1、MARCH 2024Future-Proofing Frontier AI RegulationProjecting Future Compute for Frontier AI ModelsPaul ScharreAbout the AuthorPaul Scharre is the executive vice president and director of Studies at the Center for a New American Security(CNAS).He is the award-winning author of Four Battlegrounds:Power

2、in the Age of Artificial Intelligence.His first book,Army of None:Autonomous Weapons and the Future of War,won the 2019 Colby Award,was named one of Bill Gatess top five books of 2018,and was named by The Economist as one of the top five books to read to understand modern warfare.Scharre worked in t

3、he Office of the Secretary of Defense in the Bush and Obama administrations,where he played a leading role in establishing policies on unmanned and autonomous systems and emerging weapons technologies.He led the Department of Defense(DoD)working group that drafted DoD Directive 3000.09,establishing

4、the departments policies on autonomy in weapon systems.He holds a PhD in war studies from Kings College London,an MA in political economy and public policy,and a BS in physics,cum laude,from Washington University in St.Louis.Prior to working in the Office of the Secretary of Defense,Scharre served a

5、s an infantryman,sniper,and reconnaissance team leader in the Armys 3rd Ranger Battalion and completed multiple tours in Iraq and Afghanistan.He is a graduate of the Armys Airborne,Ranger,and Sniper Schools and honor graduate of the 75th Ranger Regiments Ranger Indoctrination Program.About the Techn

6、ology&National Security ProgramThe CNAS Technology&National Security program explores the policy challenges associated with emerging technologies.A key focus of the program is bringing together the technology and policy communities to better understand these challenges and together develop solutions

7、.About the Artificial Intelligence Safety&Stability ProjectThe CNAS AI Safety&Stability Project is a multiyear,multiprogram effort that addresses the established and emerging risks associated with artificial intelligence.The work is focused on anticipating and mitigating catastrophic AI failures,imp

8、roving the U.S.Department of Defenses processes for AI test and evaluation,understanding and shaping opportunities for compute governance,and understanding Chinese and Russian decision-making on AI and stability.AcknowledgmentsI owe an enormous debt of gratitude to the researchers at Epoch,whose ana

9、lysis this report relies upon.Their study of AI trends is a valuable resource for understanding the AI revolution,and I hope this report helps to further expose policymakers to their insights.This report builds on prior research and analysis by numerous scholars to whom I am indebted,including Dario

10、 Amodei,Markus Anderljung,Tamay Besiroglu,Tom B.Brown,Ryan Carey,Ajeya Cotra,Ben Cottier,Ege Erdil,Tim Fist,Lennart Heim,Danny Hernandez,Anson Ho,Marius Hobbhahn,Saif M.Khan,Andrew J.Lohn,Alexander Mann,Micah Musser,Jaime Sevilla,and Pablo Villalobos.I am especially grateful to Markus Anderljung,Tam

11、ay Besiroglu,Ben Cottier,Lennart Heim,and Jaime Sevilla for their valuable feedback on earlier drafts of this report.CNAS Research Associate Michael Depp provided valuable background research,feedback,editing,and assistance with citations and the bibliography.Thank you to CNAS colleagues Maura McCar

12、thy,Melody Cook,Rin Rothback,and Anna Pederson for their roles in the review,production,and design of this report.Any errors are the responsibility of the author alone.This report was made possible with the generous support of Open Philanthropy.As a research and policy institution committed to the h

13、ighest standards of organizational,intellectual,and personal integrity,CNAS maintains strict intellectual independence and sole editorial direction and control over its ideas,projects,publications,events,and other research activities.CNAS does not take institutional positions on policy issues,and th

14、e content of CNAS publications reflects the views of their authors alone.In keeping with its mission and values,CNAS does not engage in lobbying activity and complies fully with all applicable federal,state,and local laws.CNAS will not engage in any representational activities or advocacy on behalf

15、of any entities or interests and,to the extent that the Center accepts funding from non-U.S.sources,its activities will be limited to bona fide scholastic,academic,and research-related activities,consistent with applicable federal law.The Center publicly acknowledges on its website annually all dono

16、rs who contribute.TABLE OF CONTENTS01 Executive Summary03 Introduction06 Cost and Access to AI Models07 Implications for Policymakers08 Understanding Cost and Compute Growth09 Related Work12 Current Best Estimates and Assumptions17 Cost and Compute Projections21 Limits on Cost Growth23 Limits on Har

17、dware Improvements27 Proliferation28 Costs for Hardware-Restricted Actors32 Compute Regulatory Threshold35 Conclusion36 Appendices44 Selected BibliographyPART I:BACKGROUNDPART II:ANALYSIS1 1 1 1 1TECHNOLOGY&NATIONAL SECURITY|MARCH 2024Future-Proofing Frontier AI Regulation:Projecting Future Compute

18、for Frontier AI Models PExecutive Summaryolicymakers should prepare for a world of signifi-cantly more powerful AI systems over the next decade.These developments could occur without fundamental breakthroughs in AI science simply by scaling up todays techniques to train larger models on more data an

19、d computation.The amount of computation(compute)used to train frontier AI models could increase significantly in the next decade.By the late 2020s or early 2030s,the amount of compute used to train frontier AI models could be approximately 1,000 times that used to train GPT-4.Accounting for algorith

20、mic progress,the amount of effec-tive compute could be approximately one million times that used to train GPT-4.There is some uncertainty about when these thresholds could be reached,but this level of growth appears possible within anticipated cost and hardware constraints.Improvements of this magni

21、tude are possible without government intervention,entirely funded by private corporations on the scale of large tech companies today.Nor do they require fundamental breakthroughs in chip manufacturing or design.Increased spending beyond the limits of private companies today or fundamentally new comp

22、uting paradigms could lead to even greater compute growth.8110321033High-low compute estimate(95%confidence interval)Compute(accounting for costand hardware constraints)202220242026202820302032203420362038204020421 millionGPT-41,000GPT-4GPT-4 estimated computeGPT-4 estimated co

23、mputeTraining Compute(FLOP)Rising costs to train frontier AI models may drive an oligopoly at the frontier of research,but capabilities are likely to proliferate rapidly.At present,algorithmic progress and hardware improvements quickly decrease the cost to train previously state-of-the-art models.Wi

24、thin five years at current trends,the cost to train a model at any given level of capability decreases roughly by a factor of 1,000,or to around 0.1 percent of the original cost,making training vastly cheaper and increasing accessibility.The U.S.government has placed export controls on advanced AI c

25、hips destined for China,and denying actors access to hardware improvements creates a growing gap in relative capability over time.Actors denied access to hardware improvements will be quickly priced out of keeping pace with frontier research.By 2027,using older,export-compliant chips could result in

26、 a roughly tenfold cost penalty for training,if export controls remain at the current technology threshold and are maximally effective.However,proliferation of any given level of AI capa-bilities will be delayed only a few additional years.At present,the cost of training models at any given level of

27、 AI capabilities declines rapidly due to algorithmic progress alone.If algorithmic improvements continue to be widely available,hardware-restricted actors will be able to train models with capabilities equivalent to once-frontier models only two to three years behind the frontier.COMPUTE USED TO TRA

28、IN A FRONTIER AI MODEL RISES OVER TIMEThe amount of compute used to train frontier AI models could increase to around 1,000 times GPT-4 by the late 2020s or early 2030s,even accounting for cost and hardware constraints.CNASDC2CNASDC2High-low effective compute estimate(95%confidence interval)Effectiv

29、e compute(accounting for cost and hardware constraints)Effective Compute(2022 FLOP equivalent)20222024202620282030203220342036203820402042GPT-4 estimated computeGPT-4 estimated compute1 millionGPT-41,000GPT-48110321033$10T$1T$100B$10B$1B$100MHigh-low cost estimate(95%confidence

30、 interval)Training cost(arbitrary tapered cost growth projection)Cost of Final Training Run(2022 dollars)U.S.GDP($23 trillion)U.S.GDP($23 trillion)$100 billion$100 billion$20 billion$20 billion20222024202620282030203220342036203820402042WITH ALGORITHMIC IMPROVEMENTS,EFFECTIVE COMPUTE GROWS OVER TIME

31、THE COST TO TRAIN A FRONTIER AI MODEL RISES OVER TIMEAccounting for algorithmic progress,the amount of effective compute used to train frontier AI models could be one million times GPT-4 by the late 2020s or early 2030s.The cost to train a frontier AI model is currently doubling approximately every

32、10 months.Cost growth is assumed to slow as costs approach the limit for private companies,currently in the tens of billions of dollars.In this projection,the cost doubling period is arbitrarily assumed to increase by 1.5 months per year,slowing the rate of cost growth.TECHNOLOGY&NATIONAL SECURITY|M

33、ARCH 2024Future-Proofing Frontier AI Regulation:Projecting Future Compute for Frontier AI Models 3Access to compute and algorithmic improvements both play a significant role in driving progress at AIs frontier and affecting how rapidly capabilities proliferate and to whom.At present,the amount of co

34、mpute used to train large AI models is doubling every seven months,due to a combination of hardware improvements and increased spending on compute.Algorithmic efficiencythe ability to achieve the same level of performance with less computeis doubling roughly every eight to nine months for large lang

35、uage models.Improved performance comes from both increased compute and algorithmic improve-ments.If compute growth slows in the 2030s due to rising costs and/or diminishing hardware performance gains,future progress in frontier models could depend heavily on algorithmic improvements.At present,fast

36、improve-ments in algorithmic efficiency enable rapid proliferation of capabilities as the amount of compute needed to train models at any given level of performance quickly declines.Recently,some leading AI labs have begun with-holding information about their most advanced models.If algorithmic impr

37、ovements slow or become less widely available,that could slow progress at AIs frontier and cause capabilities to proliferate more slowly.While there is significant uncertainty in how the future of AI develops,current trends point to a future of vastly more powerful AI systems than todays state of th

38、e art.The most advanced systems at AIs frontier will be limited initially to a small number of actors but may rapidly proliferate.Policymakers should begin to put in place today a regulatory framework to prepare for this future.Building an anticipatory regulatory framework is essential because of th

39、e disconnect in speeds between AI progress and the policymaking process,the difficulty in predicting the capabilities of new AI systems for specific tasks,and the speed with which AI models proliferate today,absent regulation.Waiting to regulate frontier AI systems until concrete harms materialize w

40、ill almost certainly result in regulation being too late.The amount of compute used to train models is likely to be a fruitful avenue for regulation if current trends continue.Massive amounts of compute are the cost of entry to train frontier AI models.Compute is likely to increase in importance ove

41、r the next 10 to 15 years as an essential input to training the most capable AI systems.However,restrictions on access to compute are likely to slow,but not halt,proliferation of capabilities,given the ability of algorithmic advances to enable training AI systems with equivalent performance on less

42、compute over time.Regulations on compute will be more effective if paired with regulations on models themselves,such as export controls on certain trained models.Introductionolicymakers and industry leaders have increased their attention on regulations for highly capable general-purpose AI models,so

43、me-times called“frontier”models.Examples of current frontier AI models include GPT-4(OpenAI),Claude 3(Anthropic),and Gemini Ultra(Google).Companies already are training larger,more capable next-gener-ation models using ever-larger amounts of data and computing hardware.The computation used to train

44、frontier AI systems is growing at an unsustainable rate.The amount of computation,or compute,used to train state-of-the-art machine learning models increased ten billionfold from 2010 to 2022 and is doubling every six months.1 For the largest models,the amount of compute used for training is doublin

45、g approximately every seven months.This rapid increase in compute exceeds the pace of hardware improvements and is in part driven by increased spending on training.Costs for training the largest models are doubling roughly every 10 months.2 Training current frontier models costs on the order of tens

46、 of millions of dollars just for the final training run.The full cost for training frontier models today,accounting for earlier training runs and exper-iments,could cost around$100 million.3 As training costs continue to rise,they could reach hundreds of millions of dollars or even billions of dolla

47、rs.In the near term,growth in large-scale training runs at AIs frontier is likely to continue.Leading AI labs already are reportedly training next-generation models or raising funds to do so.4 Nvidia is shipping hundreds of thousands of new chips,which will enable more powerful future training runs.

