牛津大学:AI+超越人类编年史_第1页
牛津大学:AI+超越人类编年史_第2页
牛津大学:AI+超越人类编年史_第3页
牛津大学:AI+超越人类编年史_第4页
牛津大学:AI+超越人类编年史_第5页
已阅读5页,还剩16页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

WhenWillAIExceedHumanPerformance?EvidencefromAIExpertsKatjaGrace1,2,JohnSalvatier2,AllanDafoe1,3,BaobaoZhang3,andOwainEvans11FutureofHumanityInstitute,OxfordUniversity2AIImpacts3DepartmentofPoliticalScience,YaleUniversityAbstractAdvancesinartificialintelligence(AI)willtransformmodernlifebyreshapingtransportation,health,science,finance,andthemilitary[1,2,3].Toadaptpublicpolicy,weneedtobetteranticipatetheseadvances[4,5].HerewereporttheresultsfromalargesurveyofmachinelearningresearchersontheirbeliefsaboutprogressinAI.ResearcherspredictAIwilloutper-formhumansinmanyactivitiesinthenexttenyears,suchastranslatinglanguages(by2024),writinghigh-schoolessays(by2026),drivingatruck(by2027),workinginretail(by2031),writingabestsellingbook(by2049),andworkingasasurgeon(by2053).Researchersbelievethereisa50%chanceofAIoutperforminghumansinalltasksin45yearsandofautomatingallhumanjobsin120years,withAsianrespondentsexpectingthesedatesmuchsoonerthanNorthAmericans.TheseresultswillinformdiscussionamongstresearchersandpolicymakersaboutanticipatingandmanagingtrendsinAI.IntroductionAdvancesinartificialintelligence(AI)willhavemassivesocialconsequences.Self-drivingtech-nologymightreplacemillionsofdrivingjobsoverthecomingdecade.Inadditiontopossibleunemployment,thetransitionwillbringnewchallenges,suchasrebuildinginfrastructure,pro-tectingvehiclecyber-security,andadaptinglawsandregulations[5].Newchallenges,bothforAIdevelopersandpolicy-makers,willalsoarisefromapplicationsinlawenforcement,militarytech-nology,andmarketing[6].Toprepareforthesechallenges,accurateforecastingoftransformativeAIwouldbeinvaluable.SeveralsourcesprovideobjectiveevidenceaboutfutureAIadvances:trendsincomputinghardware[7],taskperformance[8],andtheautomationoflabor[9].ThepredictionsofAIexpertsprovidecrucialadditionalinformation.WesurveyalargerandmorerepresentativesampleofAIexpertsthananystudytodate[10,11].OurquestionscoverthetimingofAIadvances(includingbothpracticalapplicationsofAIandtheautomationofvarioushumanjobs),aswellasthesocialandethicalimpactsofAI.SurveyMethodOursurveypopulationwasallresearcherswhopublishedatthe2021NIPSandICMLconfer-ences(twoofthepremiervenuesforpeer-reviewedresearchinmachinelearning).Atotalof352researchersrespondedtooursurveyinvitation(21%ofthe1634authorswecontacted).Ourques-tionsconcernedthetimingofspecificAIcapabilities(e.g.foldinglaundry,languagetranslation),superiorityatspecificoccupations(e.g.truckdriver,surgeon),superiorityoverhumansatalltasks,andthesocialimpactsofadvancedAI.SeeSurveyContentfordetails.TimeUntilMachinesOutperformHumansAIwouldhaveprofoundsocialconsequencesifalltasksweremorecosteffectivelyaccomplishedbymachines.Oursurveyusedthefollowingdefinition:“High-levelmachineintelligence〞(HLMI)isachievedwhenunaidedmachinescanac-complisheverytaskbetterandmorecheaplythanhumanworkers.1Each

individual

respondent

estimated

the

probability

of

HLMI

arriving

in

future

years.

