I-Microsoft iqhubekela phambili ekuzimeleni kwe-AI ngamamodeli amathathu amasha ayisisekelo

Ikhaya » Python » I-Microsoft iqhubekela phambili ekuzimeleni kwe-AI ngamamodeli amathathu amasha ayisisekelo

I-Microsoft ithatha isinyathelo esicacile esiya phambili ukuzimela okukhulu ekuhlakanipheni kokwenziwa ngokwethula amamodeli ayo amathathu ayisisekelo ahloselwe ukuqoshwa, ukwenziwa kwenkulumo kanye nokudalwa kwezithombe. Lokhu kuthutha kubonisa ukuthi inkampani ifuna inqwaba ye-AI ejulile, enezindlela eziningi elawula ngokugcwele, noma ngabe igcina ubudlelwano obuseduze bezentengiselwano ne-OpenAI busendaweni.

Lezi zinhlelo ezintsha, ezathuthukiswa ngaphansi kwe- IMicrosoft AI / MAI Superintelligence amaqembu, aklanyelwe ukuxhumana ngqo imikhiqizo efana ne-Teams ne-Azure kanye nasezinkundleni zokuhlola zangaphakathi. Empeleni, iMicrosoft ibeka isisekelo sesu lesikhathi eside lapho amamodeli ayo amboza ingxenye ekhulayo yemithwalo yemfanelo yansuku zonke, ukubhukha amamodeli angaphandle afana nalawo avela ku-OpenAI ezimweni lapho eletha khona inani elicacile nelihlukile.

Amamodeli amathathu esisekelo akhiwe yi-Microsoft okubhala umbhalo, izwi kanye nezithombe

Ukuqaliswa kugxile kumamodeli amathathu ayinhloko: I-MAI-Transcript-1 ukuze kuhunyushwe inkulumo ibe umbhalo, I-MAI-Voice-1 kombhalo ube yinkulumo kanye MAI-Image-2 ukuze kukhiqizwe okubonakalayo. Ndawonye zakha ungqimba lokuqala, olubonakala kakhulu lwendlu isitaki se-AI se-multimodal ezingaphatha umbhalo, umsindo kanye nezithombe ngaphakathi kwe-ecosystem ye-Microsoft.

Esikhundleni sokuthembela kuphela kumamodeli amakhulu, asetshenziswa kabanzi, iMicrosoft ibheja kuwo izinhlelo ezigxile emisebenzini ezishibhile futhi ezisheshayo ezimweni ezivamile zokusetshenziswa kwebhizinisi. Leyo ndlela ibaluleke kakhulu njengoba inani labasebenzisi be-Copilot kanye nezici ezisebenzisa i-AI eHhovisi, Amathimba kanye ne-Azure liqhubeka likhuphuka, ngezindleko ebezizokhula cishe ngokuhambisana nokusetshenziswa kwe-API.

Amamodeli esisekelo zalolu hlobo ziqeqeshwe kumasethi edatha amakhulu futhi ahlukahlukene ukuze kamuva zikwazi ukujwayelana nezimo eziningi ezahlukene. Lapha, lokho kusho ukunika amandla konke kusukela ekubhalweni kwesikhungo sezingcingo kanye nokufingqwa kwemihlangano kuya kumazwi okwenziwa, amathuluzi okufinyeleleka kanye nezindlela zokudala okuqukethwe okuzenzakalelayo.

I-MAI-Transcribe-1: inkulumo esheshayo, yezilimi eziningi ibe umbhalo wezilimi ezingu-25

I-MAI-Transcribe-1 iyintsha ye-Microsoft injini yokukhuluma ibe umbhalo kanye nenye yezingcezu eziyinhloko zalokhu kukhishwa. Imodeli isekela ukubhalwa kwemibhalo ku Izilimi ezahlukene ze-25 futhi ilinganiswe ngaphakathi njengesilinganiso Ngokushesha okuphindwe kabili kune-Azure Fast transcription yenkampani ekhona, okuye kwaba yindawo yokubhekisela kuyo kuphothifoliyo yayo yamanje.

