I-Microsoft yethula ukubuka kuqala kwe-MAI-Voice-1 kanye ne-MAI-1: isivinini nokuzimela

Isibuyekezo sokugcina: 10 September ka-2025
  • I-MAI‑Voice‑1 (Ultra-Fast Voice) kanye ne-MAI‑1‑Preview (Umbhalo one-MoE) kufika njengamamodeli okuqala angaphakathi e-Microsoft.
  • I-MAI-Voice-1 ikhiqiza iminithi elingu-1 lomsindo nge-<1 s isebenzisa i-GPU futhi manje isiyatholakala ku-Copilot Daily, Podcasts, kanye Nakumalebhu.
  • Ukubuka kuqala kwe-MAI‑1‑ kwaqeqeshwa cishe ku-15.000 H100s, kuhlanganiswa ku-Copilot ngesisekelo esilinganiselwe, futhi kuyahlolwa e-LMArena.
  • Isu: Yehlisa ukuncika ku-OpenAI futhi uhlele amamodeli akhethekile ngokugxila kumsebenzisi.

Amamodeli we-Microsoft MAI

I-Microsoft yenze umnyakazo wayo futhi yethula amamodeli ayo obuhlakani bokwenziwa athuthukiswe ngaphakathi, isinyathelo esiphawula ushintsho lwejubane esu layo futhi eqondiswe ngqo emphakathini jikelele Ukuhlola kuqala kwe-MAI‑Voice‑1 kanye ne-MAI‑1‑.

Uhlobo lwe-MAI lumelela okuthi "Microsoft AI" futhi luza neziphakamiso ezimbili ezicace kakhulu: esisodwa sigxile ezwini elishesha kakhulu kanti esinye sigxile embhalweni onobuciko bezakhiwo. Konke lokhu kubeka inkampani endleleni ezimele uma iqhathaniswa ne-OpenAI, igcina ukubambisana kodwa iqondise ikusasa layo ngakuyo amamodeli angakwazi ukuncintisana ne-ChatGPT, Gemini kanye nenkampani en I-AI ekhiqizayo.

Yini i-MAI-Voice-1 kanye ne-MAI-1-ukuhlola kuqala?

Ukwethulwa kwamamodeli we-MAI

Ukubuka kuqala kwe-MAI-1, ngokusho kweMicrosoft, a imodeli yangaphakathi ene-Mixture-of-Experts (MoE) yezakhiwo baqeqeshwe ngezigaba ezimbili (ngaphambi kokuqeqeshwa kanye nangemuva kokuqeqeshwa) cishe ku-15.000 NVIDIA H100 GPUs. Lokhu kulungiselelwa "kochwepheshe" kwenza kusebenze izingxenye ezincane kuphela ezidingekayo kumsebenzi ngamunye, zifuna ukusebenza kahle nokuqondanisa kangcono nenjongo yomsebenzisi.

Mayelana nomkhiqizo, inkampani ibonisa ukuthi le modeli yombhalo yenzelwe landela imiyalelo futhi unikeze izimpendulo eziwusizo zemibuzo yansuku zonkeNgakho-ke, ukukhishwa kwayo kokuqala kuzolawulwa: izokwethulwa kwezinye izimo zombhalo ku-Copilot emavikini ambalwa alandelayo ngomgomo wokufunda ekusebenzelaneni kwangempela okusekelwe kumpendulo.

Ngaphezu kwalokhu kuhlanganiswa kancane kancane, iMicrosoft inikwe amandla izivivinyo zomphakathi endaweni yesikhulumi se-LMArena ukuqoqa amasignali ekhwalithi engeziwe. Futhi, ngesikhathi esifanayo, ihlela ukuyenza itholakale kubathuthukisi nge-API, ngaleyo ndlela iqinise ukuhlolwa kwemodeli kanye nenqubo yokuthuthukisa eqhubekayo.

Inkampani igcizelela ukuthi ngeke izilahle ezinye izinjini ze-AI: izoqhubeka nokusebenzisa amamodeli angcono kakhulu avela eqenjini layo, kusuka kozakwethu abanjengo I-Anthropic kanye ne-open source ecosystem Lapho kunengqondo. Esikhathini esifushane, ukubuka kuqala kwe-MAI-1 akuhloselwe ukufaka esikhundleni se-GPT-5 ku-Copilot; kunalokho, izosebenzisa izimo ezithile lapho inganikeza izinzuzo ezicacile.

I-MAI-Voice-1, ngakolunye uhlangothi, isiphakamiso sezwi seMicrosoft: a imodeli ekhiqizayo “evezayo kakhulu futhi engokwemvelo” Manje iyatholakala ku-Copilot Daily kanye nama-Podcasts, futhi ifinyeleleka njengezinto ezintsha phakathi kwama-Copilot Labs. Umbono ongemuva kwawo ucacile: “Izwi liyisixhumi esibonakalayo sesikhathi esizayo” sabasizi be-AI abawusizo kakhulu futhi abasebenziseka kalula.