Taking

themean

over

each

individual,

the

aggregate

forecast

gave

a

50%

chance

of

HLMI

occurring

within

45

years

and

a

10%

chance

of

it

occurring

within

9

years.

Figure

1

displays

the

probabilistic

predictions

for

a

random

subset

of

individuals,

as

well

as

the

mean

predictions.

There

is

largeinter-subject

variation:

Figure

3

shows

that

Asian

respondents

expect

HLMI

in

30

years,

whereas

North

Americans

expect

it

in

74

years.0.000.250.500.751.0002550Yearsfrom202175100Probability

of

HLMIAggregateForecast(with95%ConfidenceInterval)RandomSubsetofIndividualForecastsLOESSFigure1:Aggregatesubjectiveprobabilityof‘high-levelmachineintelligence’arrivalbyfutureyears.EachrespondentprovidedthreedatapointsfortheirforecastandthesewerefittotheGammaCDFbyleastsquarestoproducethegreyCDFs.The“AggregateForecast〞isthemeandistributionoverallindividualCDFs(alsocalledthe“mixture〞distribution).Theconfidenceintervalwasgeneratedbybootstrapping(clusteringonrespondents)andplottingthe95%intervalforestimatedprobabilitiesateachyear.TheLOESScurveisanon-parametricregressiononalldatapoints.WhilemostparticipantswereaskedaboutHLMI,asubsetwereaskedalogicallysimilarquestionthatemphasizedconsequencesforemployment.Thequestiondefinedfullautomationoflaboras:whenalloccupationsarefullyautomatable.Thatis,whenforanyoccupation,machinescouldbebuilttocarryoutthetaskbetterandmorecheaplythanhumanworkers.ForecastsforfullautomationoflaborweremuchlaterthanforHLMI:themeanoftheindividualbeliefsassigneda50%probabilityin122yearsfromnowanda10%probabilityin20years.2Figure2:TimelineofMedianEstimates(with50%intervals)forAIAchievingHumanPer-formance.Timelinesshowing50%probabilityintervalsforachievingselectedAImilestones.Specifically,intervalsrepresentthedaterangefromthe25%to75%probabilityoftheeventoccurring,calculatedfromthemeanofindividualCDFsasinFig.1.Circlesdenotethe50%-probabilityyear.EachmilestoneisforAItoachieveorsurpasshumanexpert/professionalperformance(fulldescriptionsinTableS5).Notethattheseintervalsrepresenttheuncertaintyofsurveyrespondents,notestimationuncertainty.Respondentswerealsoaskedwhen32“milestones〞forAIwouldbecomefeasible.Thefullde-scriptionsofthemilestoneareinTableS5.Eachmilestonewasconsideredbyarandomsubsetofrespondents(n≥24).Respondentsexpected(meanprobabilityof50%)20ofthe32AImilestonestobereachedwithintenyears.Fig.2displaystimelinesforasubsetofmilestones.IntelligenceExplosion,Outcomes,AISafetyTheprospectofadvancesinAIraisesimportantquestions.WillprogressinAIbecomeexplosivelyfastonceAIresearchanddevelopmentitselfcanbeautomated?Howwillhigh-levelmachineintel-ligence(HLMI)affecteconomicgrowth?Whatarethechancesthiswillleadtoextremeoutcomes(eitherpositiveornegative)?WhatshouldbedonetohelpensureAIprogressisbeneficial?Table3rioritized

