牛津AI调查报告.pdf

《牛津AI调查报告.pdf》由会员分享，可在线阅读，更多相关《牛津AI调查报告.pdf（21页珍藏版）》请在淘文阁 - 分享文档赚钱的网站上搜索。

1、When Will AI Exceed Human Performance?Evidence from AI ExpertsKatja Grace1,2,John Salvatier2,Allan Dafoe1,3,Baobao Zhang3,and Owain Evans11Future of Humanity Institute,Oxford University2AI Impacts3Department of Political Science,Yale UniversityAbstractAdvances in artificial intelligence(AI)will tran

2、sform modern life by reshaping transportation,health,science,finance,and the military 1,2,3.To adapt public policy,we need to betteranticipate these advances 4,5.Here we report the results from a large survey of machinelearning researchers on their beliefs about progress in AI.Researchers predict AI

3、 will outper-form humans in many activities in the next ten years,such as translating languages(by 2024),writing high-school essays(by 2026),driving a truck(by 2027),working in retail(by 2031),writing a bestselling book(by 2049),and working as a surgeon(by 2053).Researchers believethere is a 50%chan

4、ce of AI outperforming humans in all tasks in 45 years and of automatingall human jobs in 120 years,with Asian respondents expecting these dates much sooner thanNorth Americans.These results will inform discussion amongst researchers and policymakersabout anticipating and managing trends in AI.Intro

5、ductionAdvances in artificial intelligence(AI)will have massive social consequences.Self-driving tech-nology might replace millions of driving jobs over the coming decade.In addition to possibleunemployment,the transition will bring new challenges,such as rebuilding infrastructure,pro-tecting vehicl

6、e cyber-security,and adapting laws and regulations 5.New challenges,both for AIdevelopers and policy-makers,will also arise from applications in law enforcement,military tech-nology,and marketing 6.To prepare for these challenges,accurate forecasting of transformativeAI would be invaluable.Several s

7、ources provide objective evidence about future AI advances:trends in computinghardware 7,task performance 8,and the automation of labor 9.The predictions of AI expertsprovide crucial additional information.We survey a larger and more representative sample of AIexperts than any study to date 10,11.Ou

8、r questions cover the timing of AI advances(includingboth practical applications of AI and the automation of various human jobs),as well as the socialand ethical impacts of AI.Survey MethodOur survey population was all researchers who published at the 2015 NIPS and ICML confer-ences(two of the premi

9、er venues for peer-reviewed research in machine learning).A total of 352researchers responded to our survey invitation(21%of the 1634 authors we contacted).Our ques-tions concerned the timing of specific AI capabilities(e.g.folding laundry,language translation),superiority at specific occupations(e.

10、g.truck driver,surgeon),superiority over humans at all tasks,and the social impacts of advanced AI.See Survey Content for details.Time Until Machines Outperform HumansAI would have profound social consequences if all tasks were more cost effectively accomplished bymachines.Our survey used the follow

11、ing definition:“High-level machine intelligence”(HLMI)is achieved when unaided machines can ac-complish every task better and more cheaply than human workers.1arXiv:1705.08807v1 cs.AI 24 May 2017Each individual respondent estimated the probability of HLMI arriving in future years.Taking themean over

12、 each individual,the aggregate forecast gave a 50%chance of HLMI occurring within45 years and a 10%chance of it occurring within 9 years.Figure 1 displays the probabilisticpredictions for a random subset of individuals,as well as the mean predictions.There is largeinter-subject variation:Figure 3 sh

13、ows that Asian respondents expect HLMI in 30 years,whereasNorth Americans expect it in 74 years.0.000.250.500.751.000255075100Years from 2016Probability of HLMIAggregate Forecast(with 95%Confidence Interval)Random Subset of Individual ForecastsLOESSFigure 1:Aggregate subjective probability of high-l

14、evel machine intelligence arrival byfuture years.Each respondent provided three data points for their forecast and these were fit to theGamma CDF by least squares to produce the grey CDFs.The“Aggregate Forecast”is the mean distributionover all individual CDFs(also called the“mixture”distribution).Th

15、e confidence interval was generatedby bootstrapping(clustering on respondents)and plotting the 95%interval for estimated probabilities ateach year.The LOESS curve is a non-parametric regression on all data points.While most participants were asked about HLMI,a subset were asked a logically similar q

16、uestionthat emphasized consequences for employment.The question defined full automation of labor as:when all occupations are fully automatable.That is,when for any occupation,machinescould be built to carry out the task better and more cheaply than human workers.Forecasts for full automation of labo

17、r were much later than for HLMI:the mean of the individualbeliefs assigned a 50%probability in 122 years from now and a 10%probability in 20 years.2Figure 2:Timeline of Median Estimates(with 50%intervals)for AI Achieving Human Per-formance.Timelines showing 50%probability intervals for achieving sel