48、In the long run,however,cost and possibly hardware limitations are likely to constrain future compute growth.5 The current exponential pace of compute growth cannot continue indefinitely.How long it continues,at what pace,and how much compute grows before leveling off has important implications for

49、the future of AI progress.The role of cost and access to hardware as barriers to entry for training highly capable AI systems also has policy implications,such as for export controls and some regulatory proposals.Current trends point to a future of vastly more powerful AI systems than todays state o

50、f the art.PCNASDC4Research Questions This paper aims to answer several questions about how trends in cost and compute could affect the future of AI:1.Cost and compute projections:If current trends were to continue,how would the amount of compute used to train frontier AI models and the cost of train

51、ing rise over time?Accounting for algorithmic progress,how would the amount of effective compute increase over time?2.Limits on cost growth:How much could compute increase before reaching the spending limits of private companies,and when would that occur?If the rate of cost growth slows as costs ris

52、e,how might that affect the amount of compute used for training frontier models?3.Limits on hardware improvements:How might limits on continued hardware improvements affect future compute growth?4.Proliferation:How might improvements in hardware and algorithmic efficiency affect the avail-ability of

53、 AI capabilities over time?5.Costs for hardware-restricted actors:How might constraints on hardware availability(for example,due to export controls)affect cost and compute growth for actors denied access to continued improvements in AI hardware?6.Compute regulatory threshold:How might improvements i

54、n hardware and algorithmic efficiency impact the effectiveness of training compute as a regu-latory threshold for frontier models over time?The answers to these questions have important bearing on policy-relevant decisions today,such as the anticipated effect of export controls or other proposed reg

55、ulations that would limit access to compute-intensive AI models.On the one hand,trends in rising costs are consolidating access to frontier AI models among a handful of leading AI labs.On the other hand,countervailing trends in hardware improvements and algorithmic efficiency are lowering barriers t

56、o capabilities,enabling proliferation.Some regulatory and policy interventions may be more or less feasible or desirable depending on how compute and cost change over time and the consequences for access to frontier AI models and the proliferation of capabilities.This paper aims to answer these ques

57、tions with the goal of informing policymakers understanding of possible scenarios for future AI development.ApproachUsing current trends as a baseline,this paper projects cost and compute growth under various scenarios.The paper projects compute growth due to increased spending and hardware improvem

58、ents.Additionally,it accounts for algorithmic improvements by projecting effective compute over time.The paper then estimates when training costs are projected to reach current limits for large corporations and move into the realm of what have historically been government-level expen-ditures.Additio

59、nal scenarios explore how limits in hardware improvements may affect the availability of future compute.Since the cost to train a model with any given level of capabilities will decrease over time due to improvements in hardware and algorithmic efficiency,the paper also estimates how costs will decl

60、ine over time,making capabilities more accessible to a wider array of actors,enabling greater proliferation.The paper then estimates how training costs change for actors that are restricted from continued improvements in AI hardware,such as U.S.government export controls on advanced AI chips destine

61、d for China.Finally,this paper estimates how future improvements in hardware and algorithmic efficiency may increase the accessibility of compute and capabilities relative to the U.S.governments notification threshold established in the October 2023 executive order.The paper concludes by assessing t

62、he policy impli-cations of these projections.5PART I:BACKGROUND5CNASDC6propaganda,hate speech,or misinformation at scale.Generative image models can be used to generate deep-fakes and non-consensual pornography,including of minors.Generative image models also can show gender and ethnic/racial biases

63、 in how they present images,such as sexualizing female images,generating non-con-sensual nudes,or changing skin tone.13 Additionally,large language models are known to“hallucinate”facts,presenting misleading information.The largest and most capable large language models have dual-use capabilities,in

64、cluding identifying vulnerabilities in computer code or scientific knowledge that could be used to enable chemical or biological attacks.14 The accessibility of models has important impli-cations for their potential to cause harm.For some generative models,researchers have developed filters or other

65、 safeguards to reduce the likelihood of the model generating harmful content.For example,Stability AI included in their model Stable Diffusion a content filter to prohibit the generation of harmful images.15 Additionally,some generative image models have included embedded watermarking to make it pos

66、sible to identify AI-generated images and reduce the likelihood of the model being used to generate mis-leading deepfake images.Current state-of-the-art large language models,such as OpenAIs GPT-4,Anthropics Claude 3,Metas Llama 2,and Googles Gemini Ultra,use fine-tuning with reinforcement learning

67、to reduce their likelihood of producing problematic content.16 When access to the model is limited via an API,model owners have some degree of control over the content the model produces,although many models neverthe-less still generate concerning content.17 Once models are open-sourced,however,they

68、 rapidly proliferate,and researchers can easily remove or disable filters.Following Stable Diffusions release,researchers quickly disabled the content filter and removed the watermarking.18 Similarly,open-source large language models can be easily fine-tuned at relatively low cost to remove safeguar

69、ds.It took 19 hours of training at marginal cost to create to create an“uncensored”version of Llama 2 that was then posted online for anyone to download.19 Open-sourcing models increases their accessibility to academic researchers and start-ups but also to malign actors who may use the models for ha

70、rmful applications.Cost and Access to AI Modelsising costs have important implications for deter-mining who is able to access the most capable AI models.Compute costs are already an obstacle to academic researchers,who are priced out of training the largest,state-of-the-art models.6 Training costs f

71、or current large language models are estimated to be in the tens of millions of dollars just for the final training run.Total costs,accounting for earlier training runs and experiments,could be around$100 million for the largest models to date.7 Many frontier AI labs are backed by large corporations

72、 with deep pockets.Google DeepMind is owned by Alphabet.OpenAI secured a$10 billion investment from Microsoft.Google and Amazon have invested$2 billion and$4 billion,respec-tively,in Anthropic.8 Academics and start-ups do not have the financial resources to compete at this scale.The U.S.government i

73、s launching a pilot program of the National AI Research Resource to provide compute and data resources to academics,although the effort is not yet fully funded.9 If established,a national research cloud could help mitigate the effects of,but will not fundamentally alter,the trends driving increased

74、training costs.Current trends are pushing the frontier of AI research toward an oligopoly,where only a handful of well-funded actors can afford training the most capable AI models.Since 2019,many frontier AI labs have shifted increasingly to more limited release of their models,allowing other resear

75、chers or the public to interface with the model via an application programming inter-face(API)or only releasing models to a small number of vetted researchers.This trend toward the concen-tration of AI capability in the hands of a small number of corporate actors reduces the number and diversity of

76、AI researchers able to engage with the most capable models.In response to this trend,many AI researchers and some leading companies,such as Meta,have pushed for more open-source releases of foundation models to help level the playing field and democra-tize AI capabilities to researchers who are not

77、able to afford tens of millions of dollars to train their own foundation models.10Potential Harms Research labs that have limited release of their models have pointed to potential harms from widespread prolif-eration.11 Foundation models have well-documented problems of bias and toxicity due to thei

78、r training data.12 Large language models can be prompted to generate The largest and most capable large language models have dual-use capabilities.RTECHNOLOGY&NATIONAL SECURITY|MARCH 2024Future-Proofing Frontier AI Regulation:Projecting Future Compute for Frontier AI Models 7Geopolitical Implication

79、s of Model Access Given their dual-use capabilities,access to state-of-the-art models also has important implications for geopolitical and economic power.U.S.officials have expressed concern about the Chinese Communist Partys use of AI for military development,human rights abuses,and internal repres

80、sion.In October 2022,the U.S.Commerce Department established export controls on advanced semiconductors and semiconductor manu-facturing equipment destined for China,rules that have subsequently been updated and further refined.20 The rules prohibit the export to China of the most advanced graphics

81、processing units(GPUs)used for machine learning,even when the chips are produced outside the United States.U.S.export controls also limit the transfer of U.S.semiconductor manufacturing equipment,software,and tooling to Chinese semiconductor fabrica-tion plants(fabs),restricting Chinas ability to ad

82、vance domestic chip production.Japan and the Netherlands,leading producers of semiconductor manufacturing equipment,have adopted similar export controls on chip-making equipment and tooling.These export controls aim to deny Chinese firms the ability to access the most advanced AI chips,restricting t

83、heir ability to conduct large-scale training runs.U.S.officials have stated that the threshold for banned chips will remain constant over time,even as new chips are released.21 If these controls are effective,over time they will widen the hardware gap between China and the rest of the world,as today

84、s leading-edge chips become tomorrows legacy chips.Without access to the most advanced AI chips,Chinese labs would face higher costs to train models and,in some cases,might be priced out entirely of accessing the largest and most capable models.China continues to pursue efforts to grow its indigenou

85、s chip-making capabilities,however.Recent breakthroughs suggest that export controls may merely slow Chinese indigenous chip fabrication,not stop it completely.In September 2023,Huawei announced that its latest phone,the Mate 60 Pro,contained 5G technology powered by HiSilicons new Kirin 9000S chip.

