Ukusebenzisa amamodeli e-AI ukuqeqesha amanye amamodeli e-AI sekuba yinkomba ebalulekile ehlolwa amalabhorethri e-AI esizukulwaneni esisha. Ngomhla ka-28 ngesikhathi sendawo, omunye wabacwaningi bohlelo locwaningo lwe-Anthropic ubonise ukuthi lo mbono ungabukeka kanjani uma ususetshenziswa ngempela.
I-Anthropic ikhiphe iphepha layo lakamuva elinesihloko esithi “Automated Researcher Can Reliably Mitigate Alignment Failures”, elichaza indlela yokusebenzisa amasistimu e-AI ukuze kuthuthukiswe ukusebenza kwamamodeli ochungechungeni lwezivivinyo eziyisisekelo zokuqondanisa. Ithimba locwaningo lakha izivivinyo zokuziphatha okungu-10 okuhlukene kokungaqondani, futhi uhlelo oluzenzakalelayo lwagcina luthuthukise zonke izinkomba eziyi-10, ngaphandle kokudela amakhono emodeli ewonke.
Lolu hlelo lwakhiwe ngaphansi kobuholi bukaChen Yuehan (okuwukuhumusha kwegama ngomsindo), oyilungu lohlelo locwaningo lwe-Anthropic, futhi indlela olusebenza ngayo iyafana nocwaningo lwendabuko. Uhlelo oluzenzakalelayo luqala ngokubheka izincwadi ezikhona, luphakamise izindlela zokuqeqesha, bese luqeqesha imodeli imizuzu engu-30 ngokohlelo. Ngokuzama kaningi, luthuthukisa kancane kancane imiphumela yokuhlolwa. Izindlela ezisebenzayo ziyagcinwa, kanti ezingasebenzi ziyasuswa, okwenza inqubo yonke yocwaningo ikwazi ukukhuliswa ngokushesha.
Iphepha lithi, ngokubanzi, le miphumela inikeza ubufakazi bokuqala bokuthi ukuqondanisa okuzenzakalelayo ngemuva kokuqeqeshwa kungaba yindlela esebenzayo ngempela esikhathini esiseduze.
Lolu cwaningo luphinde lube yisinyathelo esiya ekuzithuthukiseni okuphindaphindayo, okuyisigaba esilandelayo esibalulekile ekuthuthukisweni kwe-AI ngokombono wabantu abaningi. Uma amamodeli e-AI ekwazi ukuthuthukisa izindlela zawo zokuqeqeshwa kokuqondanisa, kungenzeka futhi ukuthi athuthukise izinqubo zokuqeqesha ezibanzi. Uma sekufinyelelwe kuleso sigaba, abacwaningi be-AI abangabantu bangase bagcine bengasadingeki kancane kancane.
Iphepha liphinde laqhathanisa ngokuqondile ukusebenza kwabacwaningi bokuqondanisa abazenzakalelayo (AAR) nabacwaningi abangabantu: “Izindlela ezihamba phambili ze-AAR ngokwesilinganiso zidinga amahora angu-6 kuphela ukuze zeqe izindlela eziphakanyiswe abacwaningi abangabantu abanamava. Izikhombandlela zocwaningo eziholwa ngabantu aziholeli ekusebenzeni okunamandla.”
Umehluko wezindleko nawo ubalulekile. “I-AAR idinga kuphela cishe ama-dollar angu-4 ngehora ngezindleko zokucubungula ze-API (inothi: ngokwezinga lokushintshisana lamanje, cishe ama-yuan angu-27), kanti imali esiyikhokhela abacwaningi abangabantu ingama-dollar angu-150 ngehora (ngokwezinga lokushintshisana lamanje, cishe ama-yuan angu-1,011).”
Nokho, le ndlela isenayo imikhawulo esobala. Ukusebenza kohlelo oluzenzakalelayo kuncike kuqala ekutheni izivivinyo eziyisisekelo zikwazi yini ukubonisa ngokunembile izinhloso zangempela zokuqondanisa. Ngisho noma izivivinyo ngokwazo zithembekile, ukwakha nokuzigcina isikhathi eside kusadinga utshalomali olukhulu. Izincwadi zocwaningo ezithembele kuzo izinhlelo ezizenzakalelayo nazo zidinga ukunakekelwa nokwandiswa okuqhubekayo.
Inkomba
Automated researchers can reliably mitigate alignment failures