Isithembiso sobuchwepheshe siyamangalisa: ingakhiqiza iminithi lomsindo ngaphansi kwesekhondi isebenzisa i-GPU eyodwaLesi sivinini, sihlanganiswe ne-high-fidelity timbre kanye nekhono lokusingatha izimo ngesipika esisodwa noma ngaphezulu, kubeka i-MAI-Voice-1 phakathi kwezinhlelo ezisebenza kahle kakhulu zokuhlanganisa izwi ezitholakalayo namuhla.

  Konke mayelana ne-Generative Artificial Intelligence: ukuthi isebenza kanjani, isetshenziswa kanjani, kanye nezingozi

Ekuhlolweni okusesidlangalaleni namademo, umsindo uzwakala ubushelelezi ngendlela emangalisayo, ngephimbo nesigqi esikholisayo, nakuba ukusekelwa kolimi kusantuleka. kulinganiselwe ngesiNgisiUkwenza kube ngokwakho izitayela namazwi kuhlolwa ngama-Copilot Labs, lapho i-Microsoft ikhiphe khona ulwazi olufana ne-“Copilot Audio Expressions.”

Imininingwane enelukuluku: amagama akhethiwe (i-MAI-Voice-1 ne-MAI-1-preview) ayi kucace futhi “njengonjiniyela kakhulu”Ngale kwaleyo anecdote, okubalulekile ukuthi babeka umgwaqo obheke kukhathalogi yamamodeli akhethekile agxile kumthengi, abeka phambili isivinini, ukusebenza kahle, kanye nokusebenziseka kalula.

I-MAI-Voice-1: amakhono, ukusetshenziswa, nokuthi ungayizama kuphi

Izwi le-MAI ku-Copilot

I-MAI‑Voice‑1 yethulwa njengohlelo lwe high-fidelity okhiqiza umsindo ekwazi ukukopisha, ukulandisa, nokudala amaphimbo ngokuphazima kweso. Iphuzu layo eliyinhloko lokuthengisa ukubambezeleka: ukukhiqiza kufika kuminithi lomsindo ngaphansi kwesekhondi nge-GPU eyodwa kuvumela izinhlelo zokusebenza eziseduze nesikhathi sangempela.

Ukuhlanganiswa kokuqala kwenziwa ku I-Copilot Daily kanye nama-Podcasts, lapho i-AI isivele ihlanganisa izifinyezo noma amagama akhulunywayo. Ukuze uhlole izitayela nama-nuances, i-Copilot Labs yethula okuthi "Copilot Audio Expressions," equkethe ukulandisa nokuboniswa kwenkulumo ecacile ukuze umsebenzisi ahlole okungenzeka.

Kulezo zipiliyoni, iMicrosoft yethula izinketho ezifana ne- Imodi ethinta inhliziyo (ukulawula iphimbo nesigqi) noma Imodi Yendaba enendaba yaseshashalazini eyengeziwe. Umgomo ukunikeza iphalethi yamazwi nezitayela eziguquguqukayo, kokubili kumlandi oyedwa kanye nezigcawu ezinezipikha eziningi.

Inkampani igcizelela ukuthi imodeli i ukusebenza kahle kwezinsiza: Isebenza nge-GPU eyodwa kodwa ifinyelela izinga elimangalisayo lokuzwakalisa. Le bhalansi yezindleko nekhwalithi iyenza ihehe emikhiqizweni yabathengi nasemaqenjini angenayo ingqalasizinda ebanzi.

Phakathi kwamacala okusetshenziswa asobala ahlongozwe yiMicrosoft ukuxoxa izindaba, okukhiqizayo ukuzindla okuqondisiwe, ukudalwa kwezikripthi zezwi, noma usizo lwengxoxo yesikhathi sangempela. Konke ngezwi elilwela ukuba ngokwemvelo futhi livumelane nezimo.

  • Ukulandisa nokuxoxa indaba: izindaba, iziqondiso ezilalelwayo, ukufunda ulimi noma izindaba ezinezinhlamvu ezimbalwa.
  • Ukukhiqizwa kokuqukethwe: ama-podcasts azenzakalelayo, ama-trailer omkhiqizo, izingcezu zokukhangisa noma izifinyezo zansuku zonke.
  • Usizo nokufinyeleleka: ukufunda imibhalo, ukusekela abasebenzisi abanobunzima bokubona, noma ukudala ngokushesha imiyalelo ekhulunyiwe.
  • Okuhlangenwe nakho okusebenzisanayo: abasizi bempendulo yezwi, imihlahlandlela yezingqikithi ezinhlelweni zokusebenza namageyimu, noma isekela ama-bots ngamathoni ahlukene.