86、Independent experts have assessed that the chip was produced using SMICs 7 nanometer(nm)foundry,an advanced chip fabrication process restricted under U.S.export controls.22 Questions remain about Chinas ability to produce advanced chips cost-effectively and at scale and how Chinese indigenous chip m

87、anufacturing will evolve over time.Presently,Chinese labs can access state-of-the-art AI models open source,negating the effectiveness of chip export controls.Chinese labs do not need to train their own large foundation models if they can simply download trained open-source models directly from the

88、internet.Safeguards on models can be easily fine-tuned away,undermining the U.S.governments efforts to keep advanced American AI technology from empowering the Chinese military or enabling Chinese government human rights abuses.The accessibility of future frontier AI models will have important impli

89、cations for the effective-ness of U.S.export controls.Countervailing Trends That Increase Model AccessWhile compute costs are rising,countervailing trends in hardware improvements and algorithmic efficiency are reducing the costs for training a model at any given level of capability over time.As cos

90、ts decline,an increasing number of actors can afford to train models with equiv-alent capabilities,enabling proliferation.At lower costs,it becomes increasingly likely that the cost of training a model will be affordable to an actor willing to open-source the model,as some companies such as Meta and

91、 Stability AI have done in the past.Once the model is open-sourced,it rapidly proliferates.Recent experience with generative language and image models suggests that the time lag from an initial breakthrough to an open-source version can be brief,as little as approximately seven months.23Implications

92、 for Policymakershe economic and strategic value of current state-of-the-art AI systems,such as those based on large language models or multi-modal models,is highly uncertain.Some early studies have suggested that the most capable large language models today could be used to automate a significant p

93、ortion of tasks currently done by white-collar employees.24 Increasingly capable AI systems could be used to improve productivity and accelerate scientific discovery.They also could have dangerous dual-use applications,such as enabling the development of chemical,biological,or cyber weapons.Policyma

94、kers face the difficult challenge of making AI models as accessible as possible for beneficial uses while restricting their access for harmful applications.This paper does not seek to resolve this dilemma.Rather,it aims to present policymakers with a greater under-standing of how trends in AI progre

95、ss may affect the accessibility of AI systems over time.Rising costs have Once the model is open-sourced,it rapidly proliferates.TCNASDC8important implications for the quantity and diversity of AI researchers using state-of-the-art models,prolifera-tion and potential harms,and geopolitical and econo

96、mic power.This paper projects cost and compute trends under various scenarios to better understand cost growth and when different actors may be able to access models at different levels of computational power.Ideally,as a result of this analysis,policymakers will be better able to“future-proof”the p

97、olicies they adopt today,taking into account exponential growth trends in cost,compute,and algorithmic progress.Understanding Cost and Compute Growth he amount of computation,or compute,used to train machine learning models,and the associated costs for training,have been rapidly increasing during th

98、e deep learning revolution.This revolution pairs machine learning techniques,many of which date back decades,with increased computational performance that only became more recently available to train large artificial neural networks.25 Using machine learning,algorithms are trained on data using comp

99、uting hardware.26 The result is a trained model that is a representation of patterns in the under-lying training data.This trained model has various applications,such as classifying new data or generating synthetic(AI-generated)data.While machine learning theories are decades old,for many applicatio

100、ns machine learning requires large amounts of computation in order to turn raw data into a useful trained model,and this only began to become available in the late 2000s.By refining machine learning techniques and combining them with advancements in computing hardware,algorithms,and increased data a

101、vailability,scientists have generated significant advancements in computer vision,image generation,language processing,gaming,and other areas.27Improvements in machine learning models can come from any of the three technical inputs into machine learning:the training data,the computing hardware used

102、for training,and the algorithms used for training.Progress during the deep learning revolution,which began around 2010 to 2012,has come from improve-ments in all three technical inputs.Researchers,in fact,have found remarkably predictable“scaling laws”that capture the relationship between model perf

103、or-mance and growth in model size(the size of the neural network),dataset size,and the amount of compute used to train a model.28 These empirically derived scaling laws demonstrate that model lossa measurement of model inaccuracy on test datahas an inverse relation-ship with model size,dataset size,

104、and training compute.Larger models,datasets,and training compute lead to reduced model loss,or improved accuracy on test data.This negative scaling is remarkably smooth and is not affected as much by model architecture or other factors.In short,without any fundamental advances in the science of AI o

105、r understanding of intelligence,training larger models with greater amounts of compute and more training data yields improved performance.AI researchers typically measure model perfor-mance using standardized benchmarks.The language model benchmarks Massive Multitask Language Understanding(MMLU)and

106、Beyond the Imitation Game(BIG-bench)cover a diverse array of lan-guage-based tasks,from coding to learning U.S.history.29 Owen(2023)found that while large language model performance on individual tasks was highly variable,aggregate performance on benchmarks showed“a fairly smooth relationship betwee

107、n overall performance and scale,consistent with an S-curve.”The amount of compute used to train models,adjusted for optimal scaling,was a“fairly predictable”gauge of benchmark performance.Owen concluded,“This supports the idea that higher-level model capabilities are predictable with scale,and gives

108、 support to a scaling-focused view of AI development.”30 Over time,benchmarks eventu-ally become saturated as models reach 80 or 90 percent accuracy and further attempts at improvement show diminishing returns.31 AI researchers then simply invent new benchmarks to tackle harder problems.AI researche

109、rs have used these scaling laws to train ever-larger models,with continued performance improvements.As of around 2021 to 2022,the most capable large language models had hundreds of billions of parameters and were trained on hundreds of gigabytes of data using thousands of advanced chips.32 Leading A

110、I labs such as OpenAI,Anthropic,and Google DeepMind more recently have begun to restrict publicly available information about their most capable models.Leaked information about GPT-4 suggests it is a very large model(an approximately 1.8 trillion parameter mix-ture-of-experts model),trained on a mas

111、sive dataset(approximately 13 trillion tokens of training data),Training larger models with greater amounts of compute and more training data yields improved performance.TTECHNOLOGY&NATIONAL SECURITY|MARCH 2024Future-Proofing Frontier AI Regulation:Projecting Future Compute for Frontier AI Models 9

112、9using massive amounts of computing hardware(approxi-mately 25,000 GPUs).33 Even while public details about the most advanced models have become more scarce,indi-cations are that leading AI labs are continuing to pursue ever-larger models.34 ComputeThe amount of computation used for training is a ke

113、y metric for AI models.Compute is a function of the number of chips used for training,the type of chip,the hardware utilization rate,and the amount of time the chips are used for training.The total amount of compute used for training can be captured in a single metric measured in float-ing-point ope

114、rations(FLOP).35 AI researchers can increase the amount of compute used for training by using more chips,using better chips,or increasing the training time.AI hardware continues to improve,with machine learning GPU price-performance,or performance per dollar(FLOP per second per dollar),doubling roug

115、hly every 2.1 years.These hardware improvements alone would lead to greater compute over time as AI researchers use better chips for future training runs.However,AI labs also are increasing their spending,buying tens of thousands of advanced chips for large-scale training runs.Thus,growth in compute

116、 is a function of both spending more money on chips and hardware improvements that increase the computations per dollar that chips can pute(FLOP)=expenditures(dollars)hardware improvements (FLOP per second per dollar)Future compute growth is likely to be driven by both increased spending and improve

117、d hardware.This paper will project compute growth due to both factors,as well as estimating the effects of possible limits on spending and hardware improvements.Algorithmic EfficiencyAlgorithms are also improving over time,reducing the amount of compute needed to train a model to the same level of p

118、erformance.Gains in performance can come from increases in compute and/or increases in algorithmic effi-ciency,which allows compute to be used more effectively.performance(effective compute)=compute(FLOP)algorithmic efficiencyAlgorithmic progress factors into the projections used in this paper in tw

119、o ways:(1)to decrease the amount of compute needed to train a model at any given level of capability over time;and(2)to increase performance for frontier models.36First,algorithmic progress decreases the amount of compute needed to train a model at any given level of performance over time.This effec

120、t has been measured for a variety of machine learning domains,including image classifiers,reinforcement learning algorithms,and large language models.37 Improved algorithmic effi-ciency plays a role in increasing accessibility by lowering compute(and therefore cost)as a barrier to entry to training

121、a model with any given level of capabilities.(Other factors,such as dataset availability,may still be a barrier for some actors.)As the amount of compute required to train a model at a given performance level decreases(in addition to compute per dollar increasing due to hardware improvements),costs

122、decline over time,making previously unaffordable models accessible to a wider array of actors.The Proliferation section of this paper applies improvements in algorithmic efficiency and hardware performance to estimate how training costs decline over time for any given level of perfor-mance,enabling

123、wider proliferation of capabilities.Second,overall improvements in algorithms increase the performance of frontier models.Future frontier models will use more compute due to hardware improvements and increased expendi-tures on compute.Algorithmic improvements allow researchers to use this compute mo

124、re effectively,leading to better performance.The combined effect of increased compute and algorithmic improvements is shown in this paper as“effective compute,”which is represented in 2022 FLOP equivalent(the amount of effective computational power equivalent to FLOP in 2022).38 Related Workhe amoun

125、t of computing power used to train state-of-the-art machine learning models has exploded during the deep learning revolution.As researchers have collected data on this trend,they have sometimes used this historical data to generate future projections of compute and cost growth,including estimating h

126、ow spending limits might affect compute growth.This paper builds on that prior analysis,updating projections using more recent historical esti-mates for compute and cost growth.Amodei and Hernandez(2018)assessed that the amount of compute used for training the largest AI models increased 300,000-fol

127、d from 2012 to 2018,T1010CNASDCCNASDC20222024202620282030203220342036203820402042HardwareImprovementsHardwareImprovementsIncreasedSpendingIncreasedSpending81103210331 millionGPT-41,000GPT-4GPT-4 estimated computeGPT-4 estimated computeGPT-4 estimated computeTraining Compute(FLO

128、P)20222024202620282030203220342036203820402042Algorithmic improvementsAlgorithmic improvementsEffective Compute(2022 FLOP equivalent)81103210331 millionGPT-4GPT-4 estimated computeGPT-4 estimated computeGPT-4 estimated compute1,000 GPT-41,000GPT-4IncreasedcomputeIncreasedcomput

129、eFIGURE 1.1|COMPUTE GROWS DUE TO INCREASE SPENDING AND HARDWARE IMPROVEMENTSFIGURE 1.2|PERFORMANCE IMPROVES DUE TO INCREASED COMPUTE AND ALGORITHMIC IMPROVEMENTSCNASDCCNASDC10TECHNOLOGY&NATIONAL SECURITY|MARCH 2024Future-Proofing Frontier AI Regulation:Projecting Future Compute for Frontier AI Model

130、s 11doubling every 3.4 months.39 While they assessed that cost eventually would limit compute growth,they esti-mated that this trend would continue in the short term.Subsequent analysis has attempted to more precisely project cost growth and when cost would become a limiting factor if compute were t

131、o continue to rise at current rates.Carey(2018)estimated cost growth in training compute and determined that the current rate of compute growth was not sustainable beyond a few years.40 Using a 3.5-month compute doubling period and estimating that the per unit cost of compute drops an order of magni

