WARNING: Summaries are generated by a large language model and may be inaccurate. We suggest that you use the synopsis, short and long summaries only as a loose guide to the topics discussed.
The model may attribute to the speaker or other participants views that they do not in fact hold. It may also attribute to the speaker views expressed by other participants, or vice versa. The raw transcript (see the bottom of the page) is likely to be a more accurate representation of the seminar content, except
for errors at the level of individual words during transcription.
Urban AI CEO Sam Altman predicts that the marginal cost of intelligence and energy will become close to zero in the next decade. This could have major implications for the economy, creating a phase transition in scientific research, and throwing traditional economic models into doubt. Open AI and Microsoft are developing models to give robots reasoning capabilities, while Adam proposes a new way of thinking about the economy that takes into account the physical layer of energy and materials. GDP is becoming an increasingly unreliable measure due to deflationary trends, so a new model of economics must be developed.
Sam Altman, CEO of Urban AI, predicts that the marginal cost of intelligence and energy will become close to zero in the next decade. This could lead to an exponential increase in the usage of these resources, and have major implications for the future. William Stanley Jevins first observed that as costs of resources decrease, expenditure on them increases rapidly. Digital technologies have made this even more pronounced, with the marginal cost of intelligence and energy being low or even negative. Open AI's cost of training GPT4 was 12 million dollars, but the marginal cost of each request is very low.
Access to AI systems is likely to become increasingly valuable, creating a bottleneck in deploying and widely distributing this intelligence. This could lead to a phase transition in scientific research, as those working with AGIs do all the interesting work. Markets tend to clear near the marginal cost, so when solar producers are competing to sell electricity, they can underbid other producers as their marginal costs are so low. This leads to a situation where solar is the most competitive, but this abundance of resources may not necessarily lead to a good outcome as it could be dangerous to allow everyone access to it.
In order to justify investments in capital such as a power plant, it can be difficult to expect returns when the marginal cost is zero. However, it is still possible to create and capture value even when the marginal cost of information and communications is near zero, through an internet of energy. Altman suggested that a few organizations will train the largest models, and an economy will sit on top of those models to fine-tune them to specific data sets, due to a limitation of talent and a lack of GPUs.
Open AI and Microsoft are collaborating to create models that can be adapted to different physical environments, such as robotics and drug affinity prediction. Google has developed language and vision models to give robots reasoning capabilities, but it is unclear if these models need to be continually fine-tuned. Humans have an evolutionary pressure to pass down simple ideas that can be easily taught and understood, allowing for more complex theories to be developed and shared while still ensuring the most important information is retained.
Adam proposes a new way of thinking about the economy that takes into account the physical layer of energy, materials and information flows necessary for production and services. He suggests that as civilization expands, there will be an advantage to specialization, requiring both generalist and specialized models. In a world of abundance, markets are no longer necessary and GDP is not useful as a metric for productivity, so a new model of economics must be developed.
GDP is becoming an increasingly unreliable measure of economic productivity and consumption due to technology reducing costs and increasing productivity. This deflationary trend is causing confusion and recession, inflation and deflation all at once. Positional goods may not be captured by GDP figures, but businesses may monetize them and create artificial scarcity. GDP per capita is increasing, but energy movements are not increasing proportionally. As we move towards abundance, GDP may shrink and become a less meaningful measure.
Sam Altman, CEO of Urban AI, believes that the marginal cost of intelligence and energy will reach zero in the next decade. This would mean that the current level of consumption would become exponentially cheaper, but Altman also suggests that this could lead to a dramatic increase in the usage of intelligence and energy. He believes that this could have huge implications for the future.
Adam has pointed out that as costs of resources decrease, expenditure on them increases rapidly. This has been observed for a long time, with William Stanley Jevins being the first to observe it in the 1800s. Adam's team are drawing focus to this, but the idea of costs approaching zero is a more recent phenomenon due to the digital revolution. The marginal cost of intelligence may go to zero, but access to AGIs will still be a major policy issue in the coming decades.
Marginal cost is the cost required to produce an additional unit of a good or service, and for digital technologies it is very low. In some cases, such as the energy sector, it can even be negative. Open AI's cost of training GPT4 was 12 million dollars, however the marginal cost of serving each request is very low as it only requires electricity and usage of hardware.
Abundance of resources does not necessarily lead to a good outcome, as it may be dangerous to allow everyone access to it. Sam Altman pointed out that when there is a major restructuring of cost in the economy, it triggers other sorts of change in social, economic, finance and geopolitical aspects. As an example, Jordy Williamson, a famous Australian mathematician, collaborated with Deepmind researchers to look into some problem in his field of mathematics. This opportunity to work with an AI and discover patterns is only available to a few top level mathematicians, leaving others unable to take advantage of it.
