and so when that data regulation
[01:01:20] kicks in AI models will have access to
[01:01:23] to all that information yeah on an
[01:01:26] anonymized basis of course you cannot
[01:01:28] reveal the patient's name when do
[01:01:33] that so the EU is facing a number of
[01:01:37] policy options at the moment so on hand
[01:01:40] it's obvious that uh we have comparative
[01:01:44] disadvantages we have no hyperscale
[01:01:47] infrastructure uh we have no big Tech
[01:01:49] business ecosystems we have high
[01:01:51] electricity costs uh we have regulatory
[01:01:54] disadvantages AI compliance costs are
[01:01:58] high copyright and personal data
[01:02:00] problems regulatory uncertainty that I
[01:02:03] just spoke about so all this makes it
[01:02:07] difficult and so some people say there
[01:02:09] may be a reverse Brussels effect rather
[01:02:12] than the Brussels regulation inspiring
[01:02:15] other countries the Brussels regulation
[01:02:17] inspires other countries not to do what
[01:02:19] Brussels is doing or trying to do um but
[01:02:23] the EU is trying to set up a policy
[01:02:26] response to that and you may have heard
[01:02:29] of Mario dr's report on European
[01:02:32] competitiveness that came out two months
[01:02:34] ago The Advocates uh especially that uh
[01:02:40] EU productivity grow is too slow and one
[01:02:43] of the main reasons is the slow take up
[01:02:46] and investment in a and digital
[01:02:48] Technologies and in AI so we need to do
[01:02:51] more in that respect uh there's the AI
[01:02:55] factories initiative that focuses around
[01:02:58] existing uh very large scale
[01:03:01] supercomputers in the EU uh the new
[01:03:04] commission has also announced the cloud
[01:03:06] and AI regulation to to facilitate
[01:03:10] mergers and Acquisitions maybe between
[01:03:13] small scale Cloud providers so that they
[01:03:15] can build larger scale
[01:03:17] infrastructure whether that's going to
[01:03:19] work or not is a big question mark for
[01:03:21] me I think the EU cannot afford to throw
[01:03:24] taxpayer money at reaching the AI
[01:03:27] Frontier as we saw this is extremely
[01:03:30] costly even if we throw a 100 billion of
[01:03:32] taxpayer money of that and we built one
[01:03:35] uh state-ofthe-art Computing Center a
[01:03:38] year from now it will be outdated and
[01:03:42] and unless you have a business model
[01:03:44] where you can plug this into uh where
[01:03:47] you can earn a sufficient return this is
[01:03:49] not a workable deal and we've seen that
[01:03:52] in the past with the EU where
[01:03:54] state-of-the-art Hardware or softare was
[01:03:56] produced but no business model to back
[01:03:58] it up and that doesn't work so I think
[01:04:01] collaboration with us big term big Tech
[01:04:04] firms is the only viable option in the
[01:04:06] short to medium term we should not make
[01:04:09] these big Tech firms our enemies we
[01:04:11] should be friends with them and try to
[01:04:13] find uh useful and efficient ways to
[01:04:16] collaborate with them but the other
[01:04:20] option is to prosper below the AI
[01:04:23] Frontier and it's what I said before
[01:04:27] earlier in my presentation is there are
[01:04:29] lots of derived models there lots of
[01:04:32] open-source models uh that can be used
[01:04:35] and that are freely available that
[01:04:37] requires less hardware and Computing
[01:04:40] capacity for inference purposes to use
[01:04:42] it for inference and many firms smaller
[01:04:45] firms can have strong productivity gains
[01:04:48] from user the using these smaller and
[01:04:50] derived open- Source
[01:04:53] models uh they can also upload their own
[01:04:56] proprietary data in these models and
[01:04:59] have even realtime retrieval augmented
[01:05:01] data sources in their models and and
[01:05:04] that could very well work for these
[01:05:06] companies you don't need to run a
[01:05:08] Frontier Model uh for for your company
[01:05:10] in order to increase your productivity
[01:05:12] you usually work on derived
[01:05:15] models um what we would also need is a
[01:05:18] more Pro innovation enforement of the AI
[01:05:21] act as I said the AI act has opened a
[01:05:24] long period of regulatory uncertainty on
[01:05:27] the way to go on how companies should
[01:05:30] comply with these regulation and what is
[01:05:32] expected from them and we have to close
[01:05:36] that regulatory uncertainty off as soon
[01:05:38] as
[01:05:40] possible uh finally there are
[01:05:42] initiatives in the EU to build their own
[01:05:44] large common data pools we already
[01:05:46] mentioned Health Data which is one
[01:05:48] example but there are um ongoing
[01:05:51] initiatives in other Industries and that
[01:05:53] of course it could also be very useful
[01:05:55] training data pools for AI models uh um