132、tude every 4 to 12 years due to hardware improvements,Carey estimated that the cost of the largest training runs would increase an order of mag-nitude every 1.1 to 1.4 years.Carey assumed that the maximum spending capacity of private corporations was approximately$20 billion and for governments appr

133、oximately$200 billion based on spending from the Manhattan and Apollo projects.Using a starting estimate of$10 million per training run in 2018,Carey estimated that training costs should reach the maximum spending limit for private corporations in 2021 to 2022 and for gov-ernments in 2023 to 2024.Ca

134、rey further estimated that even if new developments in AI hardware cheapened compute by 1,000 times beyond current cost-compute trends during this period,this would only add another three to four years before cost growth reached the same limits.Carey concluded that the growth rate in compute costs c

135、ould not continue beyond 3.5 to 10 years(an estimate derived in 2018).Cotra(2020)estimated compute growth over time,noting that the current growth rate is“obviously unsus-tainable in the long run”due to rising costs.41 Cotra estimated that by 2025 training costs would range from$300 million to appro

136、ximately$1 billion.Cotra further estimated that by approximately 2040,tech companies could spend hundreds of billions of dollars to train an AI model if there were sufficient economic incentives to doing so.(This equates to a doubling in cost roughly every two years.)In the long run,Cotra estimated

137、that govern-ments willingness to spend on training runs would cap out at approximately 1 percent of gross domestic product(GDP)for a major country,such as the United States.By assuming cost growth would taper over time,Cotra pro-jected that costs could continue to rise for several decades,albeit at

138、a slower pace than the current rate.Lohn and Musser(2022)arrived at a similar conclu-sion to Carey(2018),determining that current growth rates in compute were not sustainable beyond two to three years.42 Using the 3.4-month compute doubling period from Amodei and Hernandez(2018),43 and assuming that

139、 compute per dollar doubles roughly every two to four years due to hardware improvements,Lohn and Musser estimated that compute costs,on their current trajectory,would eclipse total U.S.GDP by June 2026 or May 2027 at the latest.44 Lohn and Musser further argued that in addition to cost,hardware ava

140、ilability and the engineering challenges associated with training massive models would limit compute growth.Lohn and Musser concluded that the 3.4-month compute doubling rate is not sustainable and suggested it already may be slowing.Indeed,subsequent analysis of compute trends has found slower rate

141、s of compute growth.Sevilla et al.(2022)arrived at a revised estimate for compute growth based on observations of 98 state-of-art machine learning models from 2010 to 2022.45 They updated Amodei and Hernandezs estimate and determined that compute doubled approximately every six months from 2010 to 2

142、022.They further found that around late 2015,a new trend of large-scale models emerged,with large models using two to three orders of magnitude more compute than previous state-of-the-art models but growing at a slower rate,doubling approximately every 10 months.Sevilla et al.also estimated a 20-mon

143、th doubling rate for historical machine learning models in the predeep learning era before 2010,similar to the 24-month doubling rate typically associated with Moores Law.Besiroglu et al.(2022)projected compute growth(but not cost)using the six-month doubling rate identified in Sevilla et al.(2022).

144、Based on previous work by Carey(2018)and Lohn and Musser(2022),which concluded that the current growth rate in compute was unsustain-able,Besiroglu et al.explored three different scenarios based on how long compute continued doubling every six months before reverting back to the predeep learning era

145、 doubling rate associated with Moores Law.They estimated that the six-month doubling rate could continue for 8 to 18 years,depending on hardware improvements decreasing the cost of compute.A major uncertainty in estimating cost growth trends is the rate of hardware improvement in compute per dollar.

146、Hobbhahn and Besiroglu(2022)analyzed 470 GPUs released between 2006 to 2021 and found an approximately 2.5-year doubling period for floating-point operations per second per dollar(FLOP/s per dollar),a measure of performance per dollar for GPUs.46Improvements in algorithms can affect how efficiently

147、models use compute.CNASDC12Cottier(2023)developed updated estimates of cost growth based on these new measurements of compute growth and hardware improvements.47 Relying on an updated version of the dataset on training compute for milestone AI systems in Sevilla et al.(2022),Cottier adjusted for imp

148、rovements in hardware performance via two methods.First,by simply subtracting the overall GPU price-performance,or compute per dollar,trend identified in Hobbhahn and Besiroglu from the compute growth rate.And second,by using the actual price-per-formance of the GPUs used for training specific syste

149、ms.Using the first method,Cottier identified a historical cost growth rate of 0.2 orders of magnitude per year(OOMs/yr),or a doubling of training costs approximately every 18 months.Cottier used this growth rate to project future costs,estimating that the cost of training the largest models could ec

150、lipse$200 billion by 2040.Heim(2023)projected compute costs using a similar methodology,starting with an estimated$9 million to train PaLM in 2022.48 Heim projected that at current trends,training costs would eclipse U.S.GDP in the mid-2030s and that this growth rate was not sustainable.Subsequent a

151、nalysis of hardware performance and compute growth allows further refinement of these estimates.Hobbhahn et al.(2023)assessed hardware performance growth using a dataset of nearly 2,000 GPUs from 2001 to 2021 and 47 machine learning accel-erators(GPUs and other AI chips)from 2010 to 2023.They estima

152、ted a 2.5-year doubling period for FLOP/s per dollar for general GPUs and a 2.1-year doubling period for machine learning GPUs,revising earlier estimates by Hobbhahn and Besiroglu(2022).49 Based on an updated dataset of 47 large-scale machine learning models trained from 2015 to 2023,Epoch(2023)esti

153、-mated a 7.0-month doubling period for compute growth for large models,revising the estimates in Sevilla et al.(2022).50 These new figures allow for updated cost growth estimates and projections,which this paper presents in Cost and Compute Projections.Improvements in algorithms can affect how effic

154、iently models use compute.Improved algorithmic efficiency can decrease the amount of compute needed to train a model at the same level of performance over time.Hernandez and Brown(2020)found that the amount of compute needed to train a model to the same level of performance on ImageNet reduced by 44

155、-fold from 2012 to 2019,corresponding to a doubling in algorithmic efficiency for image classifiers on ImageNet every 16 months.51 Erdil and Besiroglu(2022)similarly esti-mated algorithmic progress on ImageNet and arrived at a somewhat faster algorithmic efficiency doubling rate of every 8.95 months

156、.52 Approximately similar rates of improvement in algorithmic efficiency have been found in other machine learning domains.Dorner(2021)estimated algorithmic efficiency in deep reinforcement learning was doubling every 10 to 18 months on Atari games,every 5 to 24 months on state-based continuous cont

157、rol,and every 4 to 9 months on pixel-based contin-uous control.53 More recently,Ho et al.(forthcoming)found algorithmic efficiency for large language models was doubling approximately every 8.4 months.54 Grace(2013)found roughly similar rates of algorithmic progress across six different AI research

158、fields.55In projecting future compute resources,some analysts have included improvements in algorithmic efficiency that increase the amount of effective compute available to train models.Cotra(2020)and Hobbhahn(2022)factor this into their calculations of future compute,projecting the effective compu

159、te available for future projects.Cotra(2020)estimated algorithmic efficiency doubling every 2 to 3 years,while Hobbhahn(2022)estimated algorithmic efficiency doubling every 1.3 to 1.6 years.56 Cotra used a more conservative estimate of algorithmic efficiency than the faster rate observed for ImageNe

160、t,under the assumption that researchers had“strong feedback loops”for ImageNet but are likely to make slower progress when breaking ground on new models.57 Other compute growth projections did not include algorithmic progress.58 In addition to compute and algorithmic progress,the availability of dat

161、a is also a significant factor in scaling model performance.Recent assessments by Villalobos and Ho(2022)on the rate of growth in dataset size allow for estimates about how data availability may affect compute growth over time.59 These are discussed in Appendix A:Additional Limitations on Compute Gr

162、owth.A summary of recent observed rates of growth in relevant metrics is shown in Appendix B:Observed Growth Rates.Current Best Estimates and Assumptionshis paper projects cost,compute,and effective compute for frontier AI models using the following best estimates:Compute GrowthEpoch(2023)assessed t

163、hat compute used for training state-of-the-art machine learning models has been doubling every 6.3 months(95 percent confidence interval CI:5.5 to 7.2 months)since 2010.It assessed TTECHNOLOGY&NATIONAL SECURITY|MARCH 2024Future-Proofing Frontier AI Regulation:Projecting Future Compute for Frontier A

164、I Models 13that for the largest models,training compute has been doubling every 7.0 months(95 percent CI:5.7 to 8.6 months)since 2015.60 Since it is the cost for the largest models that is of interest,this paper uses the 7.0-month doubling period as the baseline assumption for compute growth.Hardwar

165、e Performance Chips used for training are improving over time,enabling better performance per dollar.Hobbhahn et al.(2023)estimated that price-performance for machine learning GPUs has been doubling every 2.1 years(95 percent CI:1.6 to 2.91 years).This paper starts with a baseline assumption that GP

166、U price-performance continues to double every 2.1 years(25.2 months).This assumption will be changed in an excursion scenario that explores limits on hardware improvement.Cost GrowthSince few AI papers publish cost figures,cost growth has not been directly measured.However,the rate of cost growth ca

167、n be estimated by calculating the difference between observed compute growth(doubling every 7.0 months)and improvements in compute per dollar(doubling every 2.1 years).This is the same method-ology used by Cottier(2023),updated with the most recent estimates.61 This yields an estimate of cost growth

168、 doubling every 9.7 months.62 This paper uses an estimate of training costs doubling every 9.7 months(95 percent CI:7.3 to 13.5 months)for baseline projections for cost growth.63 This assumption will be changed in excursions that explore a tapering rate of cost growth as costs approach the limits of

169、 private companies.Algorithmic EfficiencyHo et al.(forthcoming)estimated that algorithmic effi-ciency for large language models is doubling every 8.4 months(95 percent CI:5.3 to 13 months).64 Since large language models currently represent the largest,general-purpose AI models,this paper uses the 8.

