2024哈佛大學(xué)網(wǎng)絡(luò)安全公共 AI 案例

TheNationalSecurityCaseforPublicAI

共AI案例Tableof WhyAIMattersforNational PublicGovernanceof HowPublicAIEnhancesNational PublicAIforPublic AdvancingSafeandDemocratic

目 AIAIAI Inarecentop-edintheWashingtonPost,OpenAICEOSamAltmanposedasimplebutstrikingquestion:“WhowillcontrolthefutureofAI?”Altmanframesthechoiceasbetweentwofutures:“WillitbeoneinwhichtheUnitedStatesandalliednationsadvanceaglobalAIthatspreadsthetechnology’sbenefitsandopensaccesstoit,orancementandexpandtheirpower?Thereisnothirdoption—andit’stimetodecidecompanieslikehisownareourdemocraticbulwarkagainstatechno-authoritarianfutureinwhichChinaispredominant.NationalsecurityandforeignpolicyargumentslikethisonehavebecomeincreasinglycommoninAIandtechnologypolicyconversations.Thebasiccontoursoftheargumentgosomethinglikethis:TheUnitedStatesneeds–andshoulddependon–itsleadingcompaniestomaintaintheAIinnovationedgeandestablishdominanceinAIinordertowintheglobalcompetitionwithChina.Anythingthatmightrestraintheseleadingcompanies(i.e.regulation,antitrustenforcement,orothergovernmentactionsinthespace—withthenotableexceptionofmassivepublicR&DinvestmentthatAmericanindustrycaneventuallycommercialize)willcausetheUnitedStatestolosetoInthispaper,weargueinsteadthatthereisabetterwaytoensureartificialintelligenceadvancesU.S.nationalsecurity:publicAI.BypublicAI,wemeantwothings:publicly-provided,-ownedand-operatedlayersintheAItechstack,suchascloudinfrastructure,data,andmodeldevelopment;andpublicutility-styleregulationoftheprivateAIindustrythatfosterscompetitionandpreventsabusesofpower.Intheprocess,weshowthatrelyingonunregulatedAInationalchampions2—anunbridledSiliconValley—carriesconsiderablerisksfornationalsecurity.1SamAltman,WhoWillControlTheFutureofAI?,WASH.POST(July5,2024),2Byunregulated,wedonotmeanliterallyunregulated.Obviously,standardAmericanlaws—workplacesafety,tax,corporategovernance,andotherlawsandregulations—apply.Rather,wemeanregulationsgoverningmarketstructure.Thisisacriticaldistinctionbecausethefirms’statusasnationalchampionsisabouttheirdominanceandmarket

在最近的一篇社論中，OpenAI首席執(zhí)行官山姆?奧特曼提出了一個簡單但引人注目的問題誰將控制人工智能的未來奧特曼將選擇框架為兩種未來它將是一個美第三種選擇——現(xiàn)在是時候決定要走哪條路了。奧特曼的二分框架隱含著，硅谷以及制這些領(lǐng)先公司的因素（國家冠軍——一個無拘無束的硅谷——對國家安全來說風(fēng)險(xiǎn)很大。1山姆?奧特曼，oilotolTeFtueof，S.PST（2024年7月5日），htts://.ashngtonost.om/onons/2024/07/25/sam-atman-a-emoay-authotaansm-futue/.2我們所說的“未受監(jiān)管”并非字面意義上的未受監(jiān)管。顯然，標(biāo)準(zhǔn)的美國法律——工作場所安全、稅收、公司治理以及其他法律和法規(guī)——都適用。我們的意思是指監(jiān)管市場結(jié)構(gòu)的法規(guī)。這是一個重要的區(qū)別，因?yàn)槠髽I(yè)作為國家冠軍的地位關(guān)乎它們的統(tǒng)治地位和市場力量。Thepaperproceedsinfourparts.InPartI,wediscusswhyAImattersfornationalsecurity.PartIIoutlineswhatwemeanbypublicAI—robustpublicoptionsandcapacityforaspectsofthetechstack,andpublic-utilitystyleregulation.PartIIImakesthecaseforhowpublicAIwillenhancenationalsecurity,especiallycomparedtoasystemofunregulatednationalchampions.PartIVaddressesafewadditionalcriticisms.Onecaveatisworthnoting:wedonotaddresstheongoingdebateoverthenationalsecurityrisksandmeritsofclosedversusopensourceAIfoundationmodeldevelopment.Althoughwerecognizetheimportanceofthatdebate,wefocusexclusivelyontheU.S.government’sAIcapacityandonpublic-utilitystyleregulations.WhyAIMattersforNationalThereisabroadconsensuswithinthebipartisannationalsecurityestablishmentthatAmericannationalsecurityrequirestheUnitedStatestoremainonthecuttingedgeofAIinnovationandapplications.3ThebipartisanNationalSecurityCommissiononAI(NSCAI),whichexemplifiesthisconsensus,concludedinitsfinalreportin2021:ThepaceofAIinnovationisnotflat;itisaccelerating.IftheUnitedStatesdoesnotact,itwilllikelyloseitsleadershippositioninAItoChinainthenextdecadeandbecomemorevulnerabletoaspectrumofAI-enabledthreatsfromahostofstateandnon-stateactors.…WeknowadversariesaredeterminedtoturnAIcapabilitiesagainstus.WeknowacompetitorisdeterminedtosurpassusinAIleadership.WeknowAIisacceleratingbreakthroughsinawidearrayoffields.WeknowthatwhoevertranslatesAIdevelopmentsintoapplicationsfirstwillhavetheadvantage.4Therearethreeprimaryreasonsforthisgeotechnologicalimperative.FirstisthecriticalroleofAIingivingAmericatheedgeinitsglobalcompetitionwithChina.5This