Access to AGI systems is likely to become the currency of the realm, creating a bottleneck in deploying and widely distributing this intelligence. This could lead to a phase transition in scientific research, as those working with AGIs do all the interesting work. Markets tend to clear near the marginal cost, so when solar producers are competing to sell electricity, they can underbid other producers as their marginal costs are so low. This leads to a strange situation in the energy sector, where solar is the most competitive.
In order to justify investments in capital, such as a power plant, it is difficult to expect returns when the marginal cost is zero. This can lead to a saturated market, where it becomes harder to justify additional investments due to diminishing returns. This can cause adoption of new technologies to slow down and cease to grow exponentially. To capture value outside of traditional electricity units, such as information, it is important to consider how much demand there will be and how much innovation will be created to capture this value.
It is possible to create and capture value even when the marginal cost of information and Communications is near zero. This can be done through an internet of energy, where energy is a service, and services are more interesting than just buying kilowatt-hours. Altman suggested that a few organizations will train the largest models, and an economy will sit on top of those models to fine-tune them to specific data sets. This is due to a limitation of talent, as well as a lack of GPUs.
Open AI and Microsoft are working together to create models that can be fine-tuned for different applications. It is thought that this could be beneficial for robotics, as a base model could be adapted to different physical environments. This would also be useful for predicting the affinity of drugs for certain purposes. Altman was likely thinking of this when he mentioned the idea, as there are a large number of physical environments that could benefit from this, as opposed to just a few types of creativity or services.
Google has developed large language and vision models to be used in robotics applications. These models provide robots with reasoning capabilities, such as understanding instructions and how to use objects. It is unclear whether these models need to be continually fine-tuned or if a master model can be trained to an optimal level. The master model could potentially be trained to exceed the capabilities of the best Olympic athletes in every sport.
Humans have an evolutionary pressure to pass down simple ideas that can be easily taught and understood. This process of teaching and learning serves as a regularization role, ensuring that only the most useful and relevant information is passed down to the next generation. This is in contrast to a single model, where all knowledge is shared and accessible. This allows for more complex theories to be developed and shared, while still ensuring that the most important information is retained.
Our current economic metrics and measures are abstractions that don't take into account the physical layer of energy, materials and information flows that are necessary for production and services. Adam is proposing a new way of thinking about the economy that takes this layer into account. He suggests that as our civilization expands, there will always be some advantage to specialization, even if it's at a much higher level than what we are currently doing. This could mean that generalist models will expand their capabilities rapidly and there will be a need for more specialized models.
Economics is defined by scarcity, and markets are one mechanism for allocating scarce resources. Futurists generally agree that we are heading towards a world of abundance, where many goods and services are available for negligible cost. In this circumstance, markets are not necessary, as most things are not scarce. GDP is a metric used to measure productivity, but it will not be useful in a world of abundance. Instead, a different model of economics must be developed to account for the new paradigm.
GDP is a flawed measure of productivity and consumption, as technology has caused a deflationary dynamic that drastically reduces costs and increases productivity, but does not show up in GDP. As we move towards abundance, GDP may shrink and cease to be a meaningful measure, and the beginning of this journey may look like what we are seeing now.
GDP is becoming less and less legible as a measure of the economy due to the marginal cost of many goods and services approaching zero, as productivity skyrockets. This transformation of GDP may be beginning now, as the global economy is experiencing confusion, recession, inflation and deflation all at once. An example of this is the large groups of people who are opting out of the usual economy, either because they can't get jobs or because they choose not to pursue ordinary careers. The question is whether they want these things and can't get them, or they just don't want them.
GDP figures may not directly reflect the value of positional goods, but businesses may monetize them. In China, buying expensive alcohol to share with friends is seen as a sign of high status. GDP figures may not capture everything that is valuable, but there are some proxies that may show up. Artificial scarcity can be socially constructed, and beneath the layer of abstraction, production and productivity are growing. GDP per capita is increasing, but it is astonishing that the amount of energy being moved around is not increasing proportionally.
