I-DeepSeek-R1: Vula i-AI yokuqiqa iguqula imeko yetekhnoloji

EPortada » Intelligence Intelligence » I-DeepSeek-R1: Vula i-AI yokuqiqa iguqula imeko yetekhnoloji

I-DeepSeek-R1, imodeli yokuqiqa yengqondo eyenziweyo eyenziwe yilabhoratri yaseTshayina DeepSeek, inika into eninzi yokuthetha ngayo kwihlabathi lobugcisa. Lo mzekelo, odibanisa ukufikeleleka enkosi yakho MIT ilayisensi Ngomsebenzi ogqwesileyo kwiimvavanyo ezininzi eziphambili, ithembisa ukuba sesinye sezona zixhobo ziphazamisayo ngaphakathi kwe-ecosystem. vula i-AI.

Ukuqaliswa kwe-DeepSeek-R1 kubonisa inzuzo ebalulekileyo kuphuhliso lwaseTshayina kwicandelo lobuchwepheshe elilawulwa ziinkampani zaseNtshona. Ngokulingana kunye nokugqitha kuchaneka Xa kuthelekiswa neemodeli ezifana ne-OpenAI o1, i-DeepSeek-R1 ayibonisi nje amandla okwenza izinto ezintsha zabadali bayo, kodwa izisa etafileni umnikelo ofikelelekayo nofikelelekayo kubo bobabini abaphuhlisi kunye neenkampani.

Imodeli eqinileyo yemathematika, inkqubo kunye nokuqiqa okunengqiqo

Con I-671 yeebhiliyoni zeeparamithaI-DeepSeek-R1 iphakathi kwezona modeli ziphambili ze-AI zehlabathi. Ngokweemvavanyo, le modeli ifumene amanqaku angama-97,3% kwiimviwo ezifana MATH-500, ukodlula i-96,4% efunyenwe yi-OpenAI o1. Esi siganeko someleza isakhono saso soku imisebenzi enzima kwimimandla efana nemathematika, inkqubo kunye nokuqiqa okusengqiqweni, apho ukusebenza kwayo kutsale umdla wabaphuhlisi kunye nezifundiswa.

Imodeli ikwayilwe ngeendlela ezikhaphukhaphu ezaziwa ngokuba iinguqulelo distilled, eyahluka ukusuka 1,5 ibhiliyoni enye de 70 ibhiliyoni enye yeeparamitha. Ezi nguqulelo zifanelekile kubasebenzisi abane izixhobo zekhompyutha amandla angaphantsi, avumela i-DeepSeek-R1 ukuba iqhutywe ekuhlaleni ngaphandle kwesidingo sezixhobo zekhompyutha ezomeleleyo. Umzekelo, inguqulelo I-DeepSeek-R1-Distill inokuqhuba kwilaptop eqhelekileyo.

Enye indlela efikelelekayo nevulelekileyo

Enye yezona zinto zibalaseleyo kwi-DeepSeek-R1 yile inzuzo. Ngelixa i-OpenAI API ihlawulisa IRandi ye7,50 Kuzo zonke izigidi zamathokheni zokufaka, i-DeepSeek ibonelela ngemodeli yayo kancinci IRandi ye0,14 kumthamo ofanayo, ukufezekisa ukunciphisa phakathi kwe-90% kunye ne-95% kwiindleko. Ukongeza, yayo MIT ilayisensi ivumela zombini ukusetyenziswa kwezemfundo kunye nezorhwebo ngaphandle kwezithintelo, uphawu oluxabisekileyo lokuqalisa, iiyunivesithi kunye namashishini amancinci.

Imodeli ephambili kunye neenguqulelo zayo ze-distilled ziyafumaneka kumaqonga afana Ukujongana nobuso, ikwenza kube lula ukukhuphela kunye nokufikelela kubaphuhlisi kwihlabathi liphela. Ukongezelela, inokusetyenziswa njenge-API ukudibanisa ngokuthe ngqo amandla ayo kwizicelo ezahlukeneyo.

Imingeni yolawulo kunye nemiqobo yezelizwe

Ngaphandle kweenzuzo ezininzi, i-DeepSeek-R1 ayikho ngaphandle kwemingeni yayo. Njengomzekelo ophuhliswe e-China, ixhomekeke kwimimiselo eqinisekisa ukuba iimpendulo zayo "uqulathe iinqobo ezisemgangathweni zobusoshiyali”. Oku kuthetha ukuba ayisayi kuphendula imibuzo malunga nezihloko ezinobuntununtunu kwezopolitiko ezifana neTiananmen Square okanye ukuzimela kweTaiwanese, ezinokucothisa ukwamkelwa kwayo kwiimarike zamazwe ngamazwe.

