NgoMeyi ka-2020, i-OpenAI, inkampani ye-AI eyasekwa ngu-Elon Musk kunye no-Sam Altman, yapapasha i-GPT-3, emva koko yaboniswa njengenethiwekhi enkulu ye-neural okwangoku. Imodeli yolwimi lwesizukulwana esilandelayo, I-GPT-3 iquka i-175 yeebhiliyoni zeeparamitha xa kuthelekiswa ne-1,5 yeebhiliyoni zeeparamitha ze-GPT-2 eyandulelayo.
I-GPT-3 betha imodeli ye-NLG Turing (I-Microsoft's Turing Natural Language Generation) ene-17 yeebhiliyoni zeeparamitha ebezibambe irekhodi yothungelwano olukhulu lwe-neural. Imodeli yolwimi imangaliswe, yagxekwa kwaye yade yaphononongwa; Ikwafumene izicelo ezintsha nezinika umdla.
Kwaye ngoku Amahemuhemu akhululwe ukuba ukuqaliswa kwe-GPT-4, inguqulelo elandelayo yemodeli yolwimi lwe-OpenAI, inokufika kungekudala.
Nangona kunjalo Umhla wokukhupha awukabhengezwa, I-OpenAI inike izibonakaliso ezithile malunga neempawu zomlandeli we-GPT-3, kunye noko abaninzi banokulindela, ukuba i-GPT-4 ayifanele ibe mkhulu kune-GPT-3, kodwa kufuneka isebenzise izixhobo ezininzi zokubala, eziza kunciphisa impembelelo yayo yendalo.
Ngexesha leseshoni, U-Altman wayichaza loo ntoNgokuchasene neenkolelo ezithandwayo, I-GPT-4 ayisayi kuba yeyona modeli yolwimi inkulu. Imodeli ngokuqinisekileyo iya kuba nkulu kunezizukulwana zangaphambili zothungelwano lwe-neural, kodwa ubukhulu abuyi kuba luphawu lwayo.
Okokuqala, iinkampani ziye zaqaphela ukuba ukusebenzisa ubungakanani bemodeli njengeproksi yokuphucula ukusebenza ayisiyiyo yodwa okanye yeyona ndlela ilungileyo yokwenza oko. Ngo-2020, uJared Kaplan kunye noogxa bakhe e-OpenAI bagqibe kwelokuba intsebenzo iphucuka kakhulu xa ukonyuka kohlahlo lwabiwo-mali lwekhompiyutha lwabelwa ikakhulu ekwandiseni inani leeparamitha, kulandela ubudlelwane bomthetho wamandla. I-Google, i-Nvidia, iMicrosoft, i-OpenAI, i-DeepMind, kunye nezinye iinkampani eziphuhlisa imodeli yolwimi ziye zathatha ezi zikhokelo entliziyweni.
Kodwa i-MT-NLG (i-Megatron-Turing NLG, inethiwekhi ye-neural eyakhiwe yi-Nvidia kunye ne-Microsoft kunyaka odlulileyo kunye ne-530 yeebhiliyoni zeeramitha), njengoko inkulu njengoko injalo, ayiyona nto ibhetele xa ifika ekusebenzeni. Ngapha koko, ayinikwanga ireyithingi yeyona nto ingcono kulo naliphi na inqanaba lomgangatho. Iimodeli ezincinci, ezifana neGopher okanye iChinchilla (i-70 yeebhiliyoni zeeparamitha), iqhezu nje lobungakanani bazo, ziya kuba ngcono kakhulu kuneMT-NLG kuyo yonke imisebenzi. Ngaloo ndlela, kwacaca ukuba ubungakanani bemodeli ayisiyiyo kuphela into ekhokelela ekuqondeni ulwimi olungcono.
Ngokuka-Altman, iimodeli zolwimi zinengxaki yokusikelwa umda xa isiza kulungiselelo. Uqeqesho luya kuxabisa kakhulu kangangokuba iinkampani kuya kufuneka ziyeke ukuchaneka kunye neendleko. Oku kuhlala kubangela ukuba iimodeli zingalungiswanga kakuhle.
I-CEO ibike ukuba i-GPT-3 iqeqeshwe kanye kuphela, nangona ezinye iimpazamo ezinokuthi zikhokelele ekuqeqeshweni kwakhona. Ngenxa yoku, kuxelwe ukuba i-OpenAI yagqiba ngokuchasene nayo ngenxa yeendleko ezingenakufikeleleka, ezithintele abaphandi ekufumaneni eyona seti yeehyperparameters zemodeli.
Esinye isiphumo seendleko eziphezulu zoqeqesho kukuba uhlahlelo lokuziphatha kwemodeli luya kuthintelwa. Ngokutsho kwengxelo enye, xa abaphandi be-AI baphetha ukuba ubukhulu bemodeli buyinto eguquguqukayo ebaluleke kakhulu ekuphuculeni ukusebenza, abazange bacinge ngenani lamathokheni oqeqesho, oko kukuthi, inani ledatha enikezelwa kwiimodeli. Oku bekuya kufuna izixa ezingaqhelekanga zezixhobo zokusebenza zekhompyutha. Iinkampani zobuchwepheshe kuthiwa zilandele iziphumo zabaphandi kuba yayiyeyona nto ingcono kakhulu ababenayo.
UAltman yathi i-GPT-4 iya kusebenzisa izibalo ezininzi ngakumbi kuneyandulelayo. I-OpenAI ilindeleke ukuba isebenzise iingcamango ezinxulumene nokuphucula kwi-GPT-4, nangona ingakanani inqanaba elingenakuqikelelwa njengoko uhlahlo lwabiwo-mali lungaziwa.
Nangona kunjalo, iingxelo ze U-Altman ubonisa ukuba i-OpenAI kufuneka igxile ekuphuculeni izinto ezahlukeneyo ngaphandle kobungakanani bemodeli. Ukufumana eyona seti yeehyperparameters, ubungakanani bemodeli efanelekileyo, kunye nenani leeparamitha kunokukhokelela kuphuculo olumangalisayo kuzo zonke iibenchmarks.
Ngokutsho kwabahlalutyi, zonke iingqikelelo zeemodeli zolwimi ziya kubhanga ukuba ezi ndlela zidityanisiwe zibe yimodeli enye. U-Altman ukwathe abantu abanakukholelwa ukuba iimodeli ezingcono zinokuba nkulu kangakanani na ngaphandle kokuba zikhulu. Isenokuba icebisa ukuba iinzame zokulinganisa ziphelile okwangoku.
I-OpenAI kuxelwe ukuba yenze umzamo omkhulu ekusombululeni ingxaki yolungelelwaniso lwe-AI: indlela yokwenza iimodeli zolwimi zilandele iinjongo zomntu kwaye zithobele iinqobo zobuntu?
Abahlalutyi bathi le ayisiyongxaki yezibalo kuphela (senza njani ukuba i-AI iqonde ngqo le nto siyifunayo?), kodwa nakwifilosofi (akukho ndlela yehlabathi yokulungelelanisa i-AI nabantu, kuba ukuguquguquka kwemilinganiselo yabantu ukusuka kwiqela ukuya kwelinye. iqela likhulu kwaye lihlala liphikisana).
Gqibela ukuba unomdla wokwazi okungakumbi ngayoungabhekisa kwisithuba sokuqala Kule khonkco ilandelayo.