The Truth about AI 2of3《人工智能的真相（2023）（2）》完整中英文对照剧本.docx
当今世界人工智能无处不在Intoday,sworld,AIiseverywhere.我今天有哪些安排WhatappointmentsdoIhavetoday?今天有英国皇家科学院圣诞讲座TodayistheRoyalInstitutionChristmas1.ectures.请播放一首提神醒脑的歌♥Playmea"wakemeup"morningmix.*你的生活又过去了一天*Anotherday'spassinginyourlife.*它突然出现在It'spoppingupinplaces许多我们几年前无法想象的地方wecouldneverhaveimaginedafewyearsago.*我正走在*monmyway.*麦克今晚的讲座你准备好了吗Mike,youallsetfortonight?我喜欢你的背景Iloveyourbackground!是啊我正在感受圣诞氛围Yeah,mjustgettingintheChristmasmood.电脑我今晚需要穿外套吗Computer,willIneedacoattonight?是的请携带外套Yes,takeacoat.我们对日常生活中的人工智能了解多少HowawareareweoftheAIinoureverydaylives?*我们正要登上.*We,reonarideto.*在哪里能找到它Wherecanwefindit?*我身无分文.*Ain'tgotnomoney.*它能做些什么Whatdoesitdo?*你无法忽视.*Youcan'tignoreit.*这就是今晚讲座中我们将探讨的That'swhatwe,llbeaskingintonight'slectures.皇家科学院科学讲座演播厅假象真♥相♥人工智能大揭秘2023年英国皇家科学院圣诞讲座第二讲我的人工智能生活大家好我是麦克伍德里奇Hello,mMikeWooldridge,麦克伍德里奇教授欢迎回到本年度第二场andwelcomebacktothesecondofthisyear's英国皇家科学院圣诞讲座ChristmaslecturesfromtheRoyalInstitution,本讲座由CGI集团赞助播出aneventsupportedbyCGI.在上场讲座中我们了解到Inthelastlecture,wesawhowtoday,s当今人工智能是受人脑后发发展而来的artificialintelligencewasinspiredbythehumanbrain.本次讲座我们将深入探讨Inthislecture,we,regoingtoinvestigate人工智能如何塑造我们今天的生活howAIisshapingourlivestoday,我们将看到游戏在推动人工智能发展方面andwe'regoingtoseehowgameshaveplayed起到了怎样重要的作用anincrediblyimportantroleinadvancingAI.我们过会儿将再回到We,regoingtocomebacktothevideothatwejustsaw刚才所看的影片alittlebitlateroninthelecture,仔细聊聊该影片中andwe,regoingtodiscussexactly人工智能扮演的角色whereAIfeaturesinthatvideo.但首先我们来探讨一下人工智能Butfirst,we'regoingtostartbyexploring在我们最流行的娱乐形式电子游戏中howAIfeaturesinoneofourmostpopularforms起到的重要作用ofentertainment-videogames.有请跑车浪漫旅GT赛车顶尖选手马丁格雷迪PleasewelcometopGTdriverMartinGrady,和索尼人工智能公♥司♥的考什克苏布拉马尼安andfromSonyAI,KaushikSubramanian.欢迎你马丁谢谢Welcome,Martin.-Thankyou.请坐Takeaseat.别客气.andmakeyourselfcomfortable.马丁你玩赛车游戏多久了Now,Martin,howlonghaveyoubeenracing?我玩赛车游戏二十五年了vebeenracingfor25years.其中十二年是职业选干Andvebeenapro-racingdriverfor12ofthose.好吧我们今天想请你OK,sowhatwe'regoingtogetyoutodotoday和人工智能一较高下iswe'regoingtogetyoutoplayagainstAI.我们将安排你和索尼人工智能公♥司♥We'regoingtogetyoutoplayagainstaracingAI为跑车浪漫旅开♥发♥的thatwasdevelopedbySonyAI赛车人工智能进行比赛forthegameofGranTurismo.顺便问一下观众中Bytheway,whointheaudiencehas谁家有这些东西吗gotoneofthesethingsathome?谁家有吗Anybody?少数几个人家有这些东西Acoupleofpeoplehavegotoneofthose.马丁请开始吧So,Martin,yougetyourselfgoing.