- I-MAI-Voice-1 ikhiqiza izwi eliveza ukwethembeka okuphezulu elinokubambezeleka okuphansi kakhulu kusetshenziswa i-GPU eyodwa.
- I-MAI-1-preview imodeli ye-MoE eqeqeshwe nge-~15.000 H100s, egxile ekulandeleni imiyalelo.
- Kokubili kancane kancane kuhlanganiswa ku-Copilot futhi kungahlolwa kuma-Labs kanye ne-LMArena.
- Isu: I-Orchestrate amamodeli akhethekile ngenkathi uhlala ne-OpenAI's.
Isithombe esivelele
I-Microsoft yenze isinyathelo ngokwethulwa kwamamodeli ayo okuqala obuhlakani bokwenziwa ngaphansi kwesambulela se-Microsoft AI: i-MAI-Voice-1 kanye ne-MAI-1-preview . Ngalesi sinyathelo, inkampani ihlose isu lomkhiqizo lapho abasizi bayo kanye nezinsizakalo zayo bezuza khona kumamodeli akhethekile, aklanyelwe kokubili izwi kanye nombhalo.
Umgomo wenkampani yezobuchwepheshe ukwenza i-AI "indlela eya endaweni yonke yolwazi," ihlela isethi yezinhlelo ezinokwethenjelwa eziqonda umsebenzisi ngamunye. Lokhu kuhunyushwa kube izinto ezimbili ezibalulekile ezintsha: imodeli yezwi esheshayo kakhulu nevezayo, kanye nohlobo lwemodeli yombhalo oluxubile olugxile ekulandeleni imiyalelo nokunikeza izimpendulo eziwusizo zokusetshenziswa kwansuku zonke.
Iyini i-MAI futhi kungani manje?
I-MAI isifinyezo se- Microsoft AI , uhlobo lapho inkampani eseRedmond ihlanganisa khona amamodeli ayo ayisisekelo. Umongo ubalulekile: ubudlelwano ne-OpenAI buyaqhubeka, kodwa buya ngokuya buncintisana, kuze kube seqophelweni lokuthi iMicrosoft ifake i-OpenAI ohlwini lwayo lwabancintisana nabo . Noma kunjalo, inkampani iphikelela ukuthi izoqhubeka nokusebenzisa "amamodeli amahle kakhulu" avela kubalingani kanye nomphakathi ovulekile lapho kunengqondo khona.
Ngemuva kwalokhu kwethulwa kukhona uMustafa Suleyman , i-CEO ye-Microsoft AI kanye nomsunguli we-DeepMind, owenze kwacaca umgomo: ukudala "i-AI esetshenziswa njengeplatifomu yemikhiqizo." Ngamanye amazwi, amamodeli akhethekile ahlanganiswa ne-Copilot kanye nokunye okuhlangenwe nakho, ahlanganisa ngempumelelo izidingo ezithile zezwi nombhalo.
I-MAI-Voice-1: imodeli yezwi ebeka phambili isivinini nokuzwakala
Eyokuqala eyembulwa yi- MAI-Voice-1 , uhlelo lokuhlanganisa izwi olukwazi ukukhiqiza umsindo ozwakalayo, othembekile kakhulu one-latency ephansi kakhulu. Iphuzu layo elibalulekile ukuthi ingakhiqiza umzuzu owodwa womsindo ngaphansi komzuzwana owodwa , ngenkathi igcina ukuhlukahluka kwemvelo kanye nethoni.
Ngale kwejubane, elinye lamandla alo ukuzwela: imodeli isekela izitayela ezahlukene, amathoni, kanye nezincazelo ezihlukile ekukhulumeni ngezwi, okwenzelwe ukulandisa noma iziqondiso ezikhulunywayo . Ezimweni zezwi elilodwa noma eziningi, umphumela yilokho iMicrosoft ekuchaza njengezwi "elizwakalisa kakhulu futhi elingokwemvelo".
Ukusebenza kahle nakho kuyisici esibalulekile: I-MAI-Voice-1 isebenza nge -GPU eyodwa ukuze inqume, ikubeka phakathi kwezinhlelo zezwi ezisebenza kahle kakhulu ezitholakalayo njengamanje. Kuhlelo lwe-ecosystem lapho izindleko zokubala zibalulekile, lokhu kuthuthukiswa kubalulekile ekukhuliseni ukusebenza kwemikhiqizo yabathengi.
