- I-DeepSeek R1 iyimodeli ye-AI yomthombo ovulekile yaseShayina enikeza ukusebenza okuncintisanayo ngokumelene nabaholi abanjengo-OpenAI kanye ne-NVIDIA.
- Yakhiwe ngamaRandi ayizigidi ezingu-5,6 futhi yaqeqeshwa ezinyangeni ezimbili, isebenzisa i-'Expert Mix' ukuze ilungiselele ihadiwe elinganiselwe.
- Ukwethulwa kwayo kubangele ukwehla kwemakethe yamasheya ezinkampanini zobuchwepheshe futhi kungabaza isidingo sokutshalwa kwezimali kwezigidi eziningi zamaRandi kwihadiwe.
- Ukunikezwa kwelayisense kwe-MIT kanye nokuvuleleka kukhuthaza ukubambisana komhlaba wonke, kudala ukufinyelela kwentando yeningi, futhi kuphakamise imiphumela yezepolitiki mayelana nezijeziso kanye nobuholi bezobuchwepheshe.
I-DeepSeek R1 ivele emhlabeni jikelele njengemodeli yobuhlakani bokwenziwa eguqula umkhakha wezobuchwepheshe kanye nezimakethe zezezimali. Le ntuthuko, eqhutshwa yinkampani entsha yaseShayina i -DeepSeek , ihlanganisa ukusungula izinto ezintsha kanye nokufinyeleleka, izibeka njengenye indlela efanelekile esikhundleni sezinkunzi ezinkulu ezifana ne-OpenAI ne-NVIDIA. Okugqamile ngale modeli akukhona nje ukusebenza kwayo kahle kwezobuchwepheshe, kodwa futhi nendlela yayo yomthombo ovulekile kanye nekhono layo lokuletha ukusebenza okufana nabaholi bemakethe ngentengo encane kunezindleko ezivamile.
Njengoba inesabelomali esingu- $5,6 million kuphela futhi iqeqeshwe ezinyangeni ezimbili kuphela, i-DeepSeek R1 ikhombisile ukuthi intuthuko yezobuchwepheshe ayidingi njalo ukutshalwa kwezimali okukhulu. Lo modeli usebenzisa amasu anjenge- "Expert Mixing" ukuze uthuthukise ukusetshenziswa kwehadiwe , ukufeza ukusebenza okuphezulu ngisho nangezinsizakusebenza ezilinganiselwe. Endaweni ephawulwe yimikhawulo yezohwebo kanye nezijeziso ezikhawulela ukufinyelela kuma-chip athuthukile, impumelelo ye-DeepSeek R1 igcizelela ukubaluleka kokusebenza kahle kanye nesu ngaphezu kwesilinganiso sezinsizakusebenza.
Umthelela we-DeepSeek R1 ezimakethe zezimali
Ukwethulwa kwe-DeepSeek R1 akugcinanga nje ngokudonsela ukunaka komphakathi wezobuchwepheshe kodwa futhi kuthumele amagagasi okushaqisayo ezimakethe zezimali. Izinkampani ezifana ne-NVIDIA ziye zabhekana nokwehla okungaphezu kuka- 10% enanini lamasheya azo, okwehlisa ezinye izinkampani zaseNtshonalanga embonini ye-AI. Inkomba ye-Nasdaq, okuyisilinganiso sobuchwepheshe esihamba phambili, nayo ibone ukwehla okukhulu, okuthinta imikhakha ebalulekile kokubili e-Asia naseYurophu.
Lo modeli ubuza ngokungangabazeki ukusimama kokutshalwa kwezimali kwezigidigidi zamaRandi kwihadiwe eyenziwe yizikhondlakhondla ezifana ne-OpenAI ne -Microsoft . Ngenkathi i-OpenAI itshale imali engaphezu kwamaRandi ayizigidigidi ezi-5 ngo-2024, i-DeepSeek ithole imiphumela efanayo ngesabelomali esincane kakhulu. Lokhu kuholele abatshalizimali ekubuzeni isidingo sezindleko eziphakeme kangaka, okungaba uphawu lokuguquka endleleni ubuhlakani bokwenziwa obuthuthukiswa futhi buxhaswe ngayo esikhathini esizayo.