170、4-month doubling rate as the baseline assumption for improvements in algorithmic efficiency for any given model.This is used to project the amount of effective compute available for training future frontier models.It also is used to estimate improvements in algorithmic efficiency that decrease the a

171、mount of compute needed to train a model to any given level of performance over time,increasing the proliferation of capabilities.Current Costs While many details are known about recent state-of-the-art models,training costs are rarely reported.Epoch(2023)built an extensive dataset of over 500 notab

172、le machine learning systems from 1950 to 2023.65 The dataset includes,when available,parameter count,dataset size,training compute,and other relevant details.Most papers do not include cost.In some cases,costs can be estimated using the reported amount of training compute.Epoch(2023)estimated a cost

173、 of$50 million(90 percent CI:$30 million to$90 million)for the final training run for GPT-4.66 For more discussion of some of the challenges in estimating compute costs,see Appendix C:Estimating Compute Costs.This paper uses Epochs estimate of$50 million to train GPT-4 as the baseline for projecting

174、 future cost growth.Appendix D,Uncertainty in Cost Projections,shows alternate projections using different starting cost estimates.Cost projections are relatively insensitive to changes in the initial cost.A twofold change in the initial cost in either direction only leads to a twofold change in pro

175、jected cost at any point in time.By contrast,cost pro-jections are highly sensitive to errors in the rate of cost growth,since they compound over time.Current ComputeSince this paper uses GPT-4 as the starting point for projections of training cost,it relies on Epochs(2023)estimate of 2.1 1025 FLOP(

176、90 percent CI:1.1 1025 to 3.9 1025 FLOP)used to train GPT-4 as the starting point for compute for a frontier model trained in 2022,the year that GPT-4s training was completed.67 GPT-4 is an outlier,using a full order of magnitude more compute than the median trend for large models at the time it was

177、 trained.The next publicly announced model to be trained on approximately the same or greater compute was Googles Gemini Ultra,which was announced in December 2023,and which Epoch esti-mated was trained on 8 1025 FLOP.68 Using GPT-4 as a starting point for the cost and compute projections leads to p

178、rojections for the most expensive and compute-in-tensive models at any point in time,not the median for large models.AssumptionsProjecting cost and compute growth based on current trends makes several assumptions which may not hold up in reality.It assumes that current estimates of compute growth,ha

179、rdware price-performance,and model cost are at least approximately accurate.It assumes as a baseline CNASDC14that these trends continue at their current rate,although alternate projections factoring in limitations on cost and hardware improvements are also presented.It assumes no discontinuous progr

180、ess in spending,hardware performance,and algorithmic efficiency.Many of these assumptions may turn out to be incorrect.The baseline projections do not make any assump-tions about which factor(s)limit the rate of growth of compute and costs.They simply project forward cost and compute based on curren

181、t trends without an assumption of which factors are driving those trends.Many leading AI labs could afford to spend more on large-scale training runs today if they desired.The amount of compute used in large-scale training runs could be limited by several factors:available chips,the engineering chal

182、lenges associated with networking together tens of thousands of chips to train a model,and sufficient amounts of high-quality data to train a large model efficiently.For some start-ups,in addition to cost,access to the human capital needed to efficiently orchestrate large-scale training runs may be

183、a significant factor limiting their ability to train the most compute-intensive models.The baseline projections do not assume that cost is a limiting factor in compute growth.Additional scenarios incorporate cost as a limiting factor.This paper assumes that large corporations likely could marshal on

184、 the order of tens of billions of dollars in annual training costs if there was a sufficient payoff for doing so.Beyond that level,spending would likely move into the realm only historically affordable by major governments.For a dis-cussion on private sector spending limits,see Limits on Cost Growth

185、.The baseline projections also do not account for limits in continued hardware performance improve-ments.An excursion scenario in Limits on Hardware Improvements accounts for potential limits in hardware performance.These projections are not intended as predictions of how cost and compute will grow

186、over time,but rather as projections of what cost and compute growth would be if they were to continue on their current trajectories.A discussion of additional limitations on compute growth,including from hardware,data,or engineering chal-lenges,is included in Appendix A:Additional Limitations on Com

187、pute Growth.Using these current best estimates and assumptions,this paper answers the research questions posed at the beginning by projecting frontier model training cost and compute over time.1516PART II:ANALYSIS16TECHNOLOGY&NATIONAL SECURITY|MARCH 2024Future-Proofing Frontier AI Regulation:Project

188、ing Future Compute for Frontier AI Models 17$10T$1T$100B$10B$1B$100MU.S.GDP($23 trillion)U.S.GDP($23 trillion)$100 billion$20 billion20222024202620282030203220342036203820402042Cost of Final Training Run(2022 dollars)FIGURE 2.1|THE COST TO TRAIN A FRONTIER AI MODEL RISES OVER TIME (STRAIGHTFORWARD P

189、ROJECTION OF CURRENT TRENDS)The cost to train a frontier AI model is projected forward over time,starting from an initial estimate of$50 million to train GPT-4 in 2022,using a 9.7-month cost doubling rate(95 percent CI:7.3 to 13.5 months).Cost and Compute Projectionsf current trends were to continue

190、,how would the amount of compute used to train frontier AI models and the cost of training rise over time?Accounting for algorithmic progress,how would the amount of effective compute increase over time?Figures 2.1 to 2.3 show a projection of current trends in cost,compute,and effective compute,resp

191、ectively,according to the current best estimates for each growth rate.GPT-4 is used as a starting point,leading to a projection for the largest models at AIs frontier of research,rather than the median large model at any given point in time.All cost projections in this paper use 2022 constant dollar

192、s.IIn Figures 2.1 to 2.3,the y-axes are logarithmic due to the exponential growth of each variable.Each tick mark on the vertical axis represents a tenfold increase in cost,compute,or effective compute.The straight lines indicate exponential growth curves.The uncertainties for each variable are show

193、n in the shaded region,with the high and low estimates representing 95 percent confidence intervals.The dashed line for each variable represents the median estimate.The uncertainties in these projections,shown in the shaded areas,are considerable.For a discussion of uncertainties in cost estimates a

194、nd projections under different initial training cost estimates,see Appendix D:Uncertainty in Cost Projections.CNASDC0262028203020322034203620382040204281103210331 millionGPT-4Training Compute(FLOP)1,000GPT-4GPT-4 estimated computeGPT-4 estimated computeEffective Comp

195、ute(2022 FLOP equivalent)2022202420262028203020322034203620382040204281103210331 millionGPT-41,000GPT-4GPT-4 estimated computeGPT-4 estimated computeFIGURE 2.2|COMPUTE USED TO TRAIN A FRONTIER AI MODEL RISES OVER TIME (STRAIGHTFORWARD PROJECTION OF CURRENT TRENDS)FIGURE 2.3|EFF

196、ECTIVE COMPUTE RISES OVER TIME (STRAIGHTFORWARD PROJECTION OF CURRENT TRENDS)Compute is projected forward over time,starting from an initial estimate of 2.1 1025 FLOP to train GPT-4 in 2022,using a 7.0-month compute doubling rate(95 percent CI:5.7 to 8.6 months).The effective compute used to train f

197、rontier models is projected forward over time,starting from an initial estimate of 2.1 1025 FLOP to train GPT-4 in 2022 and using a 7.0-month doubling rate for compute and an 8.4-month doubling rate for algorithmic efficiency(95 percent CI:5.3 to 13 months).69 19TECHNOLOGY&NATIONAL SECURITY|MARCH 20

198、24Future-Proofing Frontier AI Regulation:Projecting Future Compute for Frontier AI Models DiscussionCompute could continue to grow several more orders of magnitude before reaching the current limit of training expenditures for major corporations,likely tens of billions of dollars.Table 1.1 shows cos

199、t and compute over time,with compute represented both in absolute numbers(FLOP)and relative to GPT-4.Accounting for algorithmic progress,the amount of effective compute used for training is even higher.Table 1.2 shows cost and effective compute,with effective compute represented both in 2022 equival

200、ent FLOP and relative to GPT-4.Current cost and compute trends are not sustain-able.On their current trajectory,training costs will TABLE 1.1|TRAINING COST AND COMPUTE OVER TIME (STRAIGHTFORWARD PROJECTION OF CURRENT TRENDS)20242027203020332036Cost of final training run in 2022 dollars(9.7-month cos

201、t doubling rate)$280M$3.6B$50B$600B$8TTraining compute in FLOP (7.0-month compute doubling)2.3 10268.0 10272.8 10291.0 10313.5 1032Compute relative to GPT-410 380 13,000 500,000 17 million FindingTraining compute could increase approximately 1,000 times above GPT-4,to around 1028 FLOP,before reachin

202、g the current spending capacity of private companies(assumed to be in the tens of billions of dollars)in the late 2020s.TABLE 1.2|TRAINING COST AND EFFECTIVE COMPUTE OVER TIME (STRAIGHTFORWARD PROJECTION OF CURRENT TRENDS)20242027203020332036Cost of final training run in 2022 dollars(9.7-month cost

203、doubling rate)$280M$3.6B$50B$600B$8TEffective compute in 2022 FLOP equivalent(8.4-month algorithmic efficiency doubling)1.6 10271.1 10307.8 10325.4 10353.7 1038Effective compute relative to GPT-480 50,000 40 million 25 billion 18 trillion Finding Accounting for algorithmic progress,the amount of eff

204、ective compute used to train frontier models could increase to around one million times GPT-4,or approximately the equivalent of 1031 FLOP in 2022,before reaching the spending capacity of private companies in the late 2020s.reach the current limits of private sector actors and move into what has bee

205、n historically the realm of government-level expenditures in the late 2020s.However,compute could increase considerably before reaching that limit.The amount of compute used to train frontier models could increase on the order of 1,000 times GPT-4,or one million times in effective compute.These stra

206、ightforward projections do not account for how rising costs might affect the rate of cost growth.As costs begin to approach the historical spending limits of corporations,cost growth might reasonably slow.The next set of projections account for the spending limits of private companies in pro-jecting

207、 cost and compute growth.TECHNOLOGY&NATIONAL SECURITY|MARCH 2024Future-Proofing Frontier AI Regulation:Projecting Future Compute for Frontier AI Models 2121Limits on Cost Growthow much could compute increase before reaching the spending limits of private companies,and when would that occur?If the ra

208、te of cost growth slows as costs rise,how might that affect the amount of compute used for training frontier models?Even the wealthiest actors have limits in their spending capacity.As costs continue to rise,it is reasonable to expect that these limits will affect compute growth.Large corporations c

209、urrently have a spending capacity in the tens of billions of dollars annually for research and/or capital expenditures.TSMCs capital expendi-tures were$36 billion in 2022 and estimated to be$32 billion in 2023.70 Metas capital expenditures were$32 billion in 2022,estimated to be$27 billion to$29 bil

210、lion in 2023,and projected to be$30 billion to$35 billion in 2024.Metas main driver for increased capital expendi-tures was an increase in AI capacity.71 Amazons capital expenditures were$59 billion in 2022 and estimated to be“slightly more than$50 billion”in 2023,although Amazons figures include fu

211、lfillment and transportation costs.72 Tech corporations have spent tens of billions of dollars on speculative research projects with no imme-diate return and no guaranteed long-term return.Meta spent$36 billion on metaverse research from 2019 to 2022.73 Large corporations likely could marshal on the

212、 order of tens of billions of dollars in annual training costs if there was a sufficient payoff for doing so.Beyond that level,spending would move into the realm historically only affordable by major governments.Governments historically have had significantly greater spending capacity than corporati

