I-DeepSeek-R1: I-AI evulekile yokucabanga eguqula isimo sobuchwepheshe

Informatec Digital » Intelligence yokufakelwa » I-DeepSeek-R1: I-AI evulekile yokucabanga eguqula isimo sobuchwepheshe

I-DeepSeek-R1, imodeli yokucabanga yokwenziwa kobuhlakani eyenziwe ilabhorethri yaseShayina I-DeepSeek, inikeza okuningi ukukhuluma ngakho emhlabeni wezobuchwepheshe. Le modeli, ehlanganisa ukufinyeleleka ngiyabonga Ilayisense ye-MIT Ngokusebenza okuphakeme ekuhlolweni okubalulekile okuningana, ithembisa ukuba ngelinye lamathuluzi aphazamisa kakhulu ngaphakathi kwe-ecosystem vula i-AI.

Ukwethulwa kwe-DeepSeek-R1 kumele inzuzo ebalulekile yentuthuko yaseShayina esigabeni esiphethwe yizinkampani zaseNtshonalanga. Ngokulingana ngisho nokudlula phakathi ukunemba Uma kuqhathaniswa namamodeli afana ne-OpenAI o1, i-DeepSeek-R1 ayibonisi nje kuphela amandla okuqamba abadali bayo, kodwa futhi iletha etafuleni umnikelo othengekayo futhi ofinyeleleka kubo bobabili abathuthukisi nezinkampani.

Imodeli eqinile yezibalo, ukuhlela nokucabanga okunengqondo

cunt 671 billion amapharamithaI-DeepSeek-R1 iphakathi kwamamodeli e-AI athuthuke kakhulu emhlabeni. Ngokusho kokuhlolwa, le modeli ithole amaphuzu angu-97,3% ezivivinyweni ezifana IZIMBALA-500, idlula ama-96,4% atholwe yi-OpenAI o1. Lesi senzakalo esiyingqopha-mlando siqinisa ikhono layo loku imisebenzi enzima ezindaweni ezifana nezibalo, izinhlelo kanye nokucabanga okunengqondo, lapho ukusebenza kwayo kudonse ukunaka konjiniyela kanye nezifundiswa.

Imodeli iphinde yaklanywa ngezinketho ezilula ezaziwa ngokuthi izinguqulo distilled, ezihluka ukusuka I-1,5 eyodwa yezigidigidi kuze kufinyelele I-70 eyodwa yezigidigidi kwamapharamitha. Lezi zinguqulo zilungele abasebenzisi abane imishini yehadiwe amandla amancane, okuvumela i-DeepSeek-R1 ukuthi isetshenziswe endaweni ngaphandle kwesidingo sezinsiza eziqinile zekhompuyutha. Ngokwesibonelo, inguqulo I-DeepSeek-R1-Distill ingasebenza kwi-laptop evamile.

Enye indlela ethengekayo nevulekile yomthombo

Okunye okugqamile kwe-DeepSeek-R1 yikho inzuzo. Ngenkathi i-OpenAI API ishaja Amadola ka-7,50 Kuwo wonke amathokheni okufakwayo ayisigidi, i-DeepSeek inikeza imodeli yayo ngemali encane nje Amadola ka-0,14 ngevolumu efanayo, ukuzuza ukwehliswa okuphakathi kuka-90% no-95% wezindleko. Ngaphezu kwalokho, yayo Ilayisense ye-MIT ivumela kokubili ukusetshenziswa kwezemfundo nokuhweba ngaphandle kwemikhawulo, isici esibalulekile sabaqalayo, amanyuvesi namabhizinisi amancane.

Imodeli eyinhloko kanye nezinguqulo zayo ze-distilled ziyatholakala kumapulatifomu afana Ubuso ObumbambayoLokhu kwenza kube lula ukulanda nokufinyelela kwayo konjiniyela emhlabeni wonke. Ngaphezu kwalokho, ingasetshenziswa njenge-API ye ukuhlanganisa ngqo amakhono abo ezinhlelweni ezahlukene.

Izinselelo zokulawula kanye nezingqinamba ze-geopolitical

Naphezu kwezinzuzo zayo eziningi, i-DeepSeek-R1 inezinselele zayo. Njengesibonelo, iyisibonelo esihle kakhulu. yathuthukiswa eShayina, ingaphansi kwemithethonqubo eqinisekisa ukuthi izimpendulo zayo "hlanganisa izimiso eziyisisekelo zobudlela-ndawonye”. Lokhu kusho ukuthi ngeke iphendule imibuzo ngezihloko ezibucayi kwezepolitiki ezifana ne-Tiananmen Square noma ukuzimela kwe-Taiwanese, okungase kubambezele ukwamukelwa kwayo ezimakethe zamazwe ngamazwe.