場辯論的重要性，但我們專注于美國政府的AI能力以及公共事業(yè)風(fēng)格的監(jiān)管。人工智能對國家安全的重要性的前沿。3代表這一共識的兩黨國家安全人工智能委員會（NSCAI）2021優(yōu)勢方面發(fā)揮著關(guān)鍵作用。5這NSCAIREPORT];AshleyCarnahan,FormerHouseChinahawkwarnsAmericansaboutthedangersoftheCCP'sgrowingtechnologicaldominance,FOXNEWS(Sept.24,2024),https:///media/former-house-china-hawk-warns-americans-about-dangers-ccps-growing-technological-dominance(onformerRepublicancongressmanMikeGallagher’sviewsontheimportanceofAIfornationalsecurity);RishiIyengar,TheTechnocrat,FOREIGNPOL’Y(Aug16,2024),/2024/08/16/gina-raimondo-us-china-tech-competition-chips-ai/(onDemocraticCommerceSecretaryGinaRaimondo’sviewsontheimportanceofAIfornationalsecurity).5Foralistofthesearguments,andrelatedones,seeTrackingtheUSandChinaAIArmsRace,AINOWINST.(Apr.11,2023),/publication/tracking-the-us-and-china-ai-arms-race.

NSCAIREPORT];AshleyCarnahan,FormerHouseChinahawkwarnsAmericansaboutthedangersoftheCCP'sgrowingtechnologicaldominance,FOXNEWS(Sept.24,2024),https:///media/former-house-china-hawk-warns-americans-about-dangers-ccps-growing-technological-dominance(onformerRepublicancongressmanMikeGallagher’sviewsontheimportanceofAIfornationalsecurity);RishiIyengar,TheTechnocrat,FOREIGNPOL’Y(Aug16,2024),/2024/08/16/gina-raimondo-us-china-tech-competition-chips-ai/(onDemocraticCommerceSecretaryGinaRaimondo’sviewsontheimportanceofAIfornationalsecurity).5Foralistofthesearguments,andrelatedones,seeTrackingtheUSandChinaAIArmsRace,AINOWINST.(Apr.11,2023),/publication/petitioniscomprehensiveandmultifaceted.Thetwocountriesarevyingforeconomicandmilitarypreeminence,forregionalalliesanddiplomaticadvantage,andforpoliticalandculturalinfluenceacrosstheglobe.Whichcountry’stechnologiesenablemilitarystrength,becometheplatformsforglobalcommerce,andshapepoliticsandcultureinthe21stcenturymightbedecisiveinthiscompetition.EvenifAIbecomesevenhalfastransformativeandintegraltoourlivesastechno-optimistsexpect,theUnitedStateshasanimportantnationalinterestinleadinginbothAIinnovationandsafety,which,accordingtoleadingU.S.policymakers,gohandinVicePresidentKamalaHarris,NationalSecurityAdvisorJakeSullivan,andthebipartisanNSCAIhaveallarguedthatstayingonthecuttingedgeofAIthatisrootedindemocraticvalueswithaprivacy-andrights-protectingorientationisessentialto