Lokhu kukhula kokusebenza kubalulekile ngoba Imithwalo yemisebenzi yokubhala ibucayi kakhulu ekubambezelekeni, ikakhulukazi ezimweni zesikhathi sangempela njengemibhalo engezansi ebukhoma, ukwesekwa kwamakhasimende noma imihlangano exubile. Ukusabalala kolimi olubanzi kuhambisana nomlando we-Microsoft womhlaba wonke, okwenza kube lula kumakhasimende asemazweni amaningi ukwenza izinto zibe zijwayelekile kumhlinzeki oyedwa esikhundleni sokuxuba amathuluzi esifunda.

Ngokombono womkhiqizo, iMicrosoft ihlela ukuxhumanisa i-MAI-Transcribe-1 ngqo ku- Amaqembu e-Microsoft ukuphatha imibhalo yemihlangano kanye namagama-ncazo abukhoma. Ngokuhamba kwesikhathi, injini efanayo kulindeleke ukuthi ivele ngaphansi kwamanye amathuluzi okukhiqiza, ukuze abasebenzisi babona isivinini esingcono kanye nezindleko eziphansi ngaphandle kokuqaphela ushintsho lomkhiqizo.

Amanani abekwe ngendlela ebukhali: I-MAI-Transcribe-1 iqala cishe cishe $0.36 ngehora lomsindo ocutshunguliwe, inani elihlose ukunciphisa okunikezwayo okufanayo okuvela ku-Google kanye ne-OpenAI ngenkathi lisasebenza kwingqalasizinda yamafu ye-Microsoft.

I-MAI-Voice-1: umbhalo-ube-inkulumo osheshayo kakhulu onamazwi enziwe ngokwezifiso

Ngasohlangothini lokukhiqizwa komsindo, I-MAI-Voice-1 iyimodeli entsha ye-Microsoft ye- ukuguqula umbhalo ube yinkulumoNgokusho kwenkampani, ingakhiqiza cishe Imizuzwana engama-60 yomsindo cishe ngomzuzwana owodwa wesikhathi sokucubungula, okuyisinyathelo esiphawulekayo ezimweni zokusetshenziswa lapho ukuphendula kubalulekile khona.

Ngale kwesivinini esingakaphekwa, isithembiso esibalulekile ukusekelwa amazwi enziwe ngokwezifiso, ahambisana nomkhiqizoIzinhlangano zizokwazi ukuchaza amazwi afanelana nobunikazi bazo noma izimo ezithile zokusetshenziswa, kusukela ezingcingweni zokusekela kanye nama-agent engxoxo kuya ezintweni zokuqeqesha, ama-podcast kanye nezici zokufinyeleleka. Lelo zinga lokulawula libaluleke kakhulu njengoba inkulumo yokwenziwa ivame kakhulu futhi izilaleli zikhula zifuna kakhulu ithoni nokucaca.

I-Microsoft ihlose ngqo i-MAI-Voice-1 ku- abathuthukisi namabhizinisi akha imikhiqizo enamandla kakhulu: izikhungo zocingo, abasizi abangaphakathi nohlelo lokusebenza, amathuluzi okufunda ulimi, amapulatifomu emidiya noma iyiphi insizakalo edinga ukulandiswa okukhuliswayo. Amanani aqala cishe $22 ngezinhlamvu eziyisigidi, imodeli ihloselwe ukuba ikwazi ukusebenza kahle ngokwezimali kokubili ngamanani amancane namakhulu kakhulu.

Kusukela ekhoneni lengqalasizinda, i-MAI-Voice-1 inikezwa ngokusebenzisa Ama-API e-Azure, I-Microsoft Foundry kanye ne-MAI Playground, okuvumela amaqembu ukuthi ahlole amazwi ngokushesha bese eqhubekela ekukhiqizeni ngaphandle kokushintsha izindawo. Umqondo uwukwenza kube lula indlela ephelele kusukela ekuhlolweni kuya ekusetshenzisweni ngaphakathi kwe-stack ye-Microsoft.