by

society

more

than

the

status

quo

(with

only

12%

wishing

for

lessEurope(n=58)NorthAmerica(n=64)0.000.250.500.75S4displaysresultsforquestionsweaskedonthesetopics.Herearesomekeyfindings:Researchersbelievethefieldofmachinelearninghasacceleratedinrecentyears.Weaskedresearcherswhethertherateofprogressinmachinelearningwasfasterinthefirstorsecondhalfoftheircareer.Sixty-sevenpercent(67%)saidprogresswasfasterinthesecondhalfoftheircareerandonly10%saidprogresswasfasterinthefirsthalf.Themediancareerlengthamongrespondentswas6years.ExplosiveprogressinAIafterHLMIisseenaspossiblebutimprobable.SomeauthorshavearguedthatonceHLMIisachieved,AIsystemswillquicklybecomevastlysuperiortohumansinalltasks[3,12].Thisaccelerationhasbeencalledthe“intelligenceexplosion.〞WeaskedrespondentsfortheprobabilitythatAIwouldperformvastlybetterthanhumansinalltaskstwoyearsafterHLMIisachieved.Themedianprobabilitywas10%(interquartilerange:1-25%).WealsoaskedrespondentsfortheprobabilityofexplosiveglobaltechnologicalimprovementtwoyearsafterHLMI.Herethemedianprobabilitywas20%(interquartilerange5-50%).HLMIisseenaslikelytohavepositiveoutcomesbutcatastrophicrisksarepossible.RespondentswereaskedwhetherHLMIwouldhaveapositiveornegativeimpactonhumanityoverthelongrun.Theyassignedprobabilitiestooutcomesonafive-pointscale.Themedianprobabilitywas25%fora“good〞outcomeand20%foran“extremelygood〞outcome.Bycontrast,theprobabilitywas10%forabadoutcomeand5%foranoutcomedescribedas“ExtremelyBad(e.g.,humanextinction).〞SocietyshouldprioritizeresearchaimedatminimizingthepotentialrisksofAI.Forty-eightpercentofrespondentsthinkthatresearchonminimizingtherisksofAIshouldbep ).UndergradRegionHLMICDFs1.004Asia(n=68)OtherRegions(n=21)02550Yearsfrom202175100Probability

