Informatec Digital » Intelligence yokufakelwa » I-INTELLECT-2: Imodeli ye-AI esabalalisiwe nevuliwe eshintsha ubuhlakani bokwenziwa

Ukuhlola I-INTELLECT-2: I-AI Ebambisene futhi Emisiwe Iguqulela Imboni

Namuhla, ubuhlakani bokwenziwa buthuthuka ngesivinini esikhulu, futhi amamodeli amasha nobuchwepheshe obuthembisa ukuguqula umkhakha buvela ngaso sonke isikhathi. Phakathi kwalezi zinto ezintsha, i-INTELLECT-2 ivelele , iphrojekthi edala umsindo omkhulu, kokubili ngendlela yayo entsha kanye nekhono layo lokuba yisishintshi ekuqeqeshweni nasekusetshenzisweni kwamamodeli olimi. Uma uke wezwa nge-INTELLECT-2 kodwa ungaqiniseki ukuthi iyini, isebenza kanjani, noma ukuthi yini eyenza ihluke kwezinye, lesi sihloko singesakho.

Ezigabeni ezimbalwa ezilandelayo, sizocubungula konke odinga ukukwazi mayelana ne-INTELLECT-2: kusukela emcabangweni wayo oyisisekelo, kuya ekwakhiweni okuyingqayizivele kobuchwepheshe ngemuva kwayo, izindlela zayo zokuqeqesha ezisebenzisanayo nezihlukaniselwe izindawo, kuya emiphumeleni ezuziwe, izinselele ezixazululiwe, kanye namathuba esikhathi esizayo ewavulayo. Lungela ukuthola ukuthi inhlanganisela yezinkulungwane zama-GPU emhlabeni wonke ikuvumela kanjani ukuthi uqeqeshe amamodeli amakhulu, anamandla ngaphandle kokuthembela ezinkampanini ezinkulu noma kwingqalasizinda evaliwe.

Iyini i-INTELLECT-2?

I-INTELLECT-2 iyimodeli yolimi enamapharamitha ayizigidigidi ezingu-32.000, eqeqeshwe kusetshenziswa uhlelo lokufunda lokuqinisa oluhlukaniswe ngokuphelele oluvulekele noma yimuphi umuntu ofuna ukunikeza amandla okusebenzisa ikhompyutha. Lo msebenzi uphawula iphuzu lokuguquka ngokwethula ithuba lokuthi noma yimuphi umuntu, inkampani, noma isikhungo sinikele ngezinsizakusebenza zaso (ikakhulukazi ama-GPU) ekuqeqeshweni komhlaba wonke kwamamodeli obuhlakani bokwenziwa esizukulwaneni esilandelayo.

Esikhundleni sokukhetha indlela yakudala ephakathi, i-INTELLECT-2 isebenzisa isakhiwo esixhumene, lapho umsebenzi usatshalaliswa khona ezinkulungwaneni zama-node. Lawa ma-node angahluka kusukela kumuntu one-GPU enamandla kuya ezikhungweni zedatha ezinkulu, zonke zihlanganiswe ngamaphrothokholi kanye nesofthiwe eyenzelwe ngqo lo msebenzi. Umgomo: ukufeza ukusebenza okuphambili emisebenzini yokucabanga, izibalo, kanye nokuhlela, ukwenza ukufinyelela kube lula kumamodeli athuthukile kanye nekhono lokuwaqeqesha.

Umsuka kanye nogqozi: yiziphi izinkinga elizixazululayo?

Enye yezinkinga ezinkulu emhlabeni we-AI ukugxila kwamandla nolwazi ezinhlanganweni ezinkulu ezimbalwa. Amamodeli amaningi amakhulu, njenge-GPT, i-Gemini, i-Llama, njll., adinga ingqalasizinda emikhulu nezinsiza ezitholakala kuphela kuma-tech giants. Lokhu kudala izithiyo zomphakathi, kuvimbe ukungafihli, futhi kukhawulele ukuhlola okuvulekile.

I-INTELLECT-2 iphakamisa enye imodeli, lapho ukuqeqeshwa kungancikile kuqoqo elimaphakathi noma ebhizinisini elilodwa, kodwa kunalokho kusekelwe emizamweni ehlangene nehlanganisiwe yabahlanganyeli abasatshalaliswe emhlabeni wonke.

