Riaan 
CNN
–
Ngosuku lokuqala ngemva kokwembulwa kwayo, I-GPT-4 imangaze abasebenzisi abaningi ezivivinyweni zakuqala kanye nedemo yenkampani enekhono layo lokubhala amacala, iphase izivivinyo ezijwayelekile kanye nokwakha iwebhusayithi esebenzayo ngomdwebo odwetshwe ngesandla.
NgoLwesibili, i-OpenAI kumenyezelwe inguqulo yesizukulwane esilandelayo yobuchwepheshe bobuhlakani bokwenziwa obusekela ithuluzi layo lokuxoxa ngegciwane, i-ChatGPT. Izithembiso ze-GPT-4 ezinamandla kakhulu ukukhipha iziphindaphindo zangaphambilini emanzini, okungenzeka kuguqule indlela esisebenzisa ngayo i-inthanethi ukusebenza, ukudlala nokudala. Kodwa futhi kungase kwengeze emibuzweni eyinselele mayelana nokuthi amathuluzi e-AI angathuthukisa kanjani umsebenzi, enze abafundi bakwazi ukukopela, futhi baguqule ubudlelwano bethu nobuchwepheshe.
I-GPT-4 inguqulo ebuyekeziwe yemodeli yolimi enkulu yenkampani, eqeqeshwe ngenani elikhulu ledatha ye-inthanethi ukuze ikhiqize inkimbinkimbi. izimpendulo kukwaziswa komsebenzisi. Manje isiyatholakala ngohlu lokulinda futhi isivele ingene emikhiqizweni yezinkampani zangaphandle, okuhlanganisa nosesho olusha lwe-Microsoft lwe-AI-powered Bing. injini. Abanye abasebenzisi abanokufinyelela kwangaphambi kwesikhathi ethuluzini babelana ngezipiliyoni zabo futhi bagqamisa ezinye zezigameko zalo zokusebenzisa eziphoqayo.
Nakhu ukuhlolisisa amandla e-GPT-4:
Emgogodleni wayo, ushintsho olukhulu ku-GPT-4 yikhono layo lokusebenza nezithombe ezilayishwa abasebenzisi.
Elinye lamacala asetshenziswa kakhulu emihlathini kuze kube manje liqhamuke idemo yevidiyo ye-OpenAI lokho kwabonisa ukuthi kanjani a umdwebo ungaguqulwa ube iwebhusayithi esebenzayo kungakapheli imizuzu. Umbonisi ulayishe isithombe ku-GPT-4 wabe esenamathisela ikhodi ewumphumela ekubukeni kuqala okubonisa ukuthi ingaba iwebhusayithi esebenzayo kanjani.
Esimemezelweni sayo, i-OpenAI nayo wabonisa indlela i-GPT-4 ecelwe ngayo ukuchaza ihlaya kusuka ochungechungeni lwezithombe – ezifaka i-smartphone eneshaja engalungile – futhi ichaze ukuthi kungani yayihlekisa. Nakuba kungase kuzwakale kuqondile, ukuhlukanisa ihlaya kuyinkimbinkimbi kakhulu kumathuluzi obuhlakani bokwenziwa ongawathatha ngenxa yomongo odingekayo.
Kokunye ukuhlola, i-New York Times wabonisa I-GPT-4 isithombe sengaphakathi lesiqandisi futhi yasishukumisela ukuba siqhamuke nokudla okusekelwe izithako.
Isici sezithombe asikabi bukhoma, kodwa i-OpenAI kulindeleke ukuthi isikhiphe emasontweni ezayo.
Abanye abasebenzisi bokuqala be-GPT-4 abanolwazi oluncane kakhulu noma olungenalo lokufaka amakhodi ngaphambili nabo baye bayisebenzisa dala kabusha imidlalo eyisakhiwo sodumo efana ne-Pong, i-Tetris noma iNyoka ngemva kokulandela imiyalelo yesinyathelo ngesinyathelo enikezwe ithuluzi ukuthi ungakwenza kanjani lokho. Abanye bazenzele ezabo imidlalo yasekuqaleni. (I-GPT-4 ingabhala ikhodi ngazo zonke izilimi ezinkulu zokuhlela, ngokusho kwe-OpenAI.)
“Amakhono olimi anamandla e-GPT-4 azosetshenziselwa yonke into kusukela ekubhaleni izindaba, ukudala abalingiswa kuya ekudalweni kokuqukethwe kwemidlalo,” kusho u-Arun Chandrasekaran, umhlaziyi we-Gartner Research. “Lokhu kungadala abahlinzeki bemidlalo abazimele ngokuzayo. Kodwa ngale komdlalo ngokwawo, i-GPT-4 namamodeli afanayo angasetshenziswa ekudaleni okuqukethwe kokumaketha okuzungeze ukubuka kuqala kwegeyimu, ukukhiqiza izindatshana zezindaba ngisho nokulinganisa amabhodi ezingxoxo amageyimu.”
Ngokufana nokudlala, i-GPT-4 ingashintsha indlela abantu abathuthukisa ngayo izinhlelo zokusebenza. Omunye umsebenzisi ku-Twitter uthe wenza uhlelo lokusebenza olulula lokudweba emizuzwini, kanti omunye uthi unawo ikhodi uhlelo lokusebenza oluncoma amamuvi amasha amahlanu nsuku zonke, kanye nokuhlinzeka ngama-trailer nemininingwane yokuthi angawabuka kuphi.
“Ukubhala amakhodi kufana nokufunda ukushayela – inqobo nje uma oqalayo ethola isiqondiso, noma ubani angabhala ikhodi,” kusho uLian Jye Su, umhlaziyi we-ABI Research. “I-AI ingaba uthisha omuhle.”
Nakuba i-OpenAI kusho isibuyekezo “asinamandla amancane” kunabantu ezimweni eziningi zomhlaba wangempela, sibonisa “ukusebenza kwezinga lomuntu” ezivivinyweni ezihlukahlukene zobungcweti nezemfundo. Le nkampani ithe i-GPT-4 isanda kuphumelela ukuhlolwa kwebha yesikole esilingisayo enamaphuzu acishe abe ngu-10% abahloli abaphezulu. Ngokuphambene, inguqulo yangaphambilini, i-GPT-3.5, ithole amaphuzu acishe abe ngu-10%. Inguqulo yakamuva iphinde yasebenza kakhulu ezivivinyweni ze-LSAT, GRE, SAT kanye neziningi ze-AP, ngokusho kwe-OpenAI.
NgoJanuwari, i-ChatGPT yaba sematheni ngekhono layo loku baphumelele izivivinyo ezisezingeni eliphezulu, njengenye yase-University of Pennsylvania’s Wharton School of Business, kodwa hhayi enamamaki aphezulu kakhulu. Le nkampani ithe ichithe izinyanga isebenzisa izifundo ezivela ohlelweni lwayo lokuhlola kanye ne-ChatGPT ukuze ithuthukise ukunemba kwesistimu nekhono lokuhlala esihlokweni.
Uma kuqhathaniswa nenguqulo yangaphambili, i-GPT-4 iyakwazi ukukhiqiza izimpendulo ezibhaliwe ezinde, ezinemininingwane eminingi nezithembeke kakhulu, ngokusho kwenkampani.
Inguqulo yakamuva manje inganikeza izimpendulo ezingafika kumagama angu-25,000, kusukela cishe kwangu-4,000 ngaphambilini, futhi inganikeza imiyalelo enemininingwane ngisho nezimo eziyingqayizivele, kusukela ekuhlanzeni ithangi lezinhlanzi ze-piranha kuya ekukhishweni kwe-DNA ye-strawberry. Omunye umsebenzisi wangaphambi kwesikhathi uthe inikeze iziphakamiso ezijulile zemigqa yokulanda esekelwe embuzweni osohlwini lwephrofayela yokuqomisana.
UJoshua Browder, oyi-CEO ye-legal services chatbot i-DoNotPay, uthe inkampani yakhe isivele ikhona ukusebenza ngokusebenzisa ithuluzi ukukhiqiza “amacala okuchofoza okukodwa” ukuze amangalele abashaya ucingo, ekukhombeni kwangaphambi kwesikhathi kwamathuba amakhulu e-GPT-4 okushintsha indlela abantu abasebenza ngayo kuzo zonke izimboni.
“Cabanga uthola ucingo, uchofoza inkinobho, [the] ikholi iyabhalwa bese kwakhiwa icala lamagama angu-1,000. I-GPT-3.5 ibingeyinhle ngokwanele, kodwa i-GPT-4 iwuphatha kahle kakhulu umsebenzi,” kusho u-Browder kuTwitter.
Khonamanjalo, uJake Kozloski, oyi-CEO ye- dating Site Keeper, uthe inkampani yakhe iyiyo usebenzisa ithuluzi ukuze ifane kangcono nabasebenzisi bayo.
Ngokuka-Su at ABI Research, kungenzeka siphinde sibone intuthuko enkulu “emotweni exhunyiwe [dashboards]ukuxilongwa okukude ekunakekelweni kwezempilo, nezinye izinhlelo zokusebenza ze-AI ebezingenzeki ngaphambilini.”
Yize inkampani yenze intuthuko enkulu kumodeli yayo ye-AI, i-GPT-4 inemikhawulo efanayo ezinguqulweni zangaphambilini. I-OpenAI ithe ubuchwepheshe ayinalo ulwazi lwezehlakalo ezenzeke ngaphambi kokuthi idatha yayo inqanyulwe (Septhemba 2021) futhi ayifundi kokuhlangenwe nakho kwayo. Ingase futhi yenze “amaphutha alula okucabanga” noma “ikhohliseke ngokweqile ekwamukeleni izitatimende ezisobala ezingelona iqiniso ezivela kumsebenzisi,” futhi hhayi umsebenzi wokuhlola kabili, kusho inkampani.
U-Chandrasekaran kaGartner uthe lokhu kubuye kubonakale kumamodeli amaningi e-AI namuhla. “Masingakhohlwa ukuthi lezi zinhlobo ze-AI aziphelele,” kusho uChandrasekaran. “Bangakwazi ukukhiqiza ulwazi olungalungile ngezikhathi ezithile futhi bangaba yibhokisi elimnyama ngokwemvelo.”
Okwamanje, i-OpenAI ithe abasebenzisi be-GPT-4 kufanele baqaphele futhi basebenzise “ukunakekela okukhulu” ikakhulukazi “ezimweni eziphakeme kakhulu.”