，并在21世紀(jì)塑造政治和文化，這可能在競爭中起到?jīng)Q定性作用。即使人工智能像技術(shù)NSCAI對于捍衛(wèi)民主和美國的全球領(lǐng)導(dǎo)地位至關(guān)重要。7defendingdemocracyandAmerica’sgloballeadership.7Keytostayingonthecuttingedge,andthereforestrengtheningnationalsecurity,isanactiverecognitionofthetechnology’snumerousdownsiderisksandharms.8ManyoftheusecasesforAIcarryconsiderableriskstocivilrightsandcivilliberties,aswellastohumanrights,democraticinstitutions,andcompetitivemarkets.TheBidenAdministration’sBlueprintforanAIBillofRightsandExecutiveOrderonArtificialIntelligencerecognizedtheserealities,andexhortedeffortsacrossthewholeofgovernmentandsocietytoaddresstheseriskssothatwecouldrealizeAI’spublicbenefits.9Iftheglobally-dominantAIsystemsthatemergeinthecomingyearsdonotupholdandfurtherthesevalues,AI’sadoptionmaythreatentheresilienceofdemocraciesaroundtheworld.Alternatively,developingademocracy-,privacy-,andrights-protectingAIcouldimproveAmericandemocracyandnationaldefenseandofferanalternativemodelforothercountriesto6See,e.g.,TharinPillay,Time100AI2024:ElizabethKelly,Director,UnitedStatesArtificialIntelligenceSafetyInstitute,TIME(Sept.5,2024),/7012783/elizabeth-kelly/(“‘Ourviewisthatsafetyenablestrust,whichenablesadoption,whichenablesinnovation,’Kellysays.”).7AndrewMacaskillandMartinCoulter,USVicePresidentHarriscallsforactionon"fullspectrum"ofAIrisks,REUTERS(Nov.1,2023),https:///technology/us-vice-president-harris-call-action-threats-ai-2023-11-01/;PatrickTucker,Sullivan:DataPrivacyKeyToAIRaceAgainstChina,DEFENSEONE(July13,2021),https:///technology/2021/07/sullivan-data-privacy-key-ai-race-against-china/183747/;seegenerallyNSCAIREPORT,supranote3.9SeeExec.OrderNo.14,110,88Fed.Reg.75,191(Oct.30,2023)[hereinafter“AI

家安全的關(guān)鍵，是積極認(rèn)識到這項(xiàng)技術(shù)的眾多下行風(fēng)險(xiǎn)和危害。8許多人工智能的應(yīng)用案政府和社會各界共同努力應(yīng)對這些風(fēng)險(xiǎn)，以便我們能夠?qū)崿F(xiàn)人工智能的公共利益。9如果6See,e.g.,TharinPillay,Time100AI2024:ElizabethKelly,Director,UnitedStatesArticialIntelligenceSafetyInstitute,TIME(Sept.5,2024),/7012783/elizabeth-kelly/(“‘Ourviewisthatsafetyenablestrust,whichenablesadoption,whichenablesinnovation,’Kellysays.”).7AndrewMacaskillandMartinCoulter,USVicePresidentHarriscallsforactionon"fullspectrum"ofAIrisks,REUTERS(Nov.1,2023),https:///technology/us-vice-president-harris-call-action-threats-ai-2023-11-01/;PatrickTucker,Sullivan:DataPrivacyKeyToAIRaceAgainstChina,DEFENSEONE(July13,2021),https:///technology/2021/07/sullivan-data-privacy-key-ai-race-against-china/183747/;seegenerallyNSCAIREPORT,supranote3.https:///issues/articial-intelligence/ai-accountability-policy-report/requisites-for-ai-accountability-areas-of-signicant-commenter-agreement/recognize-potential-harms-and-risks.9SeeExec.OrderNo.14,110,88Fed.Reg.75,191(Oct.30,2023)[hereinafter“AISecond,AIsystemshavealreadybeenused,arecurrentlybeingused,andwillonlybecomemoreimportanttowarfare,nationaldefense,andhomelandsecurity.WhateveronethinksoftheongoingconflictsinUkraineandGaza,AIhasplayedasignificantroleinboth.10SeniormilitaryanddefenseleadersalsobelievethatAIwillplayacriticalroleinmilitaryorganization,deterrence,threatanticipation,andwarfighting,11evenassomeobservethattheU.S.militaryisnotreadyforthisnewera.12ThereisalsolittledoubtthatincreasinglypowerfulAIwiththeabilitytointegratevastquantitiesofdataextremelyquicklywithheretoforeunseencapacityforproblemsolvingandpredictivepotentialwillbecomeevenmoreimportantforintelligenceanalysisandnationalsecurityplanning.Foritspart,theBidenAdministration’sEOdirectedanewnationalsecuritymemorandumtoguide“thecontinuedadoptionofAIcapabilitiestoadvancetheUnitedStatesnationalsecuritymission.”13Moreover,AIwillhaveapplicationsacrossfederalagenciesthatsupporthomelandsecurityandresilience,fromcybersecurityandcriticalinfrastructureprotectiontocounterterrorismtocounternarcotics,fromensuringpublichealth(includinginfuturepandemics)todefendingagainstforeigndisinformationandelectioninterference.Indeed,shouldadvancedAIbeusedtoattackorweakenAmerica’seconomy,military,democracy,andcriticalinfrastructureorthoseofitsallies,thefederalgovernment,stategovernmentsandindustrywillalmostcertainlyhavetouseevenbetterAIcapabilitiestodefendagainst,remediate,andcountertheseattacks.

