Ngomhla ka-25 Agasti, ngokombiko we-Wall Street Journal, amamodeli e-AI aseChina anciphise ngokushesha eminyakeni yamuva umehluko phakathi kwawo namamodeli ahamba phambili aseMelika. Lokhu akubangelwanga impumelelo yezobuchwepheshe evele ngokuzumayo esikhathini esifushane, kodwa kuwumphumela wokubambisana kwenethiwekhi yamakhono eyakhiwe isikhathi eside emanyuvesi afana neNyuvesi yaseTsinghua, ubuchwepheshe obuvulekile, ukubuyela kwamakhono ezweni kanye nokusebenza kahle okukhulu kwamandla ekhompyutha. Lo mbiko uthi iqembu labacwaningi elimelelwa uTang Jie, umsunguli obambisene we-Zhipu AI, u-Yang Zhilin, umsunguli we-Moonshot AI, kanye no-Liang Wenfeng, umsunguli we-DeepSeek, lakha amandla abalulekile ekulandeleni kwe-AI yaseChina kulo mjikelezo.
I-Wall Street Journal ithe abantu abaningi bemboni eChina naseMelika bakholelwa ukuthi umehluko wamakhono phakathi kwamamodeli e-AI angcono kakhulu aseChina namamodeli athuthuke kakhulu aseMelika manje wehle waba yizinyanga ezimbalwa. NgoJuni kulo nyaka, u-Elon Musk wabikezela ukuthi iChina izofinyelela amamodeli aphezulu e-Anthropic kuphela engxenyeni yokuqala ka-2027. UTang Jie waphendula wathi, “Ngeke kuthathe isikhathi eside kangako.”
Lo mbiko ulandelela ukwakheka kohlelo lwamakhono e-AI laseChina emuva eminyakeni engaphezu kuka-20. Ngo-2005, uYao Qizhi, owawina umklomelo we-Turing, wasungula i-“Yao Class” eNyuvesi yaseTsinghua ukuze aqeqeshe amakhono aphezulu emkhakheni wekhompyutha. Ngaleso sikhathi, uTang Jie waqhubeka nocwaningo lokumbiwa kwedatha nokufunda kwemishini eTsinghua, futhi ilabhorethri yakhe kamuva yaba ngomunye wemithombo ebalulekile yamakhono e-AI eChina. U-Yang Zhilin wake wafunda ngaphansi kukaTang Jie, wacwaninga ama-algorithm okufunda kwemishini eTsinghua, wabe eseya eCarnegie Mellon University ukuyokwenza izifundo zobudokotela. Ngemva kokubuyela eChina, wasungula i-Moonshot AI.
Ukuhamba kwabasebenzi phakathi kwamanyuvesi nezinkampani eziqalayo kwancipha kancane kancane kwaba inethiwekhi yamakhono e-AI. Ngo-2018, iChina yaphinde yathambisa imingcele yokuthi abacwaningi baqale amabhizinisi futhi benze imiphumela yocwaningo ibe ezentengiselwano. Ngonyaka olandelayo, uTang Jie wakhipha enkulisa yelabhorethri yaseTsinghua inkampani kamuva eyathuthuka yaba yi-Zhipu AI. Ngemva nje kokuba i-OpenAI ikhiphe i-GPT-3 ngo-2020, i-Zhipu yayisibeke ukuthuthukisa imodeli yezinga elifanayo njengomgomo.
Nokho, uma kuqhathaniswa nezinkampani ze-AI zaseMelika, izinkampani zaseChina isikhathi eside bezibhekene nokushoda kwezimali nama-chip athuthukile. UJefferies walinganisela ukuthi ngaleso sikhathi izinga lokutshalwa kwezimali kwezinkampani zobuchwepheshe zaseChina lalingaphansi kwengxenye yesihlanu yalabo ababencintisana nabo eMelika. Kulesi simo, ukuthuthukisa ukusebenza kahle kokusetshenziswa kwamandla ekhompyutha kwaba yindlela ebalulekile yezinkampani ze-AI zaseChina yokulandelela.
I-DeepSeek iyisibonelo salokhu. Lo mbiko uveze ukuthi i-DeepSeek yehlisa ukusetshenziswa kwememori ngesikhathi imodeli isebenza ngokusebenzisa indlela ye-multi-head latent attention (MLA), futhi yaba phakathi kwabokuqala eChina ukusebenzisa imodeli ye-mixture of experts (MoE), ukuze kuncishiswe isidingo samandla e-chip. Lobu buchwepheshe baba yisisekelo esibalulekile “sokwethuka kwe-DeepSeek” ekuqaleni kuka-2025. Ngaleso sikhathi, imodeli yayo evulekile enezindleko eziphansi yabangela ukuntengantenga kwamasheya ezobuchwepheshe aseMelika, futhi yaphoqa nabanye ababencintisana nabo eChina ukuba baphinde bahlole izindlela zabo zobuchwepheshe.
Ukuxhumana kobuchwepheshe phakathi kwezinkampani ze-AI zaseChina nakho kwashesha kakhulu ijubane lokuthuthuka okuphindaphindwayo. Amamodeli alandelayo e-Moonshot AI asebenzisa i-MoE kanye nokwehluka kwe-MLA okwakusetshenziswe noma kwaqinisekiswa yi-DeepSeek, kuyilapho i-DeepSeek nayo yasebenzisa ubuchwepheshe bokuqeqesha obathuthukiswa yi-Moonshot AI. Amaqembu ahlukahlukene afunda komunye nomunye ngamaphepha ocwaningo, amamodeli avulekile kanye nezinqubo zobuchwepheshe, akha indlela yokusabalala kobuchwepheshe ehlukile kunendlela yamamodeli avaliwe.
Ngesikhathi esifanayo, ukubuyela kwamakhono ezweni kwaba enye indlela yezobuchwepheshe. I-Tencent yaqasha abaphothule eTsinghua ababili abake basebenza e-OpenAI, kanti i-ByteDance yaletha abacwaningi baseChina abake basebenzela i-Google. Lo mbiko ukholelwa ukuthi intuthuko yezobuchwepheshe yezinkampani ze-AI zaseMelika, ukubuyela kwamakhono aphesheya kanye nenethiwekhi yocwaningo yaseChina, konke kwasheshisa ijubane izinkampani ze-AI zaseChina ezilandela ngalo.
Nokho, amandla ekhompyutha aseyisithiyo esikhulu izinkampani ze-AI zaseChina ezibhekene naso. NgoMeyi kulo nyaka, uLiang Wenfeng wathi ama-chip, hhayi amakhono, yiwona mehluko obaluleke kakhulu phakathi kwezinkampani ze-AI zaseChina naseMelika. Abacwaningi bamalabhorethri ahamba phambili afana ne-Alibaba ne-Zhipu bathi inani lama-chip asezingeni eliphezulu abawatholayo livame ukuba yingxenye yesihlanu kuphela yalawo atholwa ozakwabo e-OpenAI nase-Google.