CSET AI Policy Insights: Talent, Compute, Data

foreign assistant director at cset today we'll discuss the factors critical to Future advancements in AI such as Talent compute and data and the relative importance of each but first a brief bit of housekeeping all attendees microphones are muted if you're on a computer and experience any technical issues use the chat function at the bottom of your screen to alert us and a cset team member will try to help you out please don't use the chat for anything else just yet and now it's time to turn the mic over to our moderator Tim Wong Tim currently serves as a senior technology fellow at The Institute for Progress his research focuses on metastimes comparative studies and emerging Technologies and Grand strategy in science and technology policy he previously served as a research fellow here at cset and enrolls at the Harvard MIT ethics and governance of AI initiative and Google Tim over to you thanks Danny um thanks everybody for joining very excited to host this discussion um I think one of the things that I'm sure many of you turning in will be familiar with is um I think the degree to which sort of policy discussions around AI often fall into cliches assumptions right so um you know I'll name a few um uh AI is an arms race right or um the biggest model is the best model um and I think today I'm particularly excited to um host uh this session with Micah because I think this session is going to focus on taking on I think one of the biggest I think sacred cows of the AI policy discussion uh that I think really has dominated discourse over the last few years um and that assumption is that computing power is the most important factor um in sort of AI advantage and progression of of AI as a technology um and so uh very excited to host this discussion uh Michael will speak for about um uh 20 minutes 25 minutes or so um and then we'll use the remainder of time for Q a um to quickly give an introduction to Micah Micah is a research analyst uh at georgetown's uh cset with a cyber AI project has done a bunch of super super interesting work in the past um both on the research that it's going to present today of course uh but also on a variety of questions uh including um the questions around the application of AI for example in the in the disinformation context um like it's formerly uh of uh Georgetown um and soon to be uh in law school at NYU uh coming up uh next next year or this year actually yeah um Michael over to you all right thanks trim um really fantastic introduction as Tim said I'm happy to be here I'm very excited about this so uh real quick I'm just gonna pull up some slides here um and I'll Dive Right In because I feel like we have a ton of material exciting material hopefully to present um so I've titled this presentation somewhat provocatively how important is compute to the future of AI challenging the conventional wisdom about constraints on AI progress and Tim has already done a great job teeing up the idea that um we are challenging some sort of conventional wisdom here but just before I dive in um I'm gonna in terms of the agenda here I'm gonna briefly talk about the current policy background the set of policies on AI that we've seen recently and why it sort of suggests that there is this strong belief in the value of compute as the primary driver of AI progress I'm going to talk about some of the bases for why we have uh that many people have that belief and I'm going to talk specifically about some Trends in Computing usage among notable AI models but with a focus on what some of the limitations of the existing knowledge that we have are um and then I'll dive in and I'll talk about some recent survey work that we did here at cset to try to evaluate this issue before closing up with sort of a teaser about the policy implications I'm hoping that we'll get more into some of those in Q a but I'll largely stay away from them for the sake of my my presentation itself before we Dive In just a little bit of of level setting in terms of how we talk about AI policy there's this very common notion that has become a really useful analytical frame for thinking about targets of AI policy and that is this concept called the AI Triad this was discussed by the former director of the Cyber AI project Ben Buchanan back when he was at cset in this paper where he summed up modern AI in one sentence which is that machine Learning Systems use computing power to execute algorithms that learn from data and the basic idea here is that there's really three distinct inputs to AI progress there's various types of data either larger data sets more curated data sets application specific data sets there's improvements in algorithms which in this context we mean in a very general sense to include things like the creation of new deep learning architectures the Transformer model for instance was discovered or invented depending on how you view it in 2017 and has been underlined most of the large language models that we've seen just in the past few years and then finally the third pillar of this column I guess is computing power or compute for short and that basically has to do with the hardware the semiconductors the specialized semiconductors oftentimes abbreviated gpus which were originally developed for gaming systems but have turned out to be very useful for AI systems as well and the idea here is that all three of these factors are somehow important for driving AI progress the differential value is debatable a lot of the time that's really what we're going to be talking about here today um but that these are three distinct places where policy makers can intervene one thing this kind of leaves out though is the talent question which I want to tee up because this will be a recurrent theme and so um it's possible to say I think from a from an AI researcher's perspective algorithms in particular is very tightly coupled with Talent you need talented researchers in order to develop new algorithms but there's also a broader sense in which Talent underlies all three of these um even data even if it's not researchers themselves labeling data it's oftentimes large large Enterprises of people distributed around the world who are doing data labeling at least for some AI cases and computing power progress in semiconductors re-shoring semiconductor capabilities all of this also depends on Talent um so with that out of the way I'm going to give a brief talk about basically where is the policy agenda on AI right now my there we go and as a framing device here I'm sure many people in the audience are familiar that the National Security Commission on AI released its final report in October of 2021. in that report they outlined about a dozen high-level strategic objectives that the U.S should pursue to maintain an AI advantage um and I've pulled out a few that were specifically related to actually being competitive in AI as opposed to um sort of ensuring that ai's benefits are distributed well there's a couple here I just wanted to briefly Target um so one high-level recommendation was that the U.S should build a resilient domestic base for Designing and fabricating microelectronics and you see that very directly translate into policy with the chips and science act last year providing 50 billion dollars in funding to reassure semiconductor capacity another high level goal was to protect America's technology advantages and the nscai specifically talked about this in terms of reconstructing or reinvigorating perhaps America's export control system to better uh better track and maintain an advantage on that Hardware side and again you see a direct translation into actual policy with the October 7 export controls on Advanced gpus last fall and then the third here thing here is that the nscai called for us to accelerate AI Innovation at home and specifically it called for the government to create a national AI research resource or abbreviated as a Nair um that could provide resources for researchers domestic researchers to further Advance Ai and in January a task force that was commissioned to study this proposal for an error released its final report they they very heavily emphasize and I want to be clear that it's extremely important for the Nair to provide both computing power and data to researchers so it hits both of those columns but there's this line in that task force that does say the largest Awards should be reserved for large Computing Investments with smaller caps defined for data and Service Awards so this isn't an exclusive focus on compute but this is still somewhat of a relative focus on compute compared to other levers and then finally another high level recommendation was to win the global Talent competition and I say this because um basically the point I want to make here is that all of the nscai's recommendations that are tangibly related to Hardware have pretty quickly been translated into actual policy but it's on the Global Talent competition question where we've seen maybe the beginnings of analyzes about domestic Workforce Development issues K-12 education but no real attempt to tackle things like increasing the H1B caps for talented immigrants or those sorts of issues so the the T up here is to say to this point in time a lot of this policy making has been really really focused on this compute lever and it's possible you might say well possibly that's just because compute is the most important lever and there is a an argument a strong argument for that um the strong argument for that is my next slide um I realize this slide is a bit busy but the idea is not to look at each of these six Graphics individually but instead to to view them sort of collectively each graphic here basically shows on the x-axis we have at time years and on the y-axis we have a measure of computing intensity and each dot represents some notable model from the history of AI and uh this first the first version here came from openai in 2019 they basically all of these graphs show the same thing which is that for most of the history of AI there's a growth in Computing demands of systems that roughly tracks Moore's law but then in 2012 there's a sudden Kink and all of a sudden these these models start taking on massive amounts of computational requirements um I I show this with six different Graphics not because the differences matter but because basically uh this was first noticed and opened by open AI in 2019 and then uh significantly expanded in a really great data set compiled by some other researchers in 2022 the link at the top of the screen will take you to their data which they've done a fantastic job of compiling and making public um but all each of these Graphics is sort of taken from a different AI policy group so there's open Ai and AI lab themselves the second one is from a report that I myself co-authored about a year ago here at cset another one comes from the Stanford Hai AI index one comes from the center for the governance of AI another comes from the state of AI report and the idea here is that it's a really compelling graphic it shows that like the last decade of notable AI advances have really escalated how much Computing requirements we're using and it does tell you something really really meaningful about how important compute was over the last decade but and this is now where I'm going to start shifting I think that there's a risk that policy makers are over indexing on this type of result for two reasons the first one is that in this report I I authored last year we took the trend line as it had been analyzed at that time um and we said what happens if we just project this forward in time and we notice that very very quickly because it was such a steep exponential curve it becomes infeasible to keep scaling at this rate in fact we noticed that as early as 2026 we should be seen based on just extrapolating this trend line AI models that cost the entire U.S GDP just to train and I'll note that this trend line is assuming that prices of the cost of compute halves every two years consistent with Moore's Law and so there's a risk here of saying basically what we argued in this paper was we said um massively scaling up compute really drove a ton of the last decade of AI progress but it probably can't con it almost certainly actually can't continue at the same rate which means one of two things either we will see a general slowdown in AI advances or people are going to have to shift strategies either find more efficient algorithms or more specialized versions of these large models something might have to shift even if just on the margins to make this trend line not so unsustainable the other risk with some of this data um these this data showing the skyrocketing AI compute usage is that it's unclear how reflective it is of all AI work versus just a few models and so what I did on this slide is I took that that data set I mentioned that had been compiled by some folks at the center for the governance of AI and they had about 160 models where they had