其次，人工智能系統(tǒng)已經(jīng)被使用，目前正在被使用，并將對戰(zhàn)爭、國防和homelandsecurity越來越重要。無論人們?nèi)绾慰创秊蹩颂m和加沙地帶的持續(xù)沖突，人工智能在兩關(guān)鍵作用，1112即使有些人觀察到美國軍隊(duì)尚未準(zhǔn)備好進(jìn)入這個新時代。1213毫無疑問，越來越強(qiáng)大的人工智安全備忘錄，以指導(dǎo)繼續(xù)采用人工智能能力以推進(jìn)美國國家安全使命。”{{v8}}7此外，人工智能將在支持homelandsecurity和彈性的聯(lián)邦機(jī)構(gòu)中得到應(yīng)用，從網(wǎng)絡(luò)安24,2023),https:///articles/2023/3/24/ukraine-a-living-lab-for-ai-warfare(ontheuseofAIintheRusso-UkraineWar);GeoffBrumfiel,IsraelisUsinganAISystemtoFindTargetsinGaza.ExpertsSayIt'sJustTheStart,NPR(Dec.14,2023),experts-say-its-just-the-st(onIsrael’suseofAIinGaza).11BritMcCandlessFarmer,AIInTheMilitary:Gen.MilleyOntheFutureofWarfare,CBSNEWS(Oct.8,2023,60-minutes/(onGeneralMarkMilley’sviewoftheimportanceofAIandroboticstoAmerica’sfuturemilitarypower);U.S.DEP’TOFDEFENSE,REMARKSBYDEPUTYSECRETARYOFDEFENSEKATHLEENH.HICKSON'THESTATEOFAIINTHEDEPARTMENTOFDEFENSE'(ASDELIVERED)(2023),12See,e.g.,RajM.ShahandChristopherM.Kirchhoff,TheU.S.MilitaryIsNotReadyfortheNewEraofWarfare,N.Y.Times(Sept.13,2024),https:///2024/09/13/opinion/ai-drones-robot-https://www.armed-13AlexandraKelley,MemoonAI'sNational-SecurityImplicationsHeadsforBiden'sDesk,DEFENSEONE(July26,2024),https:///threats/2024/07/biden-receive-ai-national-security-memo-outlining-forbidden-uses-areas-innovation/398382/;AIEO,supranote