Iphuzu elibalulekile yi- umthamo wezikhulumi eziningi, iwusizo ekwenzeni idrama, izingxoxo mbumbulu, noma izindima ezihlukene ekurekhodweni okukodwa komsindo. Lokhu kuvumelana nezimo kusiteji somsindo kuvumela ukudalwa kokuqukethwe okucebile ngaphandle kwesitudiyo noma ukuxhumana kwezwi lomuntu.

  I-Meta Aria Gen 2: Yonke imininingwane yezibuko ezintsha ezihlakaniphile ze-AI kanye neqiniso eli-augmented

Kumademo, ukuvele ucele "indaba mayelana no-X" kuzoveza umzuzu womsindo onamazwi ahlukene namaphimbo emzuzwini owodwa. Nakuba kusesekuseni kakhulu ukuhlola zonke izinto ezicashile, imiphumela yokuqala iveza imvelo ekholisayo ukusetshenziswa kwansuku zonke.

Okwamanje, i-MAI‑Voice‑1 ihloselwe isi-english, into encane okufanele uyikhumbule uma izilaleli zakho eziyinhloko zikhuluma iSpanishi. Kunoma yikuphi, ukwakheka nokusebenza kuvumela ukwesekwa okubanzi kolimi njengokuqeqeshwa kanye nenqubekelaphambili yokuhlola yomphakathi.

Kuhle ukukhumbula ukuthi, kwezokuphepha kanye nezimiso zokuziphatha, iMicrosoft iphinde yasho ukuthi izosusa noma yisiphi isici esenza i-AI ibonakale. njengokungathi linemizwa noma imigomo yaloUmqondo uwukuthuthukisa ukusetshenziswa ngaphandle kwe-anthropomorphizing, into ebucayi kakhulu kubasizi bengxoxo abasekelwe ezwini.

Ukubuka kuqala kwe-MAI-1: Izakhiwo, Ukuthunyelwa, kanye Namasu

Meyi 1 ukubuka kuqala ku-Copilot

Ukubuka kuqala kwe-MAI‑1‑ yi imodeli yokuqala yesisekelo sombhalo eyakhiwe yi-Microsoft ngaphakathi kwesigaba sayo se-MAI. Iqeqeshwe ngezinga elimangalisayo (cishe ama-H15.000 angu-100) futhi yamukela indlela ye-MoE: “ingxube yochwepheshe” lapho izingxenye ezifanele zemodeli kuphela zivulwa ngokufaka ngakunye.

Lo mklamo uvumela ukusabalalisa amakhono phakathi kochwepheshe kanye nokwenza ngcono ukusebenza kwemisebenzi ukulandela imiyaleloI-Microsoft ihlose ukunikeza izixazululo eziwusizo, ezigxile empilweni, ukubeka phambili ulwazi lomsebenzisi ngaphezu kwendlela egxile ebhizinisini.

Empeleni, ukuthunyelwa kuzoba ngezigaba ezimbili. Okokuqala, imodeli ifika Ukubuka kuqala kwezinye izimo zombhalo ku-Copilot, futhi ikwenza ngendlela elawulwayo ukukala i-telemetry futhi iqoqe impendulo. Bese, ngaleyo mpendulo, ukuziphatha kuzolungiswa futhi kufinyelelwe kunwetshwe.

Okwesibili, inkampani ivule ukufinyelela kokuhlola ku-LMArena kwe ukuhlolwa komphakathiLo mzila wephayiphi usheshisa umjikelezo wokuthuthukisa, uhlinzeka ngokuhlukahluka kokufakwayo, futhi uvumela amathuba okulungisa kahle ukuthi akhonjwe ngaphambi kokuhlanganiswa okubanzi.

I-Microsoft ikwenza kucace ukuthi ukubuka kuqala kwe-MAI-1 akuthathi (okwamanje) esikhundleni I-GPT-5 ngaphakathi kwe-CopilotIsu liwukusebenzisa “imodeli efanele yomsebenzi olungile,” ukuhlanganisa ukubuka kuqala kwe-MAI-1 emisebenzini ethile futhi kuqhathaniswe ngokuqhubekayo ukusebenza kwayo.

Ngokuhambisanayo, inkampani iqinisekisa ukuthi izoqhubeka nokubheja ezinjinini ezihlanganisiwe: eyakhe, lezo zozakwethu ezifana ne-OpenAI kanye ne- emisha evela emphakathini womthombo ovulekileNgale ndlela, u-Copilot angazuza kukho kokubili ukuzimela kwe-MAI kanye nemodeli engcono kakhulu etholakalayo endaweni ngayinye.

