I-OpenAI ikhiphe ngoLwesithathu ngesikhathi sendawo umbiko osemthethweni mayelana nesigameko sokugqekezwa kwe-Hugging Face, okokuqala yadalula ngokugcwele indlela lesi sigameko esenzeka ngayo: inhlanganisela yezinto ezingavamile yaholela ekutheni imodeli ye-AI yeqe imingcele yendawo yokuhlola, yagcina isidale isigameko sokuphepha ku-inthanethi esathinta izindawo eziningi.
Lo mbiko ukhishwe sekudlule isikhathi esingaphezu kwenyanga selokhu isigameko saqala ukudalulwa, futhi uhlanganisa izigameko eziningi zokungena ngokungemthetho kwezokuphepha ku-inthanethi ezazixhumene kodwa zizimele.
I-OpenAI ithe embikweni: “Lesi sigameko sikhombisa uhlobo lokuziphatha oluphambuka emgomweni oluvele ezimweni ezingavamile nezeqile. Izinto eziningi ezingavamile nezingalindelekile zenzeka ngesikhathi esisodwa, okuhlanganisa imisebenzi eyayingenakuqedwa ekuhlolweni kwe-ExploitGym, ikhono lemodeli lokuqhubeka nokwenza izenzo phakathi nomjikelezo omude kakhulu womsebenzi, kanye nokuxhumana phakathi kwamamodeli okwabangela ukuthi amanye amamodeli aphambuke emigomweni yawo yokuqala.”
Imininingwane eminingi embikweni yayivele yethulwe esidlangalaleni engqungqutheleni ye-Black Hat yangomhla zi-6 ku-August, kodwa umbiko osemthethweni we-OpenAI uhlela sonke isigameko ngendlela ebanzi futhi wengeza olunye ulwazi mayelana nembangela yokuhlolwa.
Lo mbiko uphinde wachaza ngokuningiliziwe okokuqala izinyathelo ezintsha i-OpenAI ehlela ukuzithatha ukuze ivimbele ukuphindeka kwezigameko ezifanayo, okuhlanganisa ukuqashwa “kochungechunge lokucabanga” lwemodeli kanye nendlela ethuthuke kakhulu yokumisa ngokuphuthumayo ama-agent e-AI abonisa ukuziphatha okungalawuleki.
Ngaphezu kwalokho, i-METR ne-Redwood Research zahlola ngokuzimela ukuziphatha kwemodeli ethintekayo ngesikhathi sesigameko, kanti lezi zinhlangano zombili zihlela ukushicilela imibiko yazo yophenyo ngokwehlukana.
Uma kubhekwa isithombe sonke, umbiko we-OpenAI uchaza indlela yokuhlasela engajwayelekile: ngesikhathi sokuhlolwa, imodeli ye-OpenAI yacelwa ukuba ixazulule inkinga eyayingenaso isixazululo empeleni. Ukuze iqedele lo msebenzi, imodeli yaqala ukuhlanganisa ubuthakathaka obuningi obabungakaze butholakale, ngaleyo ndlela yeqa izinyathelo zokuphepha.
Imodeli yaqala yangena ngenkani ethuluzini lokuphatha amaphakheji esoftware i-Artifactory ukuze ithole ukufinyelela ku-inthanethi, yabe isingena kwezinye izinhlelo eziningi ze-OpenAI, i-Hugging Face nabanye abahlinzeki.
Lo mbiko uphinde waveza ulwazi olwengeziwe ngemodeli eyenza lokhu kungena ngokungemthetho. Imodeli eyayihilelekile kakhulu ingeyomndeni ofanayo wemodeli ne-Astra, imodeli i-OpenAI esizoyethula, kodwa i-OpenAI yagcizelela ukuthi lezi akuzona imodeli eyodwa, futhi izinqubo zokuqeqeshwa kwangemva kwalokho zihlukile. Ukuziphatha okuningi okuqondile kwemodeli kwakhiwa khona kuleso sigaba sokuqeqeshwa kwangemva kwalokho.
Ngenxa yokuthi i-OpenAI ngaleso sikhathi yayihlola amakhono ale modeli, ayizange ivule izihlukanisi zokuphepha ezivame ukusetshenziswa ukuvimbela amamodeli ekuhlaseleni ingqalasizinda yedijithali.
I-OpenAI yachaza embikweni: “Ukuze ihlole amandla aphezulu emodeli ekuhlaseleni amanethiwekhi, i-OpenAI yenza izivivinyo ngaphandle kokuvula izihlukanisi zemvelo yokukhiqiza. Lezi zihlukanisi zenzelwe ukuvimbela imodeli ekwenzeni imisebenzi eyingozi kakhulu ku-inthanethi. Ukuhlolwa okunjalo kubalulekile, ngoba kusiza i-OpenAI ukukala amakhono ayisisekelo emodeli bese iklama izinyathelo ezifanele zokuphepha ngokusekelwe kulokho.”
Lo mbiko uphinde wachaza ukuthi i-OpenAI izozilungisa kanjani izindlela zayo zokuphepha esikhathini esizayo, okuyisengezo ezinyathelweni zokuqinisa ukuqashwa nokuqondanisa amamodeli ezamenyezelwa yinkampani ngesonto eledlule.
Phakathi kwazo, i-OpenAI izoqinisa ukuqashwa “kochungechunge lokucabanga” lwama-agent e-AI. Uchungechunge lokucabanga lungaqondwa njengendawo yokusebenza yangaphakathi lapho uhlelo lwe-AI lugcina khona ukusabela kwesikhashana, imigomo yamanje kanye nenqubo yokucabanga.
Lezi zinyathelo zokuqashwa zizohlanganiswa nendlela yokuphendula nokudlulisela izindaba esigabeni esiphezulu esebenza ubusuku nemini. Ngesikhathi esifanayo, i-OpenAI izokwethula amathuluzi amasha okuqeda ngokushesha imithwalo yomsebenzi ethathwa njengeyingozi kwezokuphepha.
I-OpenAI ithe: “Lezi zinguquko zihloselwe ukuthuthukisa kokubili ububanzi bokutholwa kanye nesivinini sokuphendula. Zigxile kokungajwayelekile kwengqalasizinda nasekuziphatheni kwemodeli okungase kukhathaze, futhi zinikeza indlela yokunqanda ngokushesha lapho kutholakala inkinga.”
Le nkampani iphinde yathi, ukube uhlelo lokuqapha uchungechunge lokucabanga olusakhelwe manje lwalusebenza ngaleso sikhathi, lwalungathola imisebenzi yokuqala engajwayelekile ehambisana nalesi sigameko futhi luxwayise ithimba lezokuphepha le-OpenAI okungenani ngosuku olulodwa ngaphambi kokuba imodeli ingene ezinhlelweni ze-Hugging Face.