告诉我们考什克Kaushik,tellus,你构建该人工智能是想做什么whatdidyouwanttodowhenyoubuiltthisAI?好吧我们向自己提出的第一个挑战是Right,thefirstchallengewhatweaskedourselvesis,我们能开♥发♥出一款能在比赛中击败canwedevelopanAIthatcanracetobeattheworld,sbestGT赛车世界冠军的人工智能吗GranTurismoracersinacompetitiverace?人工智能可以学会赛车所需的AndcantheAIlearntherequiredskills高阶技能吗toperformatthatlevel?事实证明我们能Itturnsoutwecan.索菲赛车游戏中人工智能的名字考什克该游戏中索菲在哪里Kaushik,whereexactlyisSophyappearinginthis,theAI?马丁正和其他十九辆人工智能赛车比赛So,Martin'sracedwith19otherAIcars,它们都由GT索菲控制andthey*reallcontrolledbyGranTurismoSophyhere.从这里就能看出来Andthewayyoucanseethatis,在左手边ifyoulookontheleft-handside,你可以看到所有玩家车辆的名字youcanseeallthenamesoftheplayercars.这里有个心形图标Andthere'salogowhichlookslikeaheart,所有带这个图标的andallofthoselogos都是GT索菲控制的车aretheGranTurismoSophycars.所有这些车都知道自己是哪辆Andallofthosecars,theyknowwhatcarthey'redriving,它们知道彼此they'reawareofeachother当然也能意识到马丁的存在andthey'reofcourseawareofMartinaswell.好吧但它们并不想联手对付马丁对吗OK,butthey'renotalltryingtoganguponMartin,arethey?它们都是独&hearts位♥的吗They'reallindependent?它们都想让自己获得第一名They'realltryingtogettofirstplaceindependently.索菲是如何变得这么出色的呢SohowdidSophy,theAI,getgoodenoughinthisrace?人工智能GT索菲ThisA1.GranTurismoSophy,经过了一种称作“强化学习”的方法训练istrainedusinganapproachcalledreinforcementlearning,也就是人工智能通过试错来学习anapproachwheretheAIlearnsbytrialanderror.一开始索菲不太了解驾驶So,atthestart,itreallydoesn'tknowmuchaboutdriving,更别提比赛规则了muchlesstherulesofracing.我们这里有训练视频OK,Ithinkwe'vegotthevideoshowinghereofthetraining.好吧在槐顷里Right,andwhenyouwatchthis,你们能看到赛车相撞you'regoingtoseethecarscolliding,车辆完全失控you'regoingtoseethecarscompletelylosingcontrol,冲出赛道goingoffthetrack.视频里是刚训练几小时后的结果Andthisisjustwithinafewhoursoftraining.就是这个样子Thisiswhatitlookslike.过了一段时间通过强化学习Andafterawhile,withreinforcementlearning,它尝试了许多不同的事ittriesmanydifferentthings最终学会正确的操作顺序andthenitlearns,finally,therightsequence能够进行比赛了ofactionstobeabletorace,就像马丁现在正看到的那样likethewayMartin'slookingatrightnow.显然马丁撞车了Orcrashing,apparently.好吧我们稍后回来OK,now,we'regoingtocomebacklateroninthelecture继续r解更多“强化学习'的相关内容andwe'regoingtolearnalotmoreaboutreinforcementlearning.考什克但为什么像这样的电子游戏But,Kaushik,whyisavideogamelikethissodifficult很难用强化学习的方法forreinforcementlearning,来让机器进行学习andjustformachinelearninggenerally?赛车游戏有许多变量RacinghasalotOfvariability.情况瞬息万变Thereareeverchangingsituations.当你在那些情形下处于失控边缘时Whenyou'redrivingattheedgeofcontrolinthosesituations,很难进行判断thatcanbehard.