4/14/2015

TheAIRevolution:RoadtoSuperintelligenceWaitButWhy

about

archive

minis

theshed

dinnertable

store

supportwbw

TheAIRevolution:TheRoadto
Superintelligence

Search

Join121,034otherhumansand
haveourweeklypostemailedto
you

Email Address
SENDMENEWPOSTS

ByTimUrban

Note:ThereasonthisposttookthreeweekstofinishisthatasIdugintoresearchonArtificial
Intelligence,IcouldnotbelievewhatIwasreading.Ithitmeprettyquicklythatwhatshappeningin
theworldofAIisnotjustanimportanttopic,butbyfarTHEmostimportanttopicforourfuture.SoI
wantedtolearnasmuchasIcouldaboutit,andonceIdidthat,IwantedtomakesureIwroteapost
thatreallyexplainedthiswholesituationandwhyitmatterssomuch.Notshockingly,thatbecame
outrageouslylong,soIbrokeitintotwoparts.ThisisPart1Part2ishere.

LikeWBWOnFacebook

FollowWBWOnTwitter

193,559

24,970

RandomPost

_______________

WeareontheedgeofchangecomparabletotheriseofhumanlifeonEarth.VernorVinge

Whatdoesitfeelliketostandhere?

PopularPosts
WhyGenerationY
YuppiesAreUnhappy

Itseemslikeaprettyintenseplacetobestandingbutthenyouhavetoremembersomethingabout
whatitsliketostandonatimegraph:youcantseewhatstoyourright.Sohereshowitactuallyfeels
tostandthere:

1,099

Like

712k

WhyProcrastinators
Procrastinate

460

Like

206k

TheFermiParadox

683

Like

215k

7WaysToBeInsufferable
OnFacebook

593

http://waitbutwhy.com/2015/01/artificialintelligencerevolution1.html

Like

312k

1/26

4/14/2015

TheAIRevolution:RoadtoSuperintelligenceWaitButWhy
PuttingTimeIn
PerspectiveUPDATED

446

Like

238k

TheAIRevolution:The
RoadtoSuperintelligence

679

Like

62k

HowtoNameaBaby

217

Like

115k

TamingtheMammoth:WhyYou
ShouldStopCaringWhatOther
PeopleThink

176

Like

28k

HowtoPickYourLife
PartnerPart1

Whichprobablyfeelsprettynormal

189

Like

61k

_______________

TheFarFutureComingSoon
Imaginetakingatimemachinebackto1750atimewhentheworldwasinapermanentpower
outage,long-distancecommunicationmeanteitheryellingloudlyorfiringacannonintheair,andall

10TypesofOdd
FriendshipsYoure
ProbablyPartOf

176

Like

29k

transportationranonhay.Whenyougetthere,youretrieveadude,bringhimto2015,andthenwalk

EverythingYouDont
KnowAboutTipping

himaroundandwatchhimreacttoeverything.Itsimpossibleforustounderstandwhatitwouldbe

714

likeforhimtoseeshinycapsulesracingbyonahighway,talktopeoplewhohadbeenontheotherside

Like

oftheoceanearlierintheday,watchsportsthatwerebeingplayed1,000milesaway,hearamusical
performancethathappened50yearsago,andplaywithmymagicalwizardrectanglethathecoulduse

Religionforthe
Nonreligious

tocaptureareal-lifeimageorrecordalivingmoment,generateamapwithaparanormalmovingblue

336

dotthatshowshimwhereheis,lookatsomeonesfaceandchatwiththemeventhoughtheyreonthe
othersideofthecountry,andworldsofotherinconceivablesorcery.Thisisallbeforeyoushowhimthe
internetorexplainthingsliketheInternationalSpaceStation,theLargeHadronCollider,nuclear
weapons,orgeneralrelativity.

Like

205k

10k

FromMuhammadtoISIS:
IraqsFullStory

199

Like

29k

Thisexperienceforhimwouldntbesurprisingorshockingorevenmind-blowingthosewordsarent
bigenough.Hemightactuallydie.
Butherestheinterestingthingifhethenwentbackto1750andgotjealousthatwegottoseehis
reactionanddecidedhewantedtotrythesamething,hedtakethetimemachineandgobackthe
samedistance,getsomeonefromaroundtheyear1500,bringhimto1750,andshowhimeverything.
Andthe1500guywouldbeshockedbyalotofthingsbuthewouldntdie.Itwouldbefarlessofan
insaneexperienceforhim,becausewhile1500and1750wereverydifferent,theyweremuchless
differentthan1750to2015.The1500guywouldlearnsomemind-bendingshitaboutspaceand
physics,hedbeimpressedwithhowcommittedEuropeturnedouttobewiththatnewimperialismfad,
andhedhavetodosomemajorrevisionsofhisworldmapconception.Butwatchingeverydaylifego
byin1750transportation,communication,etc.definitelywouldntmakehimdie.
No,inorderforthe1750guytohaveasmuchfunaswehadwithhim,hedhavetogomuchfarther
backmaybeallthewaybacktoabout12,000BC,beforetheFirstAgriculturalRevolutiongaveriseto
thefirstcitiesandtotheconceptofcivilization.Ifsomeonefromapurelyhunter-gathererworldfrom
atimewhenhumanswere,moreorless,justanotheranimalspeciessawthevasthumanempiresof
1750withtheirtoweringchurches,theirocean-crossingships,theirconceptofbeinginside,andtheir
enormousmountainofcollective,accumulatedhumanknowledgeanddiscoveryhedlikelydie.
Andthenwhatif,afterdying,hegotjealousandwantedtodothesamething.Ifhewentback12,000
yearsto24,000BCandgotaguyandbroughthimto12,000BC,hedshowtheguyeverythingandthe
guywouldbelike,Okaywhatsyourpointwhocares.Forthe12,000BCguytohavethesamefun,
hedhavetogobackover100,000yearsandgetsomeonehecouldshowfireandlanguagetoforthe
firsttime.
Inorderforsomeonetobetransportedintothefutureanddiefromthelevelofshocktheyd
experience,theyhavetogoenoughyearsaheadthatadielevelofprogress,oraDieProgressUnit
http://waitbutwhy.com/2015/01/artificialintelligencerevolution1.html