Lokhu kuvula indlela yokuthi:

• Ukuhlukaniswa kwangempela kocwaningo lwe-AI, ukuvumela ulwazi kanye nekhono lokuthonya amamodeli ukuthi angagcini nje kwabambalwa.
• Ukwehliswa kwezindleko, Njengoba noma ubani enganikela nganoma imaphi amandla anawo, kusukela kumakhompyutha asekhaya anama-GPU amaningi kuya ezikhungweni zedatha.
• Ukumelana okukhulu nokucwaninga, ukukhohlisa noma ukuphahlazeka, ngoba uhlelo aluncikile endaweni eyodwa yokwehluleka.
• Kwenza kube sobala, ukusebenzisana kanye nokufinyelela okuvulekile, okhiye bokudlulela ku-AI enesimilo nenenzuzo kubo bonke.

I-athikili ehlobene:

Iyiphi i-DNS ongayisebenzisa ngo-2025: Gcwalisa Umhlahlandlela Wokukhetha Iseva Engcono Kakhulu

Isakhiwo sobuchwepheshe: Lena yindlela i-INTELLECT-2 esebenza ngayo ngaphakathi

Ngemuva kwe-INTELLECT-2 i-architecture ehlaba umxhwele yezingxenye ezenza kube nokwenzeka ukuxhumanisa ukuqeqeshwa okusabalalisiwe, ukuqinisekisa imiphumela, kanye nokuvuza ababambiqhaza. Ngezansi sichaza izici ezibalulekile:

I-PRIME-RL: Injini Yokuqeqesha Emisiwe

Inhliziyo yesistimu yi -PRIME-RL , uhlaka lomthombo ovulekile oluklanyelwe ngqo ukuphathwa kokuqeqeshwa kokuqinisa (RL) ezindaweni ezinama-node ahlukene futhi kungekho ukwethembana kwangaphambili phakathi kwawo. Yini eyenza kube okukhethekile? Ngoba kuvumela ukuhlukaniswa kwenqubo yokukhiqiza ukuqaliswa (ukuhlolwa kwemodeli ezindaweni zomsebenzi), ukuqeqeshwa kwangempela, kanye nokusatshalaliswa kwezisindo ezintsha , ukuphatha isigaba ngasinye ngendlela ehambisanayo nangokusebenza kahle okuphezulu.

Empeleni, lokhu kusho ukuthi:

• Idatha yokuqeqeshwa nolwazi luyakhiqizwa futhi luthunyelwe kusuka ezindaweni ezahlukene emhlabeni jikelele ngokuhambisana futhi kuhlukaniswe kusukela kumaseshini okuqeqesha.
• Izisindo zemodeli nama-gradient agcinwa futhi aphathwe ngeziqephu kusetshenziswa ubuchwepheshe obusezingeni eliphezulu njenge-PyTorch FSDP, futhi ukushintshisana ngolwazi kufinyelelwa ngamafayela e-Parquet nezivumelwano eziqinile.
• Uhlelo luyakwazi ukulungisa ukuhlukahluka kokuxhumana nengqalasizinda yabahlanganyeli, ngaphandle kokujezisa ukusebenza kahle noma ukuphepha kwenqubo.

I-SHARDCAST: Ukusatshalaliswa kwemodeli okusebenzayo

Ukuqinisekisa ukuthi wonke ama-node ahlala enenguqulo yakamuva yemodeli eqeqeshiwe futhi angenza iziphetho ezibuyekeziwe, kwadalwa i-SHARDCAST . Lo mtapo wolwazi uhlukanisa futhi usabalalise amafayela amakhulu (njengezindawo zokuhlola zamamodeli) kunethiwekhi yamaseva esakhiweni sesihlahla, kusetshenziswa i-HTTP. Lokhu kuvumela:

• Nciphisa kakhulu izikhathi zokusakazeka kwezisindo ezintsha, ukufinyelela bonke ababambiqhaza ngemizuzu nje embalwa.
• Sebenzisa ukungafuneki, ukulawula ubuqotho (SHA-256), kanye nobukhulu ngaphandle kokugcwalisa noma iyiphi indlela yenethiwekhi.
• Lungisa ithrafikhi emandleni e-node ngayinye (umkhawulo wezinga, ukukhethwa kokudluliselwa okunamandla okusekelwe ekusebenzeni).

I-TOPLOC: Ukuqinisekiswa Nokuphepha kokuthi Inference

Okunye okukhathazayo okuyinhloko ezinhlelweni ezisebenzisanayo nezihlukaniselwe abantu ukwethembana: siqinisekisa kanjani ukuthi iminikelo yenodi ngayinye iyasebenza futhi akukho ukungasebenzi kahle, ukukhohlisa, noma amaphutha?

