RevealingOpenAI’splantocreateAGIby2027InthisdocumentIwillberevealinginformationIhavegatheredregardingOpenAI’s(delayed)planstocreatehuman-levelAGIby2027.Notallofitwillbeeasilyverifiablebuthopefullythere’senoughevidencetoconvinceyouSummary:OpenAIstartedtraininga125trillionparametermultimodalmodelinAugustof2022.ThefirststagewasArrakisalsocalledQ*.ThemodelfinishedtraininginDecemberof2023butthelaunchwascanceledduetohighinferencecost.ThisistheoriginalGPT-5whichwasplannedforreleasein2025.Gobi(GPT-4.5)hasbeenrenamedtoGPT-5becausetheoriginalGPT-5hasbeencanceled.ThenextstageofQ*,originallyGPT-6butsincerenamedtoGPT-7(originallyforreleasein2026),hasbeenputonholdbecauseoftherecentlawsuitbyElonMuskQ*2025(GPT-8)wasplannedtobereleasedin2027achievingfullAGI...Q*2023=48IQQ*2024=96IQ(delayed)Q*2025=145IQ(delayed)ElonMuskcausedthedelaybecauseofhislawsuit.ThisiswhyI’mrevealingtheinformationnowbecausenofurtherharmcanbedoneI’veseenmanydefinitionsofAGI–artificialgeneralintelligence–butIwilldefineAGIsimplyasanartificialintelligencethatcandoanyintellectualtaskasmarthumancan.Thisishowmostpeopledefinethetermnow.2020wasthefirsttimeIwasshockedbyanAIsystem–thatwasGPT-3.GPT-3.5,anupgradedversionofGPT-3,isthemodelbehindChatGPT.WhenChatGPTwasreleased,IfeltasthoughthewiderworldwasfinallycatchinguptosomethingIwasinteractingwith2yearsprior.IusedGPT-3extensivelyin2020andwasshockedbyitsabilitytoreason.GPT-3,anditshalf-stepsuccessorGPT-3.5(whichpoweredthenowfamousChatGPT--beforeitwasupgradedtoGPT-4inMarch2023),wereamassivesteptowardsAGIinawaythatearliermodelsweren’t.Thethingtonoteis,earlierlanguagemodelslikeGPT-2(andbasicallyallchatbotssinceEliza)hadnorealabilitytorespondcoherentlyatall.SowhywasGPT-3suchamassiveleap?...ParameterCount“Deeplearning”isaconceptthatessentiallygoesbacktothebeginningofAIresearchinthe1950s.Thefirstneuralnetworkwascreatedinthe50s,andmodernneuralnetworksarejust“deeper”,meaning,theycontainmorelayers–they’remuch,muchbiggerandtrainedonlotsmoredata.MostofthemajortechniquesusedinAItodayarerootedinbasic1950sresearch,combinedwithafewminorengineeringsolutionslike“backpropogation”and“transformermodels”.TheoverallpointisthatAIresearchhasn’tfundamentallychangedin70years.So,there’sonlytworealreasonsfortherecentexplosionofAIcapabilities:sizeanddata.Agrowingnumberofpeopleinthefieldarebeginningtobelievewe’vehadthetechnicaldetailsofAGIsolvedformanydecades,butmerelydidn’thaveenoughcomputingpoweranddatatobuildituntilthe21stcentury.Obviously,21stcenturycomputersarevastlymorepowerfulthan1950scomputers.Andofcourse,theinternetiswhereallthedatacamefrom.So,whatisaparameter?Youmayalreadyknow,buttogiveabriefdigestiblesummary,it’sanalogoustoasynapseinabiologicalbrain,whichisaconnectionbetweenneurons.Eachneuroninabiologicalbrainhasroughly1000connectionstootherneurons.Obviously,digitalneuralnetworksareconceptuallyanalogoustobiologicalbrains....…So,howmanysynapses(or“parameters”)areinahumanbrain?Themostcommonlycitedfigureforsynapsecountinthebrainisroughly100trillion,whichwouldmeaneachneuron(~100billioninthehumanbrain)hasroughly1000connections.Ifeachneuroninabrainhas1000connections,thismeansacathasroughly250billionsynapses,andadoghas530billionsynapses.Synapsecountgenerallyseemstopredicthigherintelligence,withafewexceptions:forinstance,elephantstechnicallyhaveahighersynapsecountthanhumansyetdisplaylowerintelligence.Thesimplestexplanationforlargersynapsecountswithlowerintelligenceisasmalleramountofqualitydata.Fromanevolutionaryperspective,brainsare“trained”onbillionsofyearsofepigeneticdata,andhumanbrainsevolvedfromhigherqualitysocializationandcommunicationdatathanelephants,leadingtooursuperiorabilitytoreason.Regardless,synapsecountisdefinitelyimportant.Again,theexplosioninAIcapabilitiessincetheearly2010shasbeentheresultoffarmorecomputingpowerandfarmoredata.GPT-2had1.5billionconnections,whichislessthanamouse’sbrain(~10billionsynapses).GPT-3had175billionconnections,whichisgettingsomewhatclosetoacat’sbrain.Isn’titintuitivelyobviousthatanAIsystemthesizeofacat’sbrainwouldbesuperiortoanAIsystemsmallerthanamouse’sbrain?...PredictingAIPerformance…In2020,afterthereleaseofthe175billionparameterGPT-3,manyspeculatedaboutthepotentialperformanceofamodel~600timeslargerat100trillionparameters,becausethisparametercountwouldmatchthehumanbrain’ssynapsecount.Therewasnostrongindicationin2020thatanyonewasactivelyworkingonamodelofthissize,butitwasinterestingtospeculateabout.Thebigquestionis,isitpossibletopredictAIperformancebyparametercount?Asitturnsout,theanswerisyes,asyou’llseeonthenextpage.[Source:/posts/k2SNji3jXaLGhBeYP/extrapolating-gpt-n-performance][TheaboveisfromLanrian’sLessWrongpost.]…AsLanrianillustrated,extrapolationsshowthatAIperformanceinexplicablyseemstoreachhuman-levelatthesametimeashuman-levelbrainsizeismatchedwithparametercount.Hiscountforthesynapsenumberinthebrainisroughly200trillionparametersasopposedtothecommonlycited100trillionfigure,butthepointstillstands,andtheperformanceat100trillionparametersisremarkablyclosetooptimal.Bytheway–animportantthingtonoteisthatalthough100trillionisslightlysuboptimalinperformance,thereisanengineeringtechniqueOpenAIisusingtobridgethisgap.I’llexplainthistowardstheveryendofthedocumentbecauseit’scrucialtowhatOpenAIisbuilding.Lanrian’spostisoneofmanysimilarpostsonline–it’sanextrapolationofperformancebasedonthejumpbetweenpreviousmodels.OpenAIcertainlyhasmuchmoredetailedmetricsandthey’vecometothesameconclusionasLanrian,asI’llshowlaterinthisdocument.So,ifAIperformanceispredictablebasedonparametercount,and~100trillionparametersisenoughforhuman-levelperformance,whenwilla100trillionparameterAImodelbereleased?...GPT-5achievedprotoAGIinlate2023withanIQof48…Thefirstmentionofa100trillionparametermodelbeingdevelopedbyOpenAIwasinthesummerof2021,mentionedoffhandinawiredinterviewbytheCEOofCerebras(AndrewFeldman),acompanywhichSamAltmanisamajorinvestorof.SamAltman’sresponsetoAndrewFeldman,atanonlinemeetupandQ&AcalledAC10,whichtookplaceinSeptember2021.It’scrucialtonotethatSamAltmanADMITStotheirplansfora100trillionparametermodel.(Sources:/gpt-4-a-viral-case-of-ai-misinformation-c3f999c1f589/r/GPT3/comments/pj0ly6/sam_altman_gpt4_will_be_remain_textonly_will_not/TheredditpostingitselfissourcedfromaLessWrongpost,whichwasdeletedatSamAltman’srequest:/posts/aihztgJrknBdLHjd2/sam-altman-q-and-a-gpt-and-agi)…AIresearcherIgorBaikovmadetheclaim,onlyafewweekslater,thatGPT-4wasbeingtrainedandwouldbereleasedbetweenDecemberandFebruary.Again,IwillprovethatIgorreallydid

