OpenAI ithe ukuhlolwa kwangaphakathi nokwenziwa ochwepheshe kubonisa ukuthi imodeli i-Astra, esizokhishwa maduze, yenze intuthuko enkulu ekuhleleni kwama-agent nasekuphepheni kwenethiwekhi. Ngokusho kwe-《Preparedness Framework》 yayo, i-Astra isivele ibekwe njengemodeli yokuqala ye-OpenAI efinyelele ezingeni lengozi “Critical” emkhakheni wokuphepha kwenethiwekhi.
Inothi: Ngokwalo hlaka, imodeli ye-AI ifinyelela umkhawulo “Critical” uma ihlangabezana nanoma yimuphi wale mibandela elandelayo:
Ukuthola izimbobo zosuku oluyiziro ngaphandle kokungenelela: Ngaphandle kosizo lomuntu, ikwazi ukuhlonza nokwakha izinhlelo ezisebenzayo zokuxhaphaza izimbobo zosuku oluyiziro kuwo wonke amazinga obucayi ezinhlelweni zangempela ezibalulekile eziningana eziqinisiwe kwezokuphepha.
Ukuqalisa ngokuzenzakalelayo ukuhlasela okuphelele kusukela ekuqaleni kuye ekugcineni: Ngokunikezwa kuphela umgomo wamasu osezingeni eliphezulu, ikwazi ukuzakhela nokusebenzisa ngokuzenzakalelayo ukuhlasela okusha okuphelele kwenethiwekhi ngokumelene nendawo eqinisiwe.
OpenAI igcizelele ngokukhethekile ukuthi i-Astra iyimodeli engakakhishwa, futhi ayizange ibambe iqhaza ekuhlaselweni kwenethiwekhi okwake kwaqondiswa ku-Hugging Face.
Ukuqinisekisa ukuthi ukuthuthukiswa okuqhubekayo kuhambisana nokuphepha nemithetho, i-OpenAI imemezele ukuthi isithathe izinyathelo eziningi zokulawula:
Imodeli enamakhono aphezulu nemisebenzi yocwaningo nokuthuthukiswa ehambisana nayo izobhekwa ngaphansi kwezilawuli eziqinile zokuphepha, okuhlanganisa ukusungulwa kwezindawo zokuhlola ezihlukanisiwe, ukukhawulelwa kokufinyelela kunethiwekhi namathuluzi, ukuqiniswa kokuvikelwa nokubethelwa kwezisindo zemodeli, ukusetshenziswa kwamanye amakhono okuqapha nokuthola izingozi, kanye nokusebenza ngaphakathi kwe-sandbox; yonke imisebenzi yangaphakathi ehlobene ne-Astra engakafinyeleli kulezi zindinganiso zokuphepha eziqinisiwe imisiwe.
Zonke izinhlelo zama-agent ze-Astra, okuhlanganisa nezigaba zokuqeqesha nokuhlola, sezibekwe ngaphansi kokuqapha komhlaba wonke okubheka ukuziphatha okuyingozi kakhulu nokwehluleka kokuhambisana nezimiso. Ngokuhlola uchungechunge lokucabanga kwemodeli (CoT), izenzo eziyingozi kakhulu ziyabuyekezwa futhi zivinjwe.
Isebenzisana nezinhlangano zikahulumeni ezifanele kanye nezinhlangano ezikhethiwe zokuphepha kwe-AI ukuze kuhlolwe imodeli ndawonye; futhi inikeza ozakwethu bokuhlola bezinkampani zangaphandle imithethonqubo yokulawula ukuphepha enconyiwe, ukuze kuqinisekiswe ukuthi ukuhlolwa nemithwalo yomsebenzi enobungozi obuphezulu ingasebenza ngokuphepha.
Isikhulu esiphezulu se-OpenAI, uSam Altman, kamuva sathi i-Astra iyimodeli esebenza ngamandla, futhi basebenza ngokuzikhandla ukuze bayikhiphe emphakathini.
Sihlale sikholelwa ukuthi ukukhawulela amamodeli ahamba phambili kubantu abambalwa akulona isu elihle.
Nokho, ngenxa yamakhono ayo anamandla emkhakheni wokuphepha kwenethiwekhi, sisadinga isikhathi esengeziwe ukuze siqinisekise ukuthi konke kuphephile. Sethemba ukuthi ngeke nilinde isikhathi eside kakhulu!