[01:06:01] and that that that could be set up in
[01:06:03] the in the EU
[01:06:05] itself of course the risk of it this is
[01:06:08] geopolitical dependence on the US uh
[01:06:11] changes in the US
[01:06:13] Administration U there's uncertainty
[01:06:15] around this and and so do we want to
[01:06:18] take that risk on the other hand do we
[01:06:20] have much Choice can we have our own AI
[01:06:24] Frontier models H I think it's doubtful
[01:06:27] um some people may disagree but uh there
[01:06:31] can ask you so if if Europe we want to
[01:06:33] say in not be in the frontier but the
[01:06:36] frontier is moving very fast but let's
[01:06:37] suppose go two years behind the frontier
[01:06:40] could you run a Chet GPT 3.5 now and it
[01:06:43] will be fairly cheap because now the old
[01:06:45] Nvidia chips everybody can have them for
[01:06:47] free almost yes there are possibilities
[01:06:50] in that direction um as Anon Corin
[01:06:54] already showed last week uh we have
[01:06:57] these very big Frontier models but
[01:06:59] usually what happens is that very
[01:07:00] quickly afterwards these models are
[01:07:03] rerun and compressed into smaller models
[01:07:06] that have almost the same performance as
[01:07:08] their big
[01:07:09] brothers and and so what the EU could go
[01:07:13] is go for these compressed smaller
[01:07:15] models that require less infrastructure
[01:07:17] less costly to run uh you run maybe a
[01:07:21] year behind on the frontier but that's
[01:07:23] not too bad and and and you can work
[01:07:26] very well with that and so where do
[01:07:28] these firms in Europe like Mistral or
[01:07:30] black forest lab where do they stand in
[01:07:33] this so Mr is a year behind or is on the
[01:07:36] frontier as well what's their
[01:07:38] strategy um what we observe is that m is
[01:07:42] slowly moving to the US it is opening
[01:07:45] offices on the west coast it is
[01:07:47] transferring some of its staff to the
[01:07:49] West Coast it is under the same
[01:07:52] constraints as everybody else in the EU
[01:07:55] copyright uncertainty uh data access
[01:07:58] uncertainty uh High electricity costs no
[01:08:02] access to uh hyperscale Computing
[01:08:05] infrastructure or very costly and
[01:08:08] difficult and all this I think are
[01:08:10] factors that make them move towards uh
[01:08:13] the US also getting more access to
[01:08:16] Capital markets there so um I don't know
[01:08:21] how this is going to work out um for
[01:08:25] other AI startups in the EU most of them
[01:08:28] are smaller still than AI than Mist Mist
[01:08:32] has these collaboration agreements with
[01:08:34] number of big Tech firms and that keeps
[01:08:36] their head above the water iting gives
[01:08:38] them access to infrastructure so um yeah
[01:08:43] uh collaboration cooperation cooption
[01:08:47] model I think is still the only way to
[01:08:49] go for AI
[01:08:53] firms okay then just one one last slide
[01:08:56] with some
[01:08:58] thoughts behind beyond the narrow
[01:09:01] economic
[01:09:02] mindsets um what keeps AI investment
[01:09:06] going uh I think an important factor
[01:09:09] that we often Overlook is formal the
[01:09:10] fear of missing out competition between
[01:09:12] big tech companies none of these big
[01:09:15] tech companies wants to give in to the
[01:09:17] others and and and so they keep
[01:09:19] investing and they keep betting the farm
[01:09:22] ongoing for AI they cannot do otherwise
[01:09:25] as soon as you show that you're
[01:09:26] backtracking a millimeter on this stock
[01:09:29] markets punish you very
[01:09:31] heavily um so you you're forced to go on
[01:09:35] with this game the second factor is the
[01:09:37] China security Factor AI is a dual use
[01:09:40] technology can also be used for conflict
[01:09:44] uh we see it being used in armed
[01:09:46] conflict now in Ukraine in in in in the
[01:09:49] Middle East and and elsewhere no doubt
[01:09:52] is an important factor in cyber security
[01:09:56] uh and the US wants to stay on stop of
[01:09:58] China the explains the AI chip export
[01:10:01] restrictions and also a lot of
[01:10:03] encouragements for us big Tech to go
[01:10:06] Full Speed Ahead in this but I think
[01:10:09] with the Trump Administration we will
[01:10:10] see more initiative that will encourage
[01:10:12] Us big Tech to go even faster in
[01:10:15] that and last but not least the alism
[01:10:19] factor AI for good there are many people
[01:10:22] we believe who believe we should go on
[01:10:24] investing in AI because simply it
[01:10:27] generates enormous human welfare
[01:10:29] increases we see Advantage from Deep
[01:10:32] Mind for example in the health sector in
[01:10:35] medicine and
[01:10:37] pharmaceuticals uh we should keep it
[01:10:38] aligned with human values uh open AI
[01:10:42] until recently was uh the main uh driver