213、ons and have pursued expensive technology projects when suffi-ciently motivated.The Manhattan Project to build the first atomic bomb cost the U.S.government nearly$2 billion at the time,the equivalent of more than$30 billion in 2022.74 Moreover,it was willing to undertake such a large project alongs

214、ide other major technology development efforts during World War II.The B-29 bomber program cost$3 billion in total,or the equiva-lent of approximately$45 billion in 2022,making it the largest military expenditure of the war.75 The Apollo Program was even more expensive,costing nearly$3 billion per y

215、ear during peak spending in the mid-1960s,the equivalent of$25 billion annually in 2022.76 Costs in the range of tens of billions of dollars per year for strate-gically relevant technology projects are feasible for the largest governments.Governments can marshal massive financial resources when nece

216、ssary.The U.S.Defense Departments budget is over$800 billion today,in peacetime.At the start of the Cold War in the early 1950s,the United States spent over 12 percent of its GDP annually on national defense.77 During World War II,defense spending rose to 36 percent of GDP.(For comparison,36 percent

217、 of current U.S.GDP would be$9 trillion.)Governments also can dramatically surge spending in response to strategic needs.The U.S.government sent$48 billion in aid to Ukraine during the first ten months of the war in 2022.78 The U.S.Defense Department spent over$2 trillion total in direct spending on

218、 the wars in Iraq and Afghanistan and other post-9/11 military operations.79 In response to the COVID-19 pandemic,the U.S.government engaged in a swift and massive financial response,spending around$4.6 trillion in 2020 and 2021.80 This paper does not explicitly model an upper limit on government ex

219、pen-ditures,although costs in the hundreds of billions of dollars annually are possible for the largest governments.In extremis,a trillion dollars of annual spending is,in principle,possible.Tapered Cost Growth Projection Under a straightforward projection of cost growth,training costs will reach th

220、e historical limits of private companies in the late 2020s.A more sophisticated projection would account for cost growth tapering over time as costs rise.Rather than costs doubling at a rapid rate before abruptly stopping at some limit of maximum expenditures,a more reasonable assump-tion is that th

221、e rate of cost growth slows as training costs become increasingly large and begin to push the limits of affordability.To illustrate the effect that this slowing rate of growth could have on training costs,a tapered cost growth pro-jection is presented in which the cost doubling period is arbitrarily

222、 assumed to increase by 1.5 months each year.Under this projection,as costs reach$1 billion(by 2027),the rate of cost growth has slowed to a doubling every 17 months.As costs reach$10 billion(by 2033),the rate of cost growth has slowed to a doubling every 26 months(approximately 2 years).And as cost

223、s reach$100 billion(by 2041),costs are doubling every 38 months(3.2 years).Figure 3.1 shows cost growth under such a tapered cost growth projection,with the constant 9.7-month cost doubling rate shown for comparison.HLarge corporations likely could marshal on the order of tens of billions of dollars

224、 in annual training costs.CNASDC22CNASDC222222CNASDCCNASDCCNASDCCNASDC22FIGURE 3.1|FRONTIER MODEL TRAINING COSTS (ARBITRARY TAPERED COST GROWTH PROJECTION)FIGURE 3.2|TRAINING COMPUTE (TAPERED COST GROWTH PROJECTION)The amount of compute used to train frontier AI models is projected using a tapered c

225、ost growth model.The doubling period for training costs is arbitrarily assumed to increase by 1.5 months per year,slowing the rate of cost and compute growth.A constant 7.0-month compute doubling rate is shown for comparison.Training costs are projected using a tapered cost growth model.Training cos

226、ts initially double every 9.7 months,but the doubling rate is arbitrarily assumed to increase by 1.5 months per year,slowing the rate of cost growth.A constant 9.7-month cost doubling rate is shown for comparison.$10T$1T$100B$10B$1B$100MU.S.GDP($23 trillion)U.S.GDP($23 trillion)$100 billion$100 bill

227、ion$20 billion$20 billion20222024202620282030203220342036203820402042Tapered cost growthConstant cost growthCost of Final Training Run(2022 dollars)20222024202620282030203220342036203820402042With tapered cost growthConstant compute growth81103210331 millionGPT-41,000GPT-4GPT-4

228、 estimated computeGPT-4 estimated computeTraining Compute(FLOP)TECHNOLOGY&NATIONAL SECURITY|MARCH 2024Future-Proofing Frontier AI Regulation:Projecting Future Compute for Frontier AI Models 23account for the ability of AI itself to accelerate AI R&D,such as through improved chip design or algorithmi

229、c progress.Compute under a Tapered Cost Growth ProjectionUnder a tapered cost growth projection,in which the cost doubling period is arbitrarily assumed to increase by 1.5 months per year,compute grows more slowly.Figure 3.2 shows a projection of the amount of compute used to train frontier AI model

230、s under a tapered cost growth model.A straightforward projec-tion of compute under constant cost growth is included for comparison.DiscussionThe effect of slowing cost growth is to reduce the amount of compute used to train a frontier model at any given point in time.However,this merely delays by a

231、few years the time to reach various compute thresholds.Under a straightforward projection,training compute reaches 2 1028 FLOP,or approximately 1,000 times above GPT-4,in 2028 at a cost of around$10 billion.Under a tapered cost projection,this milestone is delayed by only two years,arriving in 2030.

232、In this projection,however,hardware performance con-tinues to improve,somewhat making up for the lack of spending.This assumption may not be realistic and is removed in a subsequent scenario.Limits on Hardware Improvementsow might limits on continued hardware improvements affect future compute growt

233、h?Chips are continuing to improve,which enables greater compute per dollar over time.Machine learning GPU price-performance,or performance per dollar,has been doubling about every 2.1 years.Hardware improvements may not continue indefinitely,however.Physics-based analysis from Hobbhahn and Besiroglu

234、(2022)suggested that GPU performance will stop improving between 2027 and 2035 as transistors approach the size of roughly a single silicon atom.82 It is possible that chips continue improving through new techniques.Advanced packaging techniques,more spe-cialized AI-specific chips,or entirely new co

235、mputing paradigms could enable continued growth beyond the mid-2030s.However,one plausible scenario is that hardware improvements slow dramatically or stop Under this illustrative projection of a modest tapering of the rate of cost growth over time,costs still rise initially at a very rapid clip.The

236、 effect of this modest tapering of cost growth is that training costs do not exceed the available expenditures of large corporations until the mid-2030s.Under such a projection,training costs(and compute)continue to rise and remain within the realm of private sector actors for the next 10 to 15 year

237、s.Only until the mid-to late 2030s do costs begin to exceed the level currently affordable by large corpora-tions and shift into the realm historically only possible by major governments.(For tapered cost growth projec-tions under alternate scenarios of slowing cost growth,see Appendix E:Tapered Cos

238、t Growth Projections.)This tapered cost growth projection is not presented as a prediction.The uncertainties in the rate of growth in compute are massive,and many future paths are possible.This projection is presented merely to illus-trate,in a general sense,how a slowing rate of cost growth could a

239、ffect compute projections.It is possible to envision reasonable trajectories in which training costs and compute continue to grow,albeit at a slower rate than today,for another 10 to 15 years.Alternatively,costs could continue rising beyond the current estimate for the spending limit of private comp

240、anies(tens of billions of dollars).This could occur in one of two ways:First,governments could finance large-scale training runs.Governments could do this by establishing a government project to train large models,akin to the Manhattan Project or Apollo Program.This would be a significant shift from

241、 the role of governments in AI research today,but it could be possible if governments saw sufficient strategic value in pursuing next-genera-tion models to warrant the expenditures.Alternatively,governments could provide financial support to a select group of companies training large models to help

242、offset the cost,similar to how some governments support the construction of capital-intensive fabs today.Second,private companies conceivably could fund training runs at higher levels beyond tens of billions of dollars if the revenue they generate from AI allows greater spending.Some researchers hav

243、e taken into consideration the possibility of AI accelerating growth through increased productivity.Davidson(2023)has estimated how feedback loops on increasing invest-ment and automation in AI research and development(R&D)could accelerate AI progress.81 The projec-tions presented in this paper do n

244、ot take into account increased revenue for AI companies that could raise their spending capacity.Similarly,this paper does not HCNASDC24completely as chips reach fundamental physical limits.Under one scenario,these limits could be reached relatively quickly,in roughly the next three to ten years.Alt

245、ernatively,even if these near-term anticipated limits do not materialize,other limiting factors could emerge as chips continue to improve.Ho et al.(2023)estimated fundamental limits in microprocessor energy efficiency could be reached in around 20 years,in the mid-2040s.83(For further discussion of

246、hardware lim-itations on compute growth,see Appendix A:Additional Limitations on Compute Growth.)There is a wide range of plausible scenarios for continued hardware improvements,ranging from hard limits in the relatively near term to continued improve-ments for another 20 years or more.As an illustr

247、ative example of how hardware performance limits could affect future training compute,this paper explores a scenario in which GPU performance stops improving between 2027 and 2035,per Hobbhahn and Besiroglu(2022).This is not presented as a prediction,but rather as an example of one plausible scenari

248、o for how future limits on continued hardware improvements could affect compute growth.Figure 4.1 shows compute growth due to hardware improvements alone under two scenarios:(1)a straight-forward projection of the current 2.1-year doubling rate for machine learning GPU price-performance;and(2)hardwa

249、re improvements slowing beginning in 2027 and stopping completely by 2035.(The slowing in hardware improvements used in this projection results in a final FLOP/s per dollar level roughly equivalent to an abrupt stop in approximately 2031.)This new projection of hardware improvements,slowing beginnin

250、g in 2027 and stopping completely by 2035,can be used to update projections of training compute and effective compute,accounting for poten-tial hardware limitations.Figures 4.2 to 4.4 project frontier model training cost,compute,and effective compute,respectively,accounting for hardware limits and t

251、apered cost growth.The median estimates for cost,compute,and effective compute without any cost or hardware limits are shown for comparison.DiscussionUnder this projection,costs keep rising but compute gains are slower after around 2030 as GPU price-per-formance stalls.Effective compute continues to

252、 rise,however,since algorithmic improvements are assumed to continue.In this scenario in which cost growth slows and hardware performance gains stall,algo-rithmic improvements become the dominant driver of continued performance gains in the 2030s.This does not mean that compute ceases to be importan

253、t or that the most capable models are no longer expensive.Huge costs and massive amounts of compute remain the price of entry for training the most compute-in-tensive frontier models.Merely,as compute growth from hardware improvements and increased expen-ditures slow,algorithmic improvements continu