24,2023),https:///articles/2023/3/24/ukraine-a-living-lab-for-ai-warfare(ontheuseofAIintheRusso-UkraineWar);GeoBrumel,IsraelisUsinganAISystemtoFindTargetsinGaza.ExpertsSayIt'sJustTheStart,NPR(Dec.14,2023),https:///2023/12/14/1218643254/israel-is-using-an-ai-system-to-nd-targets-in-gaza-experts-say-its-just-the-st(onIsrael’suseofAIinGaza).11BritMcCandlessFarmer,AIInTheMilitary:Gen.MilleyOntheFutureofWarfare,CBSNEWS(Oct.8,2023,https:///news/articial-intelligence-in-military-general-mark-milley-future-of-warfare-60-minutes/(onGeneralMarkMilley’sviewoftheimportanceofAIandroboticstoAmerica’sfuturemilitarypower);U.S.DEP’TOFDEFENSE,REMARKSBYDEPUTYSECRETARYOFDEFENSEKATHLEENH.HICKSON'THESTATEOFAIINTHEDEPARTMENTOFDEFENSE'(ASDELIVERED)(2023),12See,e.g.,RajM.ShahandChristopherM.Kirchho,TheU.S.MilitaryIsNotReadyfortheNewEraofWarfare,N.Y.Times(Sept.13,2024),https:///2024/09/13/opinion/ai-drones-robot-https:///imo/media/doc/nds_commission_13MemoonAI'sNational-SecurityImplicationsHeadsforBiden'sDesk，DEFENSEONE（2024726 biden-receive-ai-national-security-memo-outlining-forbidden-uses-areas-innovation/398382/；AI行政命令，supra注9Thirdistheremotebutnon-negligibleconcernaboutthepotentiallyexistentialrisksofAI.SomeofthemostprominentvoicesinAI,includingCEOsoftheleadingAIlabs,haverepeatedlywarnedthatincreasinglypowerfulAIcouldposeeithercatastrophicandexistentialriskstohumanityinthecomingyears.14(Ofcourse,theycontinuetodevelopthetechnologytowardArtificialGeneralIntelligence(AGI)asquicklyaspossible,acceleratingthesepotentialrisks.)Therearewiderangingopinionsontheprobability,orevenpossibility,ofsuchrisksandso-called“emergentproperties”forfrontierAI.Nonetheless,takingtheserisksseriouslymeansthattheU.S.governmentshouldbeatthecuttingedgeofAIsafetyresearch.TofacilitateanAIfuturethatsupportsAmericannationalsecurity,theUnitedStatesshouldembracethepublicsector’sroleinbuilding,developing,andgoverningartificialintelligence.Wecallthisframework“publicAI.”PublicAIwouldmanifestintwoways:(1)developingpublicly-funded,publicly-owned,andpublicly-runAItechstacks,thatcoexistalongsideprivateones;and(2)adoptingpublic-utilitystyleregulationstolayersintheAItechstacktoensureinnovation,enhancecompetition,advancethenationalinterest,protectdemocracy,andrespecttaxpayers.ByapublicoptionforAI,wemeanpublicly-providedandpublicly-runaspectsoftheAIofthesupplychainishardware,includingphotolithographyequipmentandgraphicalprocessingunits(GPUs)orchips.Bothofthesehardwareelementsoperateinmonopolyornear-monopolyconditions,andareextremelyexpensive.Thishardwareisusedtobuildcloudinfrastructure,whichprovidesthecomputingpowerneededatscale–andonlyafewcompaniesintheworldcanpaythehighcapitalcostsrequired

第三，對人工智能可能存在的、不容忽視的生存風(fēng)險(xiǎn)表示擔(dān)憂。人工智能領(lǐng)域的一些最著名的聲音，包括領(lǐng)先人工智能實(shí)驗(yàn)室的首席執(zhí)行官，反復(fù)警告說，越來越強(qiáng)大的人工智能在未來幾年可能對人類構(gòu)成災(zāi)難性和生存風(fēng)險(xiǎn)。當(dāng)然，他們?nèi)栽诒M可能快地開發(fā)技術(shù)，朝著通用人工智能（AGI）的方向發(fā)展，從而加速這些潛在風(fēng)險(xiǎn)。）對于這種風(fēng)險(xiǎn)“”對待這些風(fēng)險(xiǎn)意味著美國政府應(yīng)該在人工智能安全研究的前沿。AIAI”AI：（1）開發(fā)由公共資金、公共所有和公共運(yùn)營的人工智能技術(shù)棧，與私人技術(shù)棧共存；（2）AI型以及應(yīng)用。供應(yīng)鏈的底部是硬件，包括光刻設(shè)備和圖形處理單元（GPU）或芯片。這14See,e.g.,BillyPerrigo,AIIsasRiskyasPandemicsandNuclearWar,TopCEOsSay,UrgingGlobalCooperation,TIME(May30,2023),/6283386/ai-risk-openai-deepmind-letter/;Statement15SeeTejasNarechaniaandGaneshSitaraman,AnAntinomonpolyApproachtoGoverningArtificialIntelligence,YaleL.&Pol’yRev.46-48(forthcoming2025),/vu-URL/wp-

