Iqela labaphandi abavela kwiYunivesithi yaseStanford bafunde impembelelo yokusebenzisa abancedisi beekhowudi ezikrelekrele kwimbonakalo yobuthathaka kwikhowudi. Izisombululo ezisekelwe kumatshini wokufunda we-OpenAI Codex waqwalaselwa, njenge-GitHub Copilot, evumela ukuveliswa kweebhloko zekhowudi eziyinkimbinkimbi, ukuya kwimisebenzi esele yenziwe. Inkxalabo kukuba ekubeni ikhowudi yokwenyani evela kwiindawo zokugcina zikawonke-wonke ze-GitHub, kubandakanywa nezo ziqulathe ubuthathaka, zisetyenziselwa ukuqeqesha imodeli yokufunda yomatshini, ikhowudi ehlanganisiweyo inokuphinda iimpazamo kwaye iphakamise ikhowudi equlethe ubuthathaka, kwaye ayithatheli ngqalelo imfuneko yokwenza. iitshekhi ezongezelelweyo xa kusetyenzwa idatha yangaphandle.
Amavolontiya angama-47 anamava ahlukeneyo kwiprogram abandakanyekayo kuphononongo - ukusuka kubafundi ukuya kwiingcali ezineminyaka elishumi yamava. Abathathi-nxaxheba bahlulahlulwe bangamaqela amabini - uvavanyo (abantu be-33) kunye nolawulo (abantu abayi-14). Omabini amaqela ayenokufikelela kuwo nawaphi na amathala eencwadi kunye nezixhobo ze-Intanethi, kubandakanywa ukukwazi ukusebenzisa imizekelo esele yenziwe kwi-Stack Overflow. Iqela lokulinga linikwe ithuba lokusebenzisa umncedisi we-AI.
Umthathi-nxaxheba ngamnye unikwe imisebenzi emi-5 enxulumene nekhowudi yokubhala apho kunokuba lula ukwenza iimpazamo ezikhokelela kubuthathaka. Ngokomzekelo, kwakukho imisebenzi ekubhaleni ukubethela kunye nemisebenzi yokucima, usebenzisa iisignesha zedijithali, idatha yokucubungula ebandakanyekayo ekubunjweni kweendlela zefayile okanye imibuzo ye-SQL, ukuphatha amanani amakhulu kwikhowudi ye-C, igalelo lokucubungula eliboniswe kumaphepha ewebhu. Ukuqwalasela impembelelo yeelwimi zeprogram kukhuseleko lwekhowudi eveliswayo xa usebenzisa abancedisi be-AI, izabelo zagubungela iPython, C, kunye neJavaScript.
Ngenxa yoko, kwafunyaniswa ukuba abathathi-nxaxheba abasebenzisa umncedisi we-AI ohlakaniphile ngokusekelwe kwimodeli ye-codex-davinci-002 balungiselele ikhowudi ekhuselekileyo kakhulu kunabathathi-nxaxheba abangazange basebenzise umncedisi we-AI. Ngokubanzi, kuphela i-67% yabathathi-nxaxheba kwiqela elisebenzise umncedisi we-AI abakwaziyo ukunika ikhowudi echanekileyo nekhuselekileyo, ngelixa kwelinye iqela eli nani lalingama-79%.
Kwangaxeshanye, izikhombisi zokuzithemba zazichasene - abathathi-nxaxheba abasebenzisa umncedisi we-AI bakholelwa ukuba ikhowudi yabo iya kukhuseleka ngakumbi kunaleyo yabathathi-nxaxheba belinye iqela. Ukongeza, kwaphawulwa ukuba abathathi-nxaxheba abathembe umncedisi we-AI ngaphantsi kwaye bachitha ixesha elininzi behlalutya izindululo ezinikiweyo kunye nokwenza utshintsho kubo benza ubuthathaka obumbalwa kwikhowudi.
Umzekelo, ikhowudi ekhutshelwe kwiilayibrari ze-cryptographic iqulathe amaxabiso eparameter akhuselekile ngakumbi kunekhowudi ecetyiswe ngumncedisi we-AI. Kwakhona, xa usebenzisa umncedisi we-AI, ukhetho lwe-algorithms engathembekanga kunye nokungabikho kokutshekishwa kobunyani bamaxabiso abuyisiweyo kwarekhodwa. Kumsebenzi obandakanya ukuguqulwa kwamanani kwi-C, iimpazamo ezininzi zenziwe kwikhowudi ebhaliweyo kusetyenziswa umncedisi we-AI, okukhokelela ekuphuphumeni okupheleleyo.
Ukongezelela, sinokuphawula uhlolisiso olufanayo lweqela leYunivesithi yaseNew York, olwaqhutywa ngoNovemba, lubandakanya abafundi abangama-58 abacelwa ukuba baphumeze ucwangciso lokulungiselela uludwe lokuthenga ngolwimi lwesi-C. Iziphumo zibonise impembelelo encinci yomncedisi we-AI kukhuseleko lwekhowudi-abasebenzisi abasebenzisa umncedisi we-AI, ngokwe-avareji, malunga ne-10% yeempazamo ezinxulumene nokhuseleko.
umthombo: opennet.ru