Yilapho i-TOPLOC ingena khona, uhlelo lokuqinisekisa olusekelwe endaweni oluqinisekisa yonke imibono eyenziwe ngabasebenzi, ukuthola izinguquko, amaphutha, noma ukuzama ukukhwabanisa. Imisebenzi yalo ifaka:

• Hlola ukuthi imiphumela ayizange ishintshwe yizingxenyekazi zekhompuyutha noma isofthiwe ngesikhathi sokusebenza.
• Igcina ukuqina ngisho nalapho kubhekene nokunganqumi okujwayelekile kwama-GPU amanje.
• Ivumela ukuqinisekiswa ngokushesha okukhulu kunalokho obekungabiza ukuvuselela idatha, okwenza ukuqapha kwenzeke emhlabeni wonke.

I-Prime Intellect Protocol: I-Orchestration kanye Nomnotho Wokusebenzisana

Iphrothokholi yokuxhumanisa, ebizwa ngokuthi i-Prime Intellect Protocol , iyisici esivumela ukubhaliswa, ukuqinisekiswa, ukwabiwa kwemisebenzi, ukuphathwa kwezikhuthazo, kanye nokulandelela umsebenzi. Iminikelo yayo eyinhloko ifaka:

• Uhlelo lwe-ledger olunwetshiwe (eliphefumulelwe i-blockchain), elirekhoda ukusebenza kanye negalelo le-node ngayinye.
• Isofthiwe ethile yabahlanganyeli yokuxhumana ngokutholakala kwabo, ukuthumela imiphumela, ukuthola izibuyekezo, nokusebenza ezindaweni ezingazodwa besebenzisa i-Docker.
• Indlela yokukhuthaza umnotho kanye nedumela, lapho ababambiqhaza bengathola khona imivuzo (okungakabi kumathokheni wangempela ku-testnet) futhi kufanele benze inani elithile, elizohoxiswa uma kwenzeka ukuziphatha ngokungathembeki.
• Isikhathi sokuqinisekisa samahora angama-24 somsebenzi ngamunye othunyelwe, lapho uma kutholwa amaphutha noma ukukhwabanisa, izinhlawulo ezilinganayo zingasetshenziswa.

I-athikili ehlobene:

Iyini i-Blockchain futhi lobu buchwepheshe busebenza kanjani?

INTELLECT-2 Training: Updates and Results

Ukuqeqeshwa kwamamodeli kugqugquzelwa izindlela ezisezingeni eliphezulu, ezihlanganisa izixazululo zobunikazi ukuze kukhuliswe izinga lokufunda, ukuguquguquka, nokusebenza kahle. Phakathi kwezinto ezintsha ezibaluleke kakhulu yilezi:

Ukuqinisa ukufunda ngokulawula ubude

Enye yezinselelo ezinkulu zamamodeli olimi amanje ukwenza ngcono isikhathi kanye nezindleko zokuqagela ngaphandle kokubeka engcupheni ikhwalithi emisebenzini eyinkimbinkimbi. Ukuze kuxazululwe lokhu, kwaqaliswa uhlelo "lwesabelomali sokucabanga", oluvumela ukulungisa, ngemiyalo, inani lamathokheni okufanele imodeli iwasebenzise ngaphambi kokunikeza impendulo.

Ngenxa yaleli khono, izabelomali ezimfushane zingabelwa izinkinga ezilula kanye nezabelomali ezinkulu (“amathokheni okucabanga” amaningi) ezinseleleni eziyinkimbinkimbi, kuye ngokuthi ihadiwe yokucubungula inamandla kangakanani. Lokhu akugcini nje ngokuthuthukisa ukusebenza kahle kanye nesivinini kodwa futhi kuvumela ukusetshenziswa okungcono kwezinsizakusebenza zenethiwekhi ezingafani futhi kusiza ukuvumelanisa imodeli nokusetshenziswa okuguquguqukayo komhlaba wangempela.

