I-OpenAI, ilebhu esekwe eSan Francisco, eCalifornia ephuhlisa itekhnoloji yobukrelekrele bokwenziwa kubandakanywa nemifuziselo yolwimi olukhulu, ubhengeze amandla okwenza iinguqulelo eziqhelekileyo ze-GPT-3, imodeli ekwazi ukuvelisa ikhowudi efana nomntu kwisicatshulwa kunye nelizwi.
Ngale nto ukusukela ngoku Abaphuhlisi banokusebenzisa ukulungiswa kakuhle ukwenza imifuziselo ye-GPT-3 ilungiselelwe umxholo othile wezicelo kunye neenkonzo zayo, ezikhokelela kwiziphumo ezikumgangatho ophezulu kuyo yonke imisebenzi kunye nomthwalo onzima, ngokwenkampani.
Kwabo bangayazi i-GPT-3, kufuneka uyazi ukuba oku yimodeli yolwimi oluzenzekelayo esebenzisa ukufunda nzulu ukuvelisa izicatshulwa ezifana nomntu.
Nguwo imodeli yokuqikelela ulwimi yoluhlu lwesizukulwana sesithathu lwe-GPT-n oludalwe yi-OpenAI, ilabhoratri yophando yobukrelekrele esekwe eSan Francisco eyenziwe yinkampani eyenza inzuzo i-OpenAI LP kunye nenkampani engumzali, inkampani engenzi nzuzo i-OpenAI Inc.
Kuwo nawuphi na umyalezo obhaliweyo, onjengesivakalisi, i-GPT-3 ibuyisela umbhalo wolwimi lwendalo oluhambelanayo.
Abaphuhlisi Bangakwazi "iprogram" GPT-3 ngokukubonisa imizekelo embalwa okanye "i-prompts."
"Siyile i-API ukuba ibe lula kumntu wonke ukuba ayisebenzise kwaye ibhetyebhetye ngokwaneleyo ukwenza amaqela okufunda ngomatshini abe nemveliso ngakumbi," utshilo i-OpenAI ngasekupheleni kukaMatshi.
Okwangoku, izicelo ezingaphezu kwama-300 Basebenzisa i-GPT-3 kwiindidi ezahlukeneyo kunye namashishini, ukusuka kwimveliso kunye nemfundo nkqu ubuchule kunye nemidlalo.
La umthamo omtsha wokucokisa kwi-GPT-3 tuning ivumela abathengi ukuba baqeqeshe i-GPT-3 ukuqonda ipateni ethile yomthwalo womsebenzi ezifana nokuveliswa komxholo, ukuhlelwa kunye nokudityaniswa kwesicatshulwa ngaphakathi kwemida yendawo ethile.
Umthengisi oSebenzayo usebenzisa i-GPT-3 ukunceda iinkampani ukuba zisebenzise ingxelo yabathengi. Ukusebenzisa idatha engacwangciswanga, inkqubo inokuvelisa iingxelo ezishwankathela ingxelo yabathengi kunye nokusebenzisana. Ngokwenza ngokwezifiso i-GPT-3, iViable yayiza kukwazi ukwandisa ukuchaneka kweengxelo zayo ukusuka kwi-66% ukuya kwi-90%.
Okufanayo kuya kuMgcini weRhafu, isixhobo esenza lula ukubalwa kwabasebenzi bee-freelancers ngokuhlenga ngokuzenzekelayo kunye nokukhupha idatha yokuhlawula kwiingxelo zerhafu, kwibhanki okanye kwi-akhawunti yokuhlawula. Irhafu yomgcini usebenzisa i-GPT-3 ukutolika idatha yengxelo yebhanki ukufumana iindleko ezinokuthi zitsalwe kwirhafu. Inkampani iyaqhubeka nokucokisa i-GPT-3 ngedatha entsha ngeveki nganye ngokusekelwe kwintsebenzo yehlabathi yangempela yemveliso yayo, igxininise kwimizekelo apho imodeli iwele ngaphantsi komda othile wokusebenza.
Los Abaphuhlisi bongeza malunga nemizekelo emitsha engama-500 ngeveki ukucokisa imodeli. Umgcini weRhafu uthi inkqubo yokulungelelanisa ivelisa i-1% yokuphucula iveki ngeveki.
«Inye into ebesilumke kakhulu kwaye sigxininise ngayo kuphuhliso lwethu lwe-API kukwenza ukuba ifikeleleke kubaphuhlisi abangenawo amava okufunda ngoomatshini, utshilo uRachel Lim, ilungu labasebenzi bobugcisa be-OpenAI. “Indlela ezibonakalisa ngayo kukuba unokwenza ithemplate ye-GPT-3 ngokwezifiso usebenzisa i-invocation yomgca womyalelo. [Siyathemba] ukuba ngenxa yokufikeleleka kwayo, sinokufikelela kubasebenzisi abahlukeneyo abanokuthi bazise iingxaki zabo ezahlukeneyo kubuchwephesha. "
U-Lim uthi amandla okucoca i-GPT-3 nawo angakhokelela ekugcinweni kweendleko, njengoko abathengi banokulindela ukuphindaphinda okuphezulu kweziphumo ezingcono zomgangatho ukusuka kwiimodeli ezilungiswe kakuhle xa kuthelekiswa nemodeli eqhelekileyo ye-GPT-3. (Iintlawulo ze-OpenAI zokufikelela kwi-API ngokusekelwe kwinani lamathokheni , okanye amagama, iimodeli ezivelisayo.)
Ngelixa i-OpenAI ineprimiyamu kwiimodeli ezisulungekileyo, i-Lim ithi uninzi lweemodeli ezisulungekileyo zifuna ukuhanjiswa okufutshane kunye namathokheni ambalwa, anokonga imali.
I-GPT-3 API iye yafumaneka esidlangalaleni ukususela ngo-2020. Unyaka ngaphambi kokuqaliswa kwayo, abaqulunqi bayo bagqibe ekubeni bangenzi esidlangalaleni umsebenzi wophuhliso lwenguqu yangaphambili, i-GPT-2, ngokuqwalasela ukuba le nkqubo i-doped yokufunda ngomatshini ingaba yingozi ukuba yawela ezandleni zabantu abakhohlakeleyo.