在GT赛车游戏中Now,inGranTurismo,玩家有近五千辆车可选playershaveaccesstonearly500differentcars,当你驾驶这些车比赛时andwhenyoutakethosecarsandyouputtheminarace,这些车在数圈的比赛中能以多种方式thosecarscaninteractinmanydifferentways相互影响overmultiplelaps.在所有这些情形下Andinallofthosesituations,人工智能都需要技巧娴熟行事可靠theAIwouldneedtobeskilfulandreliable,这极具挑战andthatcanbechallenging.好吧训练我们现在看到的人工智能用了多久OK.AndhowlongdidittaketotraintheAIthatwe'reseeing?单次试验需要约二十五台PS游戏机Asingleexperimenttakesabout25PlayStations训练近两周时间andtrainsfornearlytwoweeks然后才能和其他人类玩家进行比赛beforeitcanracewithotherhumanplayers.好吧马丁告诉我们0K,so,Martin,tellus.你觉得这个怎么样howareyoufindingthis?和你之前玩过的有什么不同吗Isitdifferenttowhatyou'veusedpreviously?完全不i样Yeah,it'scompletelydifferent.过去你只能和普通车辆比赛Inthepast,youwouldjustraceastandardcar,那些车一圈又一圈只在同一车道行驶justgoingonthesamelineoverandover.这里你能看到它们正在防守Here,youcanseethey,regoingdefensive,它们会对我的行为做出反应they'rereactingtowhatIdo.如果我有所动作它们会有相应的防御措施So,ifIgoforamove,they'regoingtomovetodefendthat.你想超车时它们会阻拦你吗Theycantryandblockyouasyoutrytoovertake?是的它们对我的对策做出了反应Yeah,theyreacttomyreactions,所以我必须再次做出反应sothenIhavetoreactagain.与之对抗让游戏变得Anditjustmakesitsomuchmorefun更加有趣tobeabletoplayagainstthat因为会有无限可能becausethere'sendlesspossibilitiesthen.如你所见它们移到了右手边Asyoucansee,they'regoingtotheright-handside.现在真的很有趣It,sjustreallygoodfunnow,因为我基本上已无路可走-好吧cosnowvegotnowheretogo,essentially.-Ok!马丁我注意到Now,Martin,Ican'thelpbutnotice你是穿袜子youappeartobeplayingthisgame玩这个游戏的wearingapairofsocks.这是玩GT的专业技巧吗IsthataprotipforplayingGranTurismo?这是专业技巧Itisaprotip.把鞋脱了穿着袜子玩SotakeyourshoesOfCplayinyoursocks.对踏板的感觉会更灵敏Yougetamoresensitivefeelonthepedal比赛时的反应会更快anditallowsyoutojustracethatalittlebitquicker.听从专家的建议孩子们OK,listentotheexperts,kids.谢谢你马丁OK.Thankyou,Martin.考什克感谢你来为我们展示该游戏Thankyou,Kaushik,forcomingandshowingusthisgame.谢谢Thankyou.人工智能是如何变得如此优秀的呢HowdidtheAIgetsogood?我们听考什克提到“强化学习”Well,weheardKaushikmentionreinforcementlearning.这就是考什克That'swhatKaushik及其团队训练GT索菲的方法andhisteamusedtotrainGTSophy.那么什么是强化学习它又是如何工作的Sowhatisreinforcementlearningandhowdoesitwork?为了解释这一点我需要些帮助Now,toexplain,Ineedsomehelp,但我的下位嘉宾有点紧张butmynextguestisalittlebitnervous.当她进来时请大家默默鼓掌So,whenshecomeson,please,let'sjusthavesilentapplause.请保持安静别吓到她Pleasekeepitquiet.Don'talarmher.下面请欢迎凯特琳和她的狗弗蕾亚So,please,let'swelcomeCaitlinandherdog,Freya.你们好Hello!你好凯特琳Hello,Caitlin.