2/26

4/14/2015

TheAIRevolution:RoadtoSuperintelligenceWaitButWhy

(DPU)hasbeenachieved.SoaDPUtookover100,000yearsinhunter-gatherertimes,butatthepostAgriculturalRevolutionrate,itonlytookabout12,000years.Thepost-IndustrialRevolutionworldhas
movedsoquicklythata1750persononlyneedstogoforwardacouplehundredyearsforaDPUto
havehappened.
ThispatternhumanprogressmovingquickerandquickerastimegoesoniswhatfuturistRay
KurzweilcallshumanhistorysLawofAcceleratingReturns.Thishappensbecausemoreadvanced
societieshavetheabilitytoprogressatafasterratethanlessadvancedsocietiesbecausetheyre
moreadvanced.19thcenturyhumanityknewmoreandhadbettertechnologythan15thcentury
humanity,soitsnosurprisethathumanitymadefarmoreadvancesinthe19thcenturythaninthe
15thcentury15thcenturyhumanitywasnomatchfor19thcenturyhumanity.

1 1 openthese

Thisworksonsmallerscalestoo.ThemovieBacktotheFuturecameoutin1985,andthepasttook
placein1955.Inthemovie,whenMichaelJ.Foxwentbackto1955,hewascaughtoff-guardbythe
newnessofTVs,thepricesofsoda,thelackofloveforshrillelectricguitar,andthevariationinslang.It
wasadifferentworld,yesbutifthemovieweremadetodayandthepasttookplacein1985,the
moviecouldhavehadmuchmorefunwithmuchbiggerdifferences.Thecharacterwouldbeinatime
beforepersonalcomputers,internet,orcellphonestodaysMartyMcFly,ateenagerborninthelate
90s,wouldbemuchmoreoutofplacein1985thanthemoviesMartyMcFlywasin1955.
ThisisforthesamereasonwejustdiscussedtheLawofAcceleratingReturns.Theaveragerateof
advancementbetween1985and2015washigherthantheratebetween1955and1985becausethe
formerwasamoreadvancedworldsomuchmorechangehappenedinthemostrecent30yearsthan
intheprior30.
Soadvancesaregettingbiggerandbiggerandhappeningmoreandmorequickly.Thissuggests
someprettyintensethingsaboutourfuture,right?
Kurzweilsuggeststhattheprogressoftheentire20thcenturywouldhavebeenachievedinonly20
yearsattherateofadvancementintheyear2000inotherwords,by2000,therateofprogresswas
fivetimesfasterthantheaveragerateofprogressduringthe20thcentury.Hebelievesanother20th
centurysworthofprogresshappenedbetween2000and2014andthatanother20thcenturysworth
ofprogresswillhappenby2021,inonlysevenyears.Acoupledecadeslater,hebelievesa20th
centurysworthofprogresswillhappenmultipletimesinthesameyear,andevenlater,inlessthan
onemonth.Allinall,becauseoftheLawofAcceleratingReturns,Kurzweilbelievesthatthe21stcentury
willachieve1,000timestheprogressofthe20thcentury. 2
IfKurzweilandotherswhoagreewithhimarecorrect,thenwemaybeasblownawayby2030asour
1750guywasby2015i.e.thenextDPUmightonlytakeacoupledecadesandtheworldin2050
mightbesovastlydifferentthantodaysworldthatwewouldbarelyrecognizeit.
Thisisntsciencefiction.ItswhatmanyscientistssmarterandmoreknowledgeablethanyouorIfirmly
believeandifyoulookathistory,itswhatweshouldlogicallypredict.
Sothenwhy,whenyouhearmesaysomethingliketheworld35yearsfromnowmightbetotally
unrecognizable,areyouthinking,Cool.butnahhhhhhh?Threereasonswereskepticalof
outlandishforecastsofthefuture:
1)Whenitcomestohistory,wethinkinstraightlines.Whenweimaginetheprogressofthenext30
years,welookbacktotheprogressoftheprevious30asanindicatorofhowmuchwilllikelyhappen.
Whenwethinkabouttheextenttowhichtheworldwillchangeinthe21stcentury,wejusttakethe
20thcenturyprogressandaddittotheyear2000.Thiswasthesamemistakeour1750guymadewhen
hegotsomeonefrom1500andexpectedtoblowhismindasmuchashisownwasblowngoingthe
samedistanceahead.Itsmostintuitiveforustothinklinearly,whenweshouldbethinking
exponentially.Ifsomeoneisbeingmorecleveraboutit,theymightpredicttheadvancesofthenext30
yearsnotbylookingattheprevious30years,butbytakingthecurrentrateofprogressandjudging
basedonthat.Theydbemoreaccurate,butstillwayoff.Inordertothinkaboutthefuturecorrectly,
youneedtoimaginethingsmovingatamuchfasterratethantheyremovingnow.

http://waitbutwhy.com/2015/01/artificialintelligencerevolution1.html

3/26

4/14/2015

TheAIRevolution:RoadtoSuperintelligenceWaitButWhy

2)Thetrajectoryofveryrecenthistoryoftentellsadistortedstory.First,evenasteepexponential
curveseemslinearwhenyouonlylookatatinysliceofit,thesamewayifyoulookatalittlesegmentof
ahugecircleupclose,itlooksalmostlikeastraightline.Second,exponentialgrowthisnttotally
smoothanduniform.KurzweilexplainsthatprogresshappensinS-curves:

AnSiscreatedbythewaveofprogresswhenanewparadigmsweepstheworld.Thecurvegoes
throughthreephases:
1.Slowgrowth(theearlyphaseofexponentialgrowth)
2.Rapidgrowth(thelate,explosivephaseofexponentialgrowth)
3.Alevelingoffastheparticularparadigmmatures 3
Ifyoulookonlyatveryrecenthistory,thepartoftheS-curveyoureonatthemomentcanobscure
yourperceptionofhowfastthingsareadvancing.Thechunkoftimebetween1995and2007sawthe
explosionoftheinternet,theintroductionofMicrosoft,Google,andFacebookintothepublic
consciousness,thebirthofsocialnetworking,andtheintroductionofcellphonesandthensmart
phones.ThatwasPhase2:thegrowthspurtpartoftheS.But2008to2015hasbeenless
groundbreaking,atleastonthetechnologicalfront.Someonethinkingaboutthefuturetodaymight
examinethelastfewyearstogaugethecurrentrateofadvancement,butthatsmissingthebigger
picture.Infact,anew,hugePhase2growthspurtmightbebrewingrightnow.
3)Ourownexperiencemakesusstubbornoldmenaboutthefuture.Webaseourideasaboutthe
worldonourpersonalexperience,andthatexperiencehasingrainedtherateofgrowthoftherecent
pastinourheadsasthewaythingshappen.Werealsolimitedbyourimagination,whichtakesour
experienceandusesittoconjurefuturepredictionsbutoften,whatweknowsimplydoesntgiveus
2
thetoolstothinkaccuratelyaboutthefuture. Whenwehearapredictionaboutthefuturethat
http://waitbutwhy.com/2015/01/artificialintelligencerevolution1.html