[01:10:45] behind this AI for good of course then
[01:10:48] this created some tension with its
[01:10:50] commercial objectives and with the need
[01:10:52] to generate Revenue
[01:10:55] and so it's now becoming more and more a
[01:10:57] for-profit company but I think there's
[01:10:59] still a hardcore of people who who work
[01:11:02] on this alteris factor and will continue
[01:11:04] to work on
[01:11:06] that okay let me stop here thank you
[01:11:09] very much thanks a lot uh Burton
[01:11:12] so going forward you did not focus much
[01:11:15] on China I don't know whether you have a
[01:11:16] take on China it's just a final question
[01:11:19] uh you know last week we have seen from
[01:11:22] Anon that actually China is really close
[01:11:24] behind the United States would you
[01:11:27] subscribe to that so it's essentially a
[01:11:29] battle between China and us and and
[01:11:31] Europe is as you said within the
[01:11:33] frontier and that's the optimal space
[01:11:36] for Europe to
[01:11:38] take yeah uh We've indeed seen over the
[01:11:41] last two months uh a number of models
[01:11:43] coming out of China that have cognitive
[01:11:48] performance uh characteristics that are
[01:11:51] almost in line with the best models that
[01:11:53] come out of the US
[01:11:56] um theyve been able to do this despite
[01:11:59] restricted access to chips
[01:12:02] um of course they use the same data sets
[01:12:06] as us companies use for training their
[01:12:09] model these data sets are freely
[01:12:10] available for anyone to use uh they also
[01:12:14] have access to enormous volumes of
[01:12:17] Chinese social media context text Data
[01:12:21] uh other types of data so I don't think
[01:12:24] there is a data War in China at the
[01:12:27] moment um so so that gives them an
[01:12:31] advantage um yeah um and of course the
[01:12:34] Chinese government will also want to
[01:12:37] come on top of the US and so they
[01:12:39] investing everything they can in
[01:12:42] accelerating this so we'll see a
[01:12:45] technology race between those two
[01:12:47] superpowers so finally one final
[01:12:50] question normally we end with a positive
[01:12:52] note but here I would like to change
[01:12:53] this tradition let's suppose the funding
[01:12:56] is going away so Wall Street is not
[01:12:58] willing to fund all the initiatives to
[01:13:01] the same degree they funded so
[01:13:03] far uh what would be the scenario how
[01:13:06] would it play out if suddenly there
[01:13:08] couldn't access the financial markets
[01:13:10] anymore these
[01:13:11] companies um would everything smoothly
[01:13:14] slow down or things would and we just
[01:13:16] run with the old versions with the
[01:13:17] current versions of the large language
[01:13:19] models or what's your vision how would
[01:13:21] such a scenario play out yeah um
[01:13:26] as we saw with the early 2000s do com
[01:13:29] busts uh stock market can go bust but
[01:13:31] the technology moves
[01:13:33] on and it may slow down a little bit for
[01:13:37] a couple of years but eventually it
[01:13:39] comes
[01:13:39] back uh so what could happen is indeed
[01:13:44] that um the stock markets say no this is
[01:13:47] no longer feasible for big tech
[01:13:49] companies to invest so much money in AI
[01:13:52] they don't earn necessary Returns on
[01:13:54] this
[01:13:56] um and and so it would force these
[01:13:58] companies to slow down a little bit on
[01:14:01] this um they can still train models and
[01:14:06] they can still with new technologies uh
[01:14:09] uh go for more reasoning in AI models go
[01:14:12] for more uh learning memory and so on so
[01:14:17] there are ways to still increase
[01:14:19] capacity and the performance of these
[01:14:22] models even if you're financing plateau
[01:14:25] a little
[01:14:26] bit um and that would maybe also give
[01:14:30] time to the J curve to catch up uh to
[01:14:33] generate more revenues more productivity
[01:14:35] gains so as to convince Wall Street that
[01:14:39] yes we can go further ahead in this so
[01:14:42] maybe a temporary slowdown but I don't
[01:14:44] think we should expect a big
[01:14:48] slump thanks a lot Burton it was
[01:14:50] fantastic to get your insights on this
[01:14:52] and hopefully we can stay in touch and
[01:14:55] as things move keep moving very fast um
[01:14:58] you know let's keep track of what's
[01:15:01] going on and I'm surely hopefully we can
[01:15:03] rely on your scientific writings and all
[01:15:07] the other things you're doing so that
[01:15:09] the community can learn from you as more
[01:15:11] thanks again and thanks to everybody for
[01:15:13] coming next uh thing will be on AI and
[01:15:17] finance uh and then I see you again
[01:15:19] thanks for everybody for participating
[01:15:21] bye-bye thank you byebye