Masu binciken tsaro daga ƙungiyar Bincike ta Wiz sun gano bayanan da ake samu a bainar jama'a na bayanan da aka yi amfani da su a cikin ayyukan DeepSeek's AI. Saboda rashin ƙayyadaddun ƙuntatawa ga ma'ajiyar log ɗin, kowa zai iya samun bayanan sirri game da ayyukan DeepSeek. Ma'ajiyar bayanai ta ƙunshi fiye da miliyan guda, ciki har da rajistan ayyukan tare da tarihin saƙonnin mai amfani a cikin taɗi na DeepSeek AI, maɓallan samun damar API, cikakkun bayanai game da aiki na baya, da metadata da aka yi amfani da su wajen gudanar da ayyuka daban-daban.
Yayin da ake nazarin ƙananan wuraren samun damar jama'a na deepseek.com, masu binciken sun lura da runduna ooauth2callback.deepseek.com da dev.deepseek.com, akan tashoshin sadarwa na 9000 da 8123 wanda akwai sabis na ajiya dangane da ClickHouse DBMS. An yi amfani da tashar tashar sadarwa ta 9000 don haɗa aikace-aikace, kuma tashar jiragen ruwa 8123 ta ba da hanyar haɗin yanar gizo wanda ya ba da damar aika duk wata tambaya ta SQL.
Saitunan DBMS da aka fallasa sun ba da cikakken iko akan ayyuka a cikin bayanan, lokacin samun dama ba tare da tantancewa ba. A cewar masu binciken, damar da ake da ita ta isa don tsara harin da ba'a iyakance ga DBMS ba kuma an ba da damar samun dama ga abubuwan more rayuwa na DeepSeek.
Ka tuna cewa a makon da ya gabata, DeepSeek ya fito a ƙarƙashin lasisin MIT manyan nau'ikan harshe DeepSeek-R1 da DeepSeek-R1-Zero, wanda ke rufe sigogin biliyan 671. Ana ɗaukar DeepSeek-R1 a matsayin mafi girma kuma mafi girman ingancin ƙima na buɗaɗɗen tushe. A cikin gwaje-gwaje 12, ƙirar ta fi ƙarfin ƙirar mallakar ta Claude-3.5-Sonnet, OpenAI GPT-4o, da OpenAI o1, kuma a cikin gwaje-gwaje 9 ya nuna sakamako iri ɗaya.
DeepSeek-R1 ingantaccen sigar ƙirar harshen DeepSeek-V3 ne, wanda aka buga a watan Disamba 2024 ƙarƙashin lasisin MIT. Tsarin DeepSeek-V3, kamar DeepSeek-R1, yana rufe sigogin biliyan 671, yana la'akari da mahallin alamun 128 dubu, kuma yana kusa ko mafi kyau fiye da Claude-3.5-Sonnet da GPT-4o a cikin halayensa. DeepSeek-V3 da DeepSeek-R1 na iya aiki akan kayan masarufi na asali ta amfani da tsarin aiwatar da ƙirar harshe gama gari kamar vLLM, TensorRT-LLM, LMDeploy, da SGLang. Samfuran sun dace don ƙirƙirar tsarin tattaunawa, mataimakan kama-da-wane, samar da rubutu, samar da amsoshin tambayoyi, taƙaitawa da haɓaka abun ciki, bayyana ainihin ra'ayoyi da sharuɗɗan.
Baya ga DeepSeek-R1 da DeepSeek-V3, kamfanin kuma yana haɓaka samfuran Janus masu yawa na buɗe ido waɗanda zasu iya sarrafawa, fahimta, da samar da bayanan hoto da sauti. Sabon samfurin Janus Pro yana kusa da aiki zuwa OpenAI's DALL-E 3 lokacin warware matsalolin samar da hotuna daga kwatancen rubutu.
An bayyana cewa an yi amfani da katunan zane na 3 NVIDIA TESLA H2048 don horar da samfurin DeepSeek-V800, kuma farashin horon ya kai dala miliyan 5.58. Don kwatanta, an kiyasta farashin horar da samfurin GPT-4 akan dala miliyan 80-100. Wannan bambance-bambancen ya sa farashin hannayen jari na kamfanoni masu alaƙa da AI da yawa sun faɗi (farashin hannun jari na NVIDIA, alal misali, ya faɗi 17%) kuma ya haifar da hasashe cewa DeepSeek yana ɓoye ainihin farashin sa. OpenAI ya ambata cewa ya sami shaidar cewa an yi amfani da samfuran OpenAI na mallakar mallaka wajen horar da ƙirar DeepSeek, wanda ya keta manufofin sabis na OpenAI waɗanda ke hana amfani da fitowar samfurin OpenAI a cikin samfuran haɓaka samfuran da ke gasa tare da OpenAI (musamman, an taɓa shigar da OpenAI don amfani da bayanan mara izini a cikin ƙirar horo).
source: budenet.ru
