NgoMeyi 2020, i-OpenAI, inkampani ye-AI eyasungulwa ngu-Elon Musk no-Sam Altman, yashicilela i-GPT-3, yase yethulwa njengenethiwekhi enkulu ye-neural yamanje. Imodeli yolimi yesimanje, I-GPT-3 ihlanganisa amapharamitha ayizigidi eziyizinkulungwane ezingu-175 uma kuqhathaniswa namapharamitha angu-1,5 billion we-GPT-2 eyandulelayo.
GPT-3 shaya imodeli ye-NLG Turing (I-Turing Natural Language Generation) evela ku-Microsoft enamapharamitha ayizigidi eziyizinkulungwane ezingu-17 ayephethe irekhodi lenethiwekhi enkulu kunazo zonke ye-neural. Imodeli yolimi iye yamangala, yagxekwa yaze yabhekisiswa; iphinde yathola izinhlelo zokusebenza ezintsha nezithokozisayo.
Futhi manje Amahemuhemu akhululiwe ukuthi ukukhululwa kwe-GPT-4, inguqulo elandelayo yemodeli yolimi lwe-OpenAI, ingase ibe maduzane.
Noma kunjalo alukho usuku lokukhishwa olumenyezelwe okwamanje, I-OpenAI inikeze izinkomba ezithile mayelana nezici zomlandeli we-GPT-3, abaningi abangalindela ngayo, ukuthi i-GPT-4 akufanele ibe nkulu kune-GPT-3, kodwa kufanele isebenzise izinsiza eziningi zokubala, ezizokhawulela umthelela wayo kwezemvelo.
Phakathi neseshini, U-Altman ukusho lokho, ngokuphambene nenkolelo evamile, I-GPT-4 ngeke ibe imodeli yolimi enkulu kunazo zonke. Imodeli ngokungangabazeki izoba nkulu kunezizukulwane ezedlule zamanethiwekhi e-neural, kodwa usayizi ngeke ube uphawu lwayo.
Okokuqala, izinkampani ziye zaqaphela ukuthi ukusebenzisa usayizi wemodeli njengenkomba yokuthuthukisa ukusebenza akuyona ukuphela noma indlela engcono kakhulu yokwenza. Ngo-2020, u-Jared Kaplan nozakwabo e-OpenAI kubikwa ukuthi baphetha ngokuthi ukusebenza kuba ngcono kakhulu lapho ukunyuka kwesabelomali sekhompiyutha kunikezwa ngokuyinhloko ukukhulisa inani lamapharamitha, kulandela ubudlelwano bomthetho wamandla. I-Google, i-Nvidia, i-Microsoft, i-OpenAI, i-DeepMind, nezinye izinkampani ezithuthukisa amamodeli olimi zithathe le mihlahlandlela ngokwenani elibonakalayo.
Kodwa i-MT-NLG (Megatron-Turing NLG, inethiwekhi ye-neural eyakhiwe yi-Nvidia ne-Microsoft ngonyaka odlule ngamapharamitha ayizigidi eziyizinkulungwane ezingu-530), inkulu njengoba injalo, ayiyona engcono kakhulu uma kuziwa ekusebenzeni. Eqinisweni, ayilinganiselwe njengenhle kakhulu kunoma yisiphi isigaba sebhentshimakhi. Amamodeli amancane afana ne-Gopher noma i-Chinchilla (amapharamitha ayizigidi eziyizinkulungwane ezingu-70), ingxenye encane nje yosayizi wawo, angaba ngcono kakhulu kune-MT-NLG kuyo yonke imisebenzi. Ngakho-ke, kwacaca ukuthi ubukhulu bemodeli akuwona kuphela isici esiholela ekuqondeni kangcono ulimi.
Ngokuka-Altman, amamodeli ezilimi ahlushwa umkhawulo obalulekile. uma kuziwa ekwenzeni kahle. Ukuqeqesha bekuzobiza kakhulu kangangokuthi izinkampani kuzodingeka ziyekethise phakathi kokunemba kanye nezindleko. Lokhu kuvame ukuholela ekutheni amamodeli enziwe kahle.
I-CEO ibike ukuthi i-GPT-3 yaqeqeshwa kanye kuphela, naphezu kwamaphutha athile ukuthi kwezinye izimo ayengaholela ekuqeqeshweni kabusha. Ngenxa yalokhu, i-OpenAI kubikwa ukuthi inqume ngokumelene nayo ngenxa yezindleko ezingabizi, ezivimbele abacwaningi ekutholeni isethi engcono kakhulu yama-hyperparameters emodeli.
Omunye umphumela wezindleko zokuqeqesha eziphezulu ukuthi ukuhlaziya ukuziphatha okuyimodeli kuzokhawulelwa. Ngokombiko owodwa, lapho abacwaningi be-AI bephetha ngokuthi ubukhulu bemodeli buyinto eguquguqukayo efanelekile kakhulu yokuthuthukisa ukusebenza, abazange bacabangele inani lamathokheni okuqeqesha, okungukuthi, inani ledatha elinikezwe amamodeli. Lokhu bekuyodinga amanani angavamile wezinsiza zekhompiyutha. Izinkampani zezobuchwepheshe kubikwa ukuthi zilandele lokho okutholwe ngabacwaningi ngoba bekungcono kakhulu ababenakho.
I-Altman ithe i-GPT-4 izosebenzisa izibalo eziningi kuneyandulelayo. I-OpenAI kulindeleke ukuthi isebenzise imibono ehlobene nokwenza kahle ku-GPT-4, nakuba ingakanani izinga elingenakubikezelwa njengoba isabelomali sayo singaziwa.
Nokho, izitatimende ze U-Altman ukhombisa ukuthi i-OpenAI kufanele igxile ekwandiseni okuguquguqukayo ngaphandle kosayizi wemodeli.. Ukuthola isethi engcono kakhulu yama-hyperparameter, usayizi ophelele wemodeli, kanye nenani lamapharamitha kungaholela ekuthuthukisweni okumangalisayo kuwo wonke amabhentshimakhi.
Ngokwabahlaziyi, zonke izibikezelo zamamodeli olimi zizowa uma lezi zindlela zihlanganiswa zibe imodeli eyodwa. U-Altman uphinde wathi abantu ngeke bakholwe ukuthi amamodeli angaba ngcono kangakanani ngaphandle kokuthi abe makhudlwana. Kungase kuphakamise ukuthi imizamo yokukala isiphelile okwamanje.
Kubikwa ukuthi i-OpenAI yenze umzamo omkhulu ekuxazululeni inkinga yokuqondanisa kwe-AI: kwenziwa kanjani amamodeli olimi alandele izinhloso zabantu futhi ahambisane nezindinganiso zomuntu?
Abahlaziyi bathi lokhu akuyona nje inkinga enzima yezibalo (siyenza kanjani i-AI iqonde kahle esikufunayo?), kodwa futhi nefilosofi (ayikho indlela yendawo yonke yokuqondanisa i-AI nabantu, njengoba ukuhlukahluka kwe amanani omuntu ukusuka eqenjini kuya kwelinye makhulu futhi avame ukungqubuzana).
Okokugcina uma unentshisekelo yokwazi kabanzi ngakhoungabhekisela kokuthunyelwe kwangempela Kulesi sixhumanisi esilandelayo.