GDP per capita has increased exponentially over the centuries, despite fewer people actively participating in the economy. This includes young people who don't work, retirees, and people in bureaucracies who are not doing much productive work. During the pandemic, many people stopped looking for jobs and were not included in the official unemployment figures. This is known as 'laying flat' in China, or 'shut-ins' in Japan. There are also people who remain as 'children' and do not contribute to the economy until their early twenties. This shows that even with fewer people actively working, GDP per capita can still increase.
yeah what we sort of have two sets of things to talk about um which one do we want to start with I'll also flag at the outset I think we should stop at 10 minutes to the hour because there's something I want to show you before we break up this meeting so yeah should we start by talking about the Sam Altman interview or do you want to maybe get into sure yeah yeah I'm sure that's fine I'm not sure how long this this will take because we've sort of trod this ground many times before but I thought it was worth remarking on the uh the themes that Sam Altman is bringing up in this interview as as you know as being the kind of things that we've been talking about so for those of you don't know Sam Altman is the current CEO of urban AI one of the major independent research organizations doing Cutting Edge research and deep learning that's the organization that produced GPT 2 3 and Dolly and so on and this interview which I think is I mean it got I noticed that a few days ago I think the the conference where the interview took place was actually recent um so let's I made a list on Discord I'll just go through some of the things that I thought were noteworthy and by the way I guess Sam Altman is uh considered to be one of the Four Horsemen of the Apocalypse by AI safety like uh alignment people so uh he also spoke a little bit about alignment um nothing too interesting yeah so the first thing I noticed yeah sorry go ahead I I one line in particular caught my eye so you go first and then if if you didn't catch your I2 then I'll add it in okay so uh the first thing I'll write up here is I did his basic model for the next decade is that the marginal cost of intelligence and energy is is going to go to zero that's the one that's the one that's the one that caught my eye yeah and that yeah okay so that's I guess uh pretty maybe it's not obvious that he would have added the energy one in there uh given that he's the CEO of an AI company uh but yeah it's obviously relevant but I think the the follow one from that is is one that I think is not as obvious and which uh which we may find interesting to discuss a little bit which is he pointed out that as the marginal cost of both of these goes to zero you might think that well that just means that the current usage just becomes much cheaper so we stay roughly at the current level of consumption of intelligence and energy but just the the amount of costs goes exponentially down but he points out that as the marginal cost goes to zero
we expect spending on both of these resources to actually rapidly increase which is something Adam has highlighted quite a number of times as being characteristic of disruption that's right yep and I I should note that that uh that idea doesn't originate with my team we're sort of drawing Focus to it and we've we've what we've pointed out is that this has happened many times in the past um the uh and I suppose approaching a marginal cost of near zero is is a little bit more is a little bit extraordinary because that hasn't happened that is all I should say that's only happened in recent decades with the digital Revolution basically so that there were very very few instances of anything like that in the in the 19th century for example however the idea that as as um as costs of any kind to go down substantially that you can have uh a a price elasticity of demand such that you get a massive increase in expenditure around uh whatever that thing whatever the thing is who's costs fall that's that's been observed for a very long time and uh William Stanley jevins was the first to right um point to this in energy in the 1800s and that was before the marginal Revolution and before the concept of elasticity of command price elasticity of demand was was formalized so this is an idea it's been around a long time but what's weird is the idea that it costs could approach zero and that as they approach zero um you could have an explosion in in consumption of whatever those things are yeah maybe let's do it so this is debate that point a little bit is it is it he said the marginal cost of intelligence would go to zero I think he's just sort of speaking a little bit off the cuff I'm not sure if he really believes that because a little bit later in the conversation he was talking about how one of the biggest policy questions of the next few decades will be who gets access to the agis and on what terms so that hardly suggests the marginal cost is zero or at least it's not uh it's it's not it's not zero for everybody it's not it's not abundant necessarily um I mean maybe we should maybe we should clarify what one means by the marginal cost going to zero as compared to the cost going to zero I mean as right those are two quite different things in fact the marginal cost can go negative under certain circumstances um so for example for part of the time the marginal cost can be negative for certain things and you can still have a positive capital expenditure required to to you know build out the capacity to
produce whatever those things are so a marginal marginal cost is where the is the is the the cost required to produce an additional unit of a good or service and for um for digital Technologies this is very familiar so it it it's it costs you several hundred dollars to buy a digital camera or a device like a smartphone that has a digital camera in it but once you've made that Capital purchase that Capital Investments and you've got that capital in your hands you've got that piece of equipment in your hands then it costs near zero to produce one or one additional image from it whether you produce one or ten or a thousand images with your smartphone it's barely noticeable on an incremental basis so that's what we mean by the marginal cost is uh is near zero and there are some exotic Market circumstances where we can see negative marginal costs so I'll give you an example in the energy sector so because of um some perversity in in the wholesale markets and electricity there are situations where coal and nuclear power plants really really don't want to ramp their output up and down that's expensive at in their equipment and so what