4/26

4/14/2015

TheAIRevolution:RoadtoSuperintelligenceWaitButWhy

contradictsourexperience-basednotionofhowthingswork,ourinstinctisthatthepredictionmustbe
naive.IfItellyou,laterinthispost,thatyoumaylivetobe150,or250,ornotdieatall,yourinstinctwill
be,ThatsstupidiftheresonethingIknowfromhistory,itsthateverybodydies.Andyes,noonein
thepasthasnotdied.Butnooneflewairplanesbeforeairplaneswereinventedeither.
Sowhilenahhhhhmightfeelrightasyoureadthispost,itsprobablyactuallywrong.Thefactis,if
werebeingtrulylogicalandexpectinghistoricalpatternstocontinue,weshouldconcludethatmuch,
much,muchmoreshouldchangeinthecomingdecadesthanweintuitivelyexpect.Logicalsosuggests
thatifthemostadvancedspeciesonaplanetkeepsmakinglargerandlargerleapsforwardataneverfasterrate,atsomepoint,theyllmakealeapsogreatthatitcompletelyalterslifeastheyknowitand
theperceptiontheyhaveofwhatitmeanstobeahumankindoflikehowevolutionkeptmaking
greatleapstowardintelligenceuntilfinallyitmadesuchalargeleaptothehumanbeingthatit
completelyalteredwhatitmeantforanycreaturetoliveonplanetEarth.Andifyouspendsometime
readingaboutwhatsgoingontodayinscienceandtechnology,youstarttoseealotofsignsquietly
hintingthatlifeaswecurrentlyknowitcannotwithstandtheleapthatscomingnext.
_______________

TheRoadtoSuperintelligence
WhatIsAI?
Ifyourelikeme,youusedtothinkArtificialIntelligencewasasillysci-ficoncept,butlatelyyouvebeen
hearingitmentionedbyseriouspeople,andyoudontreallyquitegetit.
TherearethreereasonsalotofpeopleareconfusedaboutthetermAI:
1)WeassociateAIwithmovies.StarWars.Terminator.2001:ASpaceOdyssey.EventheJetsons.And
thosearefiction,asaretherobotcharacters.SoitmakesAIsoundalittlefictionaltous.
2)AIisabroadtopic.Itrangesfromyourphonescalculatortoself-drivingcarstosomethinginthe
futurethatmightchangetheworlddramatically.AIreferstoallofthesethings,whichisconfusing.
3)WeuseAIallthetimeinourdailylives,butweoftendontrealizeitsAI.JohnMcCarthy,who
coinedthetermArtificialIntelligencein1956,complainedthatassoonasitworks,noonecallsitAI
anymore. 4 Becauseofthisphenomenon,AIoftensoundslikeamythicalfuturepredictionmorethan
areality.Atthesametime,itmakesitsoundlikeapopconceptfromthepastthatnevercameto
fruition.RayKurzweilsayshehearspeoplesaythatAIwitheredinthe1980s,whichhecomparesto
insistingthattheInternetdiedinthedot-combustoftheearly2000s. 5
Soletsclearthingsup.First,stopthinkingofrobots.ArobotisacontainerforAI,sometimesmimicking
thehumanform,sometimesnotbuttheAIitselfisthecomputerinsidetherobot.AIisthebrain,and
therobotisitsbodyifitevenhasabody.Forexample,thesoftwareanddatabehindSiriisAI,the
womansvoicewehearisapersonificationofthatAI,andtheresnorobotinvolvedatall.
Secondly,youveprobablyheardthetermsingularityortechnologicalsingularity.Thistermhas
beenusedinmathtodescribeanasymptote-likesituationwherenormalrulesnolongerapply.Its
beenusedinphysicstodescribeaphenomenonlikeaninfinitelysmall,denseblackholeorthepoint
wewereallsquishedintorightbeforetheBigBang.Again,situationswheretheusualrulesdontapply.
In1993,VernorVingewroteafamousessayinwhichheappliedthetermtothemomentinthefuture
whenourtechnologysintelligenceexceedsourownamomentforhimwhenlifeasweknowitwillbe
foreverchangedandnormalruleswillnolongerapply.RayKurzweilthenmuddledthingsabitby
definingthesingularityasthetimewhentheLawofAcceleratingReturnshasreachedsuchanextreme
pacethattechnologicalprogressishappeningataseemingly-infinitepace,andafterwhichwellbe
livinginawholenewworld.IfoundthatmanyoftodaysAIthinkershavestoppedusingtheterm,and
itsconfusinganyway,soIwontuseitmuchhere(eventhoughwellbefocusingonthatidea
throughout).
Finally,whiletherearemanydifferenttypesorformsofAIsinceAIisabroadconcept,thecritical
categoriesweneedtothinkaboutarebasedonanAIscaliber.TherearethreemajorAIcaliber
categories:
AICaliber1)ArtificialNarrowIntelligence(ANI):SometimesreferredtoasWeakAI,ArtificialNarrow
IntelligenceisAIthatspecializesinonearea.TheresAIthatcanbeattheworldchesschampionin
chess,butthatstheonlythingitdoes.Askittofigureoutabetterwaytostoredataonaharddrive,
http://waitbutwhy.com/2015/01/artificialintelligencerevolution1.html

