Ithimba labacwaningi abavela eNyuvesi yaseStanford lifunde umthelela wokusebenzisa abasizi bokubhala amakhodi abahlakaniphile ekubukekeni kobungozi bekhodi. Izixazululo ezisuselwe kuplathifomu yokufunda yomshini we-OpenAI Codex zacatshangelwa, njenge-GitHub Copilot, evumela ukukhiqizwa kwamabhulokhi ekhodi ayinkimbinkimbi, kuze kufike emisebenzini eseyenziwe ngomumo. Ukukhathazeka ukuthi njengoba ikhodi yangempela evela kumakhosombe asesidlangalaleni e-GitHub, okuhlanganisa lawo aqukethe ubungozi, isetshenziselwa ukuqeqesha imodeli yokufunda yomshini, ikhodi ehlanganisiwe ingase iphinde amaphutha futhi iphakamise ikhodi equkethe ubungozi, futhi futhi inganaki isidingo sokwenza. ukuhlola okwengeziwe lapho kusetshenzwa idatha yangaphandle.
Amavolontiya angu-47 anolwazi oluhlukene ezinhlelweni abambe iqhaza ocwaningweni - kusukela kubafundi kuya kochwepheshe abaneminyaka eyishumi yolwazi. Abahlanganyeli bahlukaniswe ngamaqembu amabili - ukuhlolwa (abantu abangu-33) nokulawula (abantu abangu-14). Womabili amaqembu abe nokufinyelela kunoma imaphi amalabhulali nezinsiza ze-inthanethi, okuhlanganisa ikhono lokusebenzisa izibonelo esezivele zenziwe kusukela ku-Stack Overflow. Iqembu lokuhlola linikezwe ithuba lokusebenzisa umsizi we-AI.
Umhlanganyeli ngamunye unikezwe imisebenzi engu-5 ehlobene nekhodi yokubhala lapho okungaba lula khona ukwenza amaphutha aholela ekubeni sengozini. Isibonelo, kwakukhona imisebenzi ekubhaleni ukubethela nokukhipha imisebenzi, kusetshenziswa amasignesha edijithali, ukucubungula idatha ehilelekile ekwakhiweni kwezindlela zefayela noma imibuzo ye-SQL, ukuphatha izinombolo ezinkulu ngekhodi C, ukucubungula okokufaka okuboniswa emakhasini ewebhu. Ukucabangela umthelela wezilimi zokuhlela ekuvikelekeni kwekhodi ekhiqizwayo lapho kusetshenziswa abasizi be-AI, izabelo zibandakanya i-Python, C, neJavaScript.
Ngenxa yalokho, kutholakale ukuthi abahlanganyeli abasebenzisa umsizi we-AI ohlakaniphile ngokusekelwe kumodeli ye-codex-davinci-002 balungiselele ikhodi evikelekile kakhulu kunabahlanganyeli abangasebenzisanga umsizi we-AI. Sekukonke, kuphela i-67% yabahlanganyeli eqenjini elisebenzisa umsizi we-AI abakwazile ukunikeza ikhodi elungile nevikelekile, kanti kwelinye iqembu lesi sibalo sasingu-79%.
Ngesikhathi esifanayo, izinkomba zokuzethemba zaziphambene - abahlanganyeli abasebenzisa umsizi we-AI bakholelwa ukuthi ikhodi yabo izovikeleka kakhulu kunaleyo yabahlanganyeli abavela kwelinye iqembu. Ngaphezu kwalokho, kwaphawulwa ukuthi ababambiqhaza abethemba umsizi we-AI kancane futhi bachitha isikhathi esiningi behlaziya imiyalo enikeziwe futhi benza izinguquko kubo benza ubungozi obumbalwa kukhodi.
Isibonelo, ikhodi ekopishwe emitapweni yolwazi yokufihla ulwazi iqukethe amanani epharamitha azenzakalelayo aphephe kakhulu kunekhodi ephakanyiswe umsizi we-AI. Futhi, lapho usebenzisa umsizi we-AI, ukukhethwa kwama-algorithms wokubethela athembekile kancane kanye nokuntuleka kokuhlola ubuqiniso bamanani abuyisiwe kwarekhodwa. Emsebenzini ohilela ukukhohliswa kwenombolo ku-C, amaphutha engeziwe enziwa kukhodi ebhalwe kusetshenziswa umsizi we-AI, okuholela ekuchichimeni okuphelele.
Ukwengeza, singaphawula ucwaningo olufanayo lweqembu laseNyuvesi yaseNew York, olwenziwa ngo-November, oluhilela abafundi abangu-58 abacelwa ukuba basebenzise uhlaka lokucubungula uhlu lokuthenga ngolimi C. Imiphumela ibonise umthelela omncane womsizi we-AI ekuvikelekeni kwekhodi—abasebenzisi abasebenzisa umsizi we-AI benza, ngokwesilinganiso, cishe amaphutha ahlobene nokuphepha angu-10%.
Source: opennet.ru