those plants will do is they will pay people to take their electricity so that they don't have to ramp their production down and then back up again later and when they do that the cost of an additional unit of electricity on the open market can be negative and so for about for approaching 10 percent of all hours in California now the cost of electricity on the wholesale Market is actually negative the marginal cost so the cost of one additional unit of electricity and from a wholesale perspective for buyers in the wholesale Market is actually negative now that's not a negative cost for producers but it's a negative cost from the perspective of where it's consumed so we have some some those are some examples of maybe it's where we see near zero costs so the in the situation that I suppose open AI would be in maybe an example of near zero marginal cost would be costs 12 million dollars to train gpt4 but then to serve each request that a user submits to summarize a document or provide the meaning of life or whatever the the marginal cost of doing that is just the inference cost which is not zero but it's some amount of electricity and usage of Hardware somewhere it it's probably not zero but it's very low and especially low compared to the original capital cost of training the thing yep that's right that's right um yeah um so I think that's that's fair uh in
energy we're we've we've talked about that in the past with our my team's concept of superpower producing at very very low uh what we we say near zero we don't ever see Zero we say near zero marginal cost um but uh extreme you could also say just extremely low um and I I think that Sam Altman was right when he pointed out that anytime there's a major restructuring of cost throughout the economy that this triggers other sorts of change um in social change economic economics Finance change geopolitical change and so I think this is an important observation certainly Matt what was your point about security there just that the the abundance is not necessarily a good thing I guess uh can you hear me yeah yep so you said Dan you're not sure he believes in the let's even say near zero marginal costs because later he says we still need to worry about who is going to get access to it but I think that consistent you could have something that has near zero marginal cost or is abundant but that uh you don't necessarily want anyone to have access to it because it's dangerous right so if there are risks associated with resource you still need to control access even if the resource is abundant that's a great Point yep I haven't listened to the interview so I'm not sure um if that's consistent with like if that actually resolves what um was banging those two places yeah I doubt that's what he had in mind uh I maybe uh it was a very it was just a couple of sentences um I think I think he had him more in mind it was a kind of question of social justice who gets to let's take a concrete example so uh Jordy Williamson who's a famous Australian mathematician um had a recent recently did some work with deepmind where they uh deepmind researchers collaborated with Geordie to look into some problem in his field of mathematics and they they discovered some patterns and maybe managed to prove some interesting conjectures about those patterns and that's a I don't know much about it but it seems to be a substantial piece of work now not every mathematician would have been able to take advantage of that opportunity to work with the researchers at deepmind but they basically uh so right now that that resource that is work with an AI in a very deep fashion get it to look at a problem you think is interesting discover patterns make suggestions and help you discover something is an opportunity that's available to well maybe one or two or three top level mathematicians who you know are tapped on the shoulder by Deep
Mind or open AI or some other research organization with the resources to provide access to that kind of system uh suppose that Geordie goes on to do that three times a year but you know who else gets to do that so it's at the moment it seems like it's a bit of an unfair question in a way and I'm not criticizing Jordy at all for it right it's just it's cool you know they come to you and they ask you would you like to do this thing um why not but as this becomes a more routine practice it will become analogous too well if you're at a good University you have access to more compute power and in these days you have access to more gpus uh once there are agis on the scene depending on what happens who gets to collaborate with them may be the dividing line between if there's a phase transition in scientific research where those are working with the agis are doing all the interesting work which seems completely plausible if not inevitable to me then who gets to decide uh so that's that's the kind of Justice question I think he had in mind or at least that's that's how I read it so you did mention at some point that access to AGI systems is likely to be the currency of the realm which I thought was a a nice quote if the intelligence was really abundant then you wouldn't have the problem so it must be that yeah like you said there's some bottleneck um in deploying or sort of like widely Distributing the the actual access to that intelligence um I don't know if is that would that be then counted as part of the marginal cost well this brings up another another interesting another interesting issue and we were asked my team's asked about this quite a bit we don't quite have an answer yet not a complete answer at any rate which is um uh if the marginal cost approaches zero and markets 10 whole markets tend to clear near the marginal cost so when you under under circumstances of strict competitions when when you actually have a competitive market um the clearing the clearing price is often quite often typically quite near the marginal cost of course what that means is and this is one reason why we have this strange situation and the energies uh sector right now is um when the sun is shining uh and different producers are competing with one another to sell their electricity solar can under bid just about everybody else they can under bid coal they can underbid nuclear they can under underbid natural gas because their marginal costs are so low so anything that they get from selling a unit of electricity