Ngokuphathelene nokutholakala, i-MAI-Voice-1 isivele ihlanganiswe ne -Copilot Daily kanye nama-Podcast , futhi inikezwa njengolwazi olusha ngaphakathi kwe-Copilot Labs . Lapho, kungenzeka ukuzama ukukhuluma ngokulandisa kanye nokukhuluma ngendlela evezayo, ukulungisa izitayela kanye nokuhlola amazwi ahlukene ngaphakathi kwemingcele yelebhu.
Ukuthi kuzwakala kanjani nokuthi ungenzani ngakho namuhla
Labo abake bayizama bathi umsindo uzwakala ungokwemvelo kakhulu , unokulawula okuhle kwephimbo nesigqi. Nokho, kubalulekile ukuqaphela ukuthi okwamanje utholakala ngesiNgisi kuphela . I-Microsoft inikeza izibonelo ezivamile: cela "indaba ngama-dinosaur" bese uthola ngokushesha indaba yomzuzu owodwa enezinguquko zezwi ezifanele kanye namathoni.
Ngaphakathi kwe-Copilot Labs, isici esithi " Audio Expressions " sikuvumela ukuthi ukhiqize umsindo kusuka kusikripthi futhi ushintshe isitayela sokulandisa. Sihlanganisa izindlela ezifana neModi Yemizwa (yokulungisa ithoni nesigqi, noma ukwabela amazwi ahlukene) kanye neModi Yendaba ehloselwe ukulandisa okuvezayo. Umqondo uwukwenza kube lula ukuhlola kanye nokwenza isibonelo kulabo abafuna ukuhlola amathuba ngaphandle kokusetha indawo eyinkimbinkimbi.
Phakathi kwezimo zokusetshenziswa, iMicrosoft ikhuluma ngakho konke kusukela ekulandiseni izindaba nasekulandiseni kuya ekuzindleni okuqondisiwe, kanye namathuba abasizi abangochwepheshe abanengxoxo yesikhathi sangempela. Ukulibaziseka okuncishisiwe, kuhlanganiswe nokukhuluma, kuvula umnyango wokuhlangenwe nakho okubushelelezi ezimpambanweni zezwi lapho ukuba ngokwemvelo kubalulekile khona.
- Ukulandisa okunamandla ngezindaba, i-podcast noma imfundo.
- ukuzindla okuqondisiwe nokuqukethwe kwezempilo okunokuhlukahluka kwamathoni.
- Abasizi bengxoxo esikhathini (cishe) sangempela, sizwela umongo.
- I-prototyping esheshayo kuma-Copilot Labs anesitayela nokulungiswa kwezwi.
Ukusebenza kwekhompyutha: iminithi elilodwa lomsindo ngaphansi kwesekhondi elilodwa
Ukudala umsindo ozwakalayo nothembekile kudinga ikhono lobuchwepheshe, kodwa i-MAI-Voice-1 izigqaja ngokufeza lokhu nge-GPU eyodwa , kanye nokubambezeleka okuphansi kakhulu. Lesi sibalo siyamangalisa: umzuzu owodwa womsindo ngaphansi komzuzwana owodwa . Emikhiqizweni yabathengi, le nhlanganisela—ukubambezeleka okuphansi kanye nezindleko ezingabizi—ibalulekile ekuletheni okuhlangenwe nakho okungenazinkinga nokukhuliswa.
Esimweni sanamuhla, lapho izinhlelo zezwi zincintisana khona ngekhwalithi kanye nesivinini, iMicrosoft ibeka i-MAI-Voice-1 phakathi kwezinhlelo ezisebenza kahle kakhulu ezitholakalayo. Lokhu kusebenza kahle akugcini nje ngokuguqula ukonga izindleko kodwa futhi kwenza kube lula ukusebenzisa izimo ezazingasebenzi ngaphambilini ngenxa yezindleko noma ukubambezeleka.
Ungazama kuphi i-MAI-Voice-1
Namuhla, i-MAI-Voice-1 ihlolwa ngabasebenzisi nge- Copilot Daily (izifinyezo zezindaba) kanye ne -Copilot Podcasts , kanye ne- Copilot Labs . Kuma-Labs, imiboniso yenkulumo evezayo kanye neyokulandisa ibonisa izinga lokulawula isitayela, isigqi, kanye nethoni, konke ngesithembiso somphumela othembekile.
