Ukusuka ku-GPT-3 kuye ku-GPT-5: Indlela ikhwalithi yokuhumusha ye-AI eyashintsha ngayo ngempela

I-GPT-3, i-GPT-4, ne-GPT-5 zihlukene ngeminyaka emihlanu kanye nezizukulwane eziningana ze-architecture, kodwa umbuzo owusizo ekuhumusheni akusiwo othi "iyiphi ehlakaniphe kakhulu." Umncane kakhulu: yini ngempela eshintshile endleleni lawa mamodeli aphatha ngayo izingxenye zolimi ezingahumusheki igama negama, izinto ezinjengezisho, ukungacaci kahle kwencazelo, kanye nerejista? Ukuthuthuka kwenzeke ezindaweni ezithile, ezikhathini ezithile, futhi akuhambanga emgqeni oqondile kusukela kwendala kuya kwentsha.

Lokho i-GPT-3 eyakunikeza ngempela ekuhumusheni

I-OpenAI yakhipha i-GPT-3 ngo-2020, ngaphambi kokuba uhlelo lokuhlola lwe-MachineTranslation.com lube khona, ngakho-ke akukho lutho lapha oluyisilinganiso sobunikazi. Umlando obhalwe kahle, hhayi imiphumela yokuhlolwa kwangaphakathi. I-GPT-3 kwakuyimodeli yokuqala yohlobo lwayo ekwazi ukukhiqiza umbhalo oshelelayo, ozwakala njengowomuntu ngezilimi eziningi ngaphandle kokuqeqeshwa okukhethekile komsebenzi othile, okwayenza yaba yingqophamlando yangempela. Ngokukhethekile ekuhumusheni, lokho kushelela kwakha ubungozi bangempela: i-GPT-3 yayingakhiqiza umusho oqinisekisayo, owakheke kahle olimini oluhlosiwe owawuseyiphutha, ngaphandle kokuthi okukhiqiziwe kukhombise ukuthi kukhona iphutha elenzekile.

Yini eyashintsha lapho kufika i-GPT-4

I-GPT-4 yethulwa ngo-March 2023, kanti ukuhlolwa kwamamodeli kwe-MachineTranslation.com kuqala kulesi sizukulwane. Ukuthuthuka kwaba okwangempela kodwa akuzange kulingane. Ekuhloleni ibinzana lesiNgisi elingaqondakali ngqo elithi "That's sick" ngesi-Japanese, i-GPT-4.1-nano kwakuyiyo yodwa imodeli kwamayisithupha ahloliwe ehlise incazelo ekabili eqondwe yalowo musho yaba yincazelo eyodwa, kanti amanye ayisihlanu alondoloza ukungaqondi ngqo okwakuhloswe yileli binzana lokuqala. Izisho zazingubuthakathaka obukhulu: ngokuhlola isisho sesi-Spanish esithi "llevarse el gato al agua" ezinhlotsheni ezinhlanu ze-GPT kanyekanye, i-GPT-4o-mini kanye ne-GPT-4.1-nano ezincane zombili zakhiqiza ukuhumusha okufanayo kwezwi nezwi, okungalungile, zahluleka ngokuphelele ukubamba isisho.

Isimo esifanayo savela nakwezinye izindawo. Isisho sesi-German, "ich verstehe nur Bahnhof," esahlolwa ezinhlotsheni ezingaphansi eziyisishiyagalombili ze-GPT kanye ne-Claude, sahunyushwa ngezwi nezwi nangendlela engafanele yi-GPT-4o-mini ngokukhethekile, kuyilapho imodeli encane nentsha kakhulu kulokho kuhlola, i-GPT-4.1-nano, yakuthola kahle. Isizukulwane nobukhulu bemodeli akuzange kuhambisane nganoma iyiphi indlela engaguquki, okuyisimo ukuhlolwa kwezinguqulo zamamodeli kwe-MachineTranslation.com okukhuluma ngaso kabanzi.

Lokho i-GPT-5 eyakushintsha ekuhumusheni

I-OpenAI yethule i-GPT-5 njengohlelo oluhlanganisiwe oluhambisa phakathi kwemodeli ezenzakalelayo esheshayo kanye nemodeli yokucabanga ejulile kuye ngomsebenzi, okuwushintsho lwesakhiwo kusuka kumklamo wemodeli eyodwa we-GPT-4. I-GPT-5 System Card ye-OpenAI ngokwayo ibika ukuthi izinga lamaphutha amaqiniso alolu hlobo lingaphansi cishe ngama-45% kunele-GPT-4 kanti lingaphansi ngama-80% kunele-GPT-3, okuyinzuzo yokwethembeka evamile ebonakala ikakhulukazi nasekuhumusheni. Ukuhlola kwe-MachineTranslation.com kukhombisa izinzuzo ezicacile lapho amamodeli enkathi ye-GPT-4 ayezabalaza khona kakhulu: esishweni esifanayo sesi-Spanish lapho i-GPT-4o-mini ne-GPT-4.1-nano zenza amaphutha, zombili i-GPT-5.4-mini kanye ne-GPT-5.4 zithole izinguqulo ezifanele, eziyizisho, ngayinye ikubeka ngendlela ehluke kancane kodwa zombili zisiqonda kahle isisho ngokwaso.