they benefit from because it costs them so little to produce it on a marginal basis and so that in other words that makes that makes it very difficult for anybody else to compete with them and then a competitive market you know you get you can see how you would that would drive the clearing price to the you know towards zero okay well here's the problem the problem is that you still have to justify your investment in capital so if you want to build a power plant you have to have a way to expect returns out of that and if if your marginal cost is zero um then it's an open it's a it's a difficult and you know question uh how do you rationalize making the investment and how do you expect to get a return on that investment if you're you're anticipating a competitive market in which in which the price is clear near zero and so this how do you get any profit out so it's more like and it's more like a the British government investing in the Merchant Navy or something to go out and grab a fixed resource from somewhere and bring it back rather than investing in a long term I mean if you you can still make money while you're taking it off the nuclear power plant or the coal power plant right while the capacity from the solar system is still less than the total demand that's right and so this is one one uh possible contributor to the the we've talked about how these adoption of the Union Technologies tends to follow a sigmoid curve an s-curve right well one one one one reason why you might expect for example in the energy case of the energy disruption one reason why we might expect that adoption would slow down um very definitely cease to grow exponentially at some point and begin slowing down is when the market starts to get saturated with solar and then it becomes harder and harder to justify additional investment in capacity because of diminished very you know aggressively diminishing returns on that investment so but as I said my team we've not completely gotten through all of this um because the biggest question is is uh again how how much demand will there actually be and um how much Innovation will there be in in finding ways to capture value outside of the traditional sort of just cell unit of electricity so for example if you were to look at the internet naively from the from the you know for from say 1995 that you know when you look out into the future it might be difficult to imagine how you know you would continue to make money selling information and selling
Communications and selling you know related Services if the marginal cost of information and Communications was so low is near zero how would anybody make money let alone you know create trillion dollar companies like you know Google and Facebook and apple and so forth on that um and uh so but it turns out there are ways to to you know to to to continue to innovate and continue to create and capture value so it's it's we we're what we're imagining as a result just again I'm sorry to keep invoking energy here um but maybe more something more like an internet of energy um and services that are you know exotic and more interesting than just buying kilowatts hours um as we have in the past so energy is a service and and so forth maybe something along those lines we'll again my team hasn't worked all the way through these details but yeah it's definitely it's definitely a big question mark is has you know the the cost costs of the cost of the system isn't zero it's only the marginal cost of each additional unit of electricity that's going to be going to approach zero and it will be the same as Matt said with with intelligence right the cost of the cost of these models and the cost of training them is this certainly not going to be zero anytime soon even if the even if energy becomes very cheap you know we're not we don't have rep Star Trek replicators and we're not anywhere near that so we're really not approaching zero for the capital investment that's required yeah actually it was interesting to hear Altman referred to just the the roof on the number of chips that are available as a as a key constraint so he was talking at the beginning of that interview about the business models he expects to be adopted going forward so he thinks there will be I mean somewhat self-servingly of course uh a few organizations that train the largest models and I guess he has in mind orders of magnitude bigger than anything that currently exists uh there will be a few organizations that train those and then a whole sort of economy that sits on top of those models fine-tuning them to specific data sets or so somehow interfacing with them in order to serve various verticals uh and the reason he put forward for that is not only a limitation of talent I mean it takes a lot of engineering and uh you know across an increasing number of areas of computer science and engineering and and so on to run data centers and supercomputers and clusters and so on to train these models but there's also just not that many gpus
so that you can't have hundreds of companies training models at The Cutting Edge it's just not enough a100s out there so he sees that as a I suppose one of the reasons why open AI is expect thinks it will be one of those companies or at least open AI slash Microsoft um well one thing that Isis was wondering when he mentioned that in the interview one thing that came into my mind Bree immediately was um well first I thought well geez how many how many different um uh how many different you know you know specific applications would you need fine tuning for and then and so initially my mind was well you know there's there's yeah you could do art and you could do you know music maybe a few different kinds of Art and maybe you know people might have different you know and so I struggled to think of oh I think a scenario where you had where you had dozens or hundreds or thousands let alone tens of thousands or millions of of um fine-tuning jobs that you would want to do on some base mod on some handful of Base models but then then it occurred to me really quite quickly and I wonder if Altman was thinking this um if you had if you had Robotics and you had them operating in physical environments they're a very large number of physical environments in the world so never mind like Market categories or you know types of services or you know forms of creativity there are lots and lots of those but if for example you want to take a base model and train it to operate in your factory or your house or you know operate in your hospital um or you know Etc et cetera well if if you want to take a base model uh very much the way I up until you know have been for a number of years I've been thinking well you know you train a car that can drive itself and you take the same basic model and you adapt it to driving in scare quotes there um driving in you know the uh the the the parts bin at the warehouse or the you know something like that and so I I maybe it occurred to me that maybe that's the sort of an example of a very large amount of fine-tuning that would be beneficial as this sort of fine-tuning that can happen four physical spaces of which there's a there are a very very large number I think that's likely to happen uh I think you may have may have had in mind easier things just uh like fine-tuning on a protein database to make a model that is good at predicting the Affinity of drugs for certain purposes or whatever but it does it does look like robotics is going exactly the way you describe