Uma unesifiso sokuzama, indlela eqondile kakhulu ukuvula i-Copilot Labs bese ufinyelela isigaba sezwi ukuze udlale ngezindlela zokulandisa . Nakuba okulindelwe kuphezulu, kubalulekile ukukhumbula ukuthi lokhu kumane kuyiqoqo lokuqala; iMicrosoft izoqhubeka nokuphindaphinda ngempendulo yomphakathi, ilungisa amapharamitha futhi yandise amakhono ngokuhamba kwesikhathi.
Ukubuka kuqala kwe-MAI-1: imodeli yombhalo oqeqeshwe ngezinga elikhulu
Eceleni kwemodeli yezwi kuza i-MAI-1-preview , imodeli yokuqala yolimi oluyimfihlo ye-Microsoft AI eyenzelwe ukulandela imiyalelo nokuphendula imibuzo yansuku zonke ngempumelelo. Ukwakheka okukhethiwe kuyinhlanganisela yochwepheshe (i-MoE), isu lapho ochwepheshe abaningi begxile khona futhi bevuselelwa ngokukhetha kuye ngomsebenzi, okuthuthukisa ukusebenza kahle kanye nokusebenza.
Ngokuphathelene nokuqeqeshwa, iMicrosoft ichaza ukuthi i-MAI-1-preview yaqeqeshwa kusengaphambili futhi yaqeqeshwa ngemuva kusetshenziswa ama-GPU e-Nvidia H100 angaba ngu-15.000 . Lesi sikali sinikeza isisekelo esiqinile sokuqondanisa imiyalelo kanye nokunikeza izimpendulo ezisebenzayo ezimweni ezivamile zengxoxo kanye nokukhiqiza.
Indlela ye-MoE ikhumbuza izitayela zakamuva kumamodeli athuthukile: ukuhlukanisa uhlelo lube yizingxenye ezikhethekile bese kusebenze kuphela lokho okudingekayo ngombuzo ngamunye. Lolu hlelo lubonwe kumaphrojekthi afana ne -DeepSeek ne-Qwen , lapho ukusebenza kahle ngesikhathi sokuphetha kuba yinto ehlukanisayo uma kuqhathaniswa nezakhiwo ezivamile ezixineneyo.
Ukutholakala, ukuhlolwa komphakathi kanye nokusatshalaliswa okuqhubekayo
I-MAI-1-preview isiyatholakala ukuze ihlolwe epulatifomu yokuhlola umphakathi ye- LMarena . Ukuze uyihlole, vele ungene ngemvume, khetha imodi ye -Direct Chat , bese ukhetha i-“mai-1-preview” ukuze noma ubani akwazi ukuzwa ukuziphatha kwayo ezingxoxweni zangempela futhi ayiqhathanise namanye amamodeli.
I-Microsoft isiqalile futhi ukusebenzisa i-MAI-1-preview ku-Copilot kwezinye izimo zokusetshenziswa kombhalo. Umgomo ukabili: ukuqoqa impendulo yabasebenzisi kanye nokuthuthukisa imodeli endaweni yangempela, ngokubambisana nezinye izinhlelo ezisekela izici ezibalulekile zomsizi wenkampani njengamanje.
Ngesikhathi esifanayo, inkampani inike amandla ukufinyelela kwe-API kwabahloli abathembekile nabathuthukisi ekufinyeleleni kusenesikhathi. Lesi sigaba esilawulwayo sizobavumela ukuthi baqoqe impendulo ngokuqina nokusebenza ngaphambi kokuyivula kubabukeli abaningi, ngokunaka kakhulu ikhwalithi yokuphendula kanye nokuhambisana nemiyalelo.
Ingabe ithatha indawo yamamodeli we-OpenAI? Ukuhlalisana kanye nesifiso sakho siqu
I-Microsoft ikubeke kwacaca ukuthi, okwamanje, izoqhubeka nokuhlanganisa amamodeli ayo nalawo abalingani kanye nokusungula izinto ezintsha emphakathini. Empeleni, lokhu kusho ukuthi ukubuka kuqala kwe-MAI-1 akuhloselwe ukuthatha indawo yezinhlelo ezisebenzisa i-Copilot njengamanje. Eqinisweni, inkampani ihlela ukuyisebenzisela imisebenzi ethile , ukulinganisa imiphumela nokulungisa ukuthunyelwa ngokusekelwe kumpendulo.