Lelo phuzu lokugcina libalulekile: ukuhlolwa okuhlukile kwenkulumo ejwayelekile yebhizinisi, "please confirm receipt of this document," ihunyushelwa olimini lwesiJalimane, kulethe imiphumela ecishe ifane ncamashí kuwo wonke amamodeli ahloliwe, kuhlanganisa namamodeli e-GPT-4o-mini nawesizukulwane se-GPT-5. Umehluko phakathi kwezizukulwane uqondene ngqo nolimi olungokomfanekiso nolungaqondile kahle, hhayi umehluko jikelele wekhwalithi.

Ingabe isizukulwane esisha se-GPT sihlala sisho ukuhumusha okungcono?

Cha. Ekuhloleni isisho sesiHebheru kuyo yonke imindeni yamamodeli, i-GPT-5.4-mini ne-GPT-5.4 afinyelele ezincazelweni ezimbili ezihlukene, ezingaphelele zesisho esifanayo, lapho kungekho nesilodwa esiqonde ngokugcwele, kuyilapho imodeli endala, engahlobene nayo isondele kakhulu. Imodeli entsha ayibi nembayo ngokuzenzakalelayo, futhi imodeli enkulu noma yakamuva akusona isinqumo esiphephe kakhudlwana ngokuzenzakalelayo emshweni othile.

Lapho i-GPT-5 imi khona mayelana nokuhumusha namuhla

I-GPT-3 kanye ne-GPT-4 zombili sezihoxisiwe ku-ChatGPT; i-GPT-4 isatholakala kuphela nge-API ye-OpenAI ekuhlanganisweni okukhona kakade. Isizukulwane samanje, i-GPT-5.4 ne-GPT-5.5, yikho i-MachineTranslation.com ekusebenzisayo namuhla, futhi sisebenza ngendlela encintisanayo namamodeli amanje avela kwabanye abahlinzeki, nakuba singabi ngcono ngendlela efanayo kuzo zonke izilimi ezibhangqiwe. Ukuze uthole ukubhekisisa ukuthi izinguqulo ezincane zamanje ze-GPT ziqhathaniseka kanjani zodwa kanye namamodeli afana ne-Claude ne-DeepSeek, bheka ukuhlolwa kwezinguqulo ezincane ubuso nobuso kwe-MachineTranslation.com, okungena ngokujulile kunalokho okungenziwa ukubuka kabanzi kwezizukulwane.

Kusho ukuthini lokhu uma wenza i-localization ngezinga elikhulu

Okutholakele okubaluleke kakhulu ohlelweni olukhulu lwe-localization akukhona ukuthi "i-GPT-5 yehlula i-GPT-3." Kunalokho ukuthi inguqulo yemodeli, hhayi nje umndeni wemodeli, yiyona enquma ukuthi isisho esithile noma umusho ongacacile uhunyushwe ngendlela efanele yini, futhi lokho kusebenza kuzo zonke izilimi ezibhangqiwe, hhayi nje ezimbalwa eziboniswe lapha. Lokhu kubaluleke kakhulu kokuqukethwe okuncike kakhulu kuphimbo nasolimini olungokoqobo ekuqaleni, ikakhulukazi imibhalo yokumaketha nokuqukethwe okwenziwe ngabasebenzisi, lapho isisho esisodwa esihunyushwe ngokoqobo singashintsha yonke incazelo yeposi noma yesikhangiso. Uhlelo oluhumushela okuqukethwe ezilimini eziningi luyohlangabezana nalolu hlobo lolimi ncamashí kuyo ngayinye yazo. Uchungechunge lokuhlola izinguqulo ezincane kanye nokuhlolwa kokuqhathanisa amamodeli aphelele kwe-MachineTranslation.com kuhlanganisa lokhu ngokujulile. Indlela yesikhulumi yokusebenza, esebenzisa amamodeli amanje e-GPT kanye namanye angaphezu kuka-20 kukho konke ukuhumusha, ikhona ikakhulukazi ngoba umlando wenguqulo yanoma iyiphi imodeli eyodwa awusona isiqinisekiso esithembekile ngokwanele ngokwayo.

Imibuzo evame ukubuzwa