the latest work out of Google is exactly taking large language models and similar Vision models and just basically plugging them into robots with some fine tuning in order to provide the kind of reasoning capability for the for the robots so that does it does seem like in particular one of the areas where you would make use of one of these very large models would be in robotics applications just as a kind of Common Sense understanding the universe how things work understanding instructions a kind of symbolic language layer in order to reason about objects in the world and how to use them all of that has been demonstrated now at least in principle by the latest robotics work out of Google and other places so that does seem like we um having said that and this is a question I have for you guys the the last thing that that occurred to me and again it's a question is um is it is there is there is it really likely that we're going to need to fine-tune indefinitely or is it not is it simply not plausible that you could train a sort of Master multimodal model that can just be optimal with everything in other words like you know with human beings you know you you everybody sort of achieves every every neurotypical person achieves you know an approximately Baseline uh you know understanding of physics that allows them to walk across the room and and function in normal ways but you really have to do a a great deal of additional training and fine-tuning to become a professional athlete absolutely doing that in more than one oh sorry did I miss something in the chat no Matt was just saying that that's yeah we've built our economy on fine-tuning humans oh yeah I see sorry um yes yes yes exactly um so but what I wonder is I mean if if you're uh if you can if you can continue to train something you know with at uh um you train your master model um are is there going to be value to find is it in other words I guess what I'm trying to ask is is there going to be a point in a point meaning is there going to be a continued return on investment and fine training um Beyond you know uh some baseline that you can achieve with training the Master model as the Master model can do right 10 000 hours of every conceivable physical activity you know we really need wouldn't it be good enough I mean if it's say you train an AI robot to you know um you know in agility and dexterity and so forth and it becomes you know more agile than the best Olympic athletes in across every sport 10 times as good as
them you know it's had it's had a million hours of of discus throwing training in a million hours of high jump training in a million hours of Jiu Jitsu training in a million hours of whatever else are you is there any point in additional fine-tuning past that I mean I guess that presumes a static background to adapt to um go ahead Matt I think it's a it's a cool analogy the the human um version of this is like you said one uh well maybe imagine the population is like suddenly able to uh communicate Lessons Learned like a lesson that any human learns in the world suddenly you also learn that lesson and all of the lessons that you learn in the world are suddenly learned by everyone else right um and then but so so that's I mean I guess that would it's not necessarily architecturally possible with humans but it might be possible with um a larger a larger single model um one comment that I have is that I think that the the like system that humans have where like uh you sort of start from scratch every generation and you have to well not from completely from scratch because you have your kind of um instincts but um but you more or less start from scratch in terms of like you have to humans have to every human born has to like relearn culture and everything like that and then all of their skills um this plays like some kind of regularization role right so if some generation goes off and finds some really complex Theory um in in research and like understand like a single person in that in that field maybe understand some really complex idea in order to like pass it down education through education to like the Next Generation it's that's gonna like that's maybe like a useful process in terms of actually um I'm not sure exactly how to say it but but like when I learned something I I understand maybe like the high level simple ideas and um that's what gets passed down that's like a kind of regularization it's also an evolutionary generation so an evolutionary pressure because what gets passed down is usually that which you can teach to many students and they then teach students or write it down it's not just a single usually it's not sufficient to just work it out on your own and even write it down it's Can it can it actually be put into other people's brains in an active way and and that determines what gets communicated yeah you could you could imagine yeah you just accumulate yeah sorry go ahead I think I was probably going to say a similar thing you maybe you can maybe
you can just accumulate this stuff I I think it's conceivable that you could have some replacement for this regularization or evolutionary pressure in the model um but if you but you know that's that's that sounds kind of I have no idea how you would do that um it's not as simple as just remembering everything you ever learn in every situation it seems like you need something more yeah I think my take on Adam's comment is relevant here so while that seems true you can imagine a Master model being trained on every current task well what if there are new tasks there's always as the sphere of capability of our civilization expands it will there will always be some advantage marginal advantage to being paying more attention to a particular thing because the the generalist model has to absorb a lot more and will by definition be slower even if it only takes 500 milliseconds to master a skill you're still going to be faster mastering one than a hundred so I guess there's always some advantage to specialization even if it's at a much much higher level of capability than what we're currently doing I mean I think that seems to me to be what's happening with uh there is still some value to fine-tuning and in the next generation of models it won't be exactly fine tuning it'll be like you know stick some information into a very long prompt or the the precise technical nature of specialization will change but I think they will