5/26

4/14/2015

TheAIRevolution:RoadtoSuperintelligenceWaitButWhy

anditlllookatyoublankly.
AICaliber2)ArtificialGeneralIntelligence(AGI):SometimesreferredtoasStrongAI,orHumanLevelAI,ArtificialGeneralIntelligencereferstoacomputerthatisassmartasahumanacrossthe

boardamachinethatcanperformanyintellectualtaskthatahumanbeingcan.CreatingAGIisa
muchhardertaskthancreatingANI,andwereyettodoit.ProfessorLindaGottfredsondescribes
intelligenceasaverygeneralmentalcapabilitythat,amongotherthings,involvestheabilitytoreason,
plan,solveproblems,thinkabstractly,comprehendcomplexideas,learnquickly,andlearnfrom
experience.AGIwouldbeabletodoallofthosethingsaseasilyasyoucan.
AICaliber3)ArtificialSuperintelligence(ASI):OxfordphilosopherandleadingAIthinkerNick
Bostromdefinessuperintelligenceasanintellectthatismuchsmarterthanthebesthumanbrainsin
practicallyeveryfield,includingscientificcreativity,generalwisdomandsocialskills.Artificial
Superintelligencerangesfromacomputerthatsjustalittlesmarterthanahumantoonethats
trillionsoftimessmarteracrosstheboard.ASIisthereasonthetopicofAIissuchaspicymeatball
andwhythewordsimmortalityandextinctionwillbothappearinthesepostsmultipletimes.
Asofnow,humanshaveconqueredthelowestcaliberofAIANIinmanyways,anditseverywhere.
TheAIRevolutionistheroadfromANI,throughAGI,toASIaroadwemayormaynotsurvivebut
that,eitherway,willchangeeverything.
Letstakeacloselookatwhattheleadingthinkersinthefieldbelievethisroadlookslikeandwhythis
revolutionmighthappenwaysoonerthanyoumightthink:

WhereWeAreCurrentlyAWorldRunningonANI
ArtificialNarrowIntelligenceismachineintelligencethatequalsorexceedshumanintelligenceor
efficiencyataspecificthing.Afewexamples:
CarsarefullofANIsystems,fromthecomputerthatfiguresoutwhentheanti-lockbrakesshould
kickintothecomputerthattunestheparametersofthefuelinjectionsystems.Googlesselfdrivingcar,whichisbeingtestednow,willcontainrobustANIsystemsthatallowittoperceiveand
reacttotheworldaroundit.
YourphoneisalittleANIfactory.Whenyounavigateusingyourmapapp,receivetailoredmusic
recommendationsfromPandora,checktomorrowsweather,talktoSiri,ordozensofother
everydayactivities,youreusingANI.
YouremailspamfilterisaclassictypeofANIitstartsoffloadedwithintelligenceabouthowto
figureoutwhatsspamandwhatsnot,andthenitlearnsandtailorsitsintelligencetoyouasit
getsexperiencewithyourparticularpreferences.TheNestThermostatdoesthesamethingasit
startstofigureoutyourtypicalroutineandactaccordingly.
YouknowthewholecreepythingthatgoesonwhenyousearchforaproductonAmazonand
thenyouseethatasarecommendedforyouproductonadifferentsite,orwhenFacebook
somehowknowswhoitmakessenseforyoutoaddasafriend?ThatsanetworkofANIsystems,
workingtogethertoinformeachotheraboutwhoyouareandwhatyoulikeandthenusingthat
informationtodecidewhattoshowyou.SamegoesforAmazonsPeoplewhoboughtthisalso
boughtthingthatsanANIsystemwhosejobitistogatherinfofromthebehaviorofmillions
ofcustomersandsynthesizethatinfotocleverlyupsellyousoyoullbuymorethings.
GoogleTranslateisanotherclassicANIsystemimpressivelygoodatonenarrowtask.Voice
recognitionisanother,andthereareabunchofappsthatusethosetwoANIsasatagteam,
allowingyoutospeakasentenceinonelanguageandhavethephonespitoutthesamesentence
inanother.
Whenyourplanelands,itsnotahumanthatdecideswhichgateitshouldgoto.Justlikeitsnota
humanthatdeterminedthepriceofyourticket.
TheworldsbestCheckers,Chess,Scrabble,Backgammon,andOthelloplayersarenowallANI
systems.
GooglesearchisonelargeANIbrainwithincrediblysophisticatedmethodsforrankingpagesand
figuringoutwhattoshowyouinparticular.SamegoesforFacebooksNewsfeed.
Andthosearejustintheconsumerworld.SophisticatedANIsystemsarewidelyusedinsectors
andindustrieslikemilitary,manufacturing,andfinance(algorithmichigh-frequencyAItraders
accountformorethanhalfofequitysharestradedonUSmarkets 6 ),andinexpertsystemslike
thosethathelpdoctorsmakediagnosesand,mostfamously,IBMsWatson,whocontained
enoughfactsandunderstoodcoyTrebek-speakwellenoughtosoundlybeatthemostprolific
Jeopardychampions.
http://waitbutwhy.com/2015/01/artificialintelligencerevolution1.html

6/26

4/14/2015

TheAIRevolution:RoadtoSuperintelligenceWaitButWhy

ANIsystemsastheyarenowarentespeciallyscary.Atworst,aglitchyorbadly-programmedANIcan
causeanisolatedcatastrophelikeknockingoutapowergrid,causingaharmfulnuclearpowerplant
malfunction,ortriggeringafinancialmarketsdisaster(likethe2010FlashCrashwhenanANIprogram
reactedthewrongwaytoanunexpectedsituationandcausedthestockmarkettobrieflyplummet,
taking$1trillionofmarketvaluewithit,onlypartofwhichwasrecoveredwhenthemistakewas
corrected).
ButwhileANIdoesnthavethecapabilitytocauseanexistentialthreat,weshouldseethisincreasingly
largeandcomplexecosystemofrelatively-harmlessANIasaprecursoroftheworld-alteringhurricane
thatsontheway.EachnewANIinnovationquietlyaddsanotherbrickontotheroadtoAGIandASI.Or
asAaronSaenzseesit,ourworldsANIsystemsareliketheaminoacidsintheearlyEarthsprimordial
oozetheinanimatestuffoflifethat,oneunexpectedday,wokeup.