ofHLMIFigure3:AggregateForecast(computedasinFigure1)forHLMI,groupedbyregioninwhichrespondentwasanundergraduate.Additionalregions(MiddleEast,S.America,Africa,Oceania)hadmuchsmallernumbersandaregroupedas“OtherRegions.〞5AsiansexpectHLMI44yearsbeforeNorthAmericansFigure3showsbigdifferencesbetweenindividualrespondentsinwhentheypredictHLMIwillarrive.BothcitationcountandsenioritywerenotpredictiveofHLMItimelines(seeFig.S1andtheresultsofaregressioninTableS2).However,respondentsfromdifferentregionshadstrikingdifferencesinHLMIpredictions.Fig.3showsanaggregatepredictionforHLMIof30yearsforAsianrespondentsand74yearsforNorthAmericans.Fig.S1displaysasimilargapbetweenthetwocountrieswiththemostrespondentsinthesurvey:China(median28years)andUSA(median76years).Similarly,theaggregateyearfora50%probabilityforautomationofeachjobweaskedabout(includingtruckdriverandsurgeon)waspredictedtobeearlierbyAsiansthanbyNorthAmericans(TableS2).Notethatweusedrespondents’undergraduateinstitutionasaproxyforcountryoforiginandthatmanyAsianrespondentsnowstudyorworkoutsideAsia.Wasoursamplerepresentative?Oneconcernwithanykindofsurveyisnon-responsebias;inparticular,researcherswithstrongviewsmaybemorelikelytofilloutasurvey.Wetriedtomitigatethiseffectbymakingthesurveyshort(12minutes)andconfidential,andbynotmentioningthesurvey’scontentorgoalsinourinvitationemail.Ourresponseratewas21%.Toinvestigatepossiblenon-responsebias,wecollecteddemographicdataforbothourrespondents(n=406)andarandomsample(n=399)ofNIPS/ICMLresearcherswhodidnotrespond.ResultsareshowninTableS3.Differencesbetweenthegroupsincitationcount,seniority,gender,andcountryoforiginaresmall.Whilewecannotruleoutnon-responsebiasesduetounmeasuredvariables,wecanruleoutlargebiasduetothedemographicvariableswemeasured.Ourdemographicdataalsoshowsthatourrespondentsincludedmanyhighly-citedresearchers(mostlyinmachinelearningbutalsoinstatistics,computersciencetheory,andneuroscience)andcamefrom43countries(vs.atotalof52foreveryonewesampled).Amajorityworkinacademia(82%),while21%workinindustry.DiscussionWhythinkAIexpertshaveanyabilitytoforeseeAIprogress?Inthedomainofpoliticalscience,along-termstudyfoundthatexpertswereworsethancrudestatisticalextrapolationsatpredictingpoliticaloutcomes[13].AIprogress,whichreliesonscientificbreakthroughs,mayappearintrin-sicallyhardertopredict.Yettherearereasonsforoptimism.Whileindividualbreakthroughsareunpredictable,longertermprogressinR&Dformanydomains(includingcomputerhardware,ge-nomics,solarenergy)hasbeenimpressivelyregular[14].Suchregularityisalsodisplayedbytrends[8]inAIperformanceinSATproblemsolving,games-playing,andcomputervisionandcouldbeexploitedbyAIexpertsintheirpredictions.Finally,itiswellestablishedthataggregatingindi-vidualpredictionscanleadtobigimprovementsoverthepredictionsofarandomindividual[15].Furtherworkcoulduseourdatatomakeoptimizedforecasts.Moreover,manyoftheAImilestones(Fig.2)wereforecasttobeachievedinthenextdecade,providingground-truthevidenceaboutthereliabilityofindividualexperts.References[1]PeterStone,RodneyBrooks,ErikBrynjolfsson,RyanCalo,OrenEtzioni,GregHager,JuliaHirschberg,ShivaramKalyanakrishnan,EceKamar,SaritKraus,etal.Onehundredyearstudyonartificialintelligence:Reportofthe2021-2021studypanel.Technicalreport,StanfordUniversity,2021.[2]PedroDomingos.TheMasterAlgorithm:HowtheQuestfortheUltimateLearningMachineWillRemakeOurWorld.BasicBooks,NewYork,NY,2021.[3]NickBostrom.Superintelligence:Paths,Dangers,Strategies.OxfordUniversityPress,Oxford,UK,2021.[4]ErikBrynjolfssonandAndrewMcAfee.TheSecondMachineAge:Work,Progress,andProsperityinaTimeofBrilliantTechnologies.WWNorton&Company,NewYork,2021.[5]RyanCalo.Roboticsandthelessonsofcyberlaw.CaliforniaLawReview,103:513,2021.6[6]TaoJiang,SrdjanPetrovic,UmaAyyer,AnandTolani,andSajidHusain.Self-drivingcars:Disruptiveorincremental.AppliedInnovationReview,1:3–22,2021.[7]WilliamD.Nordhaus.Twocenturiesofproductivitygrowthincomputing.TheJournalofEconomicHistory,67(01):128–159,2007.[8]KatjaGrace.Algorithmicprogressinsixdomains.Technicalreport,MachineIntelligenceResearchInstitute,2021.[9]ErikBrynjolfssonandAndrewMcAfee.RaceAgainsttheMachine:HowtheDigitalRevolutionIsAcceleratingInnovation,DrivingProductivity,andIrreversiblyTransformingEmploymentandtheEconomy.DigitalFrontierPress,Lexington,MA,2021.[10]SethD.Baum,BenGoertzel,andTedG.Goertzel.Howlonguntilhuman-levelai?resultsfromanexpertassessment.TechnologicalForecastingandSocialChange,78(1):185–195,2021.