information about the model's compute usage so I took that those uh those models and I broke them out into different tiers and I said within which type of tier within each tier sorry which type of models are represented what you see here is that the most compute intensive models that we're talking about are to a large extent almost overwhelmingly actually language models um so a lot of the recent trend of this explosion and compute needs is being very very heavily driven by language models specifically um it's unclear though it's not the case that uh two-thirds three-quarters of AI researchers right now are just doing language modeling and so it's unclear what these Trends say about AI as a whole as opposed to a specific subset of one AI application area and we noticed this about a year and a half ago and we thought hey it would be really great to study this in more depths how do we go about doing it to answer that we decided that the best way to to to study this would be to run a survey a survey of AI researchers and so last summer we fielded this survey we had two target populations most of our our respondents came from basically anyone it says the slide says academics but really this does include some people at industry anyone who had published uh in a top AI conference between 2016 and 2021 and using some of cset's data we extracted about 27 000 emails associated with someone who had met that criteria and then we also to get um because that that population includes a meaningful number of industry but it is Academia dominated so we also augmented this with some sampling of Industry People based on the the LinkedIn criteria that you can see on the screen uh when we fielded this survey we got a total of 533 total results um I believe that there's been about 10 attempts or so to field surveys of AI researchers as a population and to my knowledge this is the second largest survey of that population conducted yet so we were pretty pleased with that result you can see the breakdown on the right hand side of the screen between Academia and Industry and the other thing I'll say is that based on the domains of our respondents from Academia we we mapped each University to its Qs world ranking and we have a basically an even division three ways between researchers at a top 50 University researchers at a University ranked 51 to 200 and researchers at a lower ranked University and then also a small percentage that we're affiliated with an unranked University which is why those numbers don't add up to 100. um this did I think give us a lot of a broad reach across many types of academics um it's not totally representative but it's it's pretty representative it's a much broader reach of academics than maybe policy makers are ordinarily exposed to so we fielded the survey the goal was to evaluate for each AI researcher how important is compute especially as compared to other factors as part of their research agenda and so one thing we asked researchers was if we doubled the budget for your current project what would you spend the extra money on and what we saw is that over half of our respondents said that they would spend that money on hiring either more Engineers or more researchers and only a fifth said that they would spend it on purchasing more compute and so just briefly I think that immediately this sort of raises questions about well okay there's certainly a subset of models that is exploding in compute requirements and where it's getting more and more costly to purchase enough compute to to train for instance a gpt3 a gpt4 but um this is a first indication that that might not actually be the types of projects that most researchers are working on and we also see this in other ways so we asked our researchers as well how often have you rejected a project revised a project or abandoned a project because of a lack of either compute data resources or researcher availability and again we found this result that sort of surprised us which is that in each of these three hypotheticals compute was the least cited reason for why academics are forced to change their research Plans by pretty substantial margins especially for rejected and abandoned and so again this sort of suggests that there's again a subset of models where compute is really important but focusing just on compute as a policy lever is maybe missing the needs of a potentially even broader population of AI researchers we also asked our researchers we gave them five different factors things like more compute better algorithms and so on and we asked them how strongly do you agree that this Factor drove the last decade of AI progress and then we also asked them how strongly do you agree that this Factor will drive the next decade of AI progress and what we saw is that um of all the factors compute scored the highest when researchers thought about the last decade that is to say 59 of researchers said yes I strongly agree that compute heavily drove the last decade of AI progress but it was actually the the uh Factor where the lowest percentage I strongly agreed that it would drive the next decade of AI progress and at the same time the percent saying that better algorithms would become more important Rose pretty dramatically there are a number of ways you could interpret this result for the sake of time I'm not going to dwell on it although I'm happy to come back in Q a um but again it's a sort of suggestion that hey relying on these historical Trends is maybe not the best way for conceptualizing what the needs of AI are moving forward in time um and so I'm gonna pause with with this or sorry I'm going to end with this slide here three high-level questions they're not yet drilling down into specific policies um but first of all I think one thing I I want to highlight here is that oftentimes it's just sort of said well deep learning is compute intensive and that is true but there's a lot of variation between different subfields of AI including different subfields that use deep learning and a lot of variation within subfields how compute intensive different research projects are and so when we talk about policies that impact compute we're not always talking about policies that impact all types of AI research we need to be careful about thinking about which types of applications were most heavily impacting especially because again as I mentioned language models are in some ways a unique AI application that seems uniquely compute intensive at least at this point in time it's possible that other fields will attention will pivot and other fields will become more compute intensive but right now when we talk about the really really compute intensive models we're primarily talking about language models the second high-level question here is oftentimes people also talk about compute divides between industry and Academia the inability of Academia to keep up with industry and again I think that there's an ambiguity here because people aren't always clear um are we talking about a systematic divide where industry as a whole demographic systematically has access to more compute than Academia or are we worried about the influence of a small number of big tech companies that are particularly well resourced when it comes to compute because these might actually translate into very different policy interventions which again I I'm running out of time so I'll just sort of tease that and then finally if it's the case that it's an intrinsic property of all types of deep learning research that they require tons and tons of compute then oftentimes it makes sense to think well okay then the way to get AI progress is that we just have to keep throwing more and more compute at these projects as long as we can but if it's instead the case that there's a lot of variation in what projects researchers might take on um throwing compute at the problems will end up benefiting particularly compute intensive research plans while not so much helping others and I think we ought to start asking ourselves is that actually the outcome we want instead of just assuming that everything needs more compute we maybe need to get a little more careful about asking is that do we actually want to encourage compute intensive projects or do we want to encourage the type of research that might generate more efficient ways of doing the same tasks I realize I'm leaving at pretty high level here but I'm about 21 minutes into this talk and so the way I want to the the real message that I want to end on again I open by saying AI policy is increasingly focused on this compute lever oftentimes even maybe to the um oftentimes as a substitution away from other levers we might be pooling like Talent focused ones um and I think that this is just not the right approach it's good to a certain extent it's many it's very valuable to pursue a lot of these compute focused interventions but they are not a substitution for talent focused policy um Talent policy includes lots of domestic efforts things like uh Workforce Development uh leveraging the The Talented community colleges uh K-12 education these are all topics cset has written on but it's also those are all largely medium-term interventions in the short term the shortest way to address this Talent Gap is by welcoming more immigrants to the U.S more High talented immigrants and that I think uh it's oftentimes gotten too easy to imagine that well if all AI driven I apologize stumbling over my words here a bit in the home stretch it's gotten a bit easy to say if all AI progress is just driven by compute then we don't need to think as much about Talent OR about immigration and I think that that is very much the wrong way to approach AI policy these are all levers we need to think about and we cannot use one as a substitute for handling the others so on that note I'm going to stop my my screen sharing and Tim uh jump in and and start peppering me with questions definitely uh thanks Micah that was great um so uh if you do want to get a q a or a question rather into the discussion um what you need to do is just drop it in the webinar chat um our sort of events team will be able to be collating those questions as Micah and I are talking um and I will bring them into the discussion but I guess might get to start maybe get a to get us rolling while people are kind of thinking about what they might want to ask I think you know obviously I think the position that you've put out you recognize it as such is is a little bit provocative and I think to get the discussion going I I do want to kind of take the other side a little bit you know pressure some parts of this argument think about where it works where it doesn't work where um you know uh and ultimately I think what it means for policy I think the first question I'd ask you is you know I'm kind of putting the Hat on of being like a large language model booster right where we say like it's all very well and good to say that AI is very Broad and there's lots of people doing things that don't involve any compute but look like the hugest most important breakthrough in AI um in the last few years has been language models and to the extent that we are worried about sort of the US's um advantage of the technology we need to be working on the most important thing right it's a bet it's clearly paying off and it's the most critical and so you know what do you say to someone who I think kind of takes a look at your results and says yeah this is all very well and good but from a national interest standpoint the thing that we really need to be kind of you know accelerating on right now is precisely the thing that is you know uh compute intensive yeah so um it's a it's a fantastic question I mean that's the immediate and um obvious pushback is to say well okay um sure maybe language models are the most compute intensive but they're also the most important so of course we just want to focus more there um and I suppose there are there's there's two ways I could respond to that one way is to question just how strategically valuable language models are um and I think that frankly I don't want to get down too far in that rabbit hole but uh there are reasonable questions about so for instance this arms race frame uh it's very easy to say oh well you know language models could be economically transformative in the U.S uh we're all very worried about them they're in the news we're constantly paying attention therefore it must be the case that China also views this as a critical technology and they're racing as quickly as possible to get there and I would say well first of all um the Chinese State very quickly moved to release regulations that would pretty heavily dampen the ability of companies to deploy these at least in the way that they're being deployed in the U.S and so I think that there's a sort of question of like what if the what if the Chinese government doesn't actually view this with the same level of strategic importance um and you can play that out more you can say well China's actually much more of a