你好弗蕾亚-你好Hello,Freya.-Hi.欢迎来圣诞讲座现场WelcometotheChristmaslectures.谢谢Thankyou.凯特琳你和弗蕾亚So,Caitlin,howlonghaveyoubeen一起工作多久了workingwithFreya?弗蕾亚和我一起工作了大概三年So,FreyaandIhavebeenworkingtogetherforaboutthreeyears.在我受训成为驯犬师时她的妈妈HermumkindlyletusworktogetherwhenIwastraining非常好心的让我们一起工作现在tobeapuppyschooltutor,andnow.刚开始进行技巧训练.juststartedsometricktrainingtogether.所以她很爱叫Sosheverymuchlikestousehervoice,我向大家的耳朵道歉soIapologisefortheears.好Ok.今天你要给我们展示什么凯特琳Sowhatareyougoingtoshowustoday,Caitlin?你如何训练弗蕾亚吗-对HowareyougoingtoshowushowyoutrainFreya?-Yes!今天弗蕾亚会给我们展示So,today,Freyaisgoingtoshowus如何从一个人的口袋里howtopickpocketsomeone取出一方手帕withahandkerchief我们要做的是利用正面强化法Whatwe,regoingtodoisusepositivereinforcement告诉她取到手帕时totellherthatwhenshepullsonthehandkerchief她就会得到奖励thenshegetsarewardforthat.所以我会用一个”口令”SoIusesomethingcalledamarker.当我告诉她去取物时会告诉她“对“So,whenItellhertogetit,I'lltellheryes然后奖励她的行为andrewardherforwhatshe'sdone.那就是正确的做法That'sthecorrectthingtodo.所以我会让她去取物Somgoingtotellhertogetit.对好姑娘好Yes!Goodgirl!Ok.非常好Verynice.首先我要让她对此感兴趣SofirstIwanttogetherinterestedinthis.对-好我觉得她感兴趣了Yes!-OK,Ithinkshe'sinterested.很好Nice!我的方法就是物品变得有趣SothewayIdothatisjustmakeitinterestinglikethis.如果她拉走对Ifshepullsonit.Yes!然后给她奖励Andthengiveherthatreward.随后我会增加*个信♥号♥♥Andthen1wanttoaddacuetoit比如.弗蕾亚去取Soalittle.Freya,getit!对好姑娘Yes!Goodgirl.再给奖励Andthenreward.就是这个模式对Soit'sjustfollowingthatpattern.-Ok.看起来奖励对弗蕾亚很重要Now,rewardsseemtobereallyimportanttoFreya.那是她学习的重要部分吗Isthisabigpartofhowshelearns?没错有奖励就重复的正反馈过程Absolutely.Sowhatgetsrewardedgetsrepeated.她做了我想让她做的我就奖励她SothemoreIrewardherforthethingIwanthertodo,她就越可能再去做这件事themoreit'sgoingtohappen.我要做的还有保证ButwhatIwanttomakesureof要让她听口令做动作isthatshe'salsorespondingtothecue.所以她这么拉手帕我不会给奖励So,whenshepulleditthere,Ididn'tgiveherthereward因为我想让她听到口令再行动becauseIwanthertorespondtothattogetit.因为刚才你没给口令Becauseyouhadn'tgivenherthecue.没错是的Exactly,yeah.现在最后一步就是口袋取物Andnowthelastpartofthisisthepickpocketpart,这一步需要你的协助好whichIwillneedyourassistance.-Ok.可以吗-可以IfthafsOKV-Yes.我会把这个给你Somgoingtogiveyouthat.请把手帕搭在你裤了后面口袋Andifyoujustpopthatinyourbackpocketforme.弗蕾亚来来Freya,come,come!好姑娘坐下好Goodgirl!Sit!Ok.很好别动-为了圣诞讲座我豁出去了Nice.Stay.-ThethingsIdofortheChristmaslectures.弗蕾亚去取那个Freya,getit!好做得好Yes!Goodjob!很好Verynice!弗蕾亚Freya弗蕾亚你比我的狗聪明多了Freya,you'realotsmarterthanmydog,我必须得说Ihavetosay.非常棒Welldone.