Eminye imibiko ikhombisa ukuthi ngeke ithathe indawo yezinguqulo ezithuthukisiwe zomndeni we-GPT ku-Copilot; noma kunjalo, iphuzu elibalulekile ukuthi iMicrosoft iqhubekela phambili ekuzimeleni okukhulu kwezobuchwepheshe . Ukubambisana ne-OpenAI kuyaqhubeka, kodwa kuyinkimbinkimbi kakhulu, njengoba zombili izinhlangothi manje seziqaphelana njengabancintisana emakethe ye-AI.
Lesi sahluko sihlelwe ngokwesizinda sobudlelwano obuyinkimbinkimbi ne-OpenAI. I-Microsoft itshale imali eningi, futhi okwesikhathi esithile le nhlangano yayibalulekile esu layo, kodwa ngonyaka odlule yayihlanganisa i-OpenAI njengembangi kanye ne-Amazon, i-Apple, i-Google, kanye ne-Meta. Ngesikhathi esifanayo, i-OpenAI iveze ukungabaza ngokwabelana ngobuchwepheshe obusezingeni eliphezulu besikhathi esizayo, njenge-AGI ecatshangelwayo, okugcizelela isidingo sokuba i-Microsoft ibe namakhono ayo.
Ifilosofi Yokuklama: I-AI Ewusizo, hhayi "i-digital person"
Ngaphandle kwezilinganiso kanye nezilinganiso, iMicrosoft AI ithathe isinqumo mayelana nokuhlangenwe nakho okufanele kunikezwe yi-AI yengxoxo. USuleyman uxwayise ngengozi yokwakha izinhlelo ezibonakala zinemizwa noma imigomo yazo, into engadala ukulindela okungafuneki noma ukudideka.
Umhlahlandlela uhilela ukususa izici ezenza uhlelo lube yinto yomuntu, ukuqinisa i-AI eyithuluzi eliwusizo nelinomthwalo wemfanelo, ngaphandle kwemiphumela emibi yokuzenzisa. Endabeni ye-MAI-Voice-1, lokhu kusho ukubeka phambili ukuveza kanye nokuba ngokwemvelo ezwini ngaphandle kokuwela umugqa kube ukulingisa imizwa yabantu ngokukhohlisayo .
Isu: orchestrate amamodeli akhethekile agxile emkhiqizweni
I-Microsoft igcizelela ukuthi umgomo wayo ukuhlela uhlelo lwezinhlelo zamamodeli aklanyelwe izinhloso nezimo ezithile. Empeleni, i-MAI-Voice-1 ihlanganisa isikhombikubona esikhulunywayo ngesivinini nekhwalithi, kuyilapho i-MAI-1-preview igxile emisebenzini yombhalo lapho ukulandelela imiyalelo kanye nokusetshenziswa komsebenzisi okusheshayo kukhanya khona.
Lokhu kuhlanganiswa kwenza i-Copilot nezinye izinsizakalo zihlanganise okungcono kakhulu kwemodeli ngayinye kuye ngokuthi kwenzekani, kungakhathaliseki ukuthi ilandisa isifinyezo sezindaba ngezwi elihehayo, iphendula ngokunembile umbuzo wansuku zonke, noma ilungisa isitayela sokukhishwa okukhulunywe ngaso kwe-podcast noma umhlahlandlela wokuzindla.
Sebenzisa izimo namathuba kubasebenzisi nabathuthukisi
Kumsebenzisi wokugcina, lokhu kuholela ezimpendulweni eziwusizo kakhulu kanye nokuhlangenwe nakho kwezwi okuthuthukisiwe emikhiqizweni asebevele beyisebenzisa. Kumaqembu okuqukethwe noma ama-podcast, ukuba nenjini yezwi ethembekile kakhulu, ebambezelekile kancane kusho ukukhiqiza okuqukethwe okwengeziwe nokungcono okunokuphikisana okuncane kobuchwepheshe.
Kwabathuthukisi, ukufinyelela kwangaphambi kwesikhathi nge- API ye-MAI-1-preview kuvula umnyango wokwenza amaphrothokholi nokuhlola okulawulwayo, okubavumela ukuthi bahlole ukuthi le modeli inikeza izinzuzo kuphi kunezinye. Isigaba se-LMarena siphinde sivumele ukuqhathanisa okungokomongo kanye nokuqoqwa kwamamethrikhi omhlaba wangempela ngekhwalithi yokuphendula kanye nekhono lokulandela imiyalelo.