haveaccurateinformation,andisacrediblesource.ThiswillbeimportantsoonGwernisafamousfigureintheAIworld–heisanAIresearcherandblogger.HemessagedIgorBaikovonTwitter(inSeptember2022)andthisistheresponsehereceived.Importanttoremember:“Colossalnumberofparameters”.“Text”,“audio”,“images”,“possiblyvideo”,and“multimodal”.Thiscomesfromasubredditcalled“thisisthewayitwillbe”whichisasmall,privatesubredditI’mpartof,runbyamathematicsprofessorwithaninterestinAGI.AIenthusiasts(andafewexperts)usethesubreddittodiscussAItopicsdeeperthanwhatyou’llfindinthemainstream.A“colossalnumberofparameters”?SoundslikeIgorBaikovwasreferencinga100trillionparametermodel,as500billionparametermodelsandupto1trillionparametermodelshadalreadybeentrainedmanytimesbythetimeofhistweetinsummer2022(makingmodelsofthatsizeunexceptionalandcertainlynot“colossal”).Thesetweetsfrom“rxpu”,seeminglyanAIenthusiast(?)fromTurkey,areinterestingbecausetheymakeaverysimilarclaimaboutGPT-4’sreleasewindowbeforeanyoneelsedid(trustme–Ispentmanyhours,daily,scouringtheinternetforsimilarclaims,andnooneelsemadethisspecificclaimbeforehedid).Healsomentionsa“125trillionsynapse”GPT-4–however,heincorrectlystatesGPT-3’sparametercountas1trillion.(Itseemsasthoughrxpudidhaveinsideinformation,butgotsomethingmixedupwiththeparametercounts–again,Iwillillustratethislater,andprovethatrxpuwasnotlying).…Thisisaweakerpieceofevidence,butit’sworthincludingbecause“roon”isfairlynotableasaSiliconValleyAIresearcher,andisfollowedbySamAltman,CEOofOpenAI,andotherOpenAIresearchersonTwitter.InNovember2022IreachedouttoanAIbloggernamedAlbertoRomero.HispostsseemtospreadprettyfaronlinesoIwashopingthatifIsenthimsomebasicinfoaboutGPT-4hemightdoawriteupandthewordwouldgetout.TheresultsofthisattemptwereprettyremarkableasI’llshowinthenexttwopages.AlbertoRomero’spost.Thegeneralresponsewillbeshownonthenextpage.The100trillionparameterleakwentviral,reachingmillionsofpeople,tothepointthatOpenAIemployeesincludingCEOSamAltmanhadtorespond–callingit“completebullshit”.TheVergecalledit“factuallyincorrect”.AlbertoRomeroclaimedresponsibilityfortheleakasyoucanseeontheleft.…IgorBaikov,theoriginofthe“colossalnumberofparameters”statement,alsosawtheviralspreadoftheGPT-4leak(whichwasessentiallyhisowndoing)andresponded.So,afterall,Igorreallydidmean“100trillionparameters”whenhesaid“acolossalnumberofparameters”.But,isIgorareliablesource?Arehisotherclaimsaccurate?Whataboutthemultimodality?WhatabouttheabilityforGPT-4toprocessimages,sounds,andvideos?IwillproveIgor’sreliabilityshortly.SomewherearoundOct/Nov2022IbecameconvincedthatOpenAIplannedtofirstreleasea~1-2trillionparametersubsetofGPT-4beforereleasingthefull100trillionparametermodel(“GPT-5”).Thesesourcesaren’tparticularlysolidbuttheyallsaidthesamething–includingrxpu,whoonceclaimedtherewasa125trillionparametermodelintheworks,andthenincorrectlyclaimedGPT-3was1trillion–Ibelievehegothisinformationmixedup.<(Date:2022)Thesourceshereareofvaryingcredibility(JyriandLeeorareSanFranciscoinvestorsandHarrisisanAIresearcher)buttheyallInexplicablysaythesamething--GPT-4wasbeingtestedinOct/Novof2022.AndaccordingtoUSmilitaryAIresearcherCherieMPoland,itwasdefinitely(FromOctober2022^)beingtrainedinOctober,whichagainlinesupwithIgorBaikov’sleak.…OpenAI’sofficialposition,asdemonstratedbySamAltmanhimself,isthattheideaofa100trillionparameterGPT-4is“completebullshit”.Thisishalftrue,asGPT-4isa1trillionparametersubsetofthefull100trillionparametermodel.Justtoillustratethatthe100trillionparametermodelhasn’tarrivedyetandisstillindevelopment,SemaforinMarch2023(shortlyafterthereleaseofGPT-4)claimedGPT-4is1trillionparameters.