manufacturing economy so to the extent that there is an economic payoff from something like language models it probably disproportionately benefits service economies like the US and less so manufacturing economies like China um and so there's basically where I'm going there is just you might question okay sure language model is very important um is it the case that China also sees it that way or are they focused on other types of AI research that perhaps aren't as compute intensive that alters the National Security calculus a little bit the other response I could make is that on the domestic front in terms of which type of which type of Technology do we want to encourage you could certainly say well language models are just the most important that's what we ought to be encouraging and the response I would give there is just that that's a little abnormal when it comes to that we fund science research which is that usually we try not to prejudge which types of basic research are the most valuable we Instead try to give out resources to academics pretty broadly across many many different types of applications in the hopes that eventually one of those will sort of demonstrate utility in such a way that industry picks it up and expands it at scales that Academia can't um and so from that perspective I would say like yeah you can certainly argue that like no we have enough information to conclude language models are the one that matters that's where all the government funding should be going into resources that help language models um but my pushback is just to say that usually that's not how we do science funding in the U.S and in some ways that's actually a pretty significant break from the normal process yes definitely so I see these questions that are coming in they're they're really really great and I think I would encourage you to drop your uh hat in the ring uh if you want Micah to kind of consider one of them um I think as people kind of come in I think it does make sense maybe to build on one audience question here uh so Tom dieterich uh hey Tom I think we've met many years ago it's good to hear from you um I think has a really interesting comment about the different time scales of the policy levers here right so I think the point that he made in his question was look there's a lot of things we can do for compute that will impact the Strategic environment right now right we can export control that totally changes the nature of the industry overnight whereas while we can all agree with Talent you know it's an important feature we shouldn't see this as and I don't think you're arguing that they are mutually exclusive but you know if we're worried about say let's just put it out there competitiveness with China right are we worried that Talent is maybe too slow of a lever here right and so maybe one response is to say look you know I agree with you talent's really important but compute is ultimately the thing that we need in this kind of critical moment where we want advantage and it's the tool that will get us the most impact right away um do you buy that I'm curious about how do you respond to that so um the one the one response I would give is that uh it is the case that ideally these these operate as complementary policies um I have however had multiple conversations with people who are involved in policy making and I certainly won't name names here but but conversations where I've presented the results of of this survey and I've and they've said like you know what's the one sentence summary and I've said something to the effect of like compute is great but maybe we should be spending more attention on Talent especially immigration as a short-term intervention and they've said in response oh I can't tell my boss that um and so I've I've started to wonder like in an Ideal World these are complementary but I also think that there's actually a risk that uh given political constraints we start to view them as substitutionary like oh you know great AI progress depends on compute that means I don't have to touch immigration um and I don't have to tell anyone that we need more immigrants and in a very you know heightened political environment where that is a sort of dangerous message to say just to send you get a risk there are clear incentives I think to sort of over index on the value of compute despite its its importance and its relevance and the other the other response that I would give um in terms of yeah computers may be a very short-term intervention um pay 12 education is relatively long-term Talent uh immigration is much shorter term but it's still relatively I mean you're talking a few years here um the answer to that is that I'm not convinced that we are in a crisis moment right now I think that chat gbt has certainly galvanized a ton of attention but uh I don't it's not the type of Technology where you know um chat gbt gives some sort of decisive strategic advantage to the first military that starts to ask the chat bot uh which types of military operations they should do that is not really a sort of risk that I'm concerned about I think that the payoffs have much more to do with economic automation economic improvements and in many ways it's actually fortunate that uh large language models which are somewhat more indirect of a national security risk were the type of model to generate this breakthrough the the caveat to that is that I acknowledge a lot of the disinformation risks I have personally done a lot of of research on the use of AI for disinformation and for influence operations and I'll say there that based on my assessment I'm happy to talk more I think that right now people are more in danger of over inflating the risk than they are of under inflating it but happy to to either address that in a different question or I know we're getting a lot of questions so I want to move on yeah no definitely that's all great so one other final question using My Prerogative and I think it's related to the question that Derek Slater is asking here in the chat um I am sort of really interested in not just thinking about these two things that is to say compute policies and levers and talent levers as being complementary but also in some ways like interdependent in some ways so I think there's one way of looking at it which is maybe another slice which is Micah wouldn't you agree that um you know being able to access these incredibly rare powerful Computing clusters is one reason why talented researchers might move to the US right there's also these really interesting questions about whether or not policy around compute and I'm now talking about say like Nair style approaches um is important for leveling the playing field right you could say okay well maybe immigrants and talent they really want to move to the country if they think they can do a AI startup that could really go compete with the best of them right open AI or what have you and so what we need is is you know say public research clouds to be able to kind of like open up the space for even to be worth it for us to be able to leverage talent in the first place and I guess I'm kind of curious again I'm I'm not really pressuring the argument but I am kind of want to hear sort of your view on how these two are almost kind of related variables in some ways and that there are almost ways of doing compute policy which are kind of talent policy in some sense so first level answer I would say yes absolutely um although in order for that to pay off you actually have to have a high enough cap on skilled immigrants that people can respond to that signal and decide to come to the US um but on this on this point about the The Nair sort of equalizing the playing field uh one other interesting really fascinating result from our survey I'll talk about briefly here um we do see that like AI projects have tons and tons of variation in how much compute different projects need um that is I think you know our survey supports that that's abundantly true across basically any way you try to look at it um and so there's a question of like why is there so much variation and broadly I would say there's two ways you might go go about answering that one is to say well it's access that governs how much compute people use I.E everyone would like to be doing really compute intensive work but there's tons of stratification and how much access people have and so many researchers are just forced to do other types of work the other broad types of explanation you might offer has to do with self-selection effects which is to say look different subfields different methods within subfields have a lot of variance in how much payoff you get for expanding compute and so what happens here is that researchers select into one type of of research maybe that initial selection is governed by access but then once they're in a like low compute research project it doesn't actually help them to just throw more compute at their problem um and so an interesting thing that we observed in this survey is that uh actually The respondents Who expressed the most concern about not having enough compute to do meaningful work in the future where the same respondents who reported using the most compute right now um and we got that result actually across a number of indicators so like the more compute you use the more important you thought compute was in the past the more important you think compute will be in the future the more likely you'd be to spend more money on compute Etc and this I think suggests that like people have been assuming something of the the access explanation which is like oh if we just give more compute to everyone people will use it and that will reduce the inequalities in how much compute different researchers use it might actually be the case that the reverse is somewhat true which is that the people who are most motivated to make use of new compute resources are the people already using a lot of compute um and so in that case if you just give everyone access to more compute what you actually do is you increase the stratification between how much compute different projects use maybe that's worth it maybe the more compute intensive projects as we talked about are just more valuable and so uh it's worth doing that um but I do think like this is a sort of call to the research community I think that we've been way too simplistic about just saying like is compute important isn't it compute it isn't it important I'm much more interested in asking like who benefits differentially from compute focused interventions because that I think gets us into a frame of talking about like hey certain ways of implementing these um these types of policies do create relative winners and relative losers and we actually want to be certain that that's what we are doing before we try to pursue that policy or sorry we want to be certain that that's what we want to do before we go ahead and we do it yeah no I think that makes a lot of sense um and I guess I don't know to return to I think a result that you're talking about a little bit earlier like sort of like like it or not right like we're gonna hit a ceiling on compute because people things just can't grow at the rate that they have been right um and so you know you're almost forced into a situation at some point which is okay except that we do all the compute stuff we dominate compute you know your projections would say well we just expect that things are not going to grow the way they have been like 36 months from now essentially um which I think is a really uh interesting and an important result um one kind of question maybe that you can maybe speak to now um uh Quentin um in the chat had a quick question which is now kind of flipping the analysis kind of like on the other side of the Triad to basically say okay we've been talking a lot about compute we've been talking a lot about Talent um I guess Michael do you want to sound off for a second on your views about data as a lever in the space um right because that's the that's the kind of last bit of the triangle you've not talked about yeah yeah I'm curious about like if you feel similarly around some of these issues right which is like maybe we're too obsessed with data or maybe we're not obsessed with data enough like interesting kind of just having you maybe talk a little bit about that yeah so there was certainly there was a moment maybe three or four years ago where Big Data was the solution to AI progress that we would just make data sets bigger um and uh I I think in some sense like the reason that things are framed so much in terms of compute in this presentation is because we've clearly moved on from that um at least within the policy circles that I happen to be able to to observe we're no longer talking about Big Data we're talking about controlling access to Hardware um but but data is the third leg here and so uh I do actually think data becomes especially relevant not just when we're talking about like self-super self-supervised large-scale language models