I-MAI-Image-2: ukukhiqizwa kwezithombe namavidiyo kuhlanganiswe ne-stack ye-Microsoft

Imodeli yesithathu, MAI-Image-2, igxila isithombe (futhi kwezinye izincazelo, ividiyo) ukukhiqizwa okuvela ezikhulwini zombhaloNakuba inkampani ingakadalulanga yonke imininingwane yobuchwepheshe, ibeka imodeli njengomfanekiso ohambisana nezinhlelo zayo zombhalo nomsindo, okuhloswe ngazo ukwenza kube ngokuzenzakalela ukudalwa kwezimpahla zokumaketha, izithombe zomkhiqizo, amabhodi ezindaba kanye neminye imithombo yezindaba.

Ngokuthakazelisayo, i-MAI-Image-2 yaqala ukubonakala buthule kakhulu ku- Inkundla yokudlala ye-MAI, indawo yokuhlola yeMicrosoft yamamodeli amakhulu, emuva maphakathi noMashi. Isimemezelo samanje senza indima yayo ibe semthethweni njengengxenye yohlelo olubanzi I-Foundry kanye ne-Azure i-ecosystem, lapho amabhizinisi angafinyelela khona njengengxenye ejwayelekile kunokuba kube yi-demo yocwaningo oluhlanzekile.

Amanani ahlelwe futhi ukuze ancintisane: inkampani ikhomba indawo yokungena mayelana Ama-dollar amahlanu ngesigidi ngasinye samathokheni okufaka umbhalo futhi nxazonke Ama-dollar angu-33 ngesigidi ngasinye samathokheni okukhipha izithombe ezikhiqizweLezi zinombolo zihlelwe njengezilingana, noma ngaphansi, nezigaba ezifanayo ezivela kubahlinzeki abaphikisanayo ngenkathi zizuza ekuphepheni kwebhizinisi le-Microsoft kanye ne-stack yokuthobela imithetho.

Amacala okusetshenziswa asukela ku- imisebenzi yokudala ezenzakalelayo kanye nezithombe zokumaketha ezenzelwe wena ukuze kube yi-prototyping esheshayo yomklamo womkhiqizo. Kumakhasimende amaningi asevele ejwayelene ku-Azure, iphuzu elibalulekile lokuthengisa ukuthi angazama ukukhiqiza izithombe ngaphandle kokuletha umthengisi wangaphandle owengeziwe.

Ukuhlanganiswa kuyo yonke i-Azure, i-Foundry, i-MAI Playground kanye ne-Microsoft 365

Isici esichazayo salokhu kwethulwa ukuthi amamodeli amasha ahlanganiswe kahle kangakanani Amapulatifomu e-Microsoft akhona efu kanye nokukhiqizaZonke izinhlelo ezintathu – i-MAI-Transcript-1, i-MAI-Voice-1 kanye ne-MAI-Image-2 – ziyaqaliswa ngokusebenzisa I-Microsoft Foundry, indawo yenkampani yokufinyelela nokukhulisa amamodeli esisekelo.

Abathuthukisi bangaqala ngokuthi Inkundla yokudlala ye-MAI, lapho amamodeli afanayo evezwa khona ku-interface yokuhlola kakhulu. Lokho kusetha kuhloselwe ukwehlisa isithiyo samaqembu afuna ukuzama amakhono afana nokubhala, amazwi okwenziwa noma ukukhiqiza okubonakalayo ngaphandle kokuzibophezela ekuhlanganisweni okugcwele ngokushesha.

Ngasohlangothini lomkhiqizo, iMicrosoft isivele ikhomba ku- Amaqembu e-Microsoft njengomzuzi wakuqala. I-MAI-Transcript-1 ihlelelwe ukufeza imibhalo eqoshiwe kanye namagama-ncazo, kuyilapho i-MAI-Voice-1 kanye ne-MAI-Image-2 kulindeleke ukuthi zivele ngokuhamba kwesikhathi ezinhlobonhlobo Okuhlangenwe nakho kwe-Copilot kanye ne-Microsoft 365, noma ngabe abasebenzisi bokugcina bangase bangaboni uphawu lwemodeli olucacile.