at the same time as the generalist models expand their capabilities rapidly they will be always on the frontier use for more specialized models I would think very interesting um so we move on to the other topics we wanted to discuss Adam otherwise we'll use up all our time ah sure yeah thanks Dan um I always lose track of time when we're having this these discussions um uh okay so well it's not unrelated you know I think we can make it maybe make some connections certainly to the near zero marginal cost things we were talking about um but okay here let me let me just paint the picture very quickly and and broadly the thing I've been was thinking about is um our measures of uh are measures of and metrics for understanding our economy uh today and and uh they're pretty I I so I think that it's a problem that we do a lot of our economic thinking in abstractions we abstract away from the the layer where stuff is really getting done the physical layer where there are flows of energy and materials and information and we're making stuff and Performing services and people are
getting real useful value out of that you call it utility or or use value or whatever you want okay um I think we do a lot of our reasoning and then of course policy making on top of that and decision making and planning investing and so forth um in in a in in a layer that's abstracted from from the ground reality so that's that's one piece of context there um so okay here's the here's the thought um we are our economy to date and economics itself is defined by scarcity how do you allocate scarce resources markets are one mechanism for doing that right so so markets are are one only one of a number but they're particularly effective at least when it rained in um but it's all about scarcity we don't have enough of the things goods and services that we'd like to give everybody and so there we have to allocate and we've a lot of futurists generally agree that we're heading towards a world where things are much more abundant and one can imagine just for the sake of a thought experiment you know sort of a Star Trek kind of future when there's an arbitrary large quantity of any good or service available for negligible cost you know if you have the replicator for example um uh and you have an arbitrarily large amount of energy available and arbitrarily large intelligence and information processing capacity then you would have a situation where very little was scarce in that circumstance markets are not really necessary They Don't Really function it's not for most things I suppose there are some things for which there might be artificial scarcity or a few things that might be genuinely remain scarce like you know Priceless ancient works of art or beachfront property in Malibu and these things are finite but a lot of things are going to be Super Hyper abundant and we I don't think there's a whole lot of disagreement among most futurists that that you know that the economy must look very different uh when markets sort of cease to function and we shift away from from a paradigm of scarcity to one in which abundance is much more widespread and we've already seen some major steps towards that specifically in the information and Communications domains um okay so that's all premise here's my question my question is if GDP measures productivity GDP which is a very a fundamental metric that we use right now gross domestic product so it measures uh production and and there it's not the only it's not the only thing there are lots of other things that have that that are in the
picture DDP is is very badly flawed but but for what it is right now we depended we we depend on a lot GDP is a rough measure of productivity and the consumption that is sort of concordant with that and uh we have already seen instances where technology causes sort of a a deflationary dynamic radically slashes the cost in some cases towards marginal marginal cost of near zero and as a result uh productivity increases but GDP fails to capture any of that production and it vanishes from visibility it ceases to be visible or legible to that metric because I'm specifically talking about the metric of GDP here okay so uh if it's a and a good example is your smartphone and how it has you know you can take digital pictures you can navigate you know you've got a compass you've got you know entertainment you can play music and all these things that you know 30 years or 40 years ago would have been physical products or or services that you purchased and it's not an exaggeration to say that you you can you know you can get millions of dollars worth of 1985 value out of your smartphone the smartphone this only costs 500 so let's say so so what's interesting is that all of that productivity the production of those goods and services still occurs the perfect in fact more we're producing more images than ever producing more um content more video more producing producing producing and consuming consuming consuming more in that domain again information and Communications than ever before but none of that shows up in GDP it's all invisible to GDP okay so here's the question the question is does GDP as we move towards abundance does GDP shrink does GDP actually have so much uh ephemeralized out of its visibility out of its Vision out of its ability to capture uh that as we sort of race up towards hyper productivity um GDP not only ceases to be meaningful measure but it's just it it might actually look like it's declining it does it so that's that's the first that's sort of the hypothesis number one is that is that sort of plausible given the flawed nature of GDP as a metric okay here's the hypothesis two which are much more interested and excited about hypothesis 2 is what would the beginning of that Journey look like if GDP were to start to for example basically just collapse in its usefulness and it would be a very bumpy ride would it not and what would the beginning of that look like and would it look like anything that we're seeing already that's my question can you repeat
