I-Google imemezele ukwethulwa kokuhlolwa kokuqala emhlabeni okungaziwa yizinhlangothi zombili kwamamodeli e-AI aphambili aphethwe ngasese, okubeka ukuhlolwa kwangaphandle endaweni ebethelwe “yebhokisi elimnyama” ukuze kuvinjelwe amamodeli ekutholeni kusenesikhathi ulwazi lokuhlolwa nokwenza kahle ukusebenza kwawo.
Njengoba umfundi engenakubona imibuzo yokuhlolwa kusenesikhathi ukuze izinga lakhe langempela libonakale, ukuhlolwa kwamamodeli e-AI nakho kubhekene nenkinga efanayo: uma imodeli ike yathola imibuzo yokuhlolwa kusenesikhathi (okubizwa ngokuthi “ukungcoliswa kwesilinganiso”), ukwethembeka kwemiphumela yokuhlola kuyehla kakhulu.
I-Google ibambisene ne-Singapore AI Safety Institute, i-OpenMined, i-AVERI kanye ne-MLCommons ukwenza ukuhlolwa okuyimfihlo kwemodeli ye-Gemini Flash Lite endaweni evikela ubumfihlo, ngaleyo ndlela kuthuthukiswe ubuqiniso bemiphumela yokuhlola.
Ukuhlolwa kwangaphandle okunobungozi obukhulu bekulokhu kubhekene nokukhetha okunzima: kungaba abahloli banikeze ngemiyalelo yokuhlola, okudala ingozi yokuthi abanikazi bemodeli babone imibuzo kusenesikhathi; noma abanikazi bemodeli banikeze ngezisindo zemodeli, okubeka engcupheni ukuvuza kwempahla yengqondo.
Ukuhlolwa okungaziwa yizinhlangothi zombili kuxazulula le mpikiswano. Kusetshenziswa i-Confidential Space, eyingxenye yeqoqo lemikhiqizo ye-Google Cloud yokubala okuyimfihlo, ukuqinisekiswa okubethelwe kungaqinisekisa ukuthi idatha yokuhlola yangaphandle kanye nemodeli ephethwe ngasese kuhlala kuyimfihlo kubanikazi bazo: abahloli abakwazi ukubona izisindo zemodeli ye-Gemini, kanti i-Google ayikwazi ukubona imiyalelo yokuhlola yabahloli.
Le ndlela yokubethela ingavimbela ngempumelelo ukungcoliswa kwezilinganiso, kuyilapho ivikela idatha ebucayi. Lokhu kubaluleke kakhulu ekuhlolweni okubucayi okwenziwa izinhlangano zokuphepha kwenethiwekhi noma zikahulumeni. Ukuhlolwa okungaziwa yizinhlangothi zombili kwenza izinhlangano ezizimele zikwazi ukuhlola ngokuqinile amamodeli e-AI athuthukile ngaphandle kokuphazamisa ubukhosi nokuphepha kwedatha.
I-Google ithe inethemba lokuthi lolu hlelo lokuhlola luzovula indlela entsha yokwengamela amamodeli, lusize imboni yonke yakhe izinhlelo ze-AI eziphephe kakhudlwana, ezithembeke kakhudlwana nezithenjwa kakhulu umphakathi.