Njengamanje, sesiqalile ukubona ukwanda kwezixazululo ezahlukene eziqala ukunikezwa maqondana nokukhiqizwa kwekhodi kusetshenziswa ubuhlakani bokwenziwa (AI), kanti insimu yokucubungula ulimi lwemvelo (i-NLP) ivule indlela yochungechunge lwe-AI ezikhiqiza ikhodi ezilimini ezahlukene zokuhlela.
Phakathi kwalokhu, singagqamisa i-GitHub Copilot, i-AlphaCode, kanye ne-Codex, futhi manje singangeza ikhambi elisha elivela kubacwaningi baseCarnegie Mellon University abasanda kwethula i-"PolyCoder," i-generator yamakhodi esekelwe kumodeli yolimi lwe-OpenAI ye-GPT-2 eyaqeqeshwa kusizindalwazi samakhodi esingu-249 GB ngezilimi ezingu-12 zokuhlela.
Mayelana ne-PolyCoder
Abalobi be-PolyCoder bathi iyakwazi ukubhala i-C ngokunemba okukhulu kunanoma iyiphi imodeli eyaziwayo, kufaka phakathi i-Codex.
I-AI ekhiqiza ikhodi, engabhala ikhodi yomthombo ngezilimi ezahlukene zokuhlela kusuka ku-prompt, ithembisa ukunciphisa izindleko zokuthuthukiswa kwesofthiwe ngenkathi ivumela abathuthukisi ukuthi bagxile emisebenzini yokudala nengaphindaphindi kakhulu.
I-PolyCoder yaqeqeshwa kusetshenziswa idatha evela ezindaweni eziningana zokugcina izinhlelo ze-GitHub, ehlanganisa izilimi ezingu-12 zokuhlela ezidumile: i-C, i-C#, i-C++, i-Go, i-Java, i-JavaScript, i-PHP, i-Python, i-Ruby, i-Rust, i-Scala, kanye ne-TypeScript.
Isethi yedatha engahlungiwe ifinyelele ku-631 GB yedatha kanye namafayela ayizigidi ezingu-38,9. Ithimba lithe likhethe ukuqeqesha i-PolyCoder nge-GPT-2 ngenxa yemikhawulo yesabelomali. I-PolyCoder iyatholakala njengomthombo ovulekile, futhi abacwaningi banethemba lokuthi ingakwazi ukwenza ucwaningo lube yintando yeningi emkhakheni wokukhiqiza ikhodi ye-AI, okuze kube manje obuswa yizinkampani ezixhaswe kahle.
Abacwaningi bakholelwa ukuthi i-PolyCoder isebenza kangcono kunezinye izinhlobo ekukhiqizeni ikhodi ngolimi lwe-C. Kodwa-ke, i-Codex ibilokhu isebenza kangcono kunezinye izilimi. "I-PolyCoder isebenza kahle kakhulu kune-Codex kanye nazo zonke ezinye izinhlobo ngolimi lwe-C."
“Lapho u-Copilot ephuma ku-GitHub ehlobo eledlule, kwacaca ukuthi lawa mamodeli amakhodi olimi amakhulu kakhulu angaba usizo olukhulu ekusizeni onjiniyela nokwandisa umkhiqizo wabo. Kepha ayikho imodeli esondelene naleso sikali ebitholakala esidlangalaleni,” abacwaningi batshele iVentureBeat nge-imeyili. “Ngakho-ke [i-PolyCoder] yaqala ngokuthi uVincent ezama ukuthola ukuthi iyiphi imodeli enkulu kunazo zonke engaqeqeshwa kuseva yethu yelebhu, egcine ibe amapharamitha ayizigidi eziyizinkulungwane ezingama-2700… .zazitholakala esidlangalaleni ngaleso sikhathi.”
“ Uma kuqhathaniswa amamodeli omthombo ovulekile kuphela, i-PolyCoder isebenza kangcono kunemodeli ye-GPT-Neo 2.7B enobukhulu obufanayo ku-C, JavaScript, Rust, Scala, kanye ne-TypeScript,” baphawula. “Kwezinye izilimi ezingu-11, zonke ezinye amamodeli omthombo ovulekile, okuhlanganisa neyethu, zimbi kakhulu (ziyinkimbinkimbi kakhulu) kune-Codex,” kusho abacwaningi be-CMU.
Ngalokhu, i-PolyCoder ibekwe njengesixazululo esithakazelisa kakhulu, njengoba ngenkathi amalabhorethri ocwaningo afana ne-OpenAI ka-Elon Musk kanye ne-Alphabet's DeepMind athuthukise i-AI ekhiqiza ikhodi enamandla, amasistimu amaningi aphumelela kakhulu awatholakali emthonjeni ovulekile. Izinkampani ezihola kancane azikwazi ukufinyelela kuyo futhi lesi simo sinciphisa ucwaningo lwazo kulo mkhakha.
Isibonelo, idatha yokuqeqesha evela ku-OpenAI Codex, enika amandla isici se-GitHub's Copilot, ayizange yenziwe esidlangalaleni, ivimbela abacwaningi ukuthi bacwengisise imodeli ye-AI noma ukutadisha izici ezithile zayo, njengokusebenzisana.
"Izinkampani ezinkulu zobuchwepheshe aziwakhiphi obala amamodeli azo, okuyinto ebambezela ucwaningo lwesayensi kanye nokwenza intando yeningi yamamodeli amakhulu kangaka ekhodi yolimi," kusho abacwaningi. “Ngezinga elithile, sithemba ukuthi imizamo yethu yomthombo ovulekile izoqinisekisa abanye ukuthi benze okufanayo. Kodwa isithombe esikhulu wukuthi umphakathi kumele ukwazi ukuziqeqeshela wona ngokwawo la mamodeli. Imodeli yethu yeqe umkhawulo walokho ongakwazi ukukuqeqesha kuseva eyodwa - noma yini enkulu idinga iqoqo lamaseva, anyusa kakhulu izindleko. "
Okokugcina, uma unesithakazelo sokufunda okwengeziwe ngalokhu , ungathola imininingwane kusixhumanisi esilandelayo.