TheRoadFromANItoAGI
WhyItsSoHard
Nothingwillmakeyouappreciatehumanintelligencelikelearningabouthowunbelievablychallenging
itistotrytocreateacomputerassmartasweare.Buildingskyscrapers,puttinghumansinspace,
figuringoutthedetailsofhowtheBigBangwentdownallfareasierthanunderstandingourown
brainorhowtomakesomethingascoolasit.Asofnow,thehumanbrainisthemostcomplexobjectin
theknownuniverse.
WhatsinterestingisthatthehardpartsoftryingtobuildAGI(acomputerassmartashumansin

general,notjustatonenarrowspecialty)arenotintuitivelywhatyoudthinktheyare.Buildacomputer
thatcanmultiplytwoten-digitnumbersinasplitsecondincrediblyeasy.Buildonethatcanlookata
dogandanswerwhetheritsadogoracatspectacularlydifficult.MakeAIthatcanbeatanyhumanin
chess?Done.Makeonethatcanreadaparagraphfromasix-year-oldspicturebookandnotjust
recognizethewordsbutunderstandthemeaningofthem?Googleiscurrentlyspendingbillionsof
dollarstryingtodoit.Hardthingslikecalculus,financialmarketstrategy,andlanguagetranslation
aremind-numbinglyeasyforacomputer,whileeasythingslikevision,motion,movement,and
perceptionareinsanelyhardforit.Or,ascomputerscientistDonaldKnuthputsit,AIhasbynow
succeededindoingessentiallyeverythingthatrequiresthinkingbuthasfailedtodomostofwhat
peopleandanimalsdowithoutthinking.' 7
Whatyouquicklyrealizewhenyouthinkaboutthisisthatthosethingsthatseemeasytousareactually
unbelievablycomplicated,andtheyonlyseemeasybecausethoseskillshavebeenoptimizedinus(and
mostanimals)byhundredsofmillionyearsofanimalevolution.Whenyoureachyourhanduptoward
anobject,themuscles,tendons,andbonesinyourshoulder,elbow,andwristinstantlyperformalong
seriesofphysicsoperations,inconjunctionwithyoureyes,toallowyoutomoveyourhandinastraight
linethroughthreedimensions.Itseemseffortlesstoyoubecauseyouhaveperfectedsoftwareinyour
brainfordoingit.Sameideagoesforwhyitsnotthatmalwareisdumbfornotbeingabletofigureout
theslantywordrecognitiontestwhenyousignupforanewaccountonasiteitsthatyourbrainis
superimpressiveforbeingableto.
Ontheotherhand,multiplyingbignumbersorplayingchessarenewactivitiesforbiologicalcreatures
andwehaventhadanytimetoevolveaproficiencyatthem,soacomputerdoesntneedtoworktoo
hardtobeatus.Thinkaboutitwhichwouldyouratherdo,buildaprogramthatcouldmultiplybig
numbersoronethatcouldunderstandtheessenceofaBwellenoughthatyoucouldshowitaBinany
oneofthousandsofunpredictablefontsorhandwritinganditcouldinstantlyknowitwasaB?
Onefunexamplewhenyoulookatthis,youandacomputerbothcanfigureoutthatitsarectangle
withtwodistinctshades,alternating:

http://waitbutwhy.com/2015/01/artificialintelligencerevolution1.html

7/26

4/14/2015

TheAIRevolution:RoadtoSuperintelligenceWaitButWhy

Tiedsofar.Butifyoupickuptheblackandrevealthewholeimage

youhavenoproblemgivingafulldescriptionofthevariousopaqueandtranslucentcylinders,slats,
and3-Dcorners,butthecomputerwouldfailmiserably.Itwoulddescribewhatitseesavarietyof
two-dimensionalshapesinseveraldifferentshadeswhichisactuallywhatsthere.Yourbrainisdoing
atonoffancyshittointerprettheimplieddepth,shade-mixing,androomlightingthepictureistrying
toportray. 8 Andlookingatthepicturebelow,acomputerseesatwo-dimensionalwhite,black,and
graycollage,whileyoueasilyseewhatitreallyisaphotoofanentirely-black,3-Drock:

Credit:MatthewLloyd

Andeverythingwejustmentionedisstillonlytakinginstagnantinformationandprocessingit.Tobe
human-levelintelligent,acomputerwouldhavetounderstandthingslikethedifferencebetween
subtlefacialexpressions,thedistinctionbetweenbeingpleased,relieved,content,satisfied,andglad,
andwhyBraveheartwasgreatbutThePatriotwasterrible.
Daunting.
Sohowdowegetthere?

http://waitbutwhy.com/2015/01/artificialintelligencerevolution1.html

8/26

4/14/2015

TheAIRevolution:RoadtoSuperintelligenceWaitButWhy

FirstKeytoCreatingAGI:IncreasingComputationalPower
OnethingthatdefinitelyneedstohappenforAGItobeapossibilityisanincreaseinthepowerof
computerhardware.IfanAIsystemisgoingtobeasintelligentasthebrain,itllneedtoequalthe
brainsrawcomputingcapacity.
Onewaytoexpressthiscapacityisinthetotalcalculationspersecond(cps)thebraincouldmanage,
andyoucouldcometothisnumberbyfiguringoutthemaximumcpsofeachstructureinthebrainand
thenaddingthemalltogether.
RayKurzweilcameupwithashortcutbytakingsomeonesprofessionalestimateforthecpsofone
structureandthatstructuresweightcomparedtothatofthewholebrainandthenmultiplying
proportionallytogetanestimateforthetotal.Soundsalittleiffy,buthedidthisabunchoftimeswith
variousprofessionalestimatesofdifferentregions,andthetotalalwaysarrivedinthesameballpark
around1016,or10quadrillioncps.
Currently,theworldsfastestsupercomputer,ChinasTianhe-2,hasactuallybeatenthatnumber,
clockinginatabout34quadrillioncps.ButTianhe-2isalsoadick,takingup720squaremetersofspace,
using24megawattsofpower(thebrainrunsonjust20watts),andcosting$390milliontobuild.Not
especiallyapplicabletowideusage,orevenmostcommercialorindustrialusageyet.
Kurzweilsuggeststhatwethinkaboutthestateofcomputersbylookingathowmanycpsyoucanbuy
for$1,000.Whenthatnumberreacheshuman-level10quadrillioncpsthenthatllmeanAGIcould
becomeaveryrealpartoflife.
MooresLawisahistorically-reliablerulethattheworldsmaximumcomputingpowerdoubles
approximatelyeverytwoyears,meaningcomputerhardwareadvancement,likegeneralhuman
advancementthroughhistory,growsexponentially.LookingathowthisrelatestoKurzweilscps/$1,000
metric,werecurrentlyatabout10trillioncps/$1,000,rightonpacewiththisgraphspredicted
trajectory: 9