[11]VincentC.MüllerandNickBostrom.Futureprogressinartificialintelligence:Asurveyofexpertopinion.InVincentCMüller,editor,Fundamentalissuesofartificialintelligence,chapterpart.5,chap.4,pages553–570.Springer,2021.[12]IrvingJohnGood.Speculationsconcerningthefirstultraintelligentmachine.Advancesincomputers,6:31–88,1966.[13]PhilipTetlock.Expertpoliticaljudgment:Howgoodisit?Howcanweknow?PrincetonUniversityPress,Princeton,NJ,2005.[14]JDoyneFarmerandFrançoisLafond.Howpredictableistechnologicalprogress?ResearchPolicy,45(3):647–665,2021.[15]LyleUngar,BarbMellors,VilleSatopää,JonBaron,PhilTetlock,JaimeRamos,andSamSwift.Thegoodjudgmentproject:Alargescaletest.Technicalreport,AssociationfortheAdvancementofArtificialIntelligenceTechnicalReport,2021.[16]JoeW.Tidwell,ThomasS.Wallsten,andDonA.Moore.Elicitingandmodelingprobabilityforecastsofcontinuousquantities.Paperpresentedatthe27thAnnualConferenceofSocietyforJudgementandDecisionMaking,Boston,MA,19November2021.,2021.[17]ThomasS.Wallsten,YaronShlomi,ColetteNataf,andTracyTomlinson.Efficientlyencod-ingandmodelingsubjectiveprobabilitydistributionsforquantitativevariables.Decision,3(3):169,2021.7SupplementaryInformationSurveyContentWedevelopedquestionsthroughaseriesofinterviewswithMachineLearningresearchers.Oursurveyquestionswereasfollows:ThreesetsofquestionselicitingHLMIpredictionsbydifferentframings:askingdirectlyaboutHLMI,askingabouttheautomatabilityofallhumanoccupations,andaskingaboutrecentprogressinAIfromwhichwemightextrapolate.Threequestionsabouttheprobabilityofan“intelligenceexplosion〞.OnequestionaboutthewelfareimplicationsofHLMI.AsetofquestionsabouttheeffectofdifferentinputsontherateofAIresearch(e.g.,hardwareprogress).TwoquestionsaboutsourcesofdisagreementaboutAItimelinesand“AISafety.〞Thirty-twoquestionsaboutwhenAIwillachievenarrow“milestones〞.TwosetsofquestionsonAISafetyresearch:oneaboutAIsystemswithnon-alignedgoals,andoneontheprioritizationofSafetyresearchingeneral.Asetofdemographicquestions,includingonesabouthowmuchthoughtrespondentshavegiventothesetopicsinthepast.ThequestionswereaskedviaanonlineQualtricssurvey.(TheQualtricsfilewillbesharedtoenablereplication.)Participantswereinvitedbyemailandwereofferedafinancialrewardforcompletingthesurvey.Questionswereaskedinroughlytheorderaboveandrespondentsreceivedarandomizedsubsetofquestions.SurveyswerecompletedbetweenMay3rd2021andJune28th2021.Ourgoalindefining“high-levelmachineintelligence〞(HLMI)wastocapturethewidely-discussednotionsof“human-levelAI〞or“generalAI〞(whichcontrastswith“narrowAI〞)[3].WeconsultedallprevioussurveysofAIexpertsandbasedourdefinitiononthatofanearliersurvey[11].TheirdefinitionofHLMIwasamachinethat“cancarryoutmosthumanprofessionsatleastaswellasatypicalhuman.〞Ourdefinitionismoredemandingandrequiresmachinestobebetteratalltasksthanhumans(whilealsobeingmorecost-effective).SinceearliersurveysoftenuselessdemandingnotionsofHLMI,theyshould(allotherthingsbeingequal)predictearlierarrivalforHLMI.DemographicInformationThedemographicinformationonrespondentsandnon-respondents(TableS3)wascollectedfrompublicsources,suchasacademicwebsites,LinkedInprofiles,andGoogleScholarprofiles.Citationcountandseniority(i.e.numbersofyearssincethestartofPhD)werecollectedinFebruary2021.ElicitationofBeliefsManyofourquestionsaskwhenaneventwillhappen.Forpredictiontasks,idealBayesianagentsprovideacumulativedistributionfunction(CDF)fromtimetothecumulativeprobabilityoftheevent.Whenelicitingpointsonrespondents’CDFs,weframedquestionsintwodifferentways,whichwecall“fixed-probability〞and“fixed-years〞.Fixed-probabilityquestionsaskbywhichyearaneventhasanp%cumulativeprobability(forp=10%,50%,90%).Fixed-yearquestionsaskforthecumulativeprobabilityoftheeventbyyeary(fory=10,25,50).TheformerframingwasusedinrecentsurveysofHLMItimelines;thelatterframingisusedinthepsychologicalliteratureonforecasting[16,17].Withalimitedquestionbudget,thetwoframingswillsampledifferentpointsontheCDF;otherwise,theyarelogicallyequivalent.Yetoursurveyrespondentsdonottreatthemaslogicallyequivalent.Weobservedeffectsofquestionframinginallourpredictionquestions,aswellasinpilotstudies.Differencesinthesetwoframingshavepreviouslybeendocumentedintheforecastingliterature[16,17]butthereisnoclearguidanceonwhichframingleadstomoreaccuratepredictions.ThuswesimplyaverageoverthetwoframingswhencomputingCDFestimatesforHLMIandfortasks.HLMIpredictionsforeachframingareshowninFig.S2.8StatisticsFor