but whenever we want to talk about like what are the socially valuable applications of AI that we really want oftentimes the constraint ends up being not compute which may be necessary to to train some sort of large Foundation model but in the last stage of actually getting true economic utility out of that you need a meaningful enough data set that you can fine tune that model on some socially useful purpose and a great example here is like you might have a great vision classifier right that can can work with all 1000 categories from imagenet um that is really meaningful as a researcher uh a research breakthrough if you improve that although imagenet has been saturated you know Etc um but to translate that into something like a medical diagnostic tool those thousand categories from the imagenet data set are not the right thousand categories you do need some sort of well-curated additional data and so I would say this is probably like in terms of government interventions uh that we could do to to boost research progress I do think a broad answer is to say well if there's ambiguity about what types of resources are most valuable then perhaps the best way to provide it is in the form of Grants and let researchers themselves decide how to use that money and the exception to that is well curated data sets that only the government can put together so that's that's the exception where I think I see the most utility of like a concerted whole of government effort to boost AI is to coordinate between agencies gather all the meaningful data set all the all the meaningful data sets create a system that allows structured and tiered access to it so that different research projects can work with different data sets of different levels of sensitivity and then you know basically see what happens see what kind of applications we get out of that that can be translated into having a lot of economic utility that's great yeah thank you um so for the next question I'm going to join a couple of comments and questions that have come into the chat um largely kind of drawing on rohini's comments and Tom's recent comment I think it's almost a question about how you account for the difference here right so clearly we have a lot of policy discussion in DC which is very compute focused um and uh and at the same time we have your survey which talks to a bunch of researchers and says actually we don't care that much about compute and I'm kind of curious about how you account for the difference right I think implicit in what you're arguing is well maybe the policy makers in DC have it wrong or maybe they're listening to the wrong people we should be listening to the researchers I think is maybe a subtext of what you're arguing and I think some people would say well like a look like your response rate was super low on your survey how can we trust that this is actually giving us good signal on where we we should be moving yeah and I guess I'm curious about how you reconcile like I think it's a two-part question one of the questions is like how do we end up in this place where compute is such a focus of the policy discussion and the second one would be how would you respond to people who say look this survey is just not very representative like you're getting just some weird noise and the people who happen to respond and maybe the people who are responding are so far you know in the wilderness in this space that you know they had lots of incentive to argue that the current kind of um you know Orthodoxy is wrong so I'll respond to the um the question about the survey we've gotten this this um this comment about well the response rate was too low uh a couple of times um and it's I it's worth addressing although there are there are reasons that the response rate is low that I can tell you which is for instance like I said we used DC sets data to pull tons of emails of anyone who had published in one of these top AI venues um we knew going into it that like we were confident that each of those emails was associated with someone who had done meaningful AI work at some point and we also asked a screener at the start of the survey to confirm that the respondent works on AI systems but we couldn't ensure that for instance like that's an email that is still actively being monitored um and so we went into this knowing that a lot of the emails we were going to send this to weren't there there was just no world in which people would respond um and so I think that that artificially makes the response rate seem rather low what matters much more than the response rate is the raw count of response that we got um and again I mentioned this in the in the presentation itself I think that to my knowledge this is the second largest survey that's ever been conducted of AI researchers I'm very pleased with that um we also have a meaningful you know I mentioned that we have basically equal representation across three different tiers of Academia um because of privacy issues we didn't ask a ton of personal identifying information that would allow us to exhaustively check the representativeness so there's certainly ambiguity but we're pretty confident in the main results of the survey to the other question about like why is this Gap existing one potential explanation I might give you is going back to this result that the people who use the most compute are the most concerned about compute and I think that there's a possible explanation you can give here which admittedly is speculative so I would I'm not gonna like claim that this is true but I think it's reasonably likely that the academics who happen to have the most access to compute work at relatively Elite institutions and are disproportionately likely to have connections in DC which means like to the extent that there's um a directionality of the bias I would expect policy makers in DC to be hearing more about compute than what the median AI researcher says again you might respond and just say well researchers at Elite institutions know better what type of AI research is valuable so who cares if they have a sort of bias perspective that's a totally fair uh response that someone might make but I do think that like you can see how you might get these gaps in communication creep up just based on the results of the survey that we did collect yeah that's that's a great response and I think building a little bit on that I'm going to go to a question that I think Duncan you asked quite early in the session sorry my cat um uh you asked quite early in the session um which is I think one hypothesis you can draw about why DC is most interested in compute is kind of that it's the most controllable thing in some ways right where you're like we can pull the export control lever and things can just happen and so you're kind of like well you you use the levers that you're given and that are most accessible and you sort of grab at them um and I think there has been maybe a notion that this is a particularly good lever because the greatest sort of like threat and danger comes from these big powerful models right and you can read that however you want right like you can read that in the the x-risk sense you can read that in the sense of the work that you've done in the past right which is these models seem to enable a certain kind of disinformation that we might want to be concerned about um and I guess I'm kind of curious about like if you you know that is zooming out a little bit from the point of view of competitiveness and policy to talking more about like managing risk but you you think that assumption will continue to hold right which is basically that like powerful AI systems that need to be built in large Computing clusters are where most of the risk is and that's where we need to be focusing most of our time or is that kind of historically provisional or is it not even true right now right like I'm kind of curious about how you kind of weigh those threats and you know as we kind of think about like why we might want to be focused on certain levers over others yeah absolutely so first of all I'm going to use this as my uh my point to explicitly plug the survey and say you know there's a ton of discussion in there there's a ton of results we I just don't have time for in this context and so um to this point from Duncan we start out the conclusion by saying look there's three broad reasons why compute has become really influential as a as a an anchoring an anchor basically one is the belief that it's critical to AI breakthroughs a second is the belief that there's really stratified access between different demographics and a third is like hey it's easy it's physical we can look at it we can hold it we can export control it we can do these sorts of things that it's much harder to do for data in the abstract um and we're very clear like the to the extent that we think this the survey does say things about points one and two in that that breakdown we're not saying that like somehow these survey results mean that uh the political barriers to Talent disappear um those are still very real and Rel like this is an argument why you might Focus some level of compute of attention on compute even disproportionate to the potential impact it has all well and good I'll reiterate I don't think it's a substitution for talent policy we need to also think about that but the other point is um so the other the other thing you raised is like is it actually the case that these largest scale models are the most dangerous um and in an x-risk sense I mean you might hypothetically say if we keep scaling them up we get totally unforeseen difficulties totally unforeseen capabilities that are very very worrisome um there's another way of looking at it though if we think about other types of risks like the disinformation risk I think two other pieces of recent evidence that support the narrative I'm giving you is that one we had Sam Altman say a few weeks ago uh just scaling the parameter count of these models is not a feasible Way Forward for open AI that's slightly different than saying not scale and compute is not a feasible Way Forward because there's a difference there but the other piece of evidence was this leaked um Google internal memo that was circulated the we have no moat argument um and the argument that was being Advanced there was basically like look we can train these giant Advanced language models um they can be super versatile super generalizable they maybe they have 100 billion 500 billion a trillion parameters if someone can take a 20 billion parameter and with a bit of fine tuning replicate the capability on any discrete use task of this generalizable model the economics of training those massive models gets totally blown out the window um and with the disinformation stuff one of the things that I'm working on is some cost modeling that I do things suggests for a lot of disinformation operators it might actually the preferred method of producing disinformation might be to use smaller models that are easier to run and where you can fine tune them yourselves instead of trying to access something like gpt4 behind a paywall um and so in that sense I would say I guess the the succinct way of summing that all up is um there may be these risks that are associated with the largest models but it may also be the case that other types of risks are diffusing to other types of models and so simply locking down the largest models is perhaps not a viable way to fully counteract those types of risks right um perfect answer and I think there's a lot more we can go into there um so we are essentially at time um I'd advise everybody to read the paper as Michael mentioned there's a lot more that he didn't have a chance to get into but um yeah please reach out to him with any questions and at this point I'll turn it back to Danny to close us out awesome thanks Tim for moderating today and thanks to Micah for a fascinating and provocative discussion thanks also to all of you for taking part in this discussion and for your comments and questions we're sorry we couldn't get to them all if you'd like to learn more about cset please go to cset.georgetown.edu and sign up for our newsletter and research updates we'll be in touch soon with details on our next webinar so be sure to watch your inbox until then thank you for joining us and we hope to see you again soon have a good day thank you

