I-PolyCoder, ikhowudi yomthombo ovulekileyo ovelisa i-AI enokuthi igqwese iCodex 

Umbhali: @ Laurent - Fotolia.com

Okwangoku, Siqalile ukubona ukwanda izisombululo ezahlukeneyo eziqala ukunikezelwa ngokunxulumene ukwenziwa kwekhowudi kusetyenziswa ubukrelekrele bokwenziwa (AI) kunye nentsimi yokulungiswa kolwimi lwendalo (NLP) ivule indlela yoluhlu lwee-AI ezivelisa ikhowudi kwiilwimi ezahlukeneyo zokucwangcisa.

Ngokuba Sinokugqamisa, umzekelo, iGitHub Copilot, iAlphaCode kunye neCodex kwaye ngoku sinokuthi songeze isisombululo esitsha esivela kwisandla se abaphandi kwiYunivesithi yaseCarnegie Mellon abo kutshanje yaziswa "PolyCoder", i-code generator esekelwe kwimodeli ye-OpenAI ye-GPT-2 yolwimi eyaqeqeshwa kwi-database yekhowudi ye-249 GB kwiilwimi ze-programming ze-12.

Malunga nePolyCoder

Ababhali bePolyCoder babanga ukuba kunjalo ekwaziyo ukubhala u-C ngokuchanekileyo kunayo nayiphi na imodeli eyaziwayo, kuquka iCodex.

Ikhowudi yokuvelisa i-AI, Ungabhala ikhowudi yemvelaphi kwiilwimi ezahlukeneyo zokucwangcisa ukusuka kwisaziso, ithembisa ukunciphisa iindleko zophuhliso lwesoftware ngelixa ivumela abaphuhlisi ukuba bagxile kwimisebenzi yokuyila kunye nokuphindaphinda kancinci.

I-PolyCoder yaqeqeshwa kwidatha evela kwiindawo ezininzi zokugcina ze-GitHub, ezigubungela iilwimi ezili-12 ezidumileyo zokucwangcisa: C, C #, C ++, Hamba, Java, JavaScript, PHP, Python, Ruby, Rust, Scala kunye neTypeScript.

Isethi yedatha engahluzwanga iphelele kwi-631 GB yedatha kunye neefayile ze-38,9 yezigidi. Iqela latsho oko wakhetha ukuqeqesha iPolyCoder nge-GPT-2 ngenxa yokunqongophala kwemali. I-PolyCoder ifumaneka njengomthombo ovulekileyo kwaye abaphandi bathemba ukuba banokudemokhrasi uphando kwinkalo yokuveliswa kwekhowudi ye-AI, kude kube ngoku ilawulwa ziinkampani ezixhaswa ngemali kakuhle.

Abaphandi bakholelwa ukuba iPolyCoder isebenza ngcono kunezinye iimodeli ekuveliseni ikhowudi kulwimi C. Noko ke, iCodex ibisoloko iyodlula ngezinye iilwimi. «I-PolyCoder igqwesa kakhulu iCodex kunye nazo zonke ezinye iimodeli kulwimi lwe-C.

“Xa uCopilot waphuma eGitHub kwihlobo elidlulileyo, kuye kwacaca ukuba le mifuziselo yekhowudi yolwimi mikhulu inokuba luncedo kakhulu ekuncedeni abaphuhlisi kunye nokwandisa imveliso yabo. Kodwa akukho modeli ikufutshane kweso sikali yayifumaneka esidlangalaleni, ”abaphandi baxelele iVentureBeat nge-imeyile. “Ke [iPolyCoder] yaqala ngoVincent izama ukubona ukuba yeyiphi eyona modeli inkulu inokuqeqeshwa kwiseva yethu yelebhu, ethe yaphela ibeyi-2700 yeebhiliyoni zeeparamitha… efumaneka esidlangalaleni ngelo xesha.”

Xa uthelekisa kuphela iimodeli zomthombo ovulekileyo, "I-PolyCoder igqwesa i-GPT-Neo 2.7B efana ne-C, JavaScript, Rust, Scala, kunye ne-TypeScript," balatha. "Kwezinye iilwimi ze-11, zonke ezinye iimodeli zomthombo ovulekileyo, kubandakanywa neyethu, zibi kakhulu (ukuphazamiseka okuphezulu) kuneCodex," abaphandi beCMU bongezelela.

Kungenxa yoko le nto iPolyCoder ibekwe njengesisombululo esinomdla kakhulu, kuba ngelixa iilabhoratri zophando ezifana ne-Elon Musk's OpenAI kunye ne-Alphabet's DeepMind ziye zaphuhlisa ikhowudi yokuvelisa i-AI enamandla, uninzi lweenkqubo eziphumelele kakhulu azifumaneki kumthombo ovulekileyo. Iinkampani ezinobuncwane obuphantsi azikwazi ukufikelela kuyo kwaye le meko inciphisa uphando lwabo kwintsimi.

Ngokomzekelo, idatha yoqeqesho evela kwi-OpenAI Codex, enika amandla i-GitHub's Copilot feature, ayizange yenziwe esidlangalaleni, inqanda abaphandi ekucoceni imodeli ye-AI okanye ukufunda iinkalo ezithile zayo, ezifana nokusebenzisana.

"Iinkampani ezinkulu zetekhnoloji aziyikhuphi esidlangalaleni imodeli yazo, nto leyo ibambe uphando lwezenzululwazi kunye nedemokhrasi yeemodeli ezinkulu zekhowudi yolwimi," batsho abaphandi. “Ukusa kumlinganiselo othile, sinethemba lokuba iinzame zethu zomthombo ovulekileyo ziya kuqinisekisa abanye ukuba benze okufanayo. Kodwa umfanekiso omkhulu ngowokuba uluntu kufuneka lukwazi ukuziqeqesha ngokwalo le mifuziselo. "Imodeli yethu ityhale umda wento onokuthi uyiqeqeshe kwiseva enye: nantoni na enkulu ifuna iqela leeseva, elonyusa kakhulu ixabiso."

Gqibela ukuba unomdla wokwazi okungakumbi ngayo, ungazijonga iinkcukacha kwi ukulandela ikhonkco.


Yiba ngowokuqala ukuphawula

Shiya uluvo lwakho

Idilesi yakho ye email aziyi kupapashwa. ezidingekayo ziphawulwe *

*

*

  1. Uxanduva lwedatha: UMiguel Ángel Gatón
  2. Injongo yedatha: Ulawulo lwe-SPAM, ulawulo lwezimvo.
  3. Umthetho: Imvume yakho
  4. Unxibelelwano lwedatha: Idatha ayizukuhanjiswa kubantu besithathu ngaphandle koxanduva lomthetho.
  5. Ukugcinwa kweenkcukacha
  6. Amalungelo: Ngalo naliphi na ixesha unganciphisa, uphinde uphinde ucime ulwazi lwakho.