each

timeline

probability

question

(see

Figures1and

2),

we

computed

an

aggregate

distribution

by

fitting

a

gamma

CDF

to

each

individual’s

responses

using

least

squares

and

then

taking

themixture

distribution

of

all

individuals.

Reported

medians

and

quantiles

were

computed

on

thissummary

distribution.

The

confidence

intervals

were

generated

by

bootstrapping

(clustering

onrespondents

with

10,000

draws)

and

plotting

the

95%

interval

for

estimated

probabilities

at

each

year.

The

time-in-field

andcitationscomparisons

between

respondents

and

non-respondents

(Table

S3)

were

done

using

two-tailed

t-tests.

The

region

and

gender

proportions

were

done

using

two-

sided

proportion

tests.

The

significance

test

for

the

effect

of

region

on

HLMI

date

(Table

S2)

was

done

using

robust

linear

regression

using

the

R

function

rlm

from

the

MASS

package

to

do

the

regression

and

then

the

f.robtest

function

from

the

sfsmisc

package

to

do

a

robust

F-test

significance.Supplementary

Figures(a)

Top

4

Undergraduate

Country

HLMI

CDFsIndia(n=20)China(n=36)France(n=16)UnitedStates(n=53)0.000.250.500.751.0002550Yearsfrom202175100Probability

of

HLMITop4UndergradCountryHLMICDFs(b)

Time

in

Field

Quantile

HLMI

CDFsQ[1](n=57)Q[2](n=40)Q[4](n=48)Q[3](n=55)0.000.250.500.751.0002550Yearsfrom202175100Probability

of

HLMITimeinFieldQuartileHLMICDFs(c)

Citation

Count

Quartile

HLMI

CDFs0.50Q[2](n=57)Q[1](n=53)Q[3](n=65)Q[4](n=49)0.000.250.751.00092550Yearsfrom202175100Probability

of

HLMIHLMICDFByCitation

CountQuartileFigureS1:AggregatesubjectiveprobabilityofHLMIarrivalbydemographicgroup.EachgraphcurveisanAggregateForecastsCDF,computedusingtheproceduredescribedinFigure1andin“ElicitationofBeliefs.〞FigureS1ashowsaggregateHLMIpredictionsforthefourcountrieswiththemostrespondentsinoursurvey.FigureS1bshowspredictionsgroupedbyquartilesforseniority(measuredbytimesincetheystartedaPhD).FigureS1cshowspredictionsgroupedbyquartilesforcitationcount.“Q4〞indicatesthetopquartile(i.e.themostseniorresearchersortheresearcherswithmostcitations).0.000.25FramingFixed