Ukuhlunga idatha okuhlakaniphile: ikhwalithi phezu kobuningi

Ngenkathi beklama idathasethi namasu okuqeqesha, onjiniyela bahlonze iphuzu elibalulekile: asikho isidingo sokuqeqesha imodeli ezinkingeni esezizixazulula ngokuphelele. Ngakho-ke, kusetshenziswa ukuhlunga kabili:

• Ukuhlunga okungaxhunyiwe ku-inthanethi: Izinkinga eziningi ziyahlaziywa, kutholwe izimpendulo ezimbalwa bese kugcinwa kuphela lezo ezimele inselelo yangempela yemodeli (isilinganiso sokuxazulula esilingana noma ngaphansi kwama-75% ekuhlolweni kwangaphambilini).
• Ukuhlunga ku-inthanethi: Phakathi nokuqeqeshwa, uma yonke imizamo yomsebenzi iphumela emvuzweni ofanayo, inkinga ibhekwa njengenganikezeli ukufunda futhi ayisetshenziselwa ukubuyekeza imodeli. Ngale ndlela, imisebenzi namasampula afundisa ngempela ahlala ebekwa eqhulwini, okuthuthukisa ukusetshenziswa kwezinsiza kanye nekhwalithi yokugcina yohlelo.

Amaresiphi, amabhentshimakhi kanye nokuqhathanisa namanye amamodeli

Imodeli eyisisekelo esetshenzisiwe yi-QwQ-32B, lapho kusetshenziswa khona amasu athuthukile njenge- GRPO (Generalized Reinforcement Policy Optimization) , imivuzo eqinisekisiwe evela ezizindeni zezibalo kanye nezinhlelo, kanye nezindlela zokuzinzisa ezenzelwe ngqo ukuqeqeshwa okusatshalaliswa okungahambelani.

Kusetshenziswa uhlelo olunezinhloso ezimbili: imivuzo yezimpendulo ezifanele ezinkingeni zezibalo kanye nokubhala ikhodi (0 noma 1) kanye nemivuzo yokusondela ebudeni bokucabanga okuqondiwe. Ukuze kuzinziswe isibuyekezo, ukunqunywa kwezinhlangothi ezimbili kwethulwa emathubeni ethokheni, kukhawulelwe kokubili ukwanda kanye nokwehla ngokusekelwe enzuzweni etholiwe (ngamapharamitha anjenge-delta kanye ne-epsilon ukulawula umkhawulo).

Ngokwedatha eshicilelwe, i-INTELLECT-2 idlula i-QwQ-32B ngokuphawulekayo emisebenzini yezibalo neyohlelo, nakuba ibonisa ukwehla kancane kwamabhentshimakhi afana ne-IFEval, okungenzeka ngenxa yokwakheka okuthile kwedathasethi.

Ukubamba iqhaza: Indlela noma ubani anganikela ngayo futhi azuze

Esinye sezici ezishintsha kakhulu ze-INTELLECT-2 ukuvuleka kwayo: noma ubani one-GPU efanelekile angajoyina inethiwekhi, anikele ngamandla ekhompyutha, futhi ahlanganyele ngenkuthalo ekuqeqeshweni kwamamodeli e-AI asezingeni eliphezulu. Inqubo yenzelwe ukuba ibe lula: vele ulandele imiyalelo kudeshibhodi ezinikele, ufake isofthiwe edingekayo, bese uqala ukuthola imisebenzi, ukuqinisekisa imiphumela, kanye nokuthola imivuzo.

Lokhu kusho:

• Ukufinyelela kwentando yeningi kanye nomhlaba wonke ekuthuthukisweni kwamamodeli amakhulu, ngokwesiko akuvunyelwe ezinkampanini ezinkulu.
• Ithuba lamaqembu amancane, amanyuvesi, noma onjiniyela abazimele ukuthi banikele futhi bazuze ekuthuthukisweni kobuchwepheshe okuhlangene.
• Ukudala umphakathi ozonele okwazi ukubhekana nezinkinga ngokubambisana, ukulungisa iziphazamisi, nokuthuthukisa ubuchwepheshe.

I-athikili ehlobene:

Iyini i-Scalability: Imiqondo engu-10 ebalulekile

Ukuphepha, ukuqinisekiswa, kanye nezomnotho: ukuthi ubuqotho nokwethembeka kuqinisekiswa kanjani

Ukuze uvimbele ukukhwabanisa, ukukhohlisa, kanye nemizamo "yomdlalo" isistimu (njengokuthumela idatha engamanga, izingxenyekazi zekhompuyutha ezikhohlisayo, noma imiphumela ekhuphukayo), i-INTELLECT-2 yethula izinyathelo zokuphepha eziningi:

• I-node ngayinye kufanele yenze inani elithile lamathokheni (ku-testnet okwamanje), ahoxiswayo uma kutholwa ukungaziphathi kahle.
• Ukuqinisekiswa okuqhubekayo ngamanye ama-node kanye nephrothokholi, okuhlanganisa ukuqinisekiswa kwe-TOPLOC nokubuyekezwa kwelogi.
• Amafasitela wamahora angu-24 ukuze abuyekeze umsebenzi futhi asebenzise izinhlawulo lapho kudingeka.