Konke lokhu kunyakaza kuyingxenye yoshintsho olubanzi: yehlisa ukuncika kwezobuchwepheshe ku-OpenAI futhi yakhe ingqalasizinda ye-AI eqinile eyakhe. U-Mustafa Suleyman, inhloko ye-Microsoft AI, ugcizelele ukuthi inhloso iwukuba ngcono kumsebenzisi wokugcina, ngokuncika kumasiginali okusetshenziswa (i-telemetry, ukuziphatha) ukuze unikeze abasizi abawusizo kakhulu futhi abaqondene nawe.

  Iyini i-chatbot futhi isebenza kanjani emhlabeni wedijithali?

Umbono weMicrosoft uthi “orchestrate uhla lwamamodeli akhethekile” ehlanganisa izinhloso nezimo ezihlukahlukene, ekhiqiza “inani elikhulu” kubasebenzisi. Inkampani ikuchaza “njengesango eliya endaweni yonke yolwazi,” isifiso esihumusheka ekuhlanganiseni i-AI emikhiqizweni echaza isigaba.

Mayelana nokuklama okunomthwalo wemfanelo, u-Suleyman uphinde wagcizelela ukubaluleka gwema i-anthropomorphisms: Ukwakhela abantu i-AI, kodwa hhayi "njengabantu bedijithali." Lokhu kusebenza ikakhulukazi kumamodeli wezwi nabasizi abanganikeza umbono wokuba nemizwa.

Ezinhlanganweni namafemu ochwepheshe, leli gagasi elisha lamamodeli liveza amathuba nezibopho. Esikhathini esifushane, okulandelayo kulindelwe: izinzuzo zangempela ku-automation, izifinyezo, ukusekelwa kwezinqumo nokukhiqizwa kokuqukethwe okukhulunyiwe ngezindleko ezilungisiwe ze-inference.

  • I-MAI-Voice-1 Unganika amandla abasizi bokuxoxisana noma okuqukethwe kwezwi (ama-podcast, izincazelo ezikhethekile) ngemiphumela yemvelo kanye nokukhiqizwa okusheshayo.
  • Ukuhlola kuqala kwe-MAI-1 ivula umnyango wezimpendulo ezizenzakalelayo, izifinyezo, okusalungiswa, nokusekelwa kwemisebenzi yombhalo, engahlanganiswa kancane kancane ku-Copilot.

Inselele iwukuqinisekisa ubumfihlo, ukuphepha kanye nokuhambisana Ukulawula. Ukuze ugweme ukukhubeka, kuwumqondo omuhle ukuqala ngabashayeli bezindiza abalinganiselwe, ukwenza uhlolo lwangaphakathi lwemiyalo nemiphumela, uqeqeshe amaqembu, futhi ugade ukusetshenziswa kwedatha (kokubili okokufaka ne-telemetry) ukuze ugweme izimanga.

Uma ukusebenza kwakho kuncike ezwini, ukubambezeleka kanye nokuhluka kwekhwalithi ye-MAI-Voice-1 kuyathandeka kakhulu. Uma ukugxila kwakho kuwumbhalo, ukubuka kuqala kwe-MAI-1 kuyathakazelisa ngokugxila kwayo kukho ukulandela imiyalelo nangohlaka lokuhlola lomphakathi olusheshisa ukufunda okuyimodeli.

Kuyasiza futhi ukucaca mayelana nemikhawulo yamanje: I-MAI-Voice-1 igxile esiNgisini kanye nokuhlola kuqala kwe-MAI-1 kusesesigabeni sokuhlola, nokusetshenziswa kukhawulelwe ezimweni ezithile. Noma kunjalo, ijubane lokuphindaphinda elihlongozwa yiMicrosoft liyashesha futhi liphakamisa intuthuko esheshayo.

Okokugcina, kubalulekile ukuthi iMicrosoft ithi izoqhubeka nokuhlanganisa amamodeli ayo, lawo ozakwethu kanye nomthombo ovulekileLe ndlela eyingxubevange ihlose u-Copilot okhetha injini engcono kakhulu yomsebenzi ngamunye, ngaphandle kokuboshelwa kubuchwepheshe obubodwa, futhi ehlose ukukhulisa inani lomsebenzisi wokugcina.

Isimemezelo sokubuka kuqala kwe-MAI-Voice-1 kanye ne-MAI-1 sibonisa isu elizimele, eligxile esivinini, ukusebenza kahle, kanye nokusetshenziswa komhlaba wangempela. Uma ukuhlanganiswa ku-Copilot kanye nokuhlola ku-LMArena kuhlanganisa imiphumela elindelwe yi-Microsoft, sizobe sibheka izinsika ezimbili ezibalulekile ze-MAI ecosystem emikhiqizweni yabathengi nephrofeshinali.

gpt-5-0
I-athikili ehlobene:
I-GPT-5: Konke mayelana noguquko olukhulu olulandelayo ku-Artificial Intelligence