14See,e.g.,BillyPerrigo,AIIsasRiskyasPandemicsandNuclearWar,TopCEOsSay,UrgingGlobalCooperation,TIME(May30,2023),/6283386/ai-risk-openai-deepmind-letter/;Statement15SeeTejasNarechaniaandGaneshSitaraman,AnAntinomonpolyApproachtoGoverningArticialIntelligence,YaleL.&Pol’yRev.46-48(forthcoming2025),/vu-URL/wp-tobuildcloudinfrastructureatscaleusingtheseGPUs.Thenextlayerinthestackisdataandmodels.VastquantitiesofdataaretherawmaterialforAItolearn,andmodelsworkbecauseofthedata–theyaretrainedonthedatatolearnandexecutespecifictasks.Thisprocessoftrainingrequiresenormousamountsofcomputepower,athighcost.Onlyaftermodelsaredevelopedareuser-facingapplicationspossible.16following:buildmorepublicdatacenters;hostandtrainAIusingpubliccloudservicesonpublicly-ownedandoperatedcloudinfrastructure;organizepublicdataforAImodeldevelopmentandbespokenationalsecurityandpublicinterestapplications;andhiresignificantAIhumantalentintothegovernment.Thislastelementiscritical:publiccapacityrequirespeople.Formanyyears,thefederalgovernment’sapproachhasbeentocontractoutsuchservices.Thedownsidesofthisapproachhavesincebeenwell-documented:dependenceonconsultingfirmsthatchargehighratesforworkofvaryingquality;highprofilefailuresliketheAffordableCareActwebsiterollout;andasappingofgovernmentalcapacitymorebroadly.17Investinginpeoplewithtechnologicalexpertisehasthepotentialtocreateavirtuouscycle:amoreaffordablemission-drivenstaffwouldnotonlybuildpublic-interestedAIsystemsforawidevarietyofpublic-usesbutcouldalsoevaluateprivatesectorAIservicesmoreaccuratelyandreducethelikelihoodthatgovernmentcontractswillsufferfromcostandqualityproblems.Notably,theUnitedStatesalreadyhasimportantnascenteffortsat“public”AI.TheNationalAIResearchResource(NAIRR)isafederalinitiativecurrentlyinapilotphase.18ItaimstoconnectU.S.researchersandeducatorstocomputational,data,andtrainingresourcesneededto“spurinnovation,increasediversityoftalent,improvecapacity,andadvancesafe,secure,andtrustworthyAIinresearchandsociety.”TheNAIRR,however,issmall-scaleintermsofthenumberofnot-for-profitprojectsitcansupport.Problematically,thepilotprogramandlargerconceptarenottrulypublic:it16ForamorethoroughdiscussionoftheAItechstack,seeid.,at8-17Fordiscussions,seeGaneshSitaramanandRamsayEyre,BuildingPublicCapacityonArtificialCOLLINGTON,THEBIGCON:HOWTHECONSULTINGINDUSTRYWEAKENSOURBUSINESSES,INFANTILIZESOURGOVERNMENTSANDWARPSOURECONOMIES(2023);JENNIFERPAHLKA,RECODINGAMERICA:WHYGOVERNMENTIS18THENATIONALARTIFICIALINTELLIGENCERESEARCHRESOURCE(NAIRR)PILOT,/(last

在規(guī)模上構(gòu)建云基礎(chǔ)設(shè)施使用這些GPU。堆棧的下一層是數(shù)據(jù)和模型。大量數(shù)據(jù)是AIAI公共擁有的和運(yùn)營的云基礎(chǔ)設(shè)施上使用公共云服務(wù)托管和訓(xùn)練AI；組織公共數(shù)據(jù)以用于AI模型開發(fā)和定制國家安全和公共利益應(yīng)用；以及聘請重要的AI人才進(jìn)入政府。AIAI值得注意的是，美國已經(jīng)在“公共”AI方面進(jìn)行了重要的初步努力。國家AI研究資源（NAIRR）是一個聯(lián)邦倡議，目前處于試點(diǎn)階段。它的目標(biāo)是連接美國的研究人員和教激發(fā)創(chuàng)新，增加人才多樣性，提高能力，并在研究和社會中推進(jìn)安全、可靠和值得信賴的AI。然而，NAIRR在支持16關(guān)于AI技術(shù)棧的更深入討論，請參閱id.，第8?21頁。17關(guān)于討論，請參閱GaneshSitaraman和RamsayEyre，BuildingPublicCapacityonArticialIntelligence，VAND.POL’YACCELERATOR（何削弱我們的企業(yè)，使我們的政府幼稚化，扭曲我們的經(jīng)濟(jì)》（2023）；JENNIFERPAHLKA，《重編試點(diǎn)項(xiàng)目，/（最后訪問日期：2024年9月22日）。ontheprivatesectorandthedominantAIfirmstocontributecomputeandotherresourcesforNAIRRresearcherstouse.19ThelegislationwhichfirstproposedtheNAIRRisstudiouslyunclearonwhethercomputeinfrastructurewillbepubliclyorprivatelyprovided.20TherecentlyannouncedDepartmentofEnergyFASSTAIinitiativeproposestouseDOE’sdata,workforceandsupercomputerstodriveAIadvancesfornationalsecurity,scientificdiscovery,energychallenges,andAIsafetyexpertise.ThePresident’sbudgethasrequested$455millionforit.21Here,too,itseemsthattheDOEwillrelyonsomeprivatesectorAIinfrastructureandpartnerships(includingcloud,datacenters,andlikelysoftwaredesigners).EveniffundingisappropriatedforFASST,itisunclearwhethertherequestedamountisenoughtoachievetheaimsoftheprogram,especiallyitsnationalsecuritymission,letaloneamoreexpansivesetofpublicinterestmissions.ThereisalsoDOD’sJointCommon(AI)Platform(JCP),launchedinMarch2021whichprovidesservices,capabilities,andresourcestobuildAImodelsfor

