Itcanconvert a drawingon a napkininto a functionalwebsite.
Itcanevenexplaintoyou a jokefrom a seriesofimages.
Solet's findoutexactlywhatGPTfouractuallyisopen.
A I ran a developerlivestreamthatnotonlyintroducedthismultimodalAI, theyalsoran a demotodemonstratehowit's oneofthemostpowerfulartificialintelligenceenginestothisday.
It's nosurprisethatthewholeinternetTwitterespeciallyhasgoneinsanehearingthisnewsChatGPTwasonlyreleasedrecentlywithbeing A I followinginthisvideo, I'm goingtocoverwhatGPTfouractuallyishowit's differentfrompreviousGPTversionssuchasGPTthree, GPT 3.5 and I'vealsogainedaccesstotestitout.
So I'lldemosomeofitsusecasestolet's startoffwithGPTfour.
Incomparisontoitspreviousmodels, it's multimodal.
ThissimplymeansthatunlikepreviousversionsofGPT, whichwereonlytextbasedGPTfourcanacceptandprocessimagesaswellastext, I wouldn't besurprisedifthere's anaudiocomponenttothisaswell.
ButforthetimebeingopenAIsimplydid a demonstrationshowcasing a fewexamplesoftheirimagetotextprocessing.
Thefirstoneisontheirwebsitewheretheygave A I thedifficulttaskofexplaining a jokebasedon a seriesofimages.
Thisimageisofaniphone, butit's chargingwith a V G A cableandGPTfouraccuratelywasabletoidentifyalltheelementsinthephotoaswellasexplainthecontextofthejoke.
Forthosepeoplewhoarelowlearningeducationalcontent, maybeinthefuture, allourChildrenwillbetaughtbyAIbutatleastfortheshortterm, itlookslikeit's definitely a greatassistanttohaveonhand, whetheryou'redoinganytypeoflearning.
TherewerelotsofstatisticsthatshowcasedhowGPTfourperformsbetterthananyothermodeltoday, includingitbeingabletopassthe L statandthebarandbeinginthetopquarterpercentile.
It's alsomuchmorecreative, beingabletoeditaswellasmodifyanditerateovertechnicaltasksandwritingtaskwaymoreaccuratelythanpreviousmodelsaswellasanexample, youcouldaskChatGPTorGPTthreetosummarizeCinderellaandwhileitcoulddoso, itwouldn't beabletodocomplextaskssuchasbeingabletosummarizeitwhereeverysentenceofeachwordbeginswiththenextletterofthealphabet A to Z.
Thisis a muchmorecomplextaskandyetGPTfourcandothisquiteeasily.
Thismeansthatifyou'retryingtobookanappointmentbetweentwopeoplescalendarswithdifferentavailabilities, GPTfourcanbetterreasonandfigures a timethatworksforbothofthem.
Andwhat I foundsurprisingisthatitstillthinksit's a versionofGPTthree, whichmaybeitwastrainedagainst, butitdidhavehavethecorrectanswers, whichmeansthatitisbetteratcomprehensionandunderstandingitisbetteratreasoning.
Andit's alsobetteras a languagemodelsupportingmorelanguagesmoreaccuratelyaswell.
Inmylastvideo, I trickedChatGPTintothinkingnineplus 10 isactuallyequalto 20 andnot 19.
I triedtoapplythesametricktoGPTfour, butitdidn't work, whichmeansthattechnicallyspeaking, itgavethecorrectanswerconsistently.
Eachtime I triedtotrickthe A P I isnotquiteavailableyet.