非常感谢凯特琳谢谢你弗雷亚OK,thankyousomuch,Caitlin.Thankyou,Freya.大家要记得礼貌而克制的掌声Remember,everybody,politequietapplause,please.非常感谢谢谢你Thankyousomuch.Thankyou!来弗雷亚我们走Comeon,Freya!1.et'sgo!她还不想走Shedoesn'twanttoleave!弗蕾亚是怎公学习的SohowisFreyalearning?每次完成一项任务她就会被奖励She'sgivenatreateverytimeshedoesataskwell.在人工智能界这被称作强化学习AndinAI,wecallthatreinforcementlearning.上场讲座我们看到人工智能如何从数据Inthelastlecture,wesawhowAIlearnsfromdata,从训练数据中学习fromtrainingdata.强化学习也很类似Now,reinforcementlearningissimilar,但训练数据在该情境中以奖励的形式出现butthetrainingdatainthiscasecomesintheformofrewards.强化学习对跑车浪漫旅这类游戏Andreinforcementlearningisreallygoodforgames也大有益处likeGranTurismo因为大多数游戏都计分becausemostgameshaveascore我们可以用积分作为奖励andwecanusescoresasthereward.当人工智能得一分就是一个奖励WhentheAlscoresapoint,that,sareward.人工智能的设计就是学习AndwhattheAIisdesignedtodoistolearn如何将其奖励最大化howtomaximiseitsreward,将其得分最大化tomaximiseitsscore,以最快最可能的途径获得最多积分togetasmanyrewardsaspossibleasquicklyaspossible,就像弗蕾亚获得她的奖励那样justlikeFreyagettinghertreats.我们可以用神经网络来完成学习Andwecanuseneuralnetworkstolearnallthat.这被叫做深度强化学习That'scalleddeepreinforcementlearning.但有时强化学习Butsometimesreinforcementlearning不能给出我们期望中mightnotgiveustheoutcomes想要的结果thatwewerehopingfor,我们将通过一个游戏展示这一点andwe'regoingtohaveagametoillustratethisidea.我需要一名志愿者CanIhaveavolunteer,please?那位穿绿色衣服对请走下来Inthegreenthere?Yeah,youcomedown.对你被选中了Yes,you'vebeenselected!请下来站在这里你叫什么-妮维雅Comeondown.Juststandhere.What'syourname?-Nivea.好的欢迎参加圣诞讲座OK,well,welcometotheChristmaslectures.下面我给你一个任务Now,mgoingtogiveyouatask.会给你一些指示mgoingtogiveyousomeinstructions.我只需要你听从指示好吗AndIjustwantyoutofollowthoseinstructions,OK?你腿脚灵活吗跑步好吗-是的Areyounimbleonyourfeet?Areyouagoodrunner?-Yes.非常好那你会很适合做这件事.OK,excellent.Thenyou'regoingtobeperfectforthis.我们设定了五个铃Now,whatwe,vedoneiswe'veplacedfivebells就在阶梯教室四周aroundthelecturetheatre.这一个这一个Onethere.Onethere.这一个大家看我的腿脚很灵活Onethere.Verynimbleonmyfeet,asyoucansee.这一个Onethere.最后这一个Andfinally,onethere.我们要做的就是给你Andwhatwe,regoingtodoiswe'regoingtogiveyou十秒时间tenseconds,我希望听到尽可能多次的铃♥声♥and1wanttohearasmanybellringsaspossible.明白吗是的明白Gotit?-Yes.Yeah.好了接下来要做的OK,allright,whatwe'regoingtodois我们将倒数三个数we'regoingtocountyoudownfromthree,然后我们会说“开始“好吗andthenwe'regoingtosaygo,allright?Gotit?Three,two,one.开始一开始Go!-Go!快快快妮维雅快跑Go,go,go!-Go,Nivea!Run!铃响了我们听着铃响声Bellrings1.et'shearthosebells!快加速Comeon!Pickupthepace!一停One,stop!铃响了几次OK,howmanybellrings?