Ezinkampanini, isithembiso isitaki esisodwa, esihambisanayo lapho ukubhalwa phansi, izwi nezithombe zihlala eduze kwamamodeli olimi, izinsizakalo zedatha kanye nokuhlaziya ku-Azure. Lokho kungenza kube lula ukuthobela imithetho, ukubuyekezwa kokuphepha kanye nokuphathwa kwabathengisi uma kuqhathaniswa nokuhlanganisa abahlinzeki be-AI bangaphandle abaningi.

Isu lentengo kanye nokuncintisana ne-OpenAI kanye ne-Google

Kanye nezincazelo zobuchwepheshe, iMicrosoft igxile kakhulu ku- ukuncintisana kwentengoInkampani ibeka obala la mamodeli njengezinye izindlela ezingalingana noma zinciphise okunikezwayo okufanayo okuvela ku- I-OpenAI ne-Google, ikakhulukazi ukusetshenziswa okuqhubekayo, okunomthamo ophezulu.

Amaphuzu entengo ashicilelwe – $0.36 ngehora lomsindo kwe-MAI-Transcript-1, $22 ngezigidi zezinhlamvu ye-MAI-Voice-1 kanye ne- $5 / $33 ngethokheni ngayinye eyisigidi isakhiwo se-MAI-Image-2 – akuzona nje imininingwane yobuchwepheshe. Ziyingxenye yomyalezo obanzi wokuthi iMicrosoft ifuna ukubhekwa njenge- umhlinzeki we-AI okhiqizayo osebenza kahle, osebenza kusukela ekuqaleni kuya ekugcineni kunokuba kube ngumthengisi kabusha wamamodeli abalingani kuphela.

Emakethe lapho izinhlangano eziningi zifaka khona i-AI ekusebenzeni kwansuku zonke, izindleko ngesicelo ngasinye zingaba yinguquko yesu ngokusheshaNgokuba namamodeli ayo, iMicrosoft ingalungisa ukuhwebelana phakathi kwezindleko zokubala, ubunzima bamamodeli kanye namanani abasebenzisi esikhundleni sokukhokha amamaki amakhulu kubahlinzeki bangaphandle.

Kukhona nomphumela wokubonisa: ngokugqamisa izilinganiso zayo kanye namathebula entengo, iMicrosoft itshela amakhasimende ukuthi akusadingeki ukuthi asebenzise amamodeli ezinkampani zangaphandle ukuze athole imisebenzi eyinhloko efana nokubhala phansi, inkulumo kanye nezithombe uma esezibophezele ku-Azure.

U-Mustafa Suleyman kanye nombono we-AI "ogxile kumuntu".

Amamodeli amathathu amasha avela emaqenjini aqoqwe ngaphansi IMicrosoft AI / MAI Superintelligence, eholwa ngu mustafa suleman, manje ohola iMicrosoft AI. USuleyman, owaziwa ngezindima zakhe zangaphambilini embonini ye-AI, ubelokhu eveza obala umbono awuchaza ngokuthi "i-AI yobuntu" noma ukuhlakanipha okwenziwe okugxile kumuntu.

Ekuxhumaneni kweMicrosoft mayelana nokuqaliswa, uSuleyman ugcizelela ukuthi lawa mamodeli aklanyelwe bonisa indlela abantu abaxhumana ngayo ngempela, ukubeka phambili usizo olusebenzayo kanye nokuphephaUmgomo, ngamazwi akhe, ukudala izinhlelo ezingezona amaphrojekthi ocwaningo angacacile kanye namathuluzi amaningi afanelekela imisebenzi yansuku zonke emsebenzini nasekhaya.

Uphinde waphakamisa ukuthi amamodeli amathathu amanje ukuqala nje kwephothifoliyo ebanziI-Microsoft ihlela ukuqalisa amamodeli esisekelo engeziwe nge-Foundry kanye nemikhiqizo yangaphakathi ngqo, kancane kancane yandisa amakhono ayo angaphakathi ngale kwenkulumo nezithombe ukuze ihlanganise izindlela ezengeziwe nemisebenzi ekhethekile.