(OpenAIhasrefusedtoofficiallydiscloseparametercount).SomethingelseworthnothingisthatOpenAIclaimsGPT-4was“finishedtraining”inAugust,whereasweknowthata“colossal”multimodalmodelwasbeingtrainedbetweenAugustandOctober.Oneexplanationforthisis,OpenAIlied.Anotherpossibilityisthatthe1trillionparameterGPT-4mayhavefinisheditsfirstroundoftraininginAugust,butwentthroughadditionalretrainingbetweenAugustandOctober,whichiswhenthebulkofthefull100trillionparametermodelwastrained.IwillnowprovidemyevidencethatGPT-4wasnotjusttrainedontextandimages,butwasalsotrainedonaudioandvideo.FrancisHellyerseemsmoderatelycrediblebutthispageisnotthemostsolidpieceofevidence–I’mincludingitbecauseit’sseemstocorroboratewhatothersourcesaresaying.Francisisaninvestor,entrepreneurandwriter.Theinformationhelistedinhistweetabouttheteam“runningoutofinternet”cannotbefoundinanyotherpublication,anyleak,oranyonlineposting,sohedidnot“steal”itfromsomeotherplace.Anincrediblysolidsourceonthenextpage.TheCTOofMicrosoftGermany,aweekpriortotheofficialreleaseofGPT-4,seemstohaveslippedupandrevealedthatthereexistsaGPT-4whichhastheabilitytoprocessvideos.IimaginehewasunawareofOpenAI’sdecisionnottorevealthevideocapabilitiesofthesystem.ThiscompletelyprovesthatGPT-4/5wastrainedonnotjusttextandimages,butalsovideodata,andofcoursewecaninferthataudiodatawasincludedaswell.Clearly,Igor’sclaimaboutthe100trillionparametermodelwastrue,downtoeveryprecisedetail.AnothersourcethatlinesupwithIgor’sclaim,isacredibleentrepreneurwhostated(onOct252022)thatGPT-4’sreleasedatewouldbebetweenJanuaryandFebruaryof2023:AlthoughGPT-4wasreleasedinMarch2023,slightlyoutsidetheDec-FebwindowclaimedbyIgorBaikov(whichIbelievewasdoneintentionallybyOpenAItodiscreditIgor’sleak),BingChatGPT(basedonGPT-4)wasactuallyannouncedinFebruaryof2023,clearlyshowingthatthewindowclaimedbyIgorhadvalidity,andwasprobablychangedlastminutebyapanickedOpenAI.Anoteaboutrobotics:AIresearchersarebeginningtobelievethatvisionisallthat’snecessaryforoptimalreal-world/physicalperformance.Justtogiveoneexample,Teslacompletelyditchedallsensorsandcommittedfullytovisionfortheirself-drivingcars.Thepointis,trainingahuman-brain-sizedAImodelonalltheimageandvideodataontheinternetwillclearlybemorethanenoughtohandlecomplexroboticstasks.Commonsensereasonisburiedinthevideodata,justlikeit’sburiedinthetextdata(andthetext-focusedGPT-4isstunninglygoodatcommonsensereasoning).ArecentexamplefromGoogle,ofroboticscapabilitiesbeinglearnedfromalargevision/languagemodel.(Minimalroboticsdatawasrequiredontopofthelanguageandvisiontraining,andtheknowledgefromvisualandtexttaskstransferredtotheroboticstasks.OpenAIistrainingtheir100trillionparametermodelon“allthedataontheinternet”whichwillundoubtedlyincluderoboticsdata).Palm-Eisa~500billionparametermodel–whathappenstoroboticsperformancewhenyoutraina100trillionparametermodelonallthedataavailableontheinternet?(MoreonGoogle’sPalm-Emodelonthenextpage).Anotherroboticsdevelopment–thistimefromTesla(May162023).Theytrainedtheirrobot“Optimus”tograspanobject–and“notaskspecificprogrammingwasdone”.EverythingwaslearnedfromHUMANDEMONSTRATIONS.