Ngaphezu kwalokho, ukungezwani okwandayo phakathi kweChina ne-United States emkhakheni we-AI kuholele ekutheni uhulumeni wase-US abeke imingcele eqinile, okwenza kube nzima ukufinyelela kusuka ezinkampanini zaseShayina kuya ezingxenyeni ezithile ezibalulekile zokuthuthukiswa kobuchwepheshe obuphambili. Kodwa-ke, lezi zithiyo azizange zimise i-DeepSeek-R1 ukuthi iphumelele izimbangi zaseNtshonalanga ngamabhentshimakhi amaningi.

Ukuqamba okusha kwezobuchwepheshe: Ukuqinisa ukufunda nokuqondisa

I-DeepSeek-R1 isebenzisa inhlanganisela ye imfundo yokuqinisa (RL) pure and supervised fine tuning (SFT) ukuze kuzuzwe amazinga ahlaba umxhwele ukusebenza. Le ndlela ivumela imodeli ukuthi ivumelanise amasu ayo okuxazulula izinkinga, ifunde emaphutheni ayo, futhi ihlole ezinye izixazululo ngokujula okukhulu.

Ngokwemibiko yezobuchwepheshe, phakathi nezigaba zokuqeqesha imodeli idlule ezinqubweni eziphindaphindayo ezihlanganisa ukuvota okuningi ezindaweni ezilawulwayo, okuthuthuke kakhulu ukunemba emisebenzini enzima. Isibonelo, uthole amaphuzu wokudlula ku-1 we 86,7% ezivivinyweni zokucabanga ezithuthukisiwe ezifana ISIKHATHI se-2024.

Umphumela wale ndlela uyimodeli ekwazi ukuxazulula izinkinga zesayensi, zezibalo nezobuchwepheshe nge ukuvumelana nokusheshisa lokho kukubeka phakathi kwabaholi bemboni.

Emkhakheni wezinhlelo, i-DeepSeek-R1 nayo ibonise ukusebenza kwe-stellar. Ngesikolo se 2,029 Ku-Codeforces, idlula i- 96,3% kusuka kubahleli bezinhlelo abangabantu, okuzibeka njengethuluzi eliphumelelayo lokuthuthukisa isofthiwe ethuthukisiwe kumapulatifomu alungiselelwe Amaprosesa we-AMD.

Umlingani wemikhakha eyahlukene

Ukuguquguquka kwe-DeepSeek-R1 nakho kuyenza ibe yisisombululo esikhangayo ezimbonini eziningi. Isibonelo, emkhakheni wezemfundo, izinguqulo ezihluziwe zingenza kube lula Amalebhu e-AI emanyuvesi anezinsiza ezilinganiselwe. Ngokuphathelene namabhizinisi, amamodeli e-AI afana nalawa avumela Yehlisa izindleko ngokwenza ukuhlaziya okuyinkimbinkimbi ngaphandle kokuncika emananini aphezulu ezinkampani ezinkulu.

Ngaphezu kwalokho, ukuhlanganiswa kwayo namaphrojekthi we-blockchain kanye ne-cryptocurrency kuye kwaphawuleka kakhulu. Ngenxa yekhono layo lokuhlaziya amanani amakhulu wedatha nokukhipha amaphethini awusizo, I-DeepSeek-R1 ithembisa ukuba yithuluzi elibalulekile lokuqalisa ukusebenza nalo izinkontileka smart kanye nokusebenza ku-DeFi (Decentralized Finance).

Omele i-DeepSeek uqinisekise ukuzibophezela kwelebhu ngokuthi: “Umgomo wethu uwukunikeza izixazululo ezifinyelelekayo nezivulekile, okuvumela abantu ukuthi balawule ikusasa labo lobuchwepheshe.".

Ukuvela kwe-DeepSeek-R1 kuwubufakazi obengeziwe bokuthi amamodeli e-AI avulekile avala ngokushesha igebe ngamamodeli okuhweba abiza kakhulu. Ngokugxila ku ukufinyeleleka nokusebenza, le modeli yaseShayina igqama njengebhentshimakhi ekuthuthukisweni kwamathuluzi e-AI angenamandla nje kuphela, kodwa futhi athengekayo futhi asebenzayo.

I-athikili ehlobene:

Izindaba zesitoreji: amanani, ukushoda, kanye nezinto ezintsha ezizayo