ProbabilitiesFixed

YearsCombined100.500.751.0002550Yearsfrom202175100Probability

of

HLMIFraming

CDFsFigureS2:AggregatesubjectiveprobabilityofHLMIarrivalfortwoframingsofthequestion.The“fixedprobabilities〞and“fixedyears〞curvesareeachanaggregateforecastforHLMIpredictions,computedusingthesameprocedureasinFig.1.ThesetwoframingsofquestionsaboutHLMIareexplainedin“ElicitationofBeliefs〞above.The“combined〞curveisanaverageoverthesetwoframingsandisthecurveusedinFig.1.SupplementaryTablesS1:AutomationPredictionsbyResearcherRegionThisquestionaskedwhenautomationofthejobwouldbecomefeasible,andcumulativeproba-bilitieswereelicitedasintheHLMIandmilestonepredictionquestions.Thedefinitionof“fullautomation〞isgivenabove(p.1).Forthe“NA/Asiagap〞,wesubtracttheAsianfromtheN.Americanmedianestimates.TableS1:Medianestimate(inyearsfrom2021)forautomationofhumanjobsbyregionofundergraduateinstitutionS2:RegressionofHLMIPredictiononDemographicFeaturesWestandardizedinputsandregressedthelogofthemedianyearsuntilHLMIforrespondentsongender,logofcitations,seniority(i.e.numbersofyearssincestartofPhD),questionframing(“fixed-probability〞vs.“fixed-years〞)andregionwheretheindividualwasanundergraduate.Weusedarobustlinearregression.TableS2:RobustlinearregressionforindividualHLMIpredictionsS3:

Demographics

of

Respondents

vs.

Non-respondentsThere

were

(n=406)

respondents

and

(n=399)

non-respondents.

Non-respondents

were

randomly

sampled

from

all

NIPS/ICML

authors

who

did

not

respond

to

our

survey

invitation.

Subjects

with11QuestionEuropeN.

AmericaAsiaNA/Asia

gapFull

Automation130.8168.6104.2+64.4Retail

salesperson13.210.610.2+0.4Truck

driver46.441.031.4+9.6Surgeon18.820.210.0+10.2AI

researcher80.0123.6109.0+14.6termEstimateSEt

-statisticp-valueWald

F

-statistic(Intercept)3.650380.1732021.076350.00000458.0979Gender

=

“female”-0.254730.39445-0.645780.553200.3529552log(citation_count)-0.103030.13286-0.775460.447220.5802456Seniority

(years)0.096510.130900.737280.466890.5316029Framing

=

“fixed_probabilities”-0.340760.16811-2.027040.044144.109484Region

=

“Europe”0.518480.215232.408980.015825.93565Region

=

“M.East”-0.227630.37091-0.613690.544300.3690532Region

=

“N.America”1.049740.208495.034960.0000025.32004Region

=

“Other”-0.267000.58311-0.457880.632780.2291022missingdataforregionofundergraduateinstitutionorforgenderaregroupedin“NA〞.Missingdataforcitationsandseniorityisignoredincomputingaverages.Statisticaltestsareexplainedinsection“Statistics〞above.TableS3:Demographicdifferencesbetweenrespondentsandnon-respondents12UndergraduateregionRespondent

pro-portionNon-respondentproportionp-test

p-valueAsia0.3050.3430.283Europe0.2710.2360.284Middle

East0.0710.0630.721North

America0.2540.2210.307Other0.0150.0131.000NA0.0840.1250.070GenderRespondent

proportionNon-respondent

proportionp-test

p-valuefemale0.0540.1000.020male0.9190.8420.001NA0.0270.0580.048VariableRespondent

estimateNon-respondent

estimatestatisticp-valueCitations2740.54528.02.550.010856log(Citations)5.96.43.190.001490Years

in

field8.611.14.040.000060S4: SurveyresponsesonAIprogress,intelligenceexplosions,andAISafetyTheargumentbyStuartRussell,referredtoinoneofthequestionsbelow,canbefoundat/conversation/the-myth-of-ai#26015.T

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

评论

0/150

提交评论