hypothesis one or question one is GDP so hypothesis one is Will as productivity skyrockets and um uh you know as Sam Alton was saying the the marginal cost of of intelligence energy approach zero the marginal cost or the cost in general of many other things will approach zero they will become less and less visible or legible to GDP as a measure because things will disappear out of markets they will cease to be traded on markets they will just be they'll be too cheap to bother trading on markets and so they will therefore cease to be measured um can we imagine a scenario in which this happens to so many goods and services across the economy the GDP actually depends as a measure there's just there's less and less Dimension that's hypothesis one hypothesis two is are we seeing the beginning of this now what so that I guess another way to ask the another way to ask it is a research question instead of a hypothesis per se is what would the beginning of that uh transformation of GDP look like and is does that does what we would predict it to look like align with anything that we are seeing right now there are ways in which the global economy is confusing at the moment yeah let me so we're we're by most measures we're more productive we're producing more goods and services than ever but uh uh you know we're talking about recession we're talking about inflation we're talking about a deflation at the same time where there's a lot of chaos right now and if one were to imagine GDP sort of on the edge of beginning to turn negative despite the fact that productivity is increasing and increasing increasing well maybe it would look an awful lot like the confusion I see right now so anyway this is this is on my mind at the moment I don't know does that sound crazy or is that possible isn't this what uh I'm not sure if the incels are the correct label for the category of young men who have just disappeared into video games and are not interested in taking out mortgages or getting married or having kids or just can't acquire those things even if they want them but I mean there are large groups of people um in at least in China in the US I don't know about Australia um who basically are opting out of the usual economy to a large degree right they're not getting jobs or I mean it's a question of whether they can't get them or just choose to not pursue ordinary careers and mortgages and you know having a family and the degree to which they want those I mean do they want them and can't get them or
do they you know they not want them from the beginning but uh it does seem like this the number of hours that people spend playing video games in particular is maybe an indicator of a form of I mean it's a very cheap form of entertainment compared to well okay maybe not maybe compared to watching TV it's it's kind of comparable um yeah let me switch tracks a little bit and just comment on something Matt said in the chat before I come back to that perhaps uh he mentioned positional Goods I guess positional Goods often do show up in GDP figures indirectly right so you could think about uh often businesses like bars are set up to monetize seeking positional Goods so for example just to give one example from China you know it's it's the high status thing to buy a particular alcohol that's really expensive and then you know share it with your buddies so there's the economy is hooking into positional and Status oriented desires quite a lot I guess the the positional Goods themselves depending on the form they take may not show up in GDP figures but that seems to be true of a lot of things right so I think Adam what you mean is it's not it's not like everything that is valuable to us currently shows up in GDP figures but to a lesser or greater degree it may have proxies that show up um again there are a lot of things that are going to remain that are going to remain scarce and I and and I suppose there's there are behaviors that that construct artificial scarcity so the things that are I think genuinely physically scarce the property water property in Malibu and then there are sort of positional Goods that can be that can be socially constructed so you know there's perhaps no sense in which you know a certain bottle of alcohol would fundamentally need to be scarce but we may we may for social reasons construct an artificial scarcity because of its you know because of in-servancy social dynamics um another thing maybe to just inject into the conversation we've only got a couple minutes left here is that um again just to kind of get below the layer of abstraction measures you know things like GDP and markets and supply and demand and employment and and you know um uh money itself those are all abstractions if you look if you look beneath that at the actual product production productivity like how much materials how much energy are being moved around those are all growing and so GDP per capita is is is up GDP per capita is growing but the one thing that's quite astonishing to me is that
the the pandemic just showed us how few people actually are necessary to to contribute to whose contributions are really necessary to maintain that productivity wait I'm confused about that aren't the employment isn't the employment done the employment figures uh is an unemployment very low in the US at the moment or at least it was recently it is but it's only low on the official figures because so many people have stopped pursuing jobs oh I didn't realize that so there yeah so the people are just out of the market this is the incel uh for lack of a better word uh and that's not a very kind word but um there are people who and it's not just young men there are a lot of people who simply exited the job market and are no longer looking for jobs and as a result um they're not included in those unemployment figures so there's some gaming of those of those figures sorry just making a good point that yeah we shouldn't probably use this term I do mean something more like the the Japanese shut-ins rather than I guess in cells has a kind of misogynist meaning so yeah maybe yeah I mean just that I mean maybe also shut-ins is a kind of more specific movement I mean there's a whole thing in China where they say called laying flat uh so maybe maybe I want to call it laying flat there's a then let's add that to to several other groups of people who are not in in in wealthy modern societies who who are not actively um productive their consumer these consumers in the sense that they consume goods and services but they aren't actively participating in the production so for example for example um you know children don't contribute to the economy for a very long time now I mean some in some cases we basically remain children and don't really make contributions to the economy until we're in our early 20s there's huge numbers of people in the modern society like that they're more retirees than ever before um uh so it was so there's there's a you know young people don't work old people don't work and then what we saw during the pandemic is that a lot of people in bureaucracies are doing all that much productive work right this there's quite a bit of room in many bureaucracies for people to step out and of course this is the phenomenon of uh [ __ ] jobs which is at this point pretty well documented so what's fascinating is that GDP per capita has gone up continuously and exponentially for centuries and yet fewer people are working today in wealthy modern countries than any then