- Abasebenzisi: Izwi elithuthukisiwe ku-Copilot Daily kanye nama-Podcasts; ukusebenzisana okwengeziwe kwemvelo.
- Abadali: Ukulandisa kanye nemibhalo enesitayela nokulawula isigqi.
- izinkampani: abasizi bezwi abane-latency ephansi kakhulu kanye nokuzwakala okuphezulu.
- Onjiniyela: Ukuhlolwa kwe-LMArena nokufinyelela kwangaphambi kwesikhathi kwe-API ukuze kuphindwe ngokushesha.
Ongakulindela esikhathini esifushane
Esikhathini esifushane, sizobona ukubuka kuqala kwe-MAI-1 kuhambisana namanye amamodeli ku-Copilot ezimweni zokusetshenziswa kombhalo ezikhethiwe ngenkathi i-Microsoft ihlaziya impendulo futhi ithuthukisa. Le ndlela eqhubekayo isivumela ukuthi silinganise umthelela futhi silungise ulwazi ngaphandle kokubeka engcupheni ulwazi lomsebenzisi lonke.
Ngesikhathi esifanayo, inkampani kulindeleke ukuthi ithuthukise i-MAI-Voice-1 ngaphakathi kwe-Copilot Labs kanye nemikhiqizo lapho isivele isetshenziswa khona. Ngokufaka izimo eziningi zokukhuluma kanye nezibonelo, i-Microsoft izokwazi ukuqinisekisa kangcono impendulo yomodeli ezindleleni ezahlukahlukene kanye nezinselele zokukhiqiza zangempela.
Izimpawu zendlela yesikhathi esizayo
Indlela iMicrosoft ekhuluma ngayo ayiguquki: kukhona isifiso sokufinyelela izigidigidi zabantu ngemikhiqizo yayo, ngentuthuko kumamodeli kanye namandla ekhompyutha. I-MAI-1 iyisiqalo somndeni, futhi kunengqondo ukulindela ukuhlukahluka kanye nokuphindaphinda okusha okulungisa ibhalansi phakathi kwekhwalithi, izindleko, kanye nokubambezeleka.
Ngokuncintisana, isu le-MoE kumamodeli wombhalo lisikisela ukuthi iMicrosoft izoqhubeka nokuhlola izakhiwo zokulinganisa ezisebenzayo , okuyisici esibalulekile ekukhuliseni amanani aphezulu emikhiqizweni emakethe enkulu. Icala le-MAI-Voice-1 liqinisekisa ukuthi izwi liyinto eza kuqala njengesixhumi esibonakalayo sabasizi besikhathi esizayo .
Ubudlelwano ne-ecosystem kanye nesimo somphakathi
Ubudlelwano ne-OpenAI buzohlala bubalulekile, kodwa iMicrosoft isivele ikhombisile ukuthi ayifuni ukuboshwa emthonjeni owodwa wokusungula izinto ezintsha. Izoqhubeka nokuhlanganisa okuhle kakhulu eqenjini layo , ozakwethu, kanye nomphakathi ovulekile, inqobo nje uma ingeza inani emikhiqizweni nakumakhasimende. Ngesikhathi esifanayo, ibeka uphawu lwayo (i-MAI) njengesithuthi sokufinyelela ukuzimela lapho ihlose khona.
Lesi senzo sokulinganisela—ukuhlalisana manje kanye nokuzimela okukhulu kusasa—singasheshisa ukuncintisana kwabasizi kanye nokukhiqiza. Kubasebenzisi, umphumela ofiselekayo ucacile: izinhlobo eziningi zamathuluzi kanye nokuphindaphinda okusebenzayo kulokho okuxazulula izinkinga zansuku zonke ngempela.
Indlela ye-Microsoft nge-MAI-Voice-1 kanye ne-MAI-1-preview ihlanganisa ukusebenza kahle kanye nesivinini: imodeli yezwi esebenza kahle futhi evezayo esivele inika amandla okuhlangenwe nakho ku-Copilot, kanye nemodeli yombhalo yohlobo lwe-MoE enkulu eqala ukusetshenziswa ezindaweni ezilawulwayo, ingahlolwa ku-LMarena, futhi ivulwa nge-API kubahloli abathembekile. Umgomo osanda kuvela ucacile: amamodeli akhethekile, ahlanganiswe emikhiqizweni, avumelana nezidingo zomuntu ngamunye zansuku zonke.
