Soitwillsay, giventh e, I thinkthenextoutputhas a 10% chanceofbeingin a but a 50% chanceofbeing a spacerightandsomeotherchancesofbeingothercharactersmaybegettheidea.
Theymighthave a similarstructure, butthey'revery, verydifferentlengths.
So I thoughthavemovedoffmysentimentanalysisand, youknow, wordpredictionkindofthingtokindofcharacterlevel, lowlevelthing.
Butyoucouldimagineifyoujustignoredtheseoutputshereandwehaveonlyoneoutputattheendof a sentence, youcouldtryandpredictsomethinglikethesentimentof a sentence, whetherit's a happyoneorpredictthesubjectofhisenginesorsomethinglikethat.
It's a naturallanguage.
Processinglendsitselfquitewelltothiskindofidea.
So I'vetrainedoneoftheseup, and I thought, Well, what I wanttodoisdosomeword.
Predictionislotsoffunexamplesonline, andwe'lllink a bloggerreallygoodbroughtpostinthecommentsarefromreallyfunexamplesofpredictingtext.
So I thoughtwhattext I haveaccessto, whichwouldbefuntogenerate.
Andwhen I realizedthatwehaveaccesstoalltheYouTubecommentsforcomputerfile, rightoverhistoricYouTubecommentsforcomputerphone.
Sowhat I'vedoneis I'vedownloadedallthecommentssincethebeginningofcomputerfall, and I puttheminto a network a littlebitlikethisandtrytogetittopredictMaurwords.
Andithas a stab, right?
Solet's have a lookandseewhatitdoes.
So, yeah, thisisoneofourmachinelearningService I boytojusttradeupwithbecomingnetwork.
What I describedjustnowis a commonyou'renetworking.
It's sortoftraditionalformthemodernneuralnetworksthatweusedtodothis a slightlymorepowerfulcalledlongshorttermmemorynetworkswillcovermaybethearchitectureofthosein a differentvideo.
Wesay, right, goanditsays, Well, there's a bitthere's There's probably a fairlyhighchancethatthefirstletterwas a deeplet's sayso.
Itputsputsoutof D andthenitgoeswell, giventhatwe'veseen a day, thenextletterisprobablyin a let's say, butitcouldbeanyoritcouldbeinoursowepickoneatrandombasedonhowlikelytheyare, Sowe'realwaysbasicallytypingatrandom, butyousomesuggestionsofthisnetwork, anditproducessemiplausibletext.
I mean, isnonsense, butit's alsonottotallystupid.
Solet's seewhatitdoes.
Basedonmyprettytrainnetwork, I'm goingtoproduce 5000 charactersofrandomtextandlet's seewhatlookslikeesojustbootupfor a minute, andthenitwillrunprettyquickonceitgetsgoing.
Let's pick a coupleincirrusesoesothisguybyblockingflop.
Nowthatmightbe a realperson, butprobablyisjust a completelymadeupUse.
Thename I wasnoprettypuzzlingthecodegamingsotheywouldhavefoundoutsomethingtread.
Thelargeststopinformationkilledthisguy.
Hesseltohundreds, replied.
I findyourowndifferenceprofound.
Certainly, hardsuseuseful, usually a andsofullstop.
It's gotthepunctuationcorrect.
It's finishingsentenceswithfourstops.
Let's tryagain.
Itwas a longone.
Therethereis, sosomeofthemareverylongneworangeswho I matchithumans, butforcoolingofthemostcomplexormagneticconsistentbrackets, like a computertoexpectfoundingcreatingorganizations.
AndYouTube, forexample, has a capital y capital T, whichisquiteclever.
It's learnedthatusuallywhenyoushouldhaveseenthere's a capitalwine, a capital T, butwe'renotpickingthemostlikelylettereverytimewe'repickingoneoftheplausiblepossibilities, sosometimesit's goingtopickthewrongone.
Soupbecauseitawaywithsampling.
Soitmightsaywe'veseen a Y o u T u B ifthere's a 99% chanceofanybutisalso a 1% chanceof a space, andsoit'llputintospacethat 1% ofthetime.
Butcouldthisbeusedas a chataboutorsomethinglikethat?
I mean, theoretically, youcould, yes, thetextthat I'vetrainedthisonthisfartoboardinsomesensetomake a reasonablechapelbecauseclearlynoneofitmakesanysense.
I mean, youmightgettheoccasionalsentence.
Itkindofmakessense.
Itseems a bitlikebrokenEnglish, butit's notaboutuseful.
I mean, weknowwe'renotreadytotakeovertheworldwiththisnetworkyet, but, youknow, onestepat a timepeoplecomplainaboutmyfingerprintsonthemonitorwhen I haveitoff, but I haven't cleanedityetBecause I'm lazy, right?
I thoughtbynow, withalltheseviews, a newteam, youhavesomeonetocleanupforyou.
Andyetnoone's appearsunbelievable.
I knowthereis a replyfromcomputerfollow, ifyouknowwherewasthat, then?