254、e to increase the effectiveness of the available compute,enabling better performance.Tables 2.1 and 2.2 show compute and effective compute,respectively,over time assuming limits on hardware performance and a tapered cost growth projection.These projections of effective compute assume that the rate o

255、f algorithmic efficiency improvements is constant,doubling every 8.4 months,and independent of compute growth.This assumption may not be valid.If growth in compute slows due to rising costs and/or diminishing hardware performance gains,researchers could focus more attention on algorithms,increasing

256、the rate of algorithmic improvements.Alternatively,current rapid improvements in algorithmic efficiency could be partly a function of rising compute.Researchers are able to achieve performance gains quickly by scaling compute,allowing fast feedback loops on model per-formance.Researchers then can bu

257、ild on these initial compute-driven gains by improving model efficiency through algorithmic improvements.If this is the case,then a slowdown in compute could lead to an accompa-nying slowdown in algorithmic progress.Additionally,some leading AI companies such as OpenAI,Anthropic,and Google have begu

258、n withholding details of their most advanced models.If the AI research ecosystem becomes more closed than it has been historically,and leading AI labs refrain from releasing model weights or even publishing details of their most advanced models,improvements in algorithmic efficiency for frontier mod

259、els could slow or could be confined to within leading companies.This shift may be under way already.Government regulation also could slow the rate of algo-rithmic efficiency gains,at least for frontier models,if government regulations on the most compute-intensive models prohibit or delay their rele

260、ase.84Huge costs and massive amounts of compute remain the price of entry for training the most compute-intensive frontier models.25TECHNOLOGY&NATIONAL SECURITY|MARCH 2024Future-Proofing Frontier AI Regulation:Projecting Future Compute for Frontier AI Models FIGURE 4.1|COMPUTE GROWTH DUE TO HARDWARE

261、 IMPROVEMENTS ALONE(HARDWARE PERFORMANCE LIMIT AROUND 2031)FIGURE 4.3|TRAINING COMPUTE (TAPERED COST PROJECTION AND HARDWARE LIMITS AROUND 2031)FIGURE 4.4|EFFECTIVE COMPUTE (TAPERED COST PROJECTION AND HARDWARE LIMITS AROUND 2031)FIGURE 4.2|FRONTIER MODEL TRAINING COSTS (ARBITRARY TAPERED COST GROWT

262、H PROJECTION)$10B$100B$1T$10T$1B$100M20222024202620282030203220342036203820402042Cost of Final Training Run(2022 dollars)Constant cost growthTapered cost growth$100 billionU.S.GDP($23 trillion)$20 billionConstant cost growthWith cost and hardware constraints282202420

263、2620282030203220342036203820402042Effective Compute(2022 FLOP equivalent)1 millionGPT-41,000GPT-4GPT-4 estimated compute28220242026202820302032203420362038204020421 millionGPT-41,000GPT-4GPT-4 estimated computeTraining Compute(FLOP)Hardware performance limit around20

264、31Continued hardware improvements2822024202620282030203220342036203820402042With cost and hardware constraintsConstant compute growth1 millionGPT-41,000GPT-4GPT-4 estimated computeTraining Compute(FLOP)Compute growth due to hardware improvements alone is shown under

265、two scenarios:(1)a constant 2.1-year doubling rate for machine learning GPU price-performance;and(2)hardware performance gains slowing beginning in 2027 and stopping completely by 2035,resulting in a final FLOP/s per dollar level roughly equivalent to an abrupt stop in hardware performance improveme

266、nts around 2031.Training costs are projected using a tapered cost growth model.Training costs initially double every 9.7 months,but the doubling rate is arbitrarily assumed to increase by 1.5 months per year,slowing the rate of cost growth.A constant 9.7-month cost doubling rate is shown for compari

267、son.The amount of compute used to train frontier AI models is projected using a tapered cost growth model and accounting for limits in hardware performance.The doubling period for training costs is arbitrarily assumed to increase by 1.5 months per year,slowing the rate of cost and compute growth.Har

268、dware performance gains are assumed to slow beginning in 2027 and stop completely by 2035,resulting in a final FLOP/s per dollar level roughly equivalent to an abrupt stop in hardware performance improvements around 2031.A constant 7.0-month compute doubling rate is shown for comparison.The effectiv

269、e compute used to train frontier AI models is projected using a tapered cost growth model and accounting for limits in hardware performance.A constant growth rate using a 7.0-month doubling rate for compute and an 8.4-month doubling rate for algorithmic efficiency is shown for comparison.85CNASDC26C

270、NASDC26TABLE 2.2|FRONTIER MODEL TRAINING COST AND EFFECTIVE COMPUTE OVER TIME (ASSUMING HARDWARE LIMITS AROUND 2031 AND TAPERED COST GROWTH PROJECTION,+1.5 MONTH INCREASE IN COST DOUBLING TIME PER YEAR)20242027203020332036Cost of final training run in 2022 dollars(tapered cost growth)$220M$1.2B$4.3B

271、$12B$30BEffective compute in 2022 FLOP equivalent(hardware limits around 2031,tapered cost growth)1.3 10273.7 10295.3 10314.4 10332.2 1035Effective compute relative to GPT-460 20,000 3 million 200 million 10 billion FindingAccounting for algorithmic progress,effective compute could increase by 2030

272、to approximately 1 millionfold above GPT-4,to around the equivalent of 1031 FLOP in 2022.If algorithms continue to improve,effective compute could increase by the mid-2030s to approximately 1 billionfold above GPT-4,to around the equivalent of 1034 FLOP in 2022.TABLE 2.1|FRONTIER MODEL TRAINING COST

273、 AND COMPUTE OVER TIME (ASSUMING HARDWARE LIMITS AROUND 2031 AND TAPERED COST GROWTH PROJECTION,+1.5 MONTH INCREASE IN COST DOUBLING TIME PER YEAR)20242027203020332036Cost of final training run in 2022 dollars(tapered cost growth)$220M$1.2B$4.3B$12B$30BTraining compute in FLOP (hardware limits aroun

274、d 2031,tapered cost growth)1.8 10262.6 10271.9 10288.1 10282.1 1029Compute relative to GPT-49 100 900 4,000 10,000 FindingThe amount of compute used to train frontier models continues to rise quickly for the next approximately four to five years,reaching around 100 times more than GPT-4 by 2027.By 2

275、030,however,compute growth slows significantly due to cost and hardware constraints.By 2034,compute has reached 1029 FLOP,or around 5,000 times more than GPT-4,at a cost of around$15 billion.TECHNOLOGY&NATIONAL SECURITY|MARCH 2024Future-Proofing Frontier AI Regulation:Projecting Future Compute for F

276、rontier AI Models 2727Proliferationow might improvements in hardware and algo-rithmic efficiency affect the availability of AI capabilities over time?Cost is already an obstacle to many AI researchers training the largest models.If costs continue to rise,the number of actors that can conduct the mos

277、t com-pute-intensive research will shrink further,leading to an oligopoly in frontier AI research.In the most extreme case,if costs move into the realm of major gov-ernments,the number of global actors that could afford building the most capable AI models could be very small(for example,the United S

278、tates,China,and the European Union).However,the cost to train a model at any given level of capability rapidly decreases over time due to both hardware and algorithmic improvements.At present,AI breakthroughs proliferate rapidly.It took 35 months for an open-source equivalent of AlphaFold to be rele

279、ased,14 months for an open-source equivalent of GPT-3,and seven months for an open-source equiv-alent of GPT-3.5.86 Some AI labs have switched to a limited release approach,only allowing access to their latest models through an API,slowing proliferation.As costs drop over time,however,eventually tra

280、ining costs become low enough that they become affordable to an actor willing to open-source the model.For example,the first version of Stable Diffusion,which was released open source,cost$600,000 to train.87 As model costs become low enough that they are affordable to a wider array of actors,it is

281、increasingly likely that someone releases an open-source version.Figure 5 shows how training costs for any given once-frontier model decrease over time due to improve-ments in hardware and algorithmic efficiency.Hardware improvements are assumed to lead to a doubling in GPU price-performance approxi

282、mately every 2.1 years.88 Improvements in algorithmic efficiency are assumed according to the approximately 8.4-month observed doubling rate in large language models.The cost to train a model with the equivalent capabilities of a once-frontier model decreases over time due to improvements in hardwar

283、e and algorithmic efficiency.Starting at the cost of training a frontier model in 2022($50 million),2024($300 million),2026($1.5 billion),2028($9 billion),and 2030($50 billion),the cost to train a model with equivalent capabilities is shown decreasing over time due to hardware and algorithmic improv

284、ements.Machine learning GPU price-performance is assumed to double every 2.1 years(95 percent CI:1.6 to 2.91 years)and algorithmic efficiency is assumed to double every 8.4 months(95 percent CI:5.3 to 13 months).The uncertainty in the shaded blue regions combines uncertainty for hardware and algorit

285、hmic improvements.89 H$100B$10B$1B$100M$10MCost of Final Training Run(2022 dollars)202220242026202820302032$9B$50B$1.5B$300M$50MFrontier modelFIGURE 5|CAPABILITIES OF ONCE-FRONTIER MODELS BECOME MORE ACCESSIBLE OVER TIME AS COSTS DECREASECNASDC28DiscussionCapabilities that once were limited to state

286、-of-the-art models quickly become more affordable over time.Because compute requirements halve every 8.4 months due to algorithmic improvements and hardware perfor-mance per dollar doubles every 2.1 years,the cost to train a model with equivalent performance rapidly decreases.Training a frontier mod

287、el would cost$50 million in 2022(the estimate for GPT-4);a straightforward projection of hardware and algorithmic efficiency trends predicts that training a model with the equivalent performance of GPT-4 would cost$4 million in 2024 and only$250,000 by 2026.Under this projection,training a frontier

288、model would cost$1.5 billion in 2026;training a model with the equivalent performance would cost$30 million by 2029 and$2 million by 2031.Even if models reach very high costs,exponential improvements in hardware and algorithmic efficiency make such models affordable in only a few years.A frontier mo

289、del in 2028 would cost$10 billion to train,yet the equivalent capabilities could be reached with a$160 million model in three years(2031)and a$10 million model in five years(2033).Costs for Hardware-Restricted Actorsow might constraints on hardware availability(for example,due to export controls)aff

290、ect cost and compute growth for actors denied access to con-tinued improvements in AI hardware?Access to advanced computing hardware is likely to be a limiting factor for some actors in training large models.On October 7,2022,the U.S.Commerce Department issued sweeping export controls on semiconduct

291、or technology destined for China,which the Commerce Department updated and expanded on October 17,2023.The controls limit the export to China of semiconductor manufacturing equipment and advanced GPUs,even when the chips themselves or their components are manufactured overseas,such as in Taiwan.Nvid