Ukongeza, ukunyuka kwengxwabangxwaba phakathi kweTshayina kunye ne-United States kwicandelo le-AI kukhokelele kwizithintelo ezingqongqo ngurhulumente wase-US, okwenza kube nzima fikelela ukusuka kwiinkampani zaseTshayina ukuya kumacandelo athile abalulekileyo kuphuhliso lobuchwepheshe obuphambili. Nangona kunjalo, ezi zithintelo azikhange ziyimise i-DeepSeek-R1 ekuphumezeni abakhweli baseNtshona kwiibenchmarks ezininzi.

Ubuchule obutsha: Ukomelezwa kokufunda kunye nokubeka iliso

I-DeepSeek-R1 isebenzisa indibaniselwano ye ukufunda okomeleza (RL) ulungelelwaniso olusulungekileyo nolugadiweyo (SFT) ukuphumeza amanqanaba ancomekayo ukusebenza. Le ndlela ivumela imodeli ukuba ilungelelanise izicwangciso zayo zokusombulula iingxaki, ifunde kwiimpazamo zayo, kwaye iphonononge ezinye izisombululo ngobunzulu obukhulu.

Ngokweengxelo zobuchwephesha, ngexesha lezigaba zoqeqesho imodeli ihambe kwiinkqubo eziphindaphindwayo ezibandakanya uninzi lokuvota kwindawo ezilawulwayo, eziphucule kakhulu kuchaneka kwimisebenzi enzima. Umzekelo, uphumelele i-pass@1 amanqaku 86,7% kwiimvavanyo zokuqiqa eziphambili ezifana UMNYAKA 2024.

Isiphumo sale ndlela yimodeli ekwaziyo ukusombulula iingxaki zenzululwazi, zezibalo kunye nezobuchwephesha nge ukungqinelana kunye nokukhawulezisa oko kuyibeka phakathi kweenkokheli zoshishino.

Kwindawo yenkqubo, i-DeepSeek-R1 nayo ibonise ukusebenza kwe-stellar. Ngamanqaku 2,029 KwiCodeforces, iyodlula i 96,3% yabadwelisi benkqubo abangabantu, ukuseka ngokwayo njengesixhobo esisebenzayo sophuhliso lwesoftware ephucukileyo.

Umncedisi wamacandelo ahlukeneyo

Ukuguquguquka kwe-DeepSeek-R1 kuyenza ibe sisisombululo esinomtsalane kumashishini amaninzi. Umzekelo, kwicandelo lezemfundo, iinguqulelo ezidityanisiweyo zinokwenza iilebhu ze-AI kwiidyunivesithi ezinoncedo olulinganiselweyo. Ngokuphathelele amashishini, iimodeli ze-AI ezifana nale ziyavumela Ukunciphisa iindleko ngokwenza uhlalutyo olunzima ngaphandle kokuxhomekeka kumaxabiso aphezulu eenkampani ezinkulu.

Ngapha koko, ukudityaniswa kwayo kunye ne-blockchain kunye neeprojekthi ze-cryptocurrency ziye zaphawuleka ngakumbi. Ndiyabulela ukukwazi ukuhlalutya umthamo omkhulu wedatha kunye nokukhupha iipateni eziluncedo, I-DeepSeek-R1 ithembisa ukuba sisixhobo esiphambili sokuqalisa ukusebenza kunye iimvumelwano ezifanelekileyo kunye nemisebenzi kwi-DeFi (iZimali eziBekelwe aMazwe ngaMazwe).

Ummeli we-DeepSeek uphinde waqinisekisa ukuzinikela kwelebhu ngokuthi: “Injongo yethu kukubonelela ngezisombululo ezifikelelekayo nezivulelekileyo, sivumela abantu ukuba balawule ikamva labo lobuchwephesha.".

Ukuvela kwe-DeepSeek-R1 bubungqina obongezelelweyo bokuthi iimodeli ze-AI ezivulekileyo zivala ngokukhawuleza i-gap kunye neemodeli zorhwebo zexabiso eliphezulu. Ngokugxila kwi ukufikeleleka kunye nokusebenza, le modeli yaseTshayina igqamile njengophawu kuphuhliso lwezixhobo ze-AI ezingenamandla nje kuphela, kodwa zifikeleleka kwaye zisebenza.