Umnotho ongemuva kwalolu hlelo ugqugquzelwa yi-cryptocurrencies kanye nefilosofi ye-blockchain, kodwa wajwayelaniswa neqiniso lokuqeqeshwa kwe-AI okuhlanganyelwe: izikhuthazo zomsebenzi owenziwe kahle kanye nezithiyo zokuzama ukucekela phansi noma ukukhohlisa.

Ukusetshenziswa kwamanje kanye nokusetshenziswa kwe-INTELLECT-2

Njengamanje, i-INTELLECT-2 igxile kakhulu ekucabangeni kwezibalo, ekuxazululeni izinkinga zohlelo, kanye nezilinganiso ze-logic ezithuthukisiwe. Lokhu akulona iphutha: lezi zinkundla zidinga kakhulu ukubalwa futhi zidinga amamodeli anekhono "lokucabanga" okuhlelekile nokuqhubekayo. Ukuze ufunde kabanzi mayelana nezisekelo zobuhlakani bokwenziwa kanye nezisekelo zakho, ungavakashela i- Artificial Neural Networks.

Kodwa-ke, ingqalasizinda kanye nendlela ye-INTELLECT-2 isebenza kahle kakhulu kwezinye izinkambu eziningi, kusukela ocwaningweni lwesayensi kanye nokukhiqiza okuqukethwe okuyinkimbinkimbi kuya ekwenzeni izinqumo ezindaweni ezingaqinisekile. Amandla ayo esikhathi esizayo cishe awapheli, ikakhulukazi uma umphakathi uqhubeka nokukhula nokwandisa izindawo zawo zokusetshenziswa.

Ikusasa le-AI yokuhlanganyela: izinyathelo ezilandelayo nezinselele

Abathuthukisi be-INTELLECT-2 bamemezele ukuthi ukuthunyelwa kwamanje kuseyisiqalo nje. Phakathi kwezinyathelo ezilandelayo ezihleliwe yilezi:

• Ukuqeqeshwa kokuphela kuze kube sekupheleni kwama-ejenti e-AI akwazi ukufunda ukusebenzisa amathuluzi angaphandle, njengabatoliki bekhodi noma izinhlelo zokusebenza zesayensi.
• Ukwandiswa nokuthuthukiswa kwemitapo yolwazi kanye nezindawo zokuqinisekisa, ukusiza iminikelo yomphakathi ekukhetheni nasekulawuleni ikhwalithi yezinkinga okumele zixazululwe.
• Ukuthuthukiswa okuqhubekayo kwengqalasizinda ukuze kuqhutshekwe nokunciphisa ukubambezeleka, izindleko zokucatshangelwa kanye nokuqeqeshwa, nobunzima bokugibela ababambiqhaza abasha.

Izinselele zifaka phakathi ukuqinisekisa ukukhuliswa ngaphandle kokwehlisa ukusebenza, ukuqinisa ukuphepha ekuhlaselweni okungenzeka, kanye nokugcina ukuhlanganyela okusebenzayo komphakathi ohlukahlukene ngokwengeziwe. Ukuze ufunde kabanzi mayelana nendlela ukukhuliswa kwengqalasizinda okuphathwa ngayo , sikumema ukuthi uthintane nezinsiza ezikhethekile.

I-INTELLECT-2 imele enye yezinhlelo ezifisa ukuvelela nezithembisayo zokuletha ubuhlakani bokwenziwa obusezingeni eliphezulu endaweni yokusebenzisana okuvulekile, nokubambisana komhlaba wonke. Ukuzibophezela kwayo ekwandiseni izindawo, obala, kanye nemiklomelo yomphakathi kuyenza imodeli ezizukulwaneni ezizayo ze-AI, okukhombisa ukuthi ihlanganisa ubuchwepheshe obusezingeni eliphezulu nomoya wentando yeningi nokubambisana. Uma unentshisekelo ku-AI futhi unezinsiza zekhompuyutha, ungangabazi ukuhlola ukuthi kungenzeka yini ukubamba iqhaza: ikusasa lobuhlakani bokwenziwa lingaba sezandleni zakho ngokoqobo.