“Thismeanswecannowscalequicklytomanytasks.”O(jiān)nceagain:ifhumandemonstrationsareallthatisneededforadvancedroboticsperformance,a100trillionparametermodeltrainedonallthevideoonthewebwouldcertainlybeabletoachieveastonishingroboticsperformance...Theimageontheleftshowswhatthe1trillionparameterGPT-4iscapableofintermsofimagerecognition.Theresponseisalreadyclearerandmorewellwrittenthanwhatmanyhumanswouldhavecomeupwith.So,again,whathappenswhenyoutrainamodel100timeslargerthanGPT-4,whichisthesizeofthehumanbrain,onallthedataavailableontheinternet?Theaboveimageisoneofmanyshortsamplesofthe1trillionparameterGPT-4’stextoutput.Ifthisiswhata1trillionparametermodelcanwrite,whatwilla100trillionparametermodelbeabletowrite?WhetheranAImodelcaneverbetruly“creative”isupfordebate,butanAIfakingcreativityiscertainlypossibleandisALREADYHAPPENING.Important:noticehowtheAImodelisabletogeneratemultipleanglesofthesamescenewithphysicallyaccuratelighting,andinsomecasesevenphysicallyaccuratefluidandrain.Ifyoucangenerateimagesandvideoswithaccurate,common-sensephysics,youhaveCOMMONSENSEREASONING.Ifyoucangeneratecommonsense,youUNDERSTANDcommonsense.VIDEO^IMAGE^IMAGES^VIDEO--><--VIDEO<--IMAGESExamplesofthecurrentlevelofqualityofpublicallyavailablevideo&imagegenerationAImodels.Thesemodelsarelessthan10billionparametersinsize.Whathappens,whenyoutrainamodel10,000timeslarger,onallthedataavailableontheinternet,andgiveittheabilitytogenerateimagesandvideo?(Theanswer:imagesandvideoscompletelyindistinguishablefromtherealthing,100%ofthetime,withnoexceptions,noworkarounds,nopossiblewayforanyonetotellthedifference,nomatterhowhardtheytry).-(update:SORAISFROMGPT-5Q*2023MODEL)TwopostsfromLongjumping-Sky-1971.I’mincludingthisbecauseheaccuratelypredictedthereleasedateofGPT-4weeksinadvance(nooneelsepostedthisinformationpubliclybeforehand,meaninghehadaninsidesource).Hispostsnowhavemuchmorecredibility–andheclaimedimageandaudiogenerationwouldbetrainedinQ3of2023.Ifvideogenerationtrainingissimultaneousorshortlyafter,thislinesupwithSiqiChen‘sclaimofGPT-5beingfinishedtraininginDecemberof2023.Let’stakeastrollbacktoFebruary2020,afewmonthsbeforeGPT-3wasreleased.AnarticlefromTechnologyReview,whichwasan“insidestory”aboutOpenAI,seemstosuggestthatOpenAIwasintheearlystagesofa“secret”projectinvolvinganAIsystemtrainedonimages,text,and“otherdata”,andthatleadershipatOpenAIthoughtitwasthemostpromisingwaytoreachAGI.Iwonderwhatthiscouldpossiblybereferringto.ThenextslidewillrevealsomequotesfromthePresidentofOpenAI–from2019–anditwilltellyouwhattheirplanwas.OpenAIpresidentGregBrockmanstatedin2019,followinga1billiondollarinvestmentfromMicrosoftatthetime,thatOpenAIplannedtobuildahuman-brain-sizedmodelwithinfiveyears,andthatthiswastheirplanforhowtoachieveAGI.2019+5=2024BothofthesesourcesareclearlyreferringtothesameplantoachieveAGI–ahuman-brain-sizedAImodel,trainedon“images,text,andotherdata”,duetobetrainedwithinfiveyearsof2019,so,by2024.SeemstolineupwithalltheothersourcesI’velistedinthisdocument...Source:TimeMagazine,Jan122023AsI’llshowinthesenextfewslides,AIleadersaresuddenlystartingtosoundthealarm–almostliketheyknowsomethingVERYSPECIFICthatthegeneralpublicdoesn’t.DateofNYTinterview:May12023“Ithoughtitwas30to50yearsorevenlongeraway.Obviously,Inolongerthinkthat.”Whatmadehimsuddenlychangehismind--ANDdecidetoleaveGoogletospeakaboutthedangersofAI?ShortlyafterthereleaseofGPT-4,theFutureofLifeInstitute,ahighlyinfluentialnon-profitorganizationconcernedwithmitigatingpotentialcatastrophicriskstotheworld,releasedanopenlettercallingonallAIlabstopauseAIdevelopmentforsixmonths.Why?Thefirstreleasedversionoftheletterspecificallysaid“(includingthecurrently-being-trainedGPT-5)”.Whywasthatincluded,andwhywasitremoved?Source:Wired,March292023Source:Vox,March292023SomealarmingquotesfromaninterviewandQ&AwithSamAltmanfromOctober2022--youtubelink:/watch?v=b022FECpNe8(Time:49:30)AudienceQ&Aquestion:“DowehaveenoughinformationintheinternettocreateAGI?”SamAltman’sblunt,immediateresponse,interruptingthemanaskingthequestion:“Yes.”Samelaborates:“Yeah,we’reconfidentthereis.Wethinkaboutthisandmeasureitquitealot.”Theinterviewerinterjects:“Whatgivesyouthatconfidence?”Sam’sreply:“OneofthethingsIthinkthatOpenAIhasdriveninthefieldthat’sbeenreallyhealthyisthatyoucantreatscalinglawsasascientificprediction.Youcandothisforcompute,youcandothisfordata,butyoucanmeasureatsmallscaleandyoucanpredictquiteaccuratelyhowit’sgoingtoscaleup.Howmuchdatayou’regoingtoneed,howmuchcomputeyou’regoingtoneed,howmanyparametersyou’regoingtoneed,whenthegenerateddatagetsgoodenoughtobehelpful…Andtheinternetis…there’salotofdataoutthere.There’salotofvideoouttheretoo.”MorequotesfromthisQ&Aonthenextslide.AnotherquotefromtheQ&AwithSamAltman--(Time:53:00)[Note–anAIwinterisanextendedperiodoftimewheretheAIfieldreceiveslimitedfundingandisnotgivenmuchattentionbyseriousresearchers.Thishappenedtwice--onceinthe70sand80sandagainfromthemid80suntilroughlythelate2000s.]Anotheraudiencequestion:“CouldwehaveanotherAIwinterandwhatmightcauseit?”SamAltman’sresponse:“CouldwehaveanAIwinterandwhatmightcauseit…yeah,ofcourse.Ithinkwewon’thaveoneverysoon.Becauseevenifweneverfigureoutanotherresearchidea,theeconomicvalueofthecurrentparadigmandhowmuchfurtherthatcanbepushedisgonnacarryusformanyyearstocome.Butitispossible,howeverunlikely,thatwearestillmissingthekeyideatogobeyondbehavioralcloningandthesemodelsaregonnabe,like,stuckathuman-levelforever.There’sabunchofreasonswhyIdon’tthinkthat’struebutifanyonetellsyouwecouldnotpossiblyeverhaveanotherwinterinthisresearchfieldyoushouldneverbelievethem.”IdetailwhytheseSamAltmanquotesareconcerningonthenextpage.OnSamAltman’sQ&AFirstly,SamAltmanseemshighly,highlyconfidentthatthereexistsenoughdataontheinternettotrainanAGIsystem–confidenttothepointthatitmakesonequestionifthey’vealreadydoneit,orareintheprocessofdoingit.Secondly,the“AIwinter”conceptgenerallyreferstoaperiodwhereprogressTOWARDSAGIhasbeenslowed,butSamAltmanretooledthetermtorefertoaperiodwhereprogressTOWARDSSUPERINTELLIGENCEisslowed.ThisseemstosuggestthatOpenAIhasalreadybuiltanAGIsystem,orareveryclosetoit,andAGIisnolongerthegoalbecauseitalreadyexists.AsImentionedearlierinthedocument,a100trillionparametermodelisactuallyslightlysuboptimal,butthereisanewscalingparadigmOpenAIisusingtobridgethisgap–it’sbasedonsomethingcalledthe“Chinchillascalinglaws.”ChinchillawasanAImodelunveiledbyDeepMindinearly2022.TheimplicationoftheChinchillaresearchpaperwasthatcurrentmodelsaresignificantlyundertrained,andwithfarmorecompute(meaningmoredata)wouldseeamassiveboostinperformancewithouttheneedtoincreaseparameters.Thepointis,whileanundertrained100trillionparametermodelmaybeslightlysuboptimal,ifitweretrainedonvastlymoredataitwouldeasilybeabletoEXCEED

human-levelperformance.TheChinchillaparadigmiswidelyunderstoodandacceptedinthefieldofmachinelearning,butjusttogiveaspecificexamplefromOpenAI,PresidentGregBrockmandiscussesinthisinterviewhowOpenAIrealizedtheirinitialscalinglawswereflawed,andhavesinceadjustedtotaketheChinchillalawsintoaccount:https://youtu.be/Rp3A5q9L_bg?t=1323Peoplehavesaid,“trainingacomputeoptimal100trillionparametermodelwouldcostbillionsofdollarsandjustisn’tfeasi