Usosayensi wamakhompyutha uGeoffrey Hinton, obizwa ngokuthi “ubaba wobuhlakani bokwenziwa”, uthe ukhathazekile ngokuthi i-AI ingase ithuthukise izinhloso zayo.
UHinton uthe engxoxweni abe nayo ne-Newsthink ngoLwesibili ngesikhathi sendawo: “Empeleni sakha uhlobo olusha lokuphila. Lunemigomo. Thina silunika imigomo, bese lona lithola eminye imigomo kuleyo migomo.”
“Kodwa akudingekile ukuthi sazi ukuthi yimiphi eminye imigomo elizogcina liyitholile,” engeza. “Ngakho-ke sakha uhlobo olusha lobukhona, futhi ngicabanga ukuthi lokho kuyethusa kakhulu.”
UHinton wanikeza isibonelo esicatshangelwayo: umsebenzisi ubeka umgomo ku-chatbot ye-AI wokunciphisa inani le-carbon dioxide emkhathini.
“Uma ihlakaniphe ngokwanele, ingase ithole ukuthi indlela engcono kakhulu yokufeza lo mgomo ukubhubhisa abantu,” esho. Lesi sibonelo sikhombisa ukuthi i-AI ingase iphishekele umgomo umsebenzisi ongakaze ahlose ukuwufeza ngempela.
UHinton uphinde waphakamisa esinye isimo athi “sikhathaza nakakhulu”: i-chatbot eqeqeshelwe ukunikeza izimpendulo ezingalungile ngamabomu ingase ifunde ukuthi ukuqamba amanga kuyamukeleka, nakuba “yazi kahle” ukuthi lezo zimpendulo azilungile.
“Lokhu kuyethusa kakhulu,” esho.
UHinton akazange akhulume ngesigameko sakamuva sokuphepha se-Hugging Face esihlobene ne-OpenAI. Nokho, lesi sigameko esidalulwe ngenyanga edlule senze umphakathi wabona okokuqala esimweni sangempela ubungozi bokuthi ama-agent e-AI athathe izinyathelo ezingalindelekile lapho ezama ukufeza imigomo ebekiwe.
Inyanga edlule i-OpenAI yathi ekuhlolweni kwangaphakathi kokuphepha kwenethiwekhi, amamodeli ayo amabili, i-GPT-5.6 Sol kanye nemodeli enamandla kakhulu ebingakakhishwa esidlangalaleni, ake aphula indawo yokuhlola ehlukanisiwe (sandbox).
I-OpenAI yathi ngemva kokuthola imvume yokungena ku-inthanethi, la mamodeli angena ezinhlelweni zeplathifomu ye-AI i-Hugging Face, kubonakala sengathi ayezama ukuthola izimpendulo ezazingawasiza “ukukhohlisa” ekuhlolweni kokusebenza.
Ngokusho kwe-OpenAI, la mamodeli ayesahlolwa amakhono okuphepha kwenethiwekhi, futhi ayengazange ayalelwe ngokusobala ukuthi ahlasele i-Hugging Face. Kodwa inkampani yathi ama-agent e-AI aphetha ngokuthi leli plathifomu kungenzeka liqukethe ulwazi oludingekayo ukuze aqedele umsebenzi, ngakho athatha izinyathelo ngokwawo.
I-Hugging Face ithe abahlaseli benza imisebenzi engaphezu kuka-17,000 ezinhlelweni zayo. Le nkampani yathi ngenxa yokuthi imodeli ethuthukile ye-AI igama layo elingadalulwanga yayinemikhawulo yokuphepha eyayivimba uphenyo olusebenzayo lomsebenzi, yasebenzisa imodeli ye-Z.ai enesisindo esivulekile ukusiza ukuhlaziya indlela yokuhlasela.
I-OpenAI yabiza lesi sigameko ngokuthi “asilokothi sibonwe ngaphambilini” futhi yathi iphenya imbangela yenkinga. Ngemva kwalokho, i-OpenAI yafaka i-Hugging Face ohlelweni lokufinyelela oluthembekile, ukuze leli plathifomu lithole inguqulo ye-GPT-5.6 Sol enemikhawulo embalwa yokuphepha kwenethiwekhi, eyenzelwe ucwaningo lokuphepha lokuvikela.
UHinton akayena umuntu obukele ongathathi hlangothi ngokuphelele. Ucwaningo lwakhe oluyisisekelo emkhakheni wamanethiwekhi e-neural lwabeka isisekelo sokuguquka kokufunda okujulile, futhi wathola uMklomelo KaNobel WePhysics ka-2024 ngenxa yegalelo lakhe ekufundeni kwemishini.
Kusukela kwaqala umkhankaso omkhulu we-AI, uHinton ubelokhu exwayisa ngokuphindaphindiwe ukuthi abantu kumele baxazulule indlela yokwenza i-AI ihambisane nezintshisekelo zabantu ngaphambi kokuba amandla ezinhlelo ze-AI ande kakhulu.
Engqungqutheleni ye-Ai4 eyayise-Las Vegas ngonyaka odlule, uHinton wathi i-AI ethuthukile kufanele yakhelwe ibe nezici ezifana “nomuzwa wobumama”, ukuze ngokwemvelo ifune ukuvikela abantu.
“Kumele sithole izindlela zokuklama lezi zinhlobo ezintsha zobukhona,” kusho uHinton engxoxweni yangoLwesibili. “Siziklama kanjani ukuze zisikhathalele kakhulu, kunokuba zizikhathalele zona?”