18、ected AI milestones.Specifically,intervals represent the date range from the 25%to 75%probability of the event occurring,calculated fromthe mean of individual CDFs as in Fig.1.Circles denote the 50%-probability year.Each milestone is forAI to achieve or surpass human expert/professional performance(

19、full descriptions in Table S5).Note thatthese intervals represent the uncertainty of survey respondents,not estimation uncertainty.Respondents were also asked when 32“milestones”for AI would become feasible.The full de-scriptions of the milestone are in Table S5.Each milestone was considered by a ra

20、ndom subset ofrespondents(n24).Respondents expected(mean probability of 50%)20 of the 32 AI milestonesto be reached within ten years.Fig.2 displays timelines for a subset of milestones.Intelligence Explosion,Outcomes,AI SafetyThe prospect of advances in AI raises important questions.Will progress in

21、 AI become explosivelyfast once AI research and development itself can be automated?How will high-level machine intel-ligence(HLMI)affect economic growth?What are the chances this will lead to extreme outcomes(either positive or negative)?What should be done to help ensure AI progress is beneficial?

22、Table3S4 displays results for questions we asked on these topics.Here are some key findings:1.Researchers believe the field of machine learning has accelerated in recent years.We asked researchers whether the rate of progress in machine learning was faster in thefirst or second half of their career.

23、Sixty-seven percent(67%)said progress was faster in thesecond half of their career and only 10%said progress was faster in the first half.The mediancareer length among respondents was 6 years.2.Explosive progress in AI after HLMI is seen as possible but improbable.Someauthors have argued that once H

24、LMI is achieved,AI systems will quickly become vastlysuperior to humans in all tasks 3,12.This acceleration has been called the“intelligenceexplosion.”We asked respondents for the probability that AI would perform vastly betterthan humans in all tasks two years after HLMI is achieved.The median prob

25、ability was10%(interquartile range:1-25%).We also asked respondents for the probability of explosiveglobal technological improvement two years after HLMI.Here the median probability was20%(interquartile range 5-50%).3.HLMI is seen as likely to have positive outcomes but catastrophic risks arepossibl

26、e.Respondents were asked whether HLMI would have a positive or negative impacton humanity over the long run.They assigned probabilities to outcomes on a five-pointscale.The median probability was 25%for a“good”outcome and 20%for an“extremelygood”outcome.By contrast,the probability was 10%for a bad o

27、utcome and 5%for anoutcome described as“Extremely Bad(e.g.,human extinction).”4.Society should prioritize research aimed at minimizing the potential risks of AI.Forty-eight percent of respondents think that research on minimizing the risks of AI shouldbe prioritized by society more than the status q

28、uo(with only 12%wishing for less).Asia(n=68)Europe(n=58)North America(n=64)Other Regions(n=21)0.000.250.500.751.000255075100Years from 2016Probability of HLMIUndergrad Region HLMI CDFsFigure 3:Aggregate Forecast(computed as in Figure 1)for HLMI,grouped by region inwhich respondent was an undergradua

29、te.Additional regions(Middle East,S.America,Africa,Oceania)had much smaller numbers and are grouped as“Other Regions.”4Asians expect HLMI 44 years before North AmericansFigure 3 shows big differences between individual respondents in when they predict HLMI willarrive.Both citation count and seniorit

30、y were not predictive of HLMI timelines(see Fig.S1 andthe results of a regression in Table S2).However,respondents from different regions had strikingdifferences in HLMI predictions.Fig.3 shows an aggregate prediction for HLMI of 30 years forAsian respondents and 74 years for North Americans.Fig.S1

31、displays a similar gap between thetwo countries with the most respondents in the survey:China(median 28 years)and USA(median76 years).Similarly,the aggregate year for a 50%probability for automation of each job we askedabout(including truck driver and surgeon)was predicted to be earlier by Asians th

32、an by NorthAmericans(Table S2).Note that we used respondents undergraduate institution as a proxy forcountry of origin and that many Asian respondents now study or work outside Asia.Was our sample representative?One concern with any kind of survey is non-response bias;in particular,researchers with

33、strongviews may be more likely to fill out a survey.We tried to mitigate this effect by making thesurvey short(12 minutes)and confidential,and by not mentioning the surveys content or goalsin our invitation email.Our response rate was 21%.To investigate possible non-response bias,we collected demogr

34、aphic data for both our respondents(n=406)and a random sample(n=399)of NIPS/ICML researchers who did not respond.Results are shown in Table S3.Differencesbetween the groups in citation count,seniority,gender,and country of origin are small.While wecannot rule out non-response biases due to unmeasure

35、d variables,we can rule out large bias due tothe demographic variables we measured.Our demographic data also shows that our respondentsincluded many highly-cited researchers(mostly in machine learning but also in statistics,computerscience theory,and neuroscience)and came from 43 countries(vs.a tota