Sotheworlds$1,000computersarenowbeatingthemousebrainandtheyreataboutathousandth
ofhumanlevel.Thisdoesntsoundlikemuchuntilyourememberthatwewereataboutatrillionthof
humanlevelin1985,abillionthin1995,andamillionthin2005.Beingatathousandthin2015putsus
rightonpacetogettoanaffordablecomputerby2025thatrivalsthepowerofthebrain.
Soonthehardwareside,therawpowerneededforAGIistechnicallyavailablenow,inChina,andwell
bereadyforaffordable,widespreadAGI-caliberhardwarewithin10years.Butrawcomputational
poweralonedoesntmakeacomputergenerallyintelligentthenextquestionis,howdowebring
human-levelintelligencetoallthatpower?
http://waitbutwhy.com/2015/01/artificialintelligencerevolution1.html

9/26

4/14/2015

TheAIRevolution:RoadtoSuperintelligenceWaitButWhy

SecondKeytoCreatingAGI:MakingitSmart
Thisistheickypart.Thetruthis,noonereallyknowshowtomakeitsmartwerestilldebatinghowto
makeacomputerhuman-levelintelligentandcapableofknowingwhatadogandaweird-writtenB
andamediocremovieis.Butthereareabunchoffar-fetchedstrategiesoutthereandatsomepoint,
oneofthemwillwork.HerearethethreemostcommonstrategiesIcameacross:

1)Plagiarizethebrain.
Thisislikescientiststoilingoverhowthatkidwhositsnexttotheminclassissosmartandkeepsdoing
sowellonthetests,andeventhoughtheykeepstudyingdiligently,theycantdonearlyaswellasthat
kid,andthentheyfinallydecidekfuckitImjustgonnacopythatkidsanswers.Itmakessense
werestumpedtryingtobuildasuper-complexcomputer,andtherehappenstobeaperfectprototype
foroneineachofourheads.
Thescienceworldisworkinghardonreverseengineeringthebraintofigureouthowevolutionmade
sucharadthingoptimisticestimatessaywecandothisby2030.Oncewedothat,wellknowallthe
secretsofhowthebrainrunssopowerfullyandefficientlyandwecandrawinspirationfromitandsteal
itsinnovations.Oneexampleofcomputerarchitecturethatmimicsthebrainistheartificialneural
network.Itstartsoutasanetworkoftransistorneurons,connectedtoeachotherwithinputsand
outputs,anditknowsnothinglikeaninfantbrain.Thewayitlearnsisittriestodoatask,say
handwritingrecognition,andatfirst,itsneuralfiringsandsubsequentguessesatdecipheringeach
letterwillbecompletelyrandom.Butwhenitstolditgotsomethingright,thetransistorconnectionsin
thefiringpathwaysthathappenedtocreatethatanswerarestrengthened;whenitstolditwaswrong,
thosepathwaysconnectionsareweakened.Afteralotofthistrialandfeedback,thenetworkhas,by
itself,formedsmartneuralpathwaysandthemachinehasbecomeoptimizedforthetask.Thebrain
learnsabitlikethisbutinamoresophisticatedway,andaswecontinuetostudythebrain,were
discoveringingeniousnewwaystotakeadvantageofneuralcircuitry.
Moreextremeplagiarisminvolvesastrategycalledwholebrainemulation,wherethegoalistoslicea
realbrainintothinlayers,scaneachone,usesoftwaretoassembleanaccuratereconstructed3-D
model,andthenimplementthemodelonapowerfulcomputer.Wedthenhaveacomputerofficially
capableofeverythingthebrainiscapableofitwouldjustneedtolearnandgatherinformation.If
engineersgetreallygood,theydbeabletoemulatearealbrainwithsuchexactaccuracythatthe
brainsfullpersonalityandmemorywouldbeintactoncethebrainarchitecturehasbeenuploadedtoa
computer.IfthebrainbelongedtoJimrightbeforehepassedaway,thecomputerwouldnowwakeup
asJim(?),whichwouldbearobusthuman-levelAGI,andwecouldnowworkonturningJimintoan
unimaginablysmartASI,whichhedprobablybereallyexcitedabout.
Howfararewefromachievingwholebrainemulation?Wellsofar,wevenotyetjustrecentlybeenable
toemulatea1mm-longflatwormbrain,whichconsistsofjust302totalneurons.Thehumanbrain
contains100billion.Ifthatmakesitseemlikeahopelessproject,rememberthepowerofexponential
progressnowthatweveconqueredthetinywormbrain,anantmighthappenbeforetoolong,
followedbyamouse,andsuddenlythiswillseemmuchmoreplausible.

2)Trytomakeevolutiondowhatitdidbeforebutforusthistime.
Soifwedecidethesmartkidstestistoohardtocopy,wecantrytocopythewayhestudiesforthe
testsinstead.
Heressomethingweknow.Buildingacomputeraspowerfulasthebrainispossibleourownbrains
evolutionisproof.Andifthebrainisjusttoocomplexforustoemulate,wecouldtrytoemulate
evolutioninstead.Thefactis,evenifwecanemulateabrain,thatmightbeliketryingtobuildan
airplanebycopyingabirdswing-flappingmotionsoften,machinesarebestdesignedusingafresh,
machine-orientedapproach,notbymimickingbiologyexactly.
SohowcanwesimulateevolutiontobuildAGI?Themethod,calledgeneticalgorithms,wouldwork
somethinglikethis:therewouldbeaperformance-and-evaluationprocessthatwouldhappenagain
andagain(thesamewaybiologicalcreaturesperformbylivinglifeandareevaluatedbywhether
theymanagetoreproduceornot).Agroupofcomputerswouldtrytodotasks,andthemostsuccessful
oneswouldbebredwitheachotherbyhavinghalfofeachoftheirprogrammingmergedtogetherinto
anewcomputer.Thelesssuccessfuloneswouldbeeliminated.Overmany,manyiterations,this
naturalselectionprocesswouldproducebetterandbettercomputers.Thechallengewouldbecreating
anautomatedevaluationandbreedingcyclesothisevolutionprocesscouldrunonitsown.
http://waitbutwhy.com/2015/01/artificialintelligencerevolution1.html