foreign assistant director at cset today we&#39;ll discuss the factors critical to Future advancements in AI such as Talent compute and data and the relative importance of each but first a brief bit of housekeeping all attendees microphones are muted if you&#39;re on a computer and experience any technical issues use the chat function at the bottom of your screen to alert us and a cset team member will try to help you out please don&#39;t use the chat for anything else just yet and now it&#39;s time to turn the mic over to our moderator Tim Wong Tim currently serves as a senior technology fellow at The Institute for Progress his research focuses on metastimes comparative studies and emerging Technologies and Grand strategy in science and technology policy he previously served as a research fellow here at cset and enrolls at the Harvard MIT ethics and governance of AI initiative and Google Tim over to you thanks Danny um thanks everybody for joining very excited to host this discussion um I think one of the things that I&#39;m sure many of you turning in will be familiar with is um I think the degree to which sort of policy discussions around AI often fall into cliches assumptions right so um you know I&#39;ll name a few um uh AI is an arms race right or um the biggest model is the best model um and I think today I&#39;m particularly excited to um host uh this session with Micah because I think this session is going to focus on taking on I think one of the biggest I think sacred cows of the AI policy discussion uh that I think really has dominated discourse over the last few years um and that assumption is that computing power is the most important factor um in sort of AI advantage and progression of of AI as a technology um and so uh very excited to host this discussion uh Michael will speak for about um uh 20 minutes 25 minutes or so um and then we&#39;ll use the remainder of time for Q a um to quickly give an introduction to Micah Micah is a research analyst uh at georgetown&#39;s uh cset with a cyber AI project has done a bunch of super super interesting work in the past um both on the research that it&#39;s going to present today of course uh but also on a variety of questions uh including um the questions around the application of AI for example in the in the disinformation context um like it&#39;s formerly uh of uh Georgetown um and soon to be uh in law school at NYU uh coming up uh next next year or this year actually yeah um Michael over to you all right thanks trim um really fantastic introduction as Tim said I&#39;m happy to be here I&#39;m very excited about this so uh real quick I&#39;m just gonna pull up some slides here um and I&#39;ll Dive Right In because I feel like we have a ton of material exciting material hopefully to present um so I&#39;ve titled this presentation somewhat provocatively how important is compute to the future of AI challenging the conventional wisdom about constraints on AI progress and Tim has already done a great job teeing up the idea that um we are challenging some sort of conventional wisdom here but just before I dive in um I&#39;m gonna in terms of the agenda here I&#39;m gonna briefly talk about the current policy background the set of policies on AI that we&#39;ve seen recently and why it sort of suggests that there is this strong belief in the value of compute as the primary driver of AI progress I&#39;m going to talk about some of the bases for why we have uh that many people have that belief and I&#39;m going to talk specifically about some Trends in Computing usage among notable AI models but with a focus on what some of the limitations of the existing knowledge that we have are um and then I&#39;ll dive in and I&#39;ll talk about some recent survey work that we did here at cset to try to evaluate this issue before closing up with sort of a teaser about the policy implications I&#39;m hoping that we&#39;ll get more into some of those in Q a but I&#39;ll largely stay away from them for the sake of my my presentation itself before we Dive In just a little bit of of level setting in terms of how we talk about AI policy there&#39;s this very common notion that has become a really useful analytical frame for thinking about targets of AI policy and that is this concept called the AI Triad this was discussed by the former director of the Cyber AI project Ben Buchanan back when he was at cset in this paper where he summed up modern AI in one sentence which is that machine Learning Systems use computing power to execute algorithms that learn from data and the basic idea here is that there&#39;s really three distinct inputs to AI progress there&#39;s various types of data either larger data sets more curated data sets application specific data sets there&#39;s improvements in algorithms which in this context we mean in a very general sense to include things like the creation of new deep learning architectures the Transformer model for instance was discovered or invented depending on how you view it in 2017 and has been underlined most of the large language models that we&#39;ve seen just in the past few years and then finally the third pillar of this column I guess is computing power or compute for short and that basically has to do with the hardware the semiconductors the specialized semiconductors oftentimes abbreviated gpus which were originally developed for gaming systems but have turned out to be very useful for AI systems as well and the idea here is that all three of these factors are somehow important for driving AI progress the differential value is debatable a lot of the time that&#39;s really what we&#39;re going to be talking about here today um but that these are three distinct places where policy makers can intervene one thing this kind of leaves out though is the talent question which I want to tee up because this will be a recurrent theme and so um it&#39;s possible to say I think from a from an AI researcher&#39;s perspective algorithms in particular is very tightly coupled with Talent you need talented researchers in order to develop new algorithms but there&#39;s also a broader sense in which Talent underlies all three of these um even data even if it&#39;s not researchers themselves labeling data it&#39;s oftentimes large large Enterprises of people distributed around the world who are doing data labeling at least for some AI cases and computing power progress in semiconductors re-shoring semiconductor capabilities all of this also depends on Talent um so with that out of the way I&#39;m going to give a brief talk about basically where is the policy agenda on AI right now my there we go and as a framing device here I&#39;m sure many people in the audience are familiar that the National Security Commission on AI released its final report in October of 2021. in that report they outlined about a dozen high-level strategic objectives that the U.S should pursue to maintain an AI advantage um and I&#39;ve pulled out a few that were specifically related to actually being competitive in AI as opposed to um sort of ensuring that ai&#39;s benefits are distributed well there&#39;s a couple here I just wanted to briefly Target um so one high-level recommendation was that the U.S should build a resilient domestic base for Designing and fabricating microelectronics and you see that very directly translate into policy with the chips and science act last year providing 50 billion dollars in funding to reassure semiconductor capacity another high level goal was to protect America&#39;s technology advantages and the nscai specifically talked about this in terms of reconstructing or reinvigorating perhaps America&#39;s export control system to better uh better track and maintain an advantage on that Hardware side and again you see a direct translation into actual policy with the October 7 export controls on Advanced gpus last fall and then the third here thing here is that the nscai called for us to accelerate AI Innovation at home and specifically it called for the government to create a national AI research resource or abbreviated as a Nair um that could provide resources for researchers domestic researchers to further Advance Ai and in January a task force that was commissioned to study this proposal for an error released its final report they they very heavily emphasize and I want to be clear that it&#39;s extremely important for the Nair to provide both computing power and data to researchers so it hits both of those columns but there&#39;s this line in that task force that does say the largest Awards should be reserved for large Computing Investments with smaller caps defined for data and Service Awards so this isn&#39;t an exclusive focus on compute but this is still somewhat of a relative focus on compute compared to other levers and then finally another high level recommendation was to win the global Talent competition and I say this because um basically the point I want to make here is that all of the nscai&#39;s recommendations that are tangibly related to Hardware have pretty quickly been translated into actual policy but it&#39;s on the Global Talent competition question where we&#39;ve seen maybe the beginnings of analyzes about domestic Workforce Development issues K-12 education but no real attempt to tackle things like increasing the H1B caps for talented immigrants or those sorts of issues so the the T up here is to say to this point in time a lot of this policy making has been really really focused on this compute lever and it&#39;s possible you might say well possibly that&#39;s just because compute is the most important lever and there is a an argument a strong argument for that um the strong argument for that is my next slide um I realize this slide is a bit busy but the idea is not to look at each of these six Graphics individually but instead to to view them sort of collectively each graphic here basically shows on the x-axis we have at time years and on the y-axis we have a measure of computing intensity and each dot represents some notable model from the history of AI and uh this first the first version here came from openai in 2019 they basically all of these graphs show the same thing which is that for most of the history of AI there&#39;s a growth in Computing demands of systems that roughly tracks Moore&#39;s law but then in 2012 there&#39;s a sudden Kink and all of a sudden these these models start taking on massive amounts of computational requirements um I I show this with six different Graphics not because the differences matter but because basically uh this was first noticed and opened by open AI in 2019 and then uh significantly expanded in a really great data set compiled by some other researchers in 2022 the link at the top of the screen will take you to their data which they&#39;ve done a fantastic job of compiling and making public um but all each of these Graphics is sort of taken from a different AI policy group so there&#39;s open Ai and AI lab themselves the second one is from a report that I myself co-authored about a year ago here at cset another one comes from the Stanford Hai AI index one comes from the center for the governance of AI another comes from the state of AI report and the idea here is that it&#39;s a really compelling graphic it shows that like the last decade of notable AI advances have really escalated how much Computing requirements we&#39;re using and it does tell you something really really meaningful about how important compute was over the last decade but and this is now where I&#39;m going to start shifting I think that there&#39;s a risk that policy makers are over indexing on this type of result for two reasons the first one is that in this report I I authored last year we took the trend line as it had been analyzed at that time um and we said what happens if we just project this forward in time and we notice that very very quickly because it was such a steep exponential curve it becomes infeasible to keep scaling at this rate in fact we noticed that as early as 2026 we should be seen based on just extrapolating this trend line AI models that cost the entire U.S GDP just to train and I&#39;ll note that this trend line is assuming that prices of the cost of compute halves every two years consistent with Moore&#39;s Law and so there&#39;s a risk here of saying basically what we argued in this paper was we said um massively scaling up compute really drove a ton of the last decade of AI progress but it probably can&#39;t con it almost certainly actually can&#39;t continue at the same rate which means one of two things either we will see a general slowdown in AI advances or people are going to have to shift strategies either find more efficient algorithms or more specialized versions of these large models something might have to shift even if just on the margins to make this trend line not so unsustainable the other risk with some of this data um these this data showing the skyrocketing AI compute usage is that it&#39;s unclear how reflective it is of all AI work versus just a few models and so what I did on this slide is I took that that data set I mentioned that had been compiled by some folks at the center for the governance of AI and they had about 160 models where they had information about the model&#39;s compute usage so I took that those uh those models and I broke them out into different tiers and I said within which type of tier within each tier sorry which type of models are represented what you see here is that the most compute intensive models that we&#39;re talking about are to a large extent almost overwhelmingly actually language models um so a lot of the recent trend of this