36、l of 52 for everyone wesampled).A majority work in academia(82%),while 21%work in industry.DiscussionWhy think AI experts have any ability to foresee AI progress?In the domain of political science,along-term study found that experts were worse than crude statistical extrapolations at predictingpolit

37、ical outcomes 13.AI progress,which relies on scientific breakthroughs,may appear intrin-sically harder to predict.Yet there are reasons for optimism.While individual breakthroughs areunpredictable,longer term progress in R&D for many domains(including computer hardware,ge-nomics,solar energy)has bee

38、n impressively regular 14.Such regularity is also displayed by trends8 in AI performance in SAT problem solving,games-playing,and computer vision and could beexploited by AI experts in their predictions.Finally,it is well established that aggregating indi-vidual predictions can lead to big improveme

39、nts over the predictions of a random individual 15.Further work could use our data to make optimized forecasts.Moreover,many of the AI milestones(Fig.2)were forecast to be achieved in the next decade,providing ground-truth evidence aboutthe reliability of individual experts.References1 Peter Stone,R

40、odney Brooks,Erik Brynjolfsson,Ryan Calo,Oren Etzioni,Greg Hager,JuliaHirschberg,Shivaram Kalyanakrishnan,Ece Kamar,Sarit Kraus,et al.One hundred yearstudy on artificial intelligence:Report of the 2015-2016 study panel.Technical report,StanfordUniversity,2016.2 Pedro Domingos.The Master Algorithm:Ho

41、w the Quest for the Ultimate Learning MachineWill Remake Our World.Basic Books,New York,NY,2015.3 Nick Bostrom.Superintelligence:Paths,Dangers,Strategies.Oxford University Press,Oxford,UK,2014.4 Erik Brynjolfsson and Andrew McAfee.The Second Machine Age:Work,Progress,andProsperity in a Time of Brill

42、iant Technologies.WW Norton&Company,New York,2014.5 Ryan Calo.Robotics and the lessons of cyberlaw.California Law Review,103:513,2015.56 Tao Jiang,Srdjan Petrovic,Uma Ayyer,Anand Tolani,and Sajid Husain.Self-driving cars:Disruptive or incremental.Applied Innovation Review,1:322,2015.7 William D.Nord

43、haus.Two centuries of productivity growth in computing.The Journal ofEconomic History,67(01):128159,2007.8 Katja Grace.Algorithmic progress in six domains.Technical report,Machine IntelligenceResearch Institute,2013.9 Erik Brynjolfsson and Andrew McAfee.Race Against the Machine:How the Digital Revol

44、utionIs Accelerating Innovation,Driving Productivity,and Irreversibly Transforming Employmentand the Economy.Digital Frontier Press,Lexington,MA,2012.10 Seth D.Baum,Ben Goertzel,and Ted G.Goertzel.How long until human-level ai?resultsfrom an expert assessment.Technological Forecasting and Social Cha

45、nge,78(1):185195,2011.11 Vincent C.Mller and Nick Bostrom.Future progress in artificial intelligence:A survey ofexpert opinion.In Vincent C Mller,editor,Fundamental issues of artificial intelligence,chapter part.5,chap.4,pages 553570.Springer,2016.12 Irving John Good.Speculations concerning the firs

46、t ultraintelligent machine.Advances incomputers,6:3188,1966.13 Philip Tetlock.Expert political judgment:How good is it?How can we know?PrincetonUniversity Press,Princeton,NJ,2005.14 J Doyne Farmer and Franois Lafond.How predictable is technological progress?ResearchPolicy,45(3):647665,2016.15 Lyle U

47、ngar,Barb Mellors,Ville Satop,Jon Baron,Phil Tetlock,Jaime Ramos,and SamSwift.The good judgment project:A large scale test.Technical report,Association for theAdvancement of Artificial Intelligence Technical Report,2012.16 Joe W.Tidwell,Thomas S.Wallsten,and Don A.Moore.Eliciting and modeling probab

48、ilityforecasts of continuous quantities.Paper presented at the 27th Annual Conference of Societyfor Judgement and Decision Making,Boston,MA,19 November 2016.,2013.17 Thomas S.Wallsten,Yaron Shlomi,Colette Nataf,and Tracy Tomlinson.Efficiently encod-ing and modeling subjective probability distributio

49、ns for quantitative variables.Decision,3(3):169,2016.6Supplementary InformationSurvey ContentWe developed questions through a series of interviews with Machine Learning researchers.Oursurvey questions were as follows:1.Three sets of questions eliciting HLMI predictions by different framings:asking d

50、irectlyabout HLMI,asking about the automatability of all human occupations,and asking aboutrecent progress in AI from which we might extrapolate.2.Three questions about the probability of an“intelligence explosion”.3.One question about the welfare implications of HLMI.4.A set of questions about the