Leyo mephu yendlela igcizelela inhloso yeMicrosoft yokungabonwa nje kuphela njengeplatifomu ye-AI yabanye abantu, kodwa njengomakhi wamamodeli ayo athuthukile angahambisana neminikelo evela kubalingani besikhathi eside njenge-OpenAI.

Ubudlelwano obulungisiwe kabusha ne-OpenAI kanye nomgomo wemodeli yemingcele ka-2027

Esinye sezici ezibucayi kakhulu zaleli cebo ukuthi lihlobene kanjani Ubambiswano oludumile lwe-Microsoft ne-OpenAIIzinkampani zisalokhu zixhumene kakhulu: I-Microsoft itshale imali engaphezu $ 13 billion ku-OpenAI, ibamba amamodeli ayo ku-Azure futhi ihlanganisa izinhlelo ezifana ne-GPT emikhiqizweni efana ne-Copilot.

Kodwa-ke, imibiko yakamuva ikhomba ku- ukuxoxisana kabusha kobudlelwano lokho kunikeza iMicrosoft indawo eyengeziwe yokuqhuba ucwaningo lwayo lwe-AI kanye nemigqa yemikhiqizo ngesikhathi esifanayo. USuleyman uchaze lolu shintsho njengoguquko lwemvelo, hhayi ukuqhekeka - kufana kakhulu nenkampani eklama amanye ama-chip ayo ngenkathi isathenga kubaphakeli bangaphandle.

Ngokusho kweBloomberg nezinye izitolo, iMicrosoft ihlose ukuba amamodeli ayo amakhulu, asezingeni lomngcele aqala ukusebenza cishe ngo-2027Izinhlelo ezisanda kumenyezelwa zisezingeni eliphezulu kancane kulowo mgomo: azikabekwa njengezibonelo zolimi ezisetshenziswa kabanzi, ezisezingeni eliphezulu, kodwa kunalokho zibekwe njengezibonelo zolimi ezisetshenziswa kabanzi. izingxenye ezikhethekile ezinciphisa ukuncika kuma-API abalingani emisebenzini yansuku zonke.

Empeleni, lokhu kusho ukuthi iMicrosoft ingaqhubeka nokusebenzisa amamodeli e-OpenAI afana ne-GPT-5.4 lapho enengqondo khona, kuyilapho kancane kancane ukushintshanisa amamodeli ayo noma ngabe isilinganiso sezindleko-ukusebenza noma izinto ezicatshangelwayo ezithinta ubuchwepheshe bangaphakathi. Abasebenzisi bangase baqaphele ukuthi izici ziyashesha noma zishibhile njengoba lezi zinguquko zenzeka ngemuva.

Emakethe ebanzi ye-AI, le ndlela ekabili igcizelela umkhuba ocacile: izinkampani ezinkulu zobuchwepheshe zifuna ukulinganisela phakathi kokubambisana nokwaneliseka, besebenzisa izinhlangano ukuze basheshe kodwa bakhe amakhono abo ukuze bagweme ukuboshwa ngumhlinzeki oyedwa isikhathi eside.

Ngala mamodeli amathathu, iMicrosoft itshala ifulegi ngempumelelo: ifuna ukuncintisana emazingeni amaningi e-AI stack - kusukela engqalasizinda kanye namathuluzi kuya kumamodeli ayisisekelo ngokwawo - ngenkathi isashiya isikhala kubalingani abanjengo-OpenAI lapho beletha khona amandla ahlukile. Kumakhasimende, lokho kungahunyushwa kube izinketho eziningi, amanani abukhali kanye noshintsho oluhamba kancane luye ku-AI yophawu lwe-Microsoft olusekela imikhiqizo nezinsizakalo ezijwayelekile.

I-athikili ehlobene:

La trampa de dependencia de los LLM: limites, sesgos y riesgos