explosion and compute needs is being very very heavily driven by language models specifically um it&#39;s unclear though it&#39;s not the case that uh two-thirds three-quarters of AI researchers right now are just doing language modeling and so it&#39;s unclear what these Trends say about AI as a whole as opposed to a specific subset of one AI application area and we noticed this about a year and a half ago and we thought hey it would be really great to study this in more depths how do we go about doing it to answer that we decided that the best way to to to study this would be to run a survey a survey of AI researchers and so last summer we fielded this survey we had two target populations most of our our respondents came from basically anyone it says the slide says academics but really this does include some people at industry anyone who had published uh in a top AI conference between 2016 and 2021 and using some of cset&#39;s data we extracted about 27 000 emails associated with someone who had met that criteria and then we also to get um because that that population includes a meaningful number of industry but it is Academia dominated so we also augmented this with some sampling of Industry People based on the the LinkedIn criteria that you can see on the screen uh when we fielded this survey we got a total of 533 total results um I believe that there&#39;s been about 10 attempts or so to field surveys of AI researchers as a population and to my knowledge this is the second largest survey of that population conducted yet so we were pretty pleased with that result you can see the breakdown on the right hand side of the screen between Academia and Industry and the other thing I&#39;ll say is that based on the domains of our respondents from Academia we we mapped each University to its Qs world ranking and we have a basically an even division three ways between researchers at a top 50 University researchers at a University ranked 51 to 200 and researchers at a lower ranked University and then also a small percentage that we&#39;re affiliated with an unranked University which is why those numbers don&#39;t add up to 100. um this did I think give us a lot of a broad reach across many types of academics um it&#39;s not totally representative but it&#39;s it&#39;s pretty representative it&#39;s a much broader reach of academics than maybe policy makers are ordinarily exposed to so we fielded the survey the goal was to evaluate for each AI researcher how important is compute especially as compared to other factors as part of their research agenda and so one thing we asked researchers was if we doubled the budget for your current project what would you spend the extra money on and what we saw is that over half of our respondents said that they would spend that money on hiring either more Engineers or more researchers and only a fifth said that they would spend it on purchasing more compute and so just briefly I think that immediately this sort of raises questions about well okay there&#39;s certainly a subset of models that is exploding in compute requirements and where it&#39;s getting more and more costly to purchase enough compute to to train for instance a gpt3 a gpt4 but um this is a first indication that that might not actually be the types of projects that most researchers are working on and we also see this in other ways so we asked our researchers as well how often have you rejected a project revised a project or abandoned a project because of a lack of either compute data resources or researcher availability and again we found this result that sort of surprised us which is that in each of these three hypotheticals compute was the least cited reason for why academics are forced to change their research Plans by pretty substantial margins especially for rejected and abandoned and so again this sort of suggests that there&#39;s again a subset of models where compute is really important but focusing just on compute as a policy lever is maybe missing the needs of a potentially even broader population of AI researchers we also asked our researchers we gave them five different factors things like more compute better algorithms and so on and we asked them how strongly do you agree that this Factor drove the last decade of AI progress and then we also asked them how strongly do you agree that this Factor will drive the next decade of AI progress and what we saw is that um of all the factors compute scored the highest when researchers thought about the last decade that is to say 59 of researchers said yes I strongly agree that compute heavily drove the last decade of AI progress but it was actually the the uh Factor where the lowest percentage I strongly agreed that it would drive the next decade of AI progress and at the same time the percent saying that better algorithms would become more important Rose pretty dramatically there are a number of ways you could interpret this result for the sake of time I&#39;m not going to dwell on it although I&#39;m happy to come back in Q a um but again it&#39;s a sort of suggestion that hey relying on these historical Trends is maybe not the best way for conceptualizing what the needs of AI are moving forward in time um and so I&#39;m gonna pause with with this or sorry I&#39;m going to end with this slide here three high-level questions they&#39;re not yet drilling down into specific policies um but first of all I think one thing I I want to highlight here is that oftentimes it&#39;s just sort of said well deep learning is compute intensive and that is true but there&#39;s a lot of variation between different subfields of AI including different subfields that use deep learning and a lot of variation within subfields how compute intensive different research projects are and so when we talk about policies that impact compute we&#39;re not always talking about policies that impact all types of AI research we need to be careful about thinking about which types of applications were most heavily impacting especially because again as I mentioned language models are in some ways a unique AI application that seems uniquely compute intensive at least at this point in time it&#39;s possible that other fields will attention will pivot and other fields will become more compute intensive but right now when we talk about the really really compute intensive models we&#39;re primarily talking about language models the second high-level question here is oftentimes people also talk about compute divides between industry and Academia the inability of Academia to keep up with industry and again I think that there&#39;s an ambiguity here because people aren&#39;t always clear um are we talking about a systematic divide where industry as a whole demographic systematically has access to more compute than Academia or are we worried about the influence of a small number of big tech companies that are particularly well resourced when it comes to compute because these might actually translate into very different policy interventions which again I I&#39;m running out of time so I&#39;ll just sort of tease that and then finally if it&#39;s the case that it&#39;s an intrinsic property of all types of deep learning research that they require tons and tons of compute then oftentimes it makes sense to think well okay then the way to get AI progress is that we just have to keep throwing more and more compute at these projects as long as we can but if it&#39;s instead the case that there&#39;s a lot of variation in what projects researchers might take on um throwing compute at the problems will end up benefiting particularly compute intensive research plans while not so much helping others and I think we ought to start asking ourselves is that actually the outcome we want instead of just assuming that everything needs more compute we maybe need to get a little more careful about asking is that do we actually want to encourage compute intensive projects or do we want to encourage the type of research that might generate more efficient ways of doing the same tasks I realize I&#39;m leaving at pretty high level here but I&#39;m about 21 minutes into this talk and so the way I want to the the real message that I want to end on again I open by saying AI policy is increasingly focused on this compute lever oftentimes even maybe to the um oftentimes as a substitution away from other levers we might be pooling like Talent focused ones um and I think that this is just not the right approach it&#39;s good to a certain extent it&#39;s many it&#39;s very valuable to pursue a lot of these compute focused interventions but they are not a substitution for talent focused policy um Talent policy includes lots of domestic efforts things like uh Workforce Development uh leveraging the The Talented community colleges uh K-12 education these are all topics cset has written on but it&#39;s also those are all largely medium-term interventions in the short term the shortest way to address this Talent Gap is by welcoming more immigrants to the U.S more High talented immigrants and that I think uh it&#39;s oftentimes gotten too easy to imagine that well if all AI driven I apologize stumbling over my words here a bit in the home stretch it&#39;s gotten a bit easy to say if all AI progress is just driven by compute then we don&#39;t need to think as much about Talent OR about immigration and I think that that is very much the wrong way to approach AI policy these are all levers we need to think about and we cannot use one as a substitute for handling the others so on that note I&#39;m going to stop my my screen sharing and Tim uh jump in and and start peppering me with questions definitely uh thanks Micah that was great um so uh if you do want to get a q a or a question rather into the discussion um what you need to do is just drop it in the webinar chat um our sort of events team will be able to be collating those questions as Micah and I are talking um and I will bring them into the discussion but I guess might get to start maybe get a to get us rolling while people are kind of thinking about what they might want to ask I think you know obviously I think the position that you&#39;ve put out you recognize it as such is is a little bit provocative and I think to get the discussion going I I do want to kind of take the other side a little bit you know pressure some parts of this argument think about where it works where it doesn&#39;t work where um you know uh and ultimately I think what it means for policy I think the first question I&#39;d ask you is you know I&#39;m kind of putting the Hat on of being like a large language model booster right where we say like it&#39;s all very well and good to say that AI is very Broad and there&#39;s lots of people doing things that don&#39;t involve any compute but look like the hugest most important breakthrough in AI um in the last few years has been language models and to the extent that we are worried about sort of the US&#39;s um advantage of the technology we need to be working on the most important thing right it&#39;s a bet it&#39;s clearly paying off and it&#39;s the most critical and so you know what do you say to someone who I think kind of takes a look at your results and says yeah this is all very well and good but from a national interest standpoint the thing that we really need to be kind of you know accelerating on right now is precisely the thing that is you know uh compute intensive yeah so um it&#39;s a it&#39;s a fantastic question I mean that&#39;s the immediate and um obvious pushback is to say well okay um sure maybe language models are the most compute intensive but they&#39;re also the most important so of course we just want to focus more there um and I suppose there are there&#39;s there&#39;s two ways I could respond to that one way is to question just how strategically valuable language models are um and I think that frankly I don&#39;t want to get down too far in that rabbit hole but uh there are reasonable questions about so for instance this arms race frame uh it&#39;s very easy to say oh well you know language models could be economically transformative in the U.S uh we&#39;re all very worried about them they&#39;re in the news we&#39;re constantly paying attention therefore it must be the case that China also views this as a critical technology and they&#39;re racing as quickly as possible to get there and I would say well first of all um the Chinese State very quickly moved to release regulations that would pretty heavily dampen the ability of companies to deploy these at least in the way that they&#39;re being deployed in the U.S and so I think that there&#39;s a sort of question of like what if the what if the Chinese government doesn&#39;t actually view this with the same level of strategic importance um and you can play that out more you can say well China&#39;s actually much more of a manufacturing economy so to the extent that there is an economic payoff from something like language models it probably disproportionately benefits service economies like the US and less so manufacturing economies like China um and so there&#39;s basically where I&#39;m going there is just you might question okay sure language model is very important um is it the case that China also sees it that way or are they focused on other types of AI research that perhaps aren&#39;t as compute intensive that alters the National Security calculus a little bit the other response I could make is that on the domestic front in terms of which type of which type of Technology do we want to encourage you could certainly say well language models are just the most important that&#39;s what we ought to be encouraging and the response I would give there is just that that&#39;s a little abnormal when it comes to that we fund science research which is that usually we try not to prejudge which types of basic research are the most valuable we Instead try to give out resources to academics pretty broadly across many many different types of applications in the hopes that eventually one of those will sort of demonstrate utility in such a way that industry picks it up and expands it at scales that Academia can&#39;t um and so from that perspective I would say like yeah you can certainly argue that like no we have enough information to conclude language models are the one that matters that&#39;s where all the government funding should be going into resources that help language models um but my pushback is just to say that usually that&#39;s not how we do science funding in the U.S and in some ways that&#39;s actually a pretty significant break from the normal process yes definitely so I see these questions that are coming in they&#39;re they&#39;re really really great and I think I would encourage you to drop your uh hat in the ring uh if you want Micah to kind of consider one of them um I think as people kind of come in I think it does make sense maybe to build on one audience question here uh so Tom dieterich uh hey Tom I think we&#39;ve met many years ago it&#39;s good to hear from you um I think has a really interesting comment about the different time scales of the policy levers here right so I think the point that he made in his question was look there&#39;s a lot of things we can do for compute that will impact the Strategic environment right now right we can export control that totally changes the nature of the industry overnight whereas while we can all agree with Talent you know it&#39;s an important feature we shouldn&#39;t see this as and I don&#39;t think you&#39;re arguing that they are mutually exclusive but you know if we&#39;re worried about say let&#39;s just put it out there competitiveness with China right are we worried that Talent is maybe too slow of a lever here right and so maybe one response is to say look you know I agree with you talent&#39;s really important but compute is ultimately the thing that we need in this kind of critical moment where we want advantage and it&#39;s the tool that will get us the most impact right away um do you buy that I&#39;m curious about how do you respond to that so um the one the one response I would give is that uh it is the case that ideally these these operate as complementary policies um I have however had multiple conversations with people who are involved in policy making and I certainly won&#39;t name names here but but conversations where I&#39;ve presented the results of of this survey and I&#39;ve and they&#39;ve said like you know what&#39;s the one sentence summary and I&#39;ve said something to the effect of like compute is great but maybe we should be spending more attention on Talent especially immigration as a short-term intervention and they&#39;ve said in response oh I can&#39;t tell my boss that um and so I&#39;ve I&#39;ve started to wonder like in an Ideal World these are complementary but I also think that there&#39;s actually a risk that uh given political constraints we start to view them as substitutionary like oh you know great AI progress depends on compute that means I don&#39;t have to touch immigration um and I don&#39;t have to tell anyone that we need more immigrants and in a very you know heightened political environment where that is a sort of dangerous message to say just to send you get a risk there are clear incentives I think to sort of over index on the value of compute despite its its importance and its relevance and the other the other response that I would give um in terms of yeah computers may be a very short-term intervention um pay 12 education is relatively long-term Talent uh immigration is much shorter term but it&#39;s still relatively I mean you&#39;re talking a few years here um the answer to that is that I&#39;m not convinced that we are in a crisis moment right now I think that chat gbt has certainly galvanized a ton of attention but uh I don&#39;t it&#39;s not the type of Technology where you know um chat gbt gives some sort of decisive strategic advantage to the first military that starts to ask the chat bot uh which types of military operations they should do that is not really a sort of risk that I&#39;m concerned about I think that the payoffs have much more to do with economic automation economic improvements and in many ways it&#39;s actually fortunate that uh large language models which are somewhat more indirect of a national security risk were the type of model to generate this breakthrough the the caveat to that is that I acknowledge a lot of the disinformation risks I have personally done a lot of of research on the use of AI for disinformation and for influence operations and I&#39;ll say there that based on my assessment I&#39;m happy to talk more I think that right now people are more in danger of over inflating the risk than they are of under inflating it but happy to to either address that in a different question or I know we&#39;re getting a lot of questions so I want to move on yeah no definitely that&#39;s all great so one other final question using My Prerogative and I think it&#39;s related to the question that Derek Slater is asking here in the chat um I am sort of really interested in not just thinking about these two things that is to say compute policies and levers and talent levers as being complementary but also in some ways like interdependent in some ways so I think there&#39;s one way of looking at it which is maybe another slice which is Micah wouldn&#39;t you agree that um you know being able to access these incredibly rare powerful Computing clusters is one reason why talented researchers might move to the US right there&#39;s also these really interesting questions about whether or not policy around compute and I&#39;m now talking about say like Nair style approaches um is important for leveling the playing field right you could say okay well maybe immigrants and talent they really want to move to the country if they think they can do a AI startup that could really go compete with the best of them right open AI or what have you and so what we need is is you know say public research clouds to be able to kind of like open up the space for even to be worth it for us to be able to leverage talent in the first place and I guess I&#39;m kind of curious again I&#39;m I&#39;m not really pressuring the argument but I am kind of want to hear sort of your view on how these two are almost kind of related variables in some ways and that there are almost ways of doing compute policy which are kind of talent policy in some sense so first level answer I would say yes absolutely um although in order for that to pay off you actually have to have a high enough cap on skilled immigrants that people can respond to that signal and decide to come to the US um but on this on this point about the The Nair sort of equalizing the playing field uh one other interesting really fascinating result from our survey I&#39;ll talk about briefly here um we do see that like AI projects have tons and tons of variation in how much compute different projects need um that is I think you know our survey supports that that&#39;s abundantly true across basically any way you try to look at it um and so there&#39;s a question of like why is there so much variation and broadly I would say there&#39;s two ways you might go go about answering that one is to say well it&#39;s access that governs how much compute people use I.E everyone would like to be doing really compute intensive work but there&#39;s tons of stratification and how much access people have and so many researchers are just forced to do other types of work the other broad types of explanation you might offer has to do with self-selection effects which is to say look different subfields different methods within subfields have a lot of variance in how much payoff you get for expanding compute and so what happens here is that researchers select into one type of of research maybe that initial selection is governed by access but then once they&#39;re in a like low compute research project it doesn&#39;t actually help them to just throw more compute at their problem um and so an interesting thing that we observed in this survey is that uh actually The respondents Who expressed the most concern about not having enough compute to do meaningful work in the future where the same respondents who reported using the most compute right now um and we got that result actually across a number of indicators so like the more compute you use the more important you thought compute was in the past the more important you think compute will be in the future the more likely you&#39;d be to spend more money on compute Etc and this I think suggests that like people have been assuming something of the the access explanation which is like oh if we just give more compute to everyone people will use it and that will reduce the inequalities in how much compute different researchers use it might actually be the case that the reverse is somewhat true which is that the people who are most motivated to make use of new compute resources are the people already using a lot of compute um and so in that case if you just give everyone access to more compute what you actually do is you increase the stratification between how much compute different projects use maybe that&#39;s worth it maybe the more compute intensive projects as we talked about are just more valuable and so uh it&#39;s worth doing that um but I do think like this is a sort of call to the research community I think that we&#39;ve been way too simplistic about just saying like is compute important isn&#39;t it compute it isn&#39;t it important I&#39;m much more interested in asking like who benefits differentially from compute focused interventions because that I think gets us into a frame of talking about like hey certain ways of implementing these um these types of policies do create relative winners and relative losers and we actually want to be certain that that&#39;s what we are doing before we try to pursue that policy or sorry we want to be certain that that&#39;s what we want to do before we go ahead and we do it yeah no I think that makes a lot of sense um and I guess I don&#39;t know to return to I think a result that you&#39;re talking about a little bit earlier like sort of like like it or not right like we&#39;re gonna hit a ceiling on compute because people things just can&#39;t grow at the rate that they have been right um and so you know you&#39;re almost forced into a situation at some point which is okay except that we do all the compute stuff we dominate compute you know your projections would say well we just expect that things are not going to grow the way they have been like 36 months from now essentially um which I think is a really uh interesting and an important result um one kind of question maybe that you can maybe speak to now um uh Quentin um in the chat had a quick question which is now kind of flipping the analysis kind of like on the other side of the Triad to basically say okay we&#39;ve been talking a lot about compute we&#39;ve been talking a lot about Talent um I guess Michael do you want to sound off for a second on your views about data as a lever in the space um right because that&#39;s the that&#39;s the kind of last bit of the triangle you&#39;ve not talked about yeah yeah I&#39;m curious about like if you feel similarly around some of these issues right which is like maybe we&#39;re too obsessed with data or maybe we&#39;re not obsessed with data enough like interesting kind of just having you maybe talk a little bit about that yeah so there was certainly there was a moment maybe three or four years ago where Big Data was the solution to AI progress that we would just make data sets bigger um and uh I I think in some sense like the reason that things are framed so much in terms of compute in this presentation is because we&#39;ve clearly moved on from that um at least within the policy circles that I happen to be able to to observe we&#39;re no longer talking about Big Data we&#39;re talking about controlling access to Hardware um but but data is the third leg here and so uh I do actually think data becomes especially relevant not just when we&#39;re talking about like self-super self-supervised large-scale language models but whenever we want to talk about like what are the socially valuable applications of AI that we really want oftentimes the constraint ends up being not compute which may be necessary to to train some sort of large Foundation model but in the last stage of actually getting true economic utility out of that you need a meaningful enough data set that you can fine tune that model on some socially useful purpose and a great example here is like you might have a great vision classifier right that can can work with all 1000 categories from imagenet um that is really meaningful as a researcher uh a research breakthrough if you improve that although imagenet has been saturated you know Etc um but to translate that into something like a medical diagnostic tool those thousand categories from the imagenet data set are not the right thousand categories you do need some sort of well-curated additional data and so I would say this is probably like in terms of government interventions uh that we could do to to boost research progress I do think a broad answer is to say well if there&#39;s ambiguity about what types of resources are most valuable then perhaps the best way to provide it is in the form of Grants and let researchers themselves decide how to use that money and the exception to that is well curated data sets that only the government can put together so that&#39;s that&#39;s the exception where I think I see the most utility of like a concerted whole of government effort to boost AI is to coordinate between agencies gather all the meaningful data set all the all the meaningful data sets create a system that allows structured and tiered access to it so that different research projects can work with different data sets of different levels of sensitivity and then you know basically see what happens see what kind of applications we get out of that that can be translated into having a lot of economic utility that&#39;s great yeah thank you um so for the next question I&#39;m going to join a couple of comments and questions that have come into the chat um largely kind of drawing on rohini&#39;s comments and Tom&#39;s recent comment I think it&#39;s almost a question about how you account for the difference here right so clearly we have a lot of policy discussion in DC which is very compute focused um and uh and at the same time we have your survey which talks to a bunch of researchers and says actually we don&#39;t care that much about compute and I&#39;m kind of curious about how you account for the difference right I think implicit in what you&#39;re arguing is well maybe the policy makers in DC have it wrong or maybe they&#39;re listening to the wrong people we should be listening to the researchers I think is maybe a subtext of what you&#39;re arguing and I think some people would say well like a look like your response rate was super low on your survey how can we trust that this is actually giving us good signal on where we we should be moving yeah and I guess I&#39;m curious about how you reconcile like I think it&#39;s a two-part question one of the questions is like how do we end up in this place where compute is such a focus of the policy discussion and the second one would be how would you respond to people who say look this survey is just not very representative like you&#39;re getting just some weird noise and the people who happen to respond and maybe the people who are responding are so far you know in the wilderness in this space that you know they had lots of incentive to argue that the current kind of um you know Orthodoxy is wrong so I&#39;ll respond to the um the question about the survey we&#39;ve gotten this this um this comment about well the response rate was too low uh a couple of times um and it&#39;s I it&#39;s worth addressing although there are there are reasons that the response rate is low that I can tell you which is for instance like I said we used DC sets data to pull tons of emails of anyone who had published in one of these top AI venues um we knew going into it that like we were confident that each of those emails was associated with someone who had done meaningful AI work at some point and we also asked a screener at the start of the survey to confirm that the respondent works on AI systems but we couldn&#39;t ensure that for instance like that&#39;s an email that is still actively being monitored um and so we went into this knowing that a lot of the emails we were going to send this to weren&#39;t there there was just no world in which people would respond um and so I think that that artificially makes the response rate seem rather low what matters much more than the response rate is the raw count of response that we got um and again I mentioned this in the in the presentation itself I think that to my knowledge this is the second largest survey that&#39;s ever been conducted of AI researchers I&#39;m very pleased with that um we also have a meaningful you know I mentioned that we have basically equal representation across three different tiers of Academia um because of privacy issues we didn&#39;t ask a ton of personal identifying information that would allow us to exhaustively check the representativeness so there&#39;s certainly ambiguity but we&#39;re pretty confident in the main results of the survey to the other question about like why is this Gap existing one potential explanation I might give you is going back to this result that the people who use the most compute are the most concerned about compute and I think that there&#39;s a possible explanation you can give here which admittedly is speculative so I would I&#39;m not gonna like claim that this is true but I think it&#39;s reasonably likely that the academics who happen to have the most access to compute work at relatively Elite institutions and are disproportionately likely to have connections in DC which means like to the extent that there&#39;s um a directionality of the bias I would expect policy makers in DC to be hearing more about compute than what the median AI researcher says again you might respond and just say well researchers at Elite institutions know better what type of AI research is valuable so who cares if they have a sort of bias perspective that&#39;s a totally fair uh response that someone might make but I do think that like you can see how you might get these gaps in communication creep up just based on the results of the survey that we did collect yeah that&#39;s that&#39;s a great response and I think building a little bit on that I&#39;m going to go to a question that I think Duncan you asked quite early in the session sorry my cat um uh you asked quite early in the session um which is I think one hypothesis you can draw about why DC is most interested in compute is kind of that it&#39;s the most controllable thing in some ways right where you&#39;re like we can pull the export control lever and things can just happen and so you&#39;re kind of like well you you use the levers that you&#39;re given and that are most accessible and you sort of grab at them um and I think there has been maybe a notion that this is a particularly good lever because the greatest sort of like threat and danger comes from these big powerful models right and you can read that however you want right like you can read that in the the x-risk sense you can read that in the sense of the work that you&#39;ve done in the past right which is these models seem to enable a certain kind of disinformation that we might want to be concerned about um and I guess I&#39;m kind of curious about like if you you know that is zooming out a little bit from the point of view of competitiveness and policy to talking more about like managing risk but you you think that assumption will continue to hold right which is basically that like powerful AI systems that need to be built in large Computing clusters are where most of the risk is and that&#39;s where we need to be focusing most of our time or is that kind of historically provisional or is it not even true right now right like I&#39;m kind of curious about how you kind of weigh those threats and you know as we kind of think about like why we might want to be focused on certain levers over others yeah absolutely so first of all I&#39;m going to use this as my uh my point to explicitly plug the survey and say you know there&#39;s a ton of discussion in there there&#39;s a ton of results we I just don&#39;t have time for in this context and so um to this point from Duncan we start out the conclusion by saying look there&#39;s three broad reasons why compute has become really influential as a as a an anchoring an anchor basically one is the belief that it&#39;s critical to AI breakthroughs a second is the belief that there&#39;s really stratified access between different demographics and a third is like hey it&#39;s easy it&#39;s physical we can look at it we can hold it we can export control it we can do these sorts of things that it&#39;s much harder to do for data in the abstract um and we&#39;re very clear like the to the extent that we think this the survey does say things about points one and two in that that breakdown we&#39;re not saying that like somehow these survey results mean that uh the political barriers to Talent disappear um those are still very real and Rel like this is an argument why you might Focus some level of compute of attention on compute even disproportionate to the potential impact it has all well and good I&#39;ll reiterate I don&#39;t think it&#39;s a substitution for talent policy we need to also think about that but the other point is um so the other the other thing you raised is like is it actually the case that these largest scale models are the most dangerous um and in an x-risk sense I mean you might hypothetically say if we keep scaling them up we get totally unforeseen difficulties totally unforeseen capabilities that are very very worrisome um there&#39;s another way of looking at it though if we think about other types of risks like the disinformation risk I think two other pieces of recent evidence that support the narrative I&#39;m giving you is that one we had Sam Altman say a few weeks ago uh just scaling the parameter count of these models is not a feasible Way Forward for open AI that&#39;s slightly different than saying not scale and compute is not a feasible Way Forward because there&#39;s a difference there but the other piece of evidence was this leaked um Google internal memo that was circulated the we have no moat argument um and the argument that was being Advanced there was basically like look we can train these giant Advanced language models um they can be super versatile super generalizable they maybe they have 100 billion 500 billion a trillion parameters if someone can take a 20 billion parameter and with a bit of fine tuning replicate the capability on any discrete use task of this generalizable model the economics of training those massive models gets totally blown out the window um and with the disinformation stuff one of the things that I&#39;m working on is some cost modeling that I do things suggests for a lot of disinformation operators it might actually the preferred method of producing disinformation might be to use smaller models that are easier to run and where you can fine tune them yourselves instead of trying to access something like gpt4 behind a paywall um and so in that sense I would say I guess the the succinct way of summing that all up is um there may be these risks that are associated with the largest models but it may also be the case that other types of risks are diffusing to other types of models and so simply locking down the largest models is perhaps not a viable way to fully counteract those types of risks right um perfect answer and I think there&#39;s a lot more we can go into there um so we are essentially at time um I&#39;d advise everybody to read the paper as Michael mentioned there&#39;s a lot more that he didn&#39;t have a chance to get into but um yeah please reach out to him with any questions and at this point I&#39;ll turn it back to Danny to close us out awesome thanks Tim for moderating today and thanks to Micah for a fascinating and provocative discussion thanks also to all of you for taking part in this discussion and for your comments and questions we&#39;re sorry we couldn&#39;t get to them all if you&#39;d like to learn more about cset please go to cset.georgetown.edu and sign up for our newsletter and research updates we&#39;ll be in touch soon with details on our next webinar so be sure to watch your inbox until then thank you for joining us and we hope to see you again soon have a good day thank you

Transcript for:CSET AI Policy Insights: Talent, Compute, Data

Transcript for:
CSET AI Policy Insights: Talent, Compute, Data