- I-DeepSeek-V3-0324 yethula ukuthuthuka okubalulekile ku uhlelo, ukucabanga kwezibalo nokukhiqizwa kwekhodi.
- I-China iyaqhubeka nokuthuthuka kulo mkhakha ukuhlakanipha okufakelwayo, enciphisa igebe ne-United States.
- Imodeli entsha idlula eyandulelayo ekuhlolweni okubalulekile okufana ne-MMLU-Pro, i-AIME 2024, ne-LiveCodeBench.
- Itholakala ngaphansi kwelayisense ye-MIT, evumela ukusetshenziswa okungakhawulelwe kwezohwebo.
I-DeepSeek, inkampani yaseShayina yezobunhloli bokwenziwa, yethule inguqulo entsha yemodeli yayo yolimi, ebizwa ngokuthi I-DeepSeek-V3-0324. Lesi sibuyekezo siletha ukuthuthuka okukhulu uhlelo, ukucutshungulwa kwezibalo y ukukhiqizwa kwekhodi, kanye nendlela yomncintiswano we China emjahweni wobuhlakani bokwenziwa, ukuzuza inqubekelaphambili ewusondeza kakhulu kubambangi bawo baseNtshonalanga.
Le modeli entsha ayilungisi nje kuphela amakhono ayo okuqonda ulimi, kodwa futhi ithuthukise ukucabanga kwayo nokunemba ezivivinyweni zebhentshimakhi ezihlukahlukene. Lokhu kuthuthuka kuholele ekutheni ibe ngenye yamamodeli aziwa kakhulu ezisekelweni ezifana Ubuso Obumbambayo, lapho itholakala konjiniyela nabacwaningi.
Ukweqa ngokunemba nokusebenza nge-DeepSeek-V3-0324
I-DeepSeek isebenze ekuthuthukiseni imodeli yayo yangaphambilini ye-V3, izuze ukunemba okukhulu ekuhlolweni okubalulekile ngesibuyekezo se-V3-0324. Kumbandela I-MMLU-Pro, isilinganiso esilinganisweni sokuhlola imodeli yolimi, sisuke ku-75,9% saya ku-81,2%, okubonisa umehluko obonakalayo Amandla okuphendula yohlelo. Lokhu kuphumelela kugqamisa ukubaluleka kwamamodeli avulekile endaweni yanamuhla yobuhlakani bokwenziwa.
Okunye ukuthuthuka okuphawulekayo kubandakanya ukusebenza kwabo ekuhlolweni kwezibalo ISIKHATHI se-2024, lapho isithuthuke isuka ku-39,6% yaya ku-59,4%. Lokhu kumele ukukhula cishe okungama-20%, ukuthuthuka okubalulekile mayelana nokunemba kulo mkhakha. Esivivinyweni I-LiveCodeBench, ekala ikhono lemodeli loku khiqiza ikhodi ukusebenza, imodeli entsha ikhuphuke amaphuzu ayishumi ukuze ifinyelele ku-49,2%.
Ukuthuthukisa ekukhiqizeni amakhodi nokusebenza ku-AI
Imodeli ye-V3-0324 yethula ukuthuthukiswa kwe umsebenzi call, exazulula izinkinga ezimbalwa zezinguqulo zangaphambilini ngokuya ngokunemba. Lokhu kuhumusha kokukhulu ukuthembeka ekuthuthukisweni kwesoftware kanye nezinhlelo zokusebenza ezizenzakalelayo, ukwenza lula umsebenzi we abahleli bohlelo y onjiniyela. Le mpumelelo ibalulekile kumongo wokukhiqiza ikhodi ehlanzekile.
Ngaphezu kwalokho, isibuyekezo sivumela dala izixhumanisi zewebhu ezikhanga kakhulu futhi uthuthukise ubuhle bangaphambili kwemidlalo. Esinye isici esibalulekile umthamo wayo olungiselelwe ukubhala kabusha okusebenzisanayo kwamajika amaningi, intuthuko ebalulekile yokuthuthukisa ikhwalithi yezimpendulo ku abasizi abangokoqobo y izingxoxo.
Kusukela yethulwa, ongoti abahlukahlukene baye bahlola ukusebenza kwemodeli futhi babike ukuthi isivinini sokubulawa ifinyelela ngaphezu kwamathokheni angu-20 ngomzuzwana ezinhlelweni ezifana Mac Isitudiyo esine-chip ye-M3, eqinisa ukusebenza kahle kwayo.
