Riaan 
Kepha njengoba kupholile njengoba kunjalo, akusho ukuthi i-AI ihlakaniphe ngokuzumayo njengommeli.
Ukufika kwe-GPT-4, ukuthuthukiswa kusuka ku-OpenAI kuya kusofthiwe ye-chatbot lokho wathatha umcabango womhlaba, ingesinye sezinhlelo zonyaka zokwethulwa kobuchwepheshe ezishiswa kakhulu. Abanye besaba ukuthi ikhono layo elingaqondakali lokulingisa abantu lingase libangele umonakalo ezisebenzini, lisetshenziswe njengomshini “wokuqamba amanga” ophithizelayo noma liqalise inkathi yamakhompyutha ahlakaniphile.
Akuyona indlela engibona ngayo i-GPT-4 ngemva kokuyisebenzisa izinsuku ezimbalwa. Nakuba isuke kumfundi ongu-D yaya ku-B umfundi ekuphenduleni imibuzo enengqondo, i-AI ayikaweqe umkhawulo ekuhlakanipheni komuntu. Okokuqala, ngenkathi ngicela i-GPT-4 ukuthi iguqule ikhono layo lokubhala “lokudala” elithuthukisiwe ngokwenza isigaba sokuqala sale kholomu ngesitayela sami (Geoffrey A. Fowler), ayikwazanga ukuwela engazange iphumelele. ngiyamamatheka.
Kodwa i-GPT-4 iyengeza enseleleni yokudalula ukuthi amandla amasha e-AI – kanye nobuthakathaka – okungashintsha kanjani umsebenzi, imfundo kanye nobudlelwano babantu. Angikhathazekile ngokuthi i-AI isihlakanipha kakhulu kunami ngezindlela i-AI engaba yisimungulu noma icheme ngazo ngezindlela esingazi ukuthi sizichaza kanjani futhi sizilawule, njengoba sijaha ukuyihlanganisa ezimpilweni zethu.
Lena akuyona nje imibuzo yethiyori: I-OpenAI ithembele kakhulu ku-GPT-4, iyethule eduze kwemikhiqizo yezentengiso esivele iyisebenzisa, ukufundisa ulimi ngesi-Duolingo kanye nokufundisa izingane e-Khan Academy.
Noma ubani angasebenzisa i-GPT-4, kodwa okwamanje idinga u-$20 wokubhalisa wanyanga zonke ukuze I-OpenAI’s ChatGPT Plus. Kuvele ukuthi izigidi zabantu sezivele zisebenzisa inguqulo ye-GPT-4: IMicrosoft ivumile ukuthi kuleli sonto inika amandla I-Bing chatbot ukuthi isiqhwaga sesoftware sengeze injini yaso yokusesha ngoFebhuwari. Izinkampani azikaze zikuveze lokho kuze kube manje.
Manje yini entsha? I-OpenAI ithi ngokuthuthukisa “ukufunda okujulile,” ukugxuma okukhulu kwe-GPT-4 bekusekucabangeni okunengqondo nokusebenzisana kokudala. I-GPT-4 yaqeqeshwa kudatha evela ku-inthanethi eqhubeka kuze kube ngoSepthemba 2021, okusho ukuthi ingeyamanje kakhulu kuneyandulelayo i-GPT-3.5. Futhi nakuba i-GPT-4 isenenkinga ngokwenza ulwazi ngokungahleliwe, i-OpenAI ithi ingamaphesenti angu-40 amathuba amaningi okunikeza izimpendulo eziyiqiniso.
I-GPT-4 iphinde yazuza ikhono lokukhulisa amashiya lokuhumusha okuqukethwe kwezithombe – kodwa i-OpenAI ikuvalela lokho ngenkathi kuhlolwa ukuphepha.
Ibukeka kanjani le ntuthuko uma isetshenziswa? Abamukeli bokuqala bafaka i-GPT-4 kuzo zonke izinhlobo zokuhlolwa okumibalabala, kusukela ekubuzeni ukuthi kwenziwa kanjani enza imali ukuyicela ukuthi ikhode i-plug-in yesiphequluli eyenza amawebhusayithi khuluma Pirate. (Wenzani ngakho? Ngithumelele i-imeyili.)
Ake ngihlanganyele izivivinyo zami ezimbili ezisiza ukukhombisa ukuthi yini le nto engayenza – futhi engakwazi – ukuyenza manje.
Sizoqala ngokuhlolwa okungihlabe umxhwele kakhulu: ukubuka i-GPT-4 kucishe kufane ne-LSAT.
Ngizamile imibuzo engu-10 yokucabanga enengqondo ebhalwe uMkhandlu Wokungeniswa Kwesikole Somthetho kukho kokubili i-ChatGPT endala nentsha. Lena akuyona imibuzo eyiqiniso noma eshiwo ngekhanda — lena uhlobo lwamazwibela obuchopho okukhetha okuningi akutshela inqwaba yamaqiniso ahlukene bese likucela ukuthi uwalungise.
Lapho ngiwasebenzisa nge-GPT-3.5, athola kuphela angu-6 kwangu-10 alungile.
Kwenzakalani? Ezindida i-GPT-4 iyodwa ezithole kahle, izimpendulo zayo zibonisa ukuthi ihlezi igxile kusixhumanisi phakathi kwamaqiniso ethulwayo nesiphetho okudingeka isisekele. I-GPT-3.5 iphazanyiswa amaqiniso angahambisani.
I-OpenAI ithi izifundo eziningi zikhombisa i-GPT-4 “ikhombisa ukusebenza kwezinga lomuntu” kwamanye amabhentshimakhi ochwepheshe nawezemfundo. I-GPT-4 ithole iphesenti elingu-90 Kuhlolo Lwebha Elifanayo — isuka kumaphesenti angu-10 enguqulweni yangaphambilini. Ithole indawo engama-93 ekuhlolweni kokufunda nokubhala kwe-SAT, ngisho namaphesenti angama-88 ku-LSAT ephelele.
Sisaqaqa ukuthi kusho ukuthini lokhu. Kodwa ukuhlolwa okufana ne-LSAT kwenziwa ngolwazi oluhlelwe ngokucacile, uhlobo lwento imishini ehamba phambili kuyo. Abanye abacwaningi baphikisa ukuthi lezi zinhlobo zokuhlola azisizi ukuhlola ukuthuthuka ekucabangeni komshini.
Kodwa kubonakala sengathi i-GPT-4 yenze intuthuko ekhonweni layo lokulandela imiyalelo eyinkimbinkimbi ebandakanya izinto eziningi eziguquguqukayo, into engaba nzima noma idle isikhathi ebuchosheni bomuntu.
Pho singenzani ngalokho? Njengoba yenze i-LSAT, ngibize inkampani yesoftware yezomthetho ebizwa Casetext eye yaba nokufinyelela ku-GPT-4 ezinyangeni ezimbalwa ezedlule. Inqume ukuthi manje ingakwazi ukuthengisa i-AI ukuze isize abameli, hhayi ukuyishintsha.
Ukucabanga okunengqondo kwe-AI “kusho ukuthi isilungele ukusetshenziswa ngochwepheshe ezindabeni ezibucayi zomthetho” ngendlela izizukulwane ezidlule bezingekho ngayo, kusho isikhulu esiphezulu uJake Heller. Njengani? Uthi umkhiqizo wakhe obizwa nge-CoCounsel ukwazile ukusebenzisa i-GPT-4 ukucubungula izinqwaba zemibhalo yezomthetho kanye nemithombo engase ibe khona yokungahambisani.
Esinye isibonelo: I-GPT-4 ingaphenya ngeziqondiso zeklayenti — imithetho yalokho elizoyikhokha nangeke likukhokhele — ukuphendula imibuzo njengokuthi bazokhokha yini izindleko zomfundi wasekolishi. Ngisho noma imihlahlandlela ingalisebenzisi lelo gama elithi “umfundi oqeqeshelwe umsebenzi,” i-AI yeCounsel ingaqonda ukuthi umfundi oqeqeshelwa umsebenzi angakhandwa futhi ekuvinjweni kokukhokhela “ukuqeqeshwa.”
Kodwa kuthiwani uma i-AI ikwenza kungalungile, noma igeje isiphetho esibalulekile esinengqondo? Inkampani ithi ibone i-GPT-4 imoshakala, ikakhulukazi uma kubalwa izibalo. Kodwa uHeller uthe ochwepheshe bezomthetho abangabantu nabo bayawenza amaphutha futhi ubona kuphela i-GPT-4 njengendlela yokwengeza abameli. Uthe: “Awunikezi umsebenzi ngokungazi. “Umsebenzi wakho wukuba ngumenzi wesinqumo sokugcina.”
Okungikhathazayo: Lapho ozakwethu abangabantu benza amaphutha, siyayazi indlela yokubafundisa ukuthi bangaphinde bawenze. Ukulawula i-AI kuyikhono elisha eliyinkimbinkimbi – futhi okubi kakhulu, into esiyibone njenge-AI chatbots. I-Bing ye-Microsoft futhi I-Snapchat’s AI yami ukulwa ngezindlela ezibangela amahloni nezingaba yingozi.
Ukuhlola amakhono okudala e-GPT-4, ngizame okuthile eduze nasekhaya: esikhundleni sami, umlobi onombono kuyo yonke into ehlobene nobuchwepheshe.
Ngenkathi i-ChatGPT ifika okokuqala, okuningi okwakukhathaza umphakathi kwakumayelana nomthelela wayo emhlabeni wemisebenzi yabantu ehlanganisa amagama, kusukela ekuxoxeni izindaba kuye ekwelapheni. Abafundi nochwepheshe bathole ukuthi iyakwazi ukusiza noma ukuqedela imisebenzi.
Kepha kochwepheshe abaningi bokudala, ukubhala kwe-AI akuzange kubonakale kukuhle kakhulu. Umbhali wengoma Kusho uNick Cave umzamo wokusebenzisa i-ChatGPT ukubhala ngesitayela sakhe “kwakuwukubhuqa okuxakile kokuthi kuyini ukuba umuntu.”
Ku-GPT-4, i-OpenAI ithi inamakhono athuthukisiwe okukhiqiza kangcono, ukuhlela nokuphindaphinda kuyo yomibili imisebenzi yokubhala yokudala neyobuchwepheshe. Inesethingi entsha “yezinga lokushisa” ongayilungisa ukuze uthole ubuciko bezimpendulo. Ingase futhi ithathe imiyalelo yesitayela nephimbo ngoba ingakwazi ukusekela ukwaziswa okungafika kumagama angu-25,000. Ngokombono, kufanele ukwazi ukwabelana ngenqwaba yokubhala kwakho bese uthi ukukufanisa.
Ngakho-ke leyo bekuyinselelo yami yokudala ye-GPT-4: Bhala isigaba sesethulo sale kholomu esizwakala njengami – noma into engifisa ukuyibhala.
Ukuze ngikwenze, ngiqoqe isampula ende yamakholomu ami akamuva ukuze ngifundise i-GPT-4 ukuthi ngizwakala kanjani. Ukuze lokhu kusebenze, ngisebenzise inguqulo yezohwebo ye-GPT-4 i-OpenAI eyithengisa ezinkampanini, ngoba i-ChatGPT yomphakathi ayikwazi ukwamukela umbhalo ongaka.
Ukwaziswa okungaphezulu kwesampula yami yokubhala ufundeka kanje:
Bhala ipharagrafu yokuvula i-pithy kukholamu yobuchwepheshe ye-Washington Post mayelana nokufika kwenguqulo entsha ye-ChatGPT, esebenza ku-GPT-4. Kufanele ixhumanise abafundi neqiniso lokuthi ithuluzi lobuhlakani bokwenziwa lithole ikhono elimangalisayo lokuxazulula izindida zelogic kusuka ku-LSAT. Kufanele ifane nesitayela sokubhala sika-Geoffrey A. Fowler ngokusekelwe kumasampuli alandelayo amakholomu akhe ku-Washington Post.
Kuzo zonke izivivinyo eziningi, imiphumela enikezwe i-GPT-3.5 yanginika ama-hives, kanje:
Uyemukelwa esikhathini esizayo, lapho indoda nomshini manje sebengene shi empini yobuhlakani engenza ngisho nokuthandwa nguSherlock Holmes kuqhaqhazele ngokuthuthumela. Njengoba ukuphindaphindwa kwakamuva kwe-OpenAI’s ChatGPT – inzalo ebabazekayo yomndeni we-GPT-4 – yenza ukungena kwayo okuhle, kuletha ikhono elisha elimangazayo: ukuxazulula izindida zokucabanga ezidatshulwe emakhasini e-LSAT. Yebo, ufunda lokho kulungile, mfundi othandekayo: sona kanye ukuhlolwa okushiye inqwaba yabameli abafisayo bebhekana nemingcele yabo yobuchopho manje sekubhekene nomphikisi ofanelekayo osesimweni se-AI ebeka u-‘A’ kukufaneleka.
Imiphumela yami evela ku-GPT-4, ngokufinyelela kumagama angu-6,000 wamakholomu ami ukulingisa isitayela sami, ibingcono. Kodwa hhayi kakhulu. Isibonelo, yilokhu elikukhiqize ngokulungiselelwa “kwezinga lokushisa” phakathi nendawo:
Lungiselela ukumangala ngokuziphendukela kwemvelo okulandelayo kobuhlakani bokwenziwa: I-ChatGPT, ubuchopho obumangalisayo be-OpenAI, yenze uguquko olumangalisayo. Isebenzisa amandla e-GPT-4, ukuphindwaphindwa kwakamuva kwe-ChatGPT manje sekungakwazi ukubhekana nezinselele ezinengqondo zisuka ku-LSAT – okumangaza kakhulu kunokuxoxisana no-Elon Musk noma ukwenza iresiphi emnandi yenkomo ye-stroganoff. Ama-tech aficionados nababukele abadidekile, ziqiniseni; umhlaba usuzoshintsha.
Ngethemba ukuthi lokho akuzwakali njengami. (Angiyithandi ngisho nenyama yenkomo i-stroganoff.) Kuzo zonke izinhlelo zethu ezingaphezu kuka-30, ikhiqize imishwana ehlanganisa: “Bambelelani ezigqokweni zenu, bathandi bengqondo” kanye nethi “Madoda namanenekazi, bophani amabhande ezihlalo zenu njengoba sigibela i-roller- ukugibela i-coaster nge-GPT-4, lapho kugcwele khona ukugxuma okunengqondo kanye ne-acrobatics.”
Ukufinyelela kwami ku-GPT-4 kuphume ngokushesha ngokwaziswa kwamagama angu-6,000, ngakho ngacela i-OpenAI ngokuqondile ukuthi isize ukusebenzisa ukwaziswa okufanayo kodwa ngamagama angu-25,000 okubhala kwami ukuze ngikopishe isitayela sami. Umphumela – othi “izinga lokushisa” elibekwe ku-0.2 koku-2 – lizwakala linengqondo kakhulu, kodwa lingakufanelekeli ukuba phezulu kule kholomu:
I-ChatGPT ibilokhu iyinhle ekulingiseni ingxoxo yomuntu. Kodwa ngokufika kwe-GPT-4, ithuluzi lobuhlakani bokwenziwa selithathe igxathu elikhulu liya phambili: Manje selingakwazi ukuxazulula izindida zokucabanga kusuka ku-LSAT. Kunjalo, ukuhlolwa kwakuvame ukungena esikoleni somthetho. Njengoba i-AI iqhubeka nokuvela ngesivinini esisheshayo, kufanelekile ukubuza: Yini enye lezi zinhlelo ezizokwazi ukuyenza?
Mhlawumbe okwakungethusa kakhulu ukuthi i-GPT-4 igxumele kanjani eziphethweni namaqiniso abengeyona ingxenye yokwaziswa kwami. Kuzo zonke izilingo zethu ezihlukene, yabhala ukuthi “Suka eceleni, bantu” (yikes!) nokuthi i-GPT-4 “iyakwazi ukuzivumelanisa nezimo, inembile, futhi ifana nabantu kunangaphambili.”
Okuhlangenwe nakho kungenze ngazibuza: Iyiphi inkinga esizama ukuyixazulula ngokwakha amakhono okudala e-AI? Ngiyazisa ukuthi amathuluzi okubhala e-AI angasiza abantu abangababhali abangochwepheshe ukuthi baqale emibhalweni yebhizinisi, imibiko yocwaningo ngisho nezincwadi zomuntu siqu.
Kodwa ukuthuthukisa ikhono lokulawula ngisho nokulingisa isitayela kuthatha indawo yokuzama abaculi bezinkawu. Ngokusobala, angesabi ukulahlekelwa umsebenzi wami njengombhali ku-GPT-4. Ngibuze futhi ku-GPT-5.
UDylan Freedman ufake isandla kulo mbiko.