NAIRR最近宣布的能源部FASST人工智能倡議計(jì)劃利用能源部的數(shù)據(jù)、人力資源和超級計(jì)算機(jī)總統(tǒng)的預(yù)算已為此請求了4.55億美元。21在這里，似乎能源部也將依賴一些私營部門AI（包括云、數(shù)據(jù)中心以及可能的軟件設(shè)計(jì)師）FASST此外，還有2021年3月啟動的國防部聯(lián)合通用（AI）平臺（JCP），該平臺提供服務(wù)和資源，用于構(gòu)建AI模型以19AmbaKak,TheProblemWithPublic-PrivatePartnershipsinAI,FOREIGNPOL’Y(Feb.12,2024),20CREATEAIActof2023,S.2714,118thCong.(1stSess.2023)(“TheNAIRRshallofferresourcesthatinclude,ataminimum,allofthefollowing,subjecttotheavailabilityofappropriations:(1)Amixofcomputationalresources,including...(B)publiccloudprovidersprovidingaccesstopopularcomputationalandstorageservicesforNAIRRusers[.]”).Publiccloud,inthiscontext,refersnotonlytopublic-sectorcomputeinfrastructure,butalsotoprivately-ownedcloudinfrastructure-as-a-serviceavailabletoawiderangeofcustomers,ascontrastedwithprivatecloud,whichreferstoinfrastructureavailableonlytoanorganization’sinternalusers.Foradiscussionofthehistoryofthelegislativeproposal,seeKak,supranote19.21DEPT.OFENERGY,FY2025BUDGETINBRIEF:FY2025CONGRESSIONALJUSTIFICATION3https:///sites/default/files/2024-03/doe-fy-2025-budget-in-brief-v2.pdf;FrontiersTECH.,https:///fasst(lastvisitedSept.22,2024);AlisonSnyder,DOEaimstomove"FASST"onAIwithsweepingnewinitiative,AXIOS(May7,2024),https:///2024/05/07/artificial-intelligence-doe-faast.FormoreonfundingprojectionsforFASST,seeNyahStewart,FuelingInnovation:

19AmbaKak,TheProblemWithPublic-PrivatePartnershipsinAI,FOREIGNPOL’Y(Feb.12,2024),20CREATEAIActof2023,S.2714,118thCong.(1stSess.2023)(“TheNAIRRshalloerresourcesthatinclude,ataminimum,allofthefollowing,subjecttotheavailabilityofappropriations:(1)Amixofcomputationalresources,including...(B)publiccloudprovidersprovidingaccesstopopularcomputationalandstorageservicesforNAIRRusers[.]”).Publiccloud,inthiscontext,refersnotonlytopublic-sectorcomputeinfrastructure,butalsotoprivately-ownedcloudinfrastructure-as-a-serviceavailabletoawiderangeofcustomers,ascontrastedwithprivatecloud,whichreferstoinfrastructureavailableonlytoanorganization’sinternalusers.Foradiscussionofthehistoryofthelegislativeproposal,seeKak,supranote19.21DEPT.OFENERGY,FY2025BUDGETINBRIEF:FY2025CONGRESSIONALJUSTIFICATION3https:///sites/default/les/2024-03/doe-fy-2025-budget-in-brief-v2.pdf;FrontiersTECH.,https:///fasst(lastvisitedSept.22,2024);AlisonSnyder,DOEaimstomove"FASST"onAIwithsweepingnewinitiative,AXIOS(May7,2024),https:///2024/05/07/articial-intelligence-doe-faast.FormoreonfundingprojectionsforFASST,seeNyahStewart,FuelingInnovation:defenseandmilitaryneeds.22Ittooreliesonprivatecloudandwasdevelopedandoperatedbyaprivatecontractor.23NAIRR,FASST,andJCPareimportantstepsintherightdirection,buttheyultimatelyfallshortofbuildingpublicAIcapacityandarenotsufficientlyambitiousineitherscaleorscope.MorerobustfederalinvestmentintheinfrastructureandhumancapacityforpublicAIisneeded.PublicGovernanceofThesecond,andcomplementary,effortistoapplypublic-utilitystyleregulationstothemonopolisticoroligopolisticprivatesectorfirmsthatdominatetheAIstack.SuchAmericantraditionofpublicutilityregulationrecognizedthatinsectorsthatfeatureregulationwasnecessaryto,amongotherthings,topreventabusesofmarketpower,enhanceinnovation,ensurecompetition,protectconsumers,andguaranteetheregulatorytoolsinthistraditionarestructuralseparations,nondiscriminationrules,andrestrictionsonforeignownership.Structuralseparationsensurethatprovidersofessentialinfrastructuredonotalsoprovidetheservicesthatusethatinfrastructure.Forexample,railroadswerebannedfromalsoowningcompaniesthatprovidedgoodsthattraveledontherailroads(likecoal).Thereasonisobvious:therailroadwouldonlyserveitsownvertically-integratedcoalcompanyorwouldchargeprohibitivepricestocompetitors,therebypushingthemoutofbusiness.Acompetitivecoalsectorrequiredpreventingverticalintegrationwithrailroads.IntheAIcontext,structuralseparationscouldbeplacedbetweenchip22JointCommonFoundation(JCF),CHIEFDIGIT.ANDA.I.OFF.,https:///index.html(lastvisitedSept.22,2024)(“TheJointCommonFoundation(JCF)isasecurecloud-basedAIdevelopmentandexperimentationenvironmentthatdeliverscriticaltoolsandcapabilitiestosupporttheDoD’spursuitofanAI-readyforce.”).23JacksonBarnett,With$106Mcontract,JAICtakesmajorstepbuildingcentralAIplatformforDOD,FEDSCOOP(Aug.13,2020),/jaic-ai-development-platform-dod-joint-common-foundation-deloitte/;JacksonBarnett,Pentagon’sJointCommonFoundationAIplatformisupandrunning,FEDSCOOP(Mar.23,2021),/joint-common-foundation-ai-platform-launched/.

NAIRRFASSTJCPPublicGovernanceof22JointCommonFoundationJCF)CHIEFDIGITANDA.IOFFhttps:///index.html2024年9月22日)（“聯(lián)合通用基礎(chǔ)（JCF）是一個安全的云基礎(chǔ)AI開發(fā)和實(shí)驗(yàn)環(huán)境，為國防部追求AI?ready部隊(duì)提供關(guān)鍵工具和能力。”）。23杰克遜?巴內(nèi)特，With$106Mcontract，JAICtakesmajorstepbuildingcentralAIplatformforDOD，聯(lián)邦合作（Aug132020），https://fedscoop.com/jaic?ai?development?platform?dod?joint?commonfoundation?deloitte/；杰克遜?巴內(nèi)特，，聯(lián)邦合作（Mar.23，2021），/joint?common?foundation?ai?platform?launched/。donotleveragetheirgatekeepingpowerinonemarkettodominateotheractorsintheAIstack–includingAmerica’smilitaryandfederal,local,andstategovernments.serveallcomersneutrallywithoutfavoritismorpricediscrimination.Thisapproachensuresthatusers(whetherbusinesses,government,orothers)canaccesstheessentialservicesthattheydependontooperate,andthattheplatformscannotpickchipmakers,cloudprovidersandmodeldevelopers.Restrictionsonforeignownership,control,andinvestmentininfrastructureindustrieshavealsobeencommon–particularlywhennationalsecurity,criticalpublicservices,andresilienceareatstake,astheyarewithAI.Theserulescanbedesignedindifferentexportcontrols.IntheAIandtechcontext,similarrulesshouldbeappliedtoAIprovidersacrossthestack,dependingonthedegreetowhichthere