10/26

4/14/2015

TheAIRevolution:RoadtoSuperintelligenceWaitButWhy

Thedownsideofcopyingevolutionisthatevolutionlikestotakeabillionyearstodothingsandwe
wanttodothisinafewdecades.
Butwehavealotofadvantagesoverevolution.First,evolutionhasnoforesightandworksrandomly
itproducesmoreunhelpfulmutationsthanhelpfulones,butwewouldcontroltheprocesssoitwould
onlybedrivenbybeneficialglitchesandtargetedtweaks.Secondly,evolutiondoesntaimforanything,
includingintelligencesometimesanenvironmentmightevenselectagainsthigherintelligence(since
itusesalotofenergy).We,ontheotherhand,couldspecificallydirectthisevolutionaryprocesstoward
increasingintelligence.Third,toselectforintelligence,evolutionhastoinnovateinabunchofother
waystofacilitateintelligencelikerevampingthewayscellsproduceenergywhenwecanremove
thoseextraburdensandusethingslikeelectricity.Itsnodoubtwedbemuch,muchfasterthan
evolutionbutitsstillnotclearwhetherwellbeabletoimproveuponevolutionenoughtomakethis
aviablestrategy.

3)Makethiswholethingthecomputersproblem,notours.
Thisiswhenscientistsgetdesperateandtrytoprogramthetesttotakeitself.Butitmightbethemost
promisingmethodwehave.
TheideaisthatwedbuildacomputerwhosetwomajorskillswouldbedoingresearchonAIand
codingchangesintoitselfallowingittonotonlylearnbuttoimproveitsownarchitecture.Wedteach
computerstobecomputerscientistssotheycouldbootstraptheirowndevelopment.Andthatwould
betheirmainjobfiguringouthowtomakethemselvessmarter.Moreonthislater.

AllofThisCouldHappenSoon
Rapidadvancementsinhardwareandinnovativeexperimentationwithsoftwarearehappening
simultaneously,andAGIcouldcreepuponusquicklyandunexpectedlyfortwomainreasons:
1)Exponentialgrowthisintenseandwhatseemslikeasnailspaceofadvancementcanquicklyrace
upwardsthisGIFillustratesthisconceptnicely:

Source

2)Whenitcomestosoftware,progresscanseemslow,butthenoneepiphanycaninstantlychangethe
rateofadvancement(kindoflikethewayscience,duringthetimehumansthoughttheuniversewas
geocentric,washavingdifficultycalculatinghowtheuniverseworked,butthenthediscoverythatitwas
heliocentricsuddenlymadeeverythingmucheasier).Or,whenitcomestosomethinglikeacomputer
thatimprovesitself,wemightseemfarawaybutactuallybejustonetweakofthesystemawayfrom
havingitbecome1,000timesmoreeffectiveandzoomingupwardtohuman-levelintelligence.

TheRoadFromAGItoASI
Atsomepoint,wellhaveachievedAGIcomputerswithhuman-levelgeneralintelligence.Justabunch
ofpeopleandcomputerslivingtogetherinequality.
Ohactuallynotatall.
Thethingis,AGIwithanidenticallevelofintelligenceandcomputationalcapacityasahumanwould
http://waitbutwhy.com/2015/01/artificialintelligencerevolution1.html

11/26

4/14/2015

TheAIRevolution:RoadtoSuperintelligenceWaitButWhy

stillhavesignificantadvantagesoverhumans.Like:

Hardware:
Speed.Thebrainsneuronsmaxoutataround200Hz,whiletodaysmicroprocessors(whichare
muchslowerthantheywillbewhenwereachAGI)runat2GHz,or10milliontimesfasterthan
ourneurons.Andthebrainsinternalcommunications,whichcanmoveatabout120m/s,are
horriblyoutmatchedbyacomputersabilitytocommunicateopticallyatthespeedoflight.
Sizeandstorage.Thebrainislockedintoitssizebytheshapeofourskulls,anditcouldntget
muchbiggeranyway,orthe120m/sinternalcommunicationswouldtaketoolongtogetfrom
onebrainstructuretoanother.Computerscanexpandtoanyphysicalsize,allowingfarmore
hardwaretobeputtowork,amuchlargerworkingmemory(RAM),andalongtermmemory(hard
drivestorage)thathasbothfargreatercapacityandprecisionthanourown.
Reliabilityanddurability.Itsnotonlythememoriesofacomputerthatwouldbemoreprecise.
Computertransistorsaremoreaccuratethanbiologicalneurons,andtheyrelesslikelyto
deteriorate(andcanberepairedorreplacediftheydo).Humanbrainsalsogetfatiguedeasily,
whilecomputerscanrunnonstop,atpeakperformance,24/7.

Software:
Editability,upgradability,andawiderbreadthofpossibility.Unlikethehumanbrain,
computersoftwarecanreceiveupdatesandfixesandcanbeeasilyexperimentedon.The
upgradescouldalsospantoareaswherehumanbrainsareweak.Humanvisionsoftwareis
superblyadvanced,whileitscomplexengineeringcapabilityisprettylow-grade.Computerscould
matchthehumanonvisionsoftwarebutcouldalsobecomeequallyoptimizedinengineeringand
anyotherarea.
Collectivecapability.Humanscrushallotherspeciesatbuildingavastcollectiveintelligence.
Beginningwiththedevelopmentoflanguageandtheformingoflarge,densecommunities,
advancingthroughtheinventionsofwritingandprinting,andnowintensifiedthroughtoolslike
theinternet,humanityscollectiveintelligenceisoneofthemajorreasonswevebeenabletoget
sofaraheadofallotherspecies.Andcomputerswillbewaybetteratitthanweare.Aworldwide
networkofAIrunningaparticularprogramcouldregularlysyncwithitselfsothatanythingany
onecomputerlearnedwouldbeinstantlyuploadedtoallothercomputers.Thegroupcouldalso
takeononegoalasaunit,becausetherewouldntnecessarilybedissentingopinionsand
motivationsandself-interest,likewehavewithinthehumanpopulation. 10
AI,whichwilllikelygettoAGIbybeingprogrammedtoself-improve,wouldntseehuman-level
intelligenceassomeimportantmilestoneitsonlyarelevantmarkerfromourpointofviewand
wouldnthaveanyreasontostopatourlevel.Andgiventheadvantagesoverusthatevenhuman
intelligence-equivalentAGIwouldhave,itsprettyobviousthatitwouldonlyhithumanintelligencefor
abriefinstantbeforeracingonwardstotherealmofsuperior-to-humanintelligence.
Thismayshocktheshitoutofuswhenithappens.Thereasonisthatfromourperspective,A)whilethe
intelligenceofdifferentkindsofanimalsvaries,themaincharacteristicwereawareofaboutany
animalsintelligenceisthatitsfarlowerthanours,andB)weviewthesmartesthumansasWAY
smarterthanthedumbesthumans.Kindoflikethis:

http://waitbutwhy.com/2015/01/artificialintelligencerevolution1.html

12/26

4/14/2015

TheAIRevolution:RoadtoSuperintelligenceWaitButWhy

SoasAIzoomsupwardinintelligencetowardus,wellseeitassimplybecomingsmarter,forananimal.
Then,whenithitsthelowestcapacityofhumanityNickBostromusesthetermthevillageidiot
wellbelike,Ohwow,itslikeadumbhuman.Cute!Theonlythingis,inthegrandspectrumof
intelligence,allhumans,fromthevillageidiottoEinstein,arewithinaverysmallrangesojustafter
hittingvillageidiot-levelandbeingdeclaredtobeAGI,itllsuddenlybesmarterthanEinsteinandwe
wontknowwhathitus:

Andwhathappensafterthat?

AnIntelligenceExplosion
Ihopeyouenjoyednormaltime,becausethisiswhenthistopicgetsunnormalandscary,andits
gonnastaythatwayfromhereforward.IwanttopauseheretoremindyouthateverysinglethingIm
goingtosayisrealrealscienceandrealforecastsofthefuturefromalargearrayofthemost
respectedthinkersandscientists.Justkeeprememberingthat.
Anyway,asIsaidabove,mostofourcurrentmodelsforgettingtoAGIinvolvetheAIgettingthereby
self-improvement.AndonceitgetstoAGI,evensystemsthatformedandgrewthroughmethodsthat
didntinvolveself-improvementwouldnowbesmartenoughtobeginself-improvingiftheywantedto.
3

Andhereswherewegettoanintenseconcept:recursiveself-improvement.Itworkslikethis
http://waitbutwhy.com/2015/01/artificialintelligencerevolution1.html

13/26

4/14/2015

TheAIRevolution:RoadtoSuperintelligenceWaitButWhy

AnAIsystematacertainlevelletssayhumanvillageidiotisprogrammedwiththegoalof
50.4k

improvingitsownintelligence.Onceitdoes,its
smartermaybeatthispointitsatEinsteinslevelso
Shares
nowwhenitworkstoimproveitsintelligence,withanEinstein-levelintellect,ithasaneasiertimeandit
canmakebiggerleaps.Theseleapsmakeit
muchsmarterthananyhuman,allowingittomakeeven
29k

biggerleaps.Astheleapsgrowlargerandhappenmorerapidly,theAGIsoarsupwardsinintelligence
11
andsoonreachesthesuperintelligentlevelofanASIsystem.ThisiscalledanIntelligenceExplosion,
10.2k
anditstheultimateexampleofTheLawofAcceleratingReturns.
ThereissomedebateabouthowsoonAIwillreachhuman-levelgeneralintelligencethemedianyear
onasurveyofhundredsofscientistsaboutwhentheybelievedwedbemorelikelythannottohave
reachedAGIwas2040 12 thatsonly25yearsfromnow,whichdoesntsoundthathugeuntilyou
considerthatmanyofthethinkersinthisfieldthinkitslikelythattheprogressionfromAGItoASI
happensveryquickly.Likethiscouldhappen:

IttakesdecadesforthefirstAIsystemtoreachlow-levelgeneralintelligence,butitfinallyhappens.A
computerisabletounderstandtheworldarounditaswellasahumanfour-year-old.Suddenly,within
anhourofhittingthatmilestone,thesystempumpsoutthegrandtheoryofphysicsthatunifies
generalrelativityandquantummechanics,somethingnohumanhasbeenabletodefinitivelydo.90
minutesafterthat,theAIhasbecomeanASI,170,000timesmoreintelligentthanahuman.
Superintelligenceofthatmagnitudeisnotsomethingwecanremotelygrasp,anymorethana
bumblebeecanwrapitsheadaroundKeynesianEconomics.Inourworld,smartmeansa130IQand
stupidmeansan85IQwedonthaveawordforanIQof12,952.
WhatwedoknowisthathumansutterdominanceonthisEarthsuggestsaclearrule:withintelligence

comespower.WhichmeansanASI,whenwecreateit,willbethemostpowerfulbeinginthehistoryof
lifeonEarth,andalllivingthings,includinghumans,willbeentirelyatitswhimandthismighthappen
inthenextfewdecades.
Ifourmeagerbrainswereabletoinventwifi,thensomething100or1,000or1billiontimessmarter
thanweareshouldhavenoproblemcontrollingthepositioningofeachandeveryatomintheworldin
anywayitlikes,atanytimeeverythingweconsidermagic,everypowerweimagineasupremeGodto
havewillbeasmundaneanactivityfortheASIasflippingonalightswitchisforus.Creatingthe
technologytoreversehumanaging,curingdiseaseandhungerandevenmortality,reprogramming
theweathertoprotectthefutureoflifeonEarthallsuddenlypossible.Alsopossibleistheimmediate
endofalllifeonEarth.Asfaraswereconcerned,ifanASIcomestobeing,thereisnowanomnipotent
GodonEarthandtheall-importantquestionforusis:

WillitbeaniceGod?

ThatsthetopicofPart2ofthispost.
___________

SourcesatthebottomofPart2.

RelatedWaitButWhyPosts
TheFermiParadoxWhydontweseeanysignsofalienlife?
PuttingTimeinPerspectiveAvisuallookatthehistoryoftimesincetheBigBang
Orforsomethingtotallydifferentandyetsomehowrelated,WhyProcrastinatorsProcrastinate
AndheresYear1ofWaitButWhyonanebook.

Like

Share

62k

Tweet

7,321

1point

http://waitbutwhy.com/2015/01/artificialintelligencerevolution1.html

105

12753

StumbleUpon

14/26