292、ias A100 chip,which has been used to train several milestone machine learning models,is now banned for export to China,as is Nvidias most recent H100 chip.The 2023 update additionally covers Nvidias A800 and H800 chips.If these export controls remain in place and are maximally effective,they could h

293、ave a significant effect on slowing Chinese frontier AI development over time.Chinese AI labs could attempt to circumvent these export controls through several means.Some Chinese suppliers reportedly are smuggling banned chips into China,although the scale of such efforts is unclear.92 Additionally,

294、Chinese AI researchers can still access compute through cloud pro-viders,which are not captured under current export controls,although recent moves by the U.S.government suggest it may be looking to address this gap.93 Additionally,while Chinese-headquartered companies and their overseas subsidiarie

295、s can no longer purchase export-controlled chips,creative corpo-rate restructuring may be able to circumvent these rules.94 For U.S.export controls to be effective in restraining Chinese labs ability to train frontier AI models,the U.S.government would need to put in place additional measures to cut

296、 down on large-scale smuggling of advanced chips and deny Chinese AI labs access to large-scale compute through cloud providers or data centers outside China.The immediate effect of U.S.export controls on Chinese AI development is likely marginal.Some now-banned A100 chips already were sold to Chine

297、se firms before the export controls took effect.95 Additionally,Nvidia responded to U.S.export controls by releasing the A800 and H800 chips,export-compliant versions of the banned A100 and H100,respectively.96 While the A800 and H800 now are banned based on updated rules issued by the U.S.governmen

298、t in October 2023,approximately 100,000 have been sold to Chinese firms.97 Using 800-series chips,which have lower interconnect bandwidth than the A100 and H100,would make large-scale training runs more inefficient and costly.HFINDINGEven if costs rise to the point where they severely limit access t

299、o only the wealthiest global actors,any given level of capability is still likely to proliferate rapidly within a few years as training costs decline if hardware and algorithmic advancements remain widely available.Wealthy actors training models at AIs frontier are likely to retain a relative advant

300、age over competitors,however,if they continue to train ever-larger and more compute-intensive models.Under these projections,the capabilities of once-fron-tier models become widely accessible in only a few years as training costs for any given level of compute rapidly fall due to improvements in har

301、dware and algorithmic efficiency.One or both of these assumptions may not be correct.Hardware improvements could slow or stall completely,a scenario this projection does not include.Some actors may not have access to the most advanced chips,for example,due to export controls.Algorithmic efficiency g

302、ains may not be widely available if frontier labs restrict information about their models,as some are doing already.On the other hand,even if frontier labs restrict information,the open-source community can still improve algorithms using smaller models.90 And some information about frontier models m

303、ay leak out.91 As in other areas of projecting AI trends,there is tremen-dous uncertainty about how rapidly training costs might fall in the future,enabling proliferation.TECHNOLOGY&NATIONAL SECURITY|MARCH 2024Future-Proofing Frontier AI Regulation:Projecting Future Compute for Frontier AI Models 29

304、29Even if the immediate effect is small,if U.S.export controls remain in place and can be effectively enforced,they are likely to severely restrict Chinas access to leading-edge GPUs in the coming years.U.S.Commerce Department officials have said that they anticipate the current threshold for restri

305、cted GPUs will remain in place,even as chips continue to advance,preventing Chinese labs from accessing future,more efficient GPUs.98 Moreover,because U.S.export controls also target the manufacturing equipment needed for advanced Chinese fabs,China also will be hindered in its ability to manufactur

306、e its own advanced chips.U.S.export controls currently prohibit the sale of U.S.-origin manufacturing equipment to leading manufacturing nodes in China.Japan and the Netherlands have adopted similar restrictions,effectively shutting China out of manufacturing its own advanced chips in the near term.

307、99(The United States,the Netherlands,and Japan collec-tively control 90 percent of the global semiconductor manufacturing equipment market.)In the long run,U.S.chip export controls will not remain effective forever.China is working to develop its own indigenous tooling capacity for manufacturing adv

308、anced chips,although it faces significant technolog-ical hurdles to doing so.China has seen some success with chips produced at SMICs 7 nm node,which are in Huaweis Mate 60 Pro phone.These chips represent a significant improvement over previous Chinese attempts to indigenously produce advanced chips

309、.100 However,there remain challenges to producing these chips in large quantities using Chinas currently available equipment(deep ultraviolet lithography),and these chips remain substantially behind TSMCs in quality.101 In addition to Chinese efforts to improve indigenous chip production,U.S.export

310、controls create incentives for global semi-conductor manufacturers to design-out U.S.components from their fabs to circumvent U.S.restrictions to sell to the Chinese market.However,U.S.export controls could hinder Chinas access to large-scale compute for at least several years.If the United States i

311、s successful in denying China access to leading-edge chips,Chinese AI research labs would be forced to use older,less advanced GPUs for training,effectively increasing the cost of large-scale training runs.Khan and Mann(2020)estimated that denying an actor access to leading-edge chips could quickly

312、make training large models“economically pro-hibitive.”102 In addition to increased costs,using older chips also is likely to increase the engineering challenges in training larger numbers of chips in parallel.It also may slow AI research by forcing longer training runs and slowing iteration on model

313、 development.The effects of denying actors access to advanced chips is likely to be complex and multifaceted.Hardware-restricted actors certainly will not be able to overcome all these hurdles simply by paying more in compute.However,one effect is likely to increase compute costs for accessing state

314、-of-the-art models.Figure 6.1 projects costs for“hardware-restricted actors”who are denied access to continued hardware improvements,such as due to export controls.This projection assumes that export controls are maxi-mally effective,denying hardware-restricted actors from any further advances in ch

315、ips.The projection assumes hardware-restricted actors are unable to access hardware improvements beyond 2020,since the now-banned Nvidia A100 was released in 2020.As such,this projection estimates an upper bound on cost for hardware-restricted actors.In practice,improvements in Chinese indigenous ch

316、ip fabrication could enable China to close the cost gap somewhat.These projec-tions estimate the cost for hardware-restricted actors to attempt to keep pace with frontier models,using the naive assumption that hardware-restricted actors can access larger amounts of compute by simply paying more,usin

317、g more older model chips for longer training runs.This simplification assumes away significant engi-neering challenges and limitations in scaling compute using older model chips and likely underestimates the engineering challenges for hardware-restricted actors.Under a scenario of maximally effectiv

318、e export controls,the cost for hardware-restricted actors to keep pace with frontier AI model development quickly becomes unfea-sible.The uncertainty in the projection,indicated in the shaded blue region,is due to uncertainty in the rate of hardware improvements,which hardware-restricted actors are

319、denied.Advances in algorithmic efficiency will make training models with the equivalent capabilities of once-frontier models affordable within only a few years.CNASDC30CNASDC30CNASDC30FIGURE 6.1|RESTRICTIONS ON HARDWARE MAKE IT UNAFFORDABLE TO KEEP PACE WITH FRONTIER MODELS The cost for hardware-res

320、tricted actors to train a frontier AI model without access to any hardware improvements after 2022(e.g.,due to export controls)is shown,with the cost to train a frontier AI model using the latest hardware included for comparison.The uncertainty in the shaded blue region is solely from uncertainty in

321、 the rate of growth in hardware improvements(95 percent CI),which hardware-restricted actors are denied.$100B$10B$1B$100M$10MCost of Final Training Run(2022 dollars)202220242026202820302032Cost for hardware-restricted actorsFrontier model(using latest hardware)TECHNOLOGY&NATIONAL SECURITY|MARCH 2024

322、Future-Proofing Frontier AI Regulation:Projecting Future Compute for Frontier AI Models 3131FIGURE 6.2|CAPABILITIES PROLIFERATE TO HARDWARE-RESTRICTED ACTORS DUE TO ALGORITHMIC IMPROVEMENTSEven without access to hardware improvements after 2022,hardware-restricted actors are assumed to be able to be

323、nefit from algorithmic progress.The cost to train a model with the equivalent capabilities of a once-frontier model decreases over time due to improvements in algorithmic efficiency.Starting at the cost for hardware-restricted actors to train a frontier model in 2022($100M),2024($1B),2026($10B),and

324、2028($120B),the cost to train a model with equivalent capabilities is shown decreasing over time due to algorithmic improvements.Algorithmic efficiency is assumed to double every 8.4 months(95 percent CI:5.3 to 13 months).$100B$10B$1B$100M$10M202220242026202820302032$10B$10B$1B$1B$100M$100M$120B$120

325、BCost of Final Training Run(2022 dollars)Cost for hardware-restricted actorsFrontier model(using latest hardware)DiscussionUsing the simplistic assumption that hardware-restricted actors cannot access hardware improvements after 2020,this projection estimates a modest present-day twofold cost penalt

326、y for hardware-restricted actors to attempt to keep pace with present-day frontier AI models.This is not presented as an accurate assessment of the current cost to train a GPT-4 level model with export-compliant chips and does not reflect the practical engineering challenges in doing so.Rather,this

327、projection shows that,even if hardware-restricted actors attempted to keep pace with frontier AI models,and even if the engineering challenges in doing so were assumed away,costs quickly would become unaffordable.By 2024,the effect of denying an actor access to all hardware improvements after 2020 i

328、s a nearly fourfold increase in cost.A frontier model that normally would cost approximately$280 million to train in 2024 would cost a hardware-restricted actor approximately$1 billion.By 2025,the difference is a fivefold increase in cost.An approximately$650 million frontier model in 2025 would cos

329、t a hardware-restricted actor more than$3 billion.As costs rise,a more likely effect of hardware restric-tions is only to delay the time when hardware-restricted actors are able to access capabilities until improvements in algorithmic efficiency make models more affordable.Even though hardware-restr

330、icted actors cannot afford to keep pace with frontier models(assuming export controls are maximally effective),capabilities neverthe-less proliferate rapidly due to algorithmic improvements.Figure 6.2 estimates how costs for any given level of capabilities fall over time due to algorithmic improve-m

331、ents,enabling proliferation of capabilities to hardware-restricted actors.The cost to train a model with equivalent perfor-mance to a once-state-of-the-art$50 million model in 2022(i.e.,GPT-4 level model)rapidly declines to$7 million by 2024 due to algorithmic improvements alone.To train the equival

332、ent of a$1.5 billion frontier model in 2026 would cost a hardware-restricted actor initially over$10 billion,even assuming they could solve the associated engineering challenges.Yet the cost to train a model with equivalent capabilities declines to around$50 million in only four years,by 2030.Even i

333、f rising costs force hardware-restricted actors to fall behind the frontier of AI research for the largest models,the time delay is likely to be only a few years.Even costs for very expensive models decline rapidly.Training the equivalent of an approximately$10 billion frontier model in 2028 initially would cost a hardware-re-stricted actor a whopping$120 billion.However,CNASDC32CNASDC32CNASDC32al