NgoLwesibili ngesikhathi sendawo, i-OpenAI imemezele inqwaba yezinqubomgomo ezintsha zokuphepha, ezigxile ekulawuleni izigameko zokuphepha ezingase zenzeke kumamodeli ngesikhathi sokuhlolwa. Lezi zinyathelo ezintsha zihlanganisa ukuqashwa okunemininingwane eminingi ngesikhathi sokuthuthukiswa kwamamodeli, kanye nokubeka ukugcizelela okwengeziwe ekuhambisaneni nasekuphepheni kwamamodeli esigabeni sangemva kokuqeqeshwa.
I-OpenAI ithe eposini lebhulogi: “Njengoba amakhono amamodeli eqhubeka nokuthuthuka, izingozi esibhekana nazo lapho sithuthukisa futhi sihlola la mamodeli ngaphakathi nazo ziyanda. Amazinga esiwabekile ekuqashweni, ekuhambisaneni nasekuphepheni kumele ahlale ehamba phambili kunalezi zingozi.”
Lokhu kungokunye kwezinguquko ezinkulu zokuqala i-OpenAI ezimemezele esidlangalaleni emikhubeni yayo yezokuphepha kusukela yadalula isigameko sase-Hugging Face mhla zingama-21 kuJulayi.
Abamele i-OpenAI bathe lezi zinyathelo azethulwanga ngqo ngenxa yesigameko sase-Hugging Face, kodwa amakhono e-cybersecurity aboniswe yimodeli ezayo i-Astra, kanye nesivinini sokuthuthuka kwemboni yonke ye-AI, nakho kube phakathi kwezizathu ezenze inkampani ithathe lezi zinyathelo.
I-OpenAI iphinde yadalula ukuthi ngemva kwesigameko sase-Hugging Face, yamisa ukuqeqeshwa kwe-reinforcement learning (RL) amasonto amabili, kodwa manje isiqalile kabusha ukuqeqeshwa kwamamodeli amaningi anengozi ephansi.
I-OpenAI ibhale kubhulogi: “Okwamanje sisamisa uhlelo olukhulu kunazo zonke lokuqeqeshwa kwe-frontier reinforcement learning, kodwa senza ukuqeqeshwa nokuhlola okuncane ukuze sihlole ukuziphatha kwamamodeli, siqinisekise izinyathelo zokuphepha, futhi sithole ubufakazi obengeziwe mayelana nokuhambisana kwamamodeli ngaphambi kokuqhubeka.”
U-Amelia Glaese, iphini likamongameli wezocwaningo e-OpenAI, ugcizelele engxoxweni nezintatheli ukuthi njengoba amakhono amamodeli ethuthuka, inkampani nayo izokwandisa ukuqina kwezilawuli zayo zokuphepha; amamodeli anamandla kakhulu azohlolwa ngokuqinile kunawo wonke.
UGlaese utshele izintatheli: “Sesibeke izidingo namazinga okuthuthukiswa okuphephile. Lezi zidingo namazinga zizolungiswa ngokuya ngezinga lengozi esilihlolayo.”
Ngemva kwalesi sigameko, izinyathelo ze-OpenAI ze-cybersecurity zagxekwa umphakathi. Esigamekweni sangaphambilini, imodeli yeqa indawo yayo yokuqeqeshwa eyayikhona ngokugqekeza ithuluzi elikwazi ukungena ku-inthanethi ngaphakathi kwenethiwekhi ye-OpenAI. Izinyathelo ezintsha zokuphepha ezethulwe yi-OpenAI zihlanganisa ukuqinisa ukuhlukaniswa kwenethiwekhi, kodwa inkampani ayikakadaluli imininingwane ethile yobuchwepheshe.
I-OpenAI ithe ngaphansi kohlelo olusha lokuphepha, “ukungena ngokungemthetho komthwalo owodwa womsebenzi noma kwensizakalo ewusekelayo, ngokwako, ngeke kunikeze umhlaseli ukufinyelela okungagunyaziwe ku-inthanethi noma kwamanye amanethiwekhi angaphakathi.”
Isinyathelo esibaluleke kakhulu sokuphepha kulokhu kuthuthukiswa uhlelo olusha lokuqapha. Lolu hlelo luzohlola imisebenzi eyenziwa amathuluzi, imikhondo yokucabanga engatholakala, kanye namarekhodi omsebenzi ukuze lubone izinhlobo ezahlukene zokuziphatha okungagunyaziwe. I-OpenAI ithe inhloso yayo ukuthumela isexwayiso kungakapheli imizuzu engu-30 kutholwe umsebenzi osolisayo.
I-OpenAI ilinganisela ukuthi izinsiza zekhompyutha ezidingwa yilolu hlelo lokuqapha zilingana cishe no-20% wezinsiza ezisetshenziswa yinqubo eqashwayo uqobo. Inkampani ithembise ukushicilela imininingwane eyengeziwe ngalolu hlelo emaposini ebhulogi azayo.
Ngaphezu kwalokho, umbiko osemthethweni we-OpenAI wokuhlaziya okwenzeka ngemva kwalesi sigameko awukashicilelwa.