Uguquko lomthombo ovulekile
I-DeepSeek R1 iyazihlukanisa ngokusobala nabancintisani bayo baseNtshonalanga ngenxa yendlela yayo evulekile . Njengoba inelayisensi ngaphansi kwelayisensi ye-MIT, le modeli ivumela abacwaningi nabathuthukisi emhlabeni wonke ukufinyelela, ukuguqula, nokuphinda basebenzise isakhiwo sayo, okukhuthaza ukubambisana komhlaba wonke okungakaze kubonwe ngaphambili emkhakheni we-AI.
Le ndlela ayigcini nje ngokuvumela ukufinyelela kubuchwepheshe obuthuthukisiwe kodwa futhi yehlisa izithiyo zamabhizinisi amancane kanye nabathuthukisi ngabanye, futhi iletha izingozi nezinsongo ze-AI phambili . Ngaphezu kwalokho, iqokomisa umkhuba okhulayo wokubonisa obala emkhakheni ophethwe ngokomlando ngamamodeli avaliwe futhi anelungelo lobunikazi eliphakeme. Ochwepheshe abanjengoYann LeCun bagcizelele ukuthi lokhu kuvuleleka kungamelela kanjani ikusasa lobuhlakani bokwenziwa, kusuke kumamodeli azimele ukuze kuthuthukiswe intuthuko ebandakanya wonke umuntu futhi eqhubekayo.
Izici Zobuchwepheshe ze-DeepSeek R1
Imodeli ye-DeepSeek R1 isebenzisa amasu anjengokufunda okuqiniswayo kanye nokwenza ngcono ihadiwe kusetshenziswa ama-GPU alinganiselwe, njenge-NVIDIA H800, eyenzelwe ngqo imakethe yaseShayina. Le mikhuba iyivumela ukuthi ifinyelele imiphumela efana namamodeli aseMelika afana ne-OpenAI, kodwa ngentengo encane kakhulu. Izici zayo eziyinhloko zifaka:
- I-Scalability nokusebenza kahle: Ingaqeqeshwa ngedatha nezinsiza ezincane, inciphise izindleko kakhulu.
- Imodeli efinyelelekayo: Isakhiwo sayo somthombo ovulekile siyenza itholakale kunoma yimuphi umsebenzisi onentshisekelo noma inkampani.
- Intengo yokuncintisana: Cishe 10 amasenti ngesigidi samathokheni, uma kuqhathaniswa ne-$15 ye-OpenAI.
Imithelela yezepolotiki neyesikhathi esizayo
Impumelelo ye-DeepSeek R1 nayo inemiphumela ebalulekile kwezepolitiki . Ikhono le-China lokuthuthukisa imodeli yokuncintisana enjalo isebenzisa ihadiwe elinganiselwe libonisa ukuthi izijeziso zase-US azizange zisebenze kahle njengoba bekulindelwe. Eqinisweni, le mikhawulo ibonakala sengathi isebenze njengesisusa sokusungula izinto ezintsha zaseShayina, iphoqa izinkampani ezintsha ukuthi zifune izindlela ezisebenza kahle futhi ezingabizi kakhulu.
Ngakolunye uhlangothi, ithonya elikhulayo le-DeepSeek liletha izinselelo ekubuseni kwezobuchwepheshe baseNtshonalanga emkhakheni wobuhlakani bokwenziwa. Amamodeli asebenza kahle njenge-R1 angaba yizinto ezishintsha umdlalo, okuphoqa izinkampani zaseNtshonalanga ukuthi zicabange kabusha amasu azo kanye nobukhosi bezobuchwepheshe ukuze zihlale zincintisana.
I-DeepSeek R1 ayiyona nje imodeli ye-AI; Kuwuphawu lokuthi izinto ezintsha zingaphumelela kanjani nangaphansi kwezimo ezimbi. Indlela yayo efinyelelekayo nephumelelayo ichaza kabusha ikusasa lobuhlakani bokwenziwa, ingenisa inkathi entsha lapho ubuchwepheshe buba phakathi kwabantu futhi bube nentando yeningi.
