Tun lokacin da DeepSeek ta fitar da samfurinta na tunani mai zurfi, DeepSeek-R1, a watan Janairun da ya gabata, babu wani sabon sabuntawa da aka samu. Jita-jitar cewa sabuwar samfurin DeepSeek AI wani lokaci tana haifar da hayaniya a yanar gizo, amma da alama wata mai zuwa za a fitar da samfurin DeepSeek-V4 mai yawan modal, wanda zai nuna ci gaba mai mahimmanci fiye da sigar da ta gabata.
A cewar majiyar, tsawon watanni shida da suka gabata, ƙungiyar haɓaka DeepSeek, ƙarƙashin jagorancin Liang Wenfeng, ta yi aiki don magance gazawar DeepSeek a fannin sarrafa abubuwan gani da kuma inganta ƙwarewar binciken AI. Kamfanin ya kuma yi ƙoƙarin haɓaka ƙwarewar ƙirƙirar lambar ƙirar AI tare da faɗaɗa taga mahallinsa. Don cimma waɗannan manufofin, DeepSeek ya fara haɗin gwiwa da Baidu a bara.
Masu amfani da dandamali na ƙwararru a faɗin duniya suna ta fafutukar gano alamun sabuwar sigar DeepSeek. Kwanaki kaɗan da suka gabata, an fitar da algorithms na Healer Alpha da Hunter Alpha akan OpenRouter, babban mai haɗa API don samfuran AI. Healer Alpha samfurin harshe ne mai yawa wanda ke iya fahimtar bayanan gani da na ji, yin tunani tsakanin hanyoyin, da kuma aiwatar da ayyuka da yawa tare da babban daidaito. Tagar mahallin algorithm shine alamun 260 kawai. An tsara Hunter Alpha musamman don aikace-aikacen da suka dogara da wakili. Yana da tiriliyan sigogi da taga mahallin alamu miliyan 1. Bayanin ya bayyana cewa algorithm yana da kyau a cikin tsari na dogon lokaci, tunani mai rikitarwa, da kuma ci gaba da aiwatar da ayyuka da yawa. Yana iya bin umarni daidai, wanda yake da mahimmanci lokacin aiki tare da tsarin kamar OpenClaw, wanda ke ba ku damar ƙirƙirar wakilan AI.
Bayan fitowar waɗannan samfuran harsuna guda biyu, tattaunawa a shafin sada zumunta na X game da fitowar DeepSeek-V4 da ke tafe ta sake tashi. Duk da haka, idan aka yi la'akari da sanarwar da aka yi a baya game da DeepSeek-V4, samfurin yana da dubban ma'auni, taga mahallin alamun miliyan 1, da kuma ikon fahimta da samar da bayanai masu yawa, wato, sarrafawa da ƙirƙirar rubutu, hotuna, da bidiyo. Wannan yana nufin cewa halayen algorithms da aka saki kwanan nan akan OpenRouter ba su cika daidai da sigogin da ake tsammani na DeepSeek-V4 ba.
A cewar majiyar, alkiblar ci gaban sigar DeepSeek ta gaba ta mayar da hankali ne kan inganta ƙwaƙwalwar ajiya ta dogon lokaci, wacce aka yi la'akari da ita a matsayin ɗaya daga cikin mahimman halaye na samfuran harshe. A cikin watanni shida da suka gabata, Liang Wenfeng ya rubuta takardu uku na bincike tare, ciki har da waɗanda suka shafi bincika yuwuwar faɗaɗa ƙwaƙwalwar ajiya ta dogon lokaci na samfuran harshe.
Sakamakon binciken Wenfeng da tawagarsa sun kuma nuna wata hanya bayyananna ta juyin halittar fasaha. Bayan sun kafa tsarin ilmantarwa mai ƙarfafawa don tunani a cikin DeepSeek-R1, masu haɓaka suna binciken ƙarin sabbin abubuwa a cikin tsarin gine-ginen da ke ƙasa. Musamman, ta hanyar sabbin kayayyaki kamar "ƙwaƙwalwar sharaɗi," suna ƙoƙarin inganta aikin algorithm ta hanyar magance matsalolin da aka sani game da gine-ginen gargajiya dangane da ƙwaƙwalwa da ƙarfin kwamfuta. Wannan aikin kuma yana aiki azaman shirye-shiryen fasaha don ƙaddamar da DeepSeek-V4. Bugu da ƙari, sabon algorithm zai dace da masu haɓaka AI na China kuma zai iya zama samfurin AI na farko da ke aiki gaba ɗaya a cikin "tsarin tsarin kwamfuta na gida."
Ana sa ran watan Afrilu zai ga gagarumin ci gaba a kasuwar samfurin AI ta China. Baya ga sabon sigar DeepSeek, ana sa ran Tencent zai ƙaddamar da wani samfurin AI mai sigogi biliyan 30.
source:
source: 3dnews.ru
