- I-Hybrid Architecture: ukwenziwa kwasendaweni okunama-enclave futhi kugxumele efwini elivaliwe (TIE phezu kwama-TPU) anobufakazi obukude.
- Ubumfihlo obungaqinisekiswa: idatha esetshenziselwa umsebenzi ngamunye kuphela, ukubethela ngasemaphethelweni, i-IP edluliselwe ngamathokheni angaboni, kanye nokuhlolwa kwangaphandle.
- Okwenzeka emhlabeni wangempela ku-Pixel: I-Magic Cue, Isiqophi Sezilimi Eziningi, Umbuthano Wosesho, ukuhumusha nokunye, ngokulawula komsebenzisi.
Ubuhlakani bokwenziwa sebuyingxenye ebalulekile yempilo yethu yansuku zonke, kodwa kanye nabo kuye kwafika ukukhathazeka mayