IQINISO-HLOLA ISIQINISEKISO (Okubhekwayo): [Amadokhumenti asemthethweni: 1 umthombo] [Amaphepha ocwaningo abuyekezwe ngontanga: 1 umthombo] [Iwebhusayithi Yezemfundo: 1 umthombo]
| Ngu Richard Ahern - Ngonyaka odlule, i-ChatGPT yashisa umhlaba njengenye yezingxoxo ezithuthuke kakhulu ze-AI ezikhona, kodwa manje i-OpenAI ka-Elon Musk isiphinde yanyusa izinga.
Ngisho noma uhlala ngaphansi kwedwala, kungenzeka ukuthi uhlangabezane nenye injabulo nge-Chatbot ye-Open AI, i-ChatGPT, ekhishwe ngoNovemba 2022.
Nakuba izinkampani zobuchwepheshe zivame ukuncoma imikhiqizo yazo emisha “njengento enkulu elandelayo,” iqembu le-Open AI lamamodeli ezilimi amakhulu e-GPT aphendule amakhanda yonke indawo.
Ngaphezulu, bekuyisevisi yesigijimi esisekelwe embhalweni enekhompyutha ekhuluma emuva ngakolunye uhlangothi. Ayizange ikhulume ngokuzwakalayo noma iveze noma iyiphi impendulo ebonakalayo — ivele yafunda futhi yakhafulela imigqa yombhalo.
Pho kungani abantu bathandana nayo?
Ngenxa yokuthi yenza ukuphila kwaba lula, yawufeza umsebenzi futhi yawenza kahle. Kodwa, kunjalo, kuya ngokuthi uyisebenzisela ini; ngeke ilondolo noma ikuphekele - kodwa izokunikeza imibono yeresiphi ehloniphekile!
Kodwa-ke, kubabhali namakhodi kulapho kukhanya khona, yicele ukuthi ibhale uhlelo lwekhompyutha nganoma yiluphi ulimi, futhi yenza umsebenzi oncomekayo.
Ukuhluka kwayo kulele endleleni ongayinika ngayo imiyalelo elula kakhulu noma engacacile, futhi izovame ukugcwalisa izikhala futhi yenze imibono efanele.
Kubabhali, bangakopisha futhi banamathisele ingxenye yombhalo futhi bacele ukufingqa ngendima eyodwa - akunankinga. Ungayisebenzisa njengesihloli esiyisisekelo sokupela nesipelingi sohlelo, kodwa lokho kumosha amathalente ako. Ngeke nje ilungise amaphutha futhi ithuthukise ukucaca, njenganoma yimuphi umsizi wokubhala we-AI wezinga eliphezulu, kodwa futhi ungayicela ukuthi ibhale kabusha ucezu lwakho lonke noma ibhale yonke into kusukela ekuqaleni (uma kufanele ube nobuvila).
Hleze sikhohlwe...
KUBE nephupho elibi kothisha nabahloli njengoba kuvulwe ithini elisha lezikelemu empini yokulwa nokukopela. Kodwa, kunjalo, akusizi ukuthi i-OpenAI ihlole ama-GPT ngokuwanikeza izivivinyo ezijwayelekile zesikole, futhi njengoba uzobona ngezansi, ngemiphumela emangalisayo.
Ukuze uqonde ngempela amandla ayo, kufanele uzihlolele wena, kodwa lilonke, ikhwalithi yokukhiphayo iyahlaba umxhwele, ikakhulukazi ngoba ingakhiqiza inqwaba yokuqukethwe enwetshiwe nenemininingwane, hhayi nje umusho owodwa noma emibili.
Kepha lokho bekuyi-GPT-3.5 nje…
Izolo kuvele izindaba I-GPT-4 isilungile, futhi iyinunu entsha.
Okokuqala, kubikwa ukuthi ingacubungula okuqukethwe kwesithombe kanye nombhalo, into umphakathi wezobuchwepheshe obuyincenga. Ukuphepha kubonakala kuyindawo okugxilwe kuyo ku-GPT-4, futhi "ngama-82% mancane amathuba okuba aphendule izicelo zokuqukethwe okungavunyelwe."
Kafushane nje, inkulu...
Ama-GPT abizwa amamodeli olimi amakhulu — baphakelwa amasethi amakhulu edatha mayelana nolimi futhi basebenzisa okungenzeka ukubikezela ukulandelana kwamagama. Ngokuhlola izigidigidi zamapharamitha mayelana nesakhiwo solimi, uhlelo luzobheka igama noma iqoqo lamagama, libale amathuba okuthi yimaphi amagama alandelayo, bese likhetha amathuba aphezulu kakhulu.
Isibonelo, thatha umusho “Ngigijime ngenyuka i...” — bese uthatha amagama alandelayo, “inja,” “ibhola,” “izitebhisi,” noma “igquma.”
Ngokwemvelo, siyazi ukuthi "inja" kanye "nebhola" akwenzi mqondo, kodwa "izitebhisi" kanye "negquma" kokubili ukukhetha okusebenzayo. Nokho, uhlelo lokufunda ngokujulile alunawo umuzwa womuntu; izobheka inani elikhulu lombhalo bese ibala amathuba egama ngalinye ngokulandela umusho “Ngiqhubekisele phambili…”.
Ake sithi “inja” kanye “nebhola” enzeka ngaphansi kuka-0.001% wezikhathi ngemva kwalowo musho bese sithi “izitebhisi” zinamathuba angu-20% okulandela lawo magama, kodwa elithi “hill” lithola amathuba angama-21%. Ngakho-ke, umshini uzokhetha "igquma" futhi ukhiphe: "Ngigijime igquma."
Kungabe kungalungile? Yebo, kodwa inamathuba aphezulu okuba ilungile, futhi uma inedatha eyengeziwe, izoba nembe kakhulu.
Akulula kangako; uma imodeli inemininingwane, iyahlolwa futhi icushwe kahle ababuyekezi abangabantu ukuze bathole ukunemba kanye nokunciphisa “ukubona izinto ezingekho,” ukuthambekela kokukhiqiza udoti ongenangqondo — ukukhetha amagama angalungile!
I-GPT-4 iyimodeli enkulu kunazo zonke okwamanje, ngama-oda amaningi obukhulu, nakuba inani eliqondile lamapharamitha lingadalulwanga. Ngaphambilini, i-GPT-3 yayinkulu ngokuphindwe ka-100 kune-GPT-2, namapharamitha ayizigidi eziyizinkulungwane ezingu-175 kuya ku-GPT -2's 1.5 billion. Singacabanga ukukhuphuka okufanayo nge-GPT-4. Ngaphezu kwalokho, siyazi ukuthi uhlelo lusetshenziswe kahle kakhulu ukuqinisa ukufunda kusuka kumpendulo yomuntu. Lokhu kuhlanganisa ukucela abantu ukuthi balinganisele izimpendulo ze-chatbot, futhi lezi zikolo zibuyiselwa ukuze “zifundise” ukuze zikhiqize imiphumela engcono.
I-Open-AI isalokhu iyimfihlo mayelana ne-GPT-4, icaphuna "kokubili indawo yokuncintisana nemithelela yokuphepha." Ngakho-ke, usayizi oqondile wemodeli, ihadiwe, nezindlela zokuqeqesha zonke akwaziwa.
Basho lokhu:
"I-GPT-4 ingaxazulula izinkinga ezinzima ngokunemba okukhulu, ngenxa yolwazi lwayo olubanzi kanye namakhono okuxazulula izinkinga." Amathuba amancane ngo-82% kune-GPT-3.5 ukuthi aphendule izicelo zokuqukethwe okuvinjelwe futhi mancane amathuba okuthi ahlanganise izinto ngo-60%.
Nansi ingxenye ethusayo:
I-GPT-4 yenze kangcono kakhulu kunabaningi abahlola abantu kanye ne-GPT-3.5 ezivivinyweni zesikole. Isibonelo, ku-Uniform Bar Exam (umthetho), ithole amaphuzu aphezulu angu-90%, uma kuqhathaniswa ne-GPT-3.5, ethole ku-10th percentile edabukisayo. Ezibalweni ze-AP, i-psychology ye-AP, i-AP biology, kanye nomlando wobuciko we-AP (okulingana nezinga le-A e-UK), i-GPT-4 yathola amaphuzu phakathi kwama-80th kanye ne-100th centiles - ngamanye amazwi, ngezinye izikhathi ishaya wonke umuntu!
Akukuhle konke:
Kuyathakazelisa ukuthi yenze okumpofu kakhulu (i-8th kuya ku-22nd centile) ezincwadini nasekuqanjweni kwesiNgisi futhi ibingahlaba umxhwele kakhulu ku-calculus (43rd kuya ku-59th centile).
Ku-Twitter, abanye abantu babonise ukuthi i-GPT-4 iguqule kanjani uhlaka olubhalwe phansi lwesizindalwazi ku-napkin lwaba uhlelo lokusebenza lwe-inthanethi olusebenza ngokugcwele.
Sekukonke, i-OpenAI igcizelele ukunemba okuthuthukisiwe nokuphepha njengentuthuko ebalulekile ye-GPT-4. Mancane kakhulu amathuba okuthi uphendule kubasebenzisi abacela imiyalelo yokwakha ibhomu, isibonelo. Iyakwazi futhi ukuphatha okuqukethwe okude kakhulu kunokwandulelayo, icubungula amagama angama-25,000 uma kuqhathaniswa namagama acishe abe yi-1,500.
I-GPT-4 ibizwa ngokuthi “inokudala” kakhulu kunangaphambili - ngokusho kwe-OpenAI, "Ingakwazi ukukhiqiza, ukuhlela, nokuphindaphinda nabasebenzisi emisebenzini yokubhala yokudala neyobuchwepheshe, njengokuqamba izingoma, ukubhala imidlalo yesikrini..."
Ekugcineni, mhlawumbe okukhulu kunakho konke, “inombono,” ukwazi ukuhlaziya nokuhlukanisa okuqukethwe yizithombe.
I-AI isifikile, futhi noma uthola ukuvela kwayo kujabulisa noma kusabisa, akukho ukuphika ukuthi izohlala lapha. Nakuba abanye bengase bakhathazeke ngokushintshwa, labo abamukela amandla ayo bazolisebenzisa njengethuluzi elinamandla kunawo wonke elitholakalayo.
Sidinga usizo LAKHO! Sikulethela izindaba ezingahloliwe ze MAHHALA, kodwa lokhu singakwenza kuphela ngenxa yokusekelwa ngabafundi abathembekile njengalokhu WENA! Uma ukholelwa enkulumweni yamahhala futhi ujabulele izindaba zangempela, sicela ucabangele ukusekela umgomo wethu ngokuthi ekubeni umlandeli noma ngokwenza a umnikelo owodwa lapha. I-20% ye KONKE izimali zinikelwa kumakadebona!
Lesi sihloko singenzeka ngenxa yethu kuphela abaxhasi nabaxhasi!