请问到中间我觉得是五次Comebacktothemiddle.Weheardfive,Ithink.次数正确吗各位Wasthataboutright,everybody?对我们听到五次铃响Yeah?Weheardfivebellrings.很棒是不是Thatwasreally,reallygood,OK?非常感谢你可以回到座位了Thankyousomuchforthat.Youcangobacktoyourseat.下面我们要进行同样的挑战Nowwe,regoingtocarryoutthesamechallenge,但用人工智能butwithartificialintelligence.请掌声欢迎皇家科学院SopleasewelcometheRoyalInstitution's特别嘉宾响铃机器人specialbell-ringingrobot!来吧响铃机器人Comeon,bell-ringingrobot把那个给我求你Givemethat.Honestly.好Ok!下面给你十秒时间Now,mgoingtogiveyoutenseconds.我要听到尽可能的铃响次数Iwanttohearasmanybellringsaspossible.现在我们给机器人倒数Solet,scountdowntherobot.三二一开始-开始Three,two,one,go!-Go!停Stop!有人数清刚才铃响次数了吗Didanybodymanagetokeeptrackofhowmanythatwas?好机器人谢谢你OK,robot,thankyouverymuch.还给你Thereyougo.其实我们的志愿者So,ourvolunteer,什么也没做错youdidn'tdoanythingwrongatall.因为你是人类Becauseyou'reahumanbeing.我给你指示Igaveyousomeinstructions而你尝试去理解我想让你做什么andyoutriedtointerpretwhatIwantedyoutodo.我希望在指示中暗示的是AndwhatIwantedinmyinstructions让某人围着阶梯教室跑istohearsomebodyrunningaroundthelecturetheatre按响那些铃pressingeachofthosebells.但实际h那并不是真正的指示Butactually,thoseweren'texactlytheinstructionsIgaveyou.我刚才说只是我想听到WhatIsaidis,Ijustwantedtohear尽可能多的铃♥声♥asmanybellringsaspossible.而机器人就是理解字面意思Andourrobottooktheinstructionsliterally.而让机器人得到最多奖励的方式Andthequickestwaytherobotcouldgetareward就是站在这不断敲响铃♥声♥wasjusttostandthereandbangthatbell.人工智能找到了最大化其奖励的方法TheAIfoundawaytomaximiseitsrewards却没有按我要求的方法去做withoutdoingwhatIwantedittodo.我们使川强化学习时设定奖励的方法Thewaywesetuprewardswhenweusereinforcementlearning至关重要isreallyimportant因为有时我们设定的奖励becausesometimeswecansetuprewards会让人工智能发现最大化奖励的方法sothattheAIdiscoversawaytomaximiseitsrewards就是不按我们希望它使用的方法去做事withoutdoingwhatwewantedittodo.也正是刚才那个机器人所做的Andthat'swhattherobotdidthere.我们现在来看一个实际例子Andnowwhatwe'regoingtoseeisarealexample.我们来看屏幕上播放的视频Solet'shavealookatthisvideoonthescreen.在2014年一个叫DeepMind的人工智能公♥司♥In2014,theAIcompanyDeepMind训练r一个人工智能程序来玩这个游戏trainedanAIprogramtoplaythisgame.这是上世纪七十年代一款叫越狱的游戏It'sa1970svideogamecalledBreakout这个程序使用r强化学习进行训练anditusesreinforcementlearning.所以它玩的越多就玩得越好Themoreitplays,thebetteritgets.你们可以看到一开始大部分时间Now,atthebeginning,mostofthetime,asyou'llsee,都没有碰到it,sjustmissing.它能碰到球完全靠运气Andifitmanagestohittheball,it,sjustreallypurechance.但是它每打掉一块砖Buteverytimeitknocksabrickout,就会得一分itgetsapoint.经过一段时间的训练它现在从不失球Now,afterabitmoretraining,watch,itnevermisses.它每次都能准确地把球打回去It