I-DeepSeek-V3-0324 kanye nesu lobuhlakani bokwenziwa lase-China
Ukwethulwa kwemodeli kuyingxenye ye-a isu elibanzi kusuka eChina ukuze ihlanganise isikhundla sayo njengomholi kwezobuhlakani bokwenziwa futhi inciphise igebe nezinkampani zaseNtshonalanga ezifana I-OpenAI kanye ne-Anthropic. Esinye sezihluthulelo zalokhu kukhula silele ekusetshenzisweni kahle kwezinsiza: I-DeepSeek ikwazile ukuncintisana ezingeni eliphezulu ngokusetshenziswa okuphansi amakhadi wehluzo kanye nezindleko eziphansi zokubala, into eye yadonsa ukunaka kochwepheshe bezimboni.
Imodeli itholakala ngaphansi Ilayisense ye-MIT, evumela ukusetshenziswa kwayo ngaphandle kwemikhawulo yezohwebo, ngaleyo ndlela ikhulise ukufinyeleleka kwayo kanye namandla okutholwa ezimbonini ezahlukene. Lokhu kuyayihlukanisa kumamodeli avaliwe njenge GPT-4.5, enikeza enye indlela enamandla nevulekile ye onjiniyela y izinkampani zobuchwepheshe. Lokhu kuvulwa kuyisisekelo ekuziphendukeleni kwemvelo IA esifundeni
Imodeli eqhubekayo
Nakuba inqubekelaphambili eyenziwe nge-DeepSeek-V3-0324 ibalulekile, inkampani isasebenzela ukuthuthukisa uhlelo lwayo. Kunamahemuhemu okuthi inguqulo entsha, okungenzeka ukuthi ibizwa I-DeepSeek R2, ingase yethulwe ezinyangeni ezizayo, ezobonisa ukuguquguquka okuqhubekayo kwamakhono ayo.
Ochwepheshe bezobunhloli bokwenziwa bagqamise ukuthi le modeli izuze ukunemba okukhulu ukubhala imibhalo emide nobude obumaphakathi, ngaphezu kokuthuthukiswa kwe ikhwalithi yokuhumusha y ukubhala. Ukubusa kwakhe kwe Ulimi Chinese Iphinde yathuthukiswa, yazihlanganisa nesitayela sohlobo lwe-R1, okunye okugqama kakhulu kwenkampani.
Ukuthuthuka kwe-DeepSeek kumelela inselelo ezinkampanini zaseNtshonalanga, okuzodingeka ziqhubeke nokusungula izinto ezintsha ukuze zigcine ubuholi bazo kulo mkhakha. Ngekhono layo lokukhiqiza ikhodi yekhwalithi kanye nokwenza kahle kwayo ekucabangeni isazi sezibalo, le modeli ibekwe njengethuluzi elibalulekile esikhathini esizayo sobuhlakani bokwenziwa.
Ukuthuthukiswa kwalezi zinhlobo zamamodeli nakho kuyingxenye yomongo we-geopolitical wokuncintisana okukhulu kwezobuchwepheshe. Ngenkathi i-United States igcina imikhawulo ekuthengisweni kwamanye amazwe kwe hardware kuthuthukile maqondana neChina, izinkampani zaseShayina ziyaqhubeka nokuthola izindlela zokuthuthuka nokuncintisana kulo mkhakha. Lokhu kwethulwa akubonisi nje intuthuko ngokobuchwepheshe, kodwa futhi kusabela ku-a isidingo samasu lokuzibusa kubuhlakani bokwenziwa.
Ukuziphendukela kwemvelo okuqhubekayo kwe-DeepSeek kukhombisa ukuthi i-China yenza inqubekelaphambili ebalulekile ekuthuthukiseni amamodeli e-AI avulekile futhi ancintisanayo. Ngokuphindaphinda ngakunye, igebe neNtshonalanga liyancipha, okubonisa ukuncintisana okukhulu nakakhulu eminyakeni ezayo.
Umbhali oshisekayo ngomhlaba wamabhayithi nobuchwepheshe ngokujwayelekile. Ngiyathanda ukwabelana ngolwazi lwami ngokubhala, futhi yilokho engizokwenza kule bhulogi, ngikubonise zonke izinto ezithakazelisayo kakhulu ngamagajethi, isofthiwe, ihadiwe, izitayela zobuchwepheshe, nokuningi. Inhloso yami ukukusiza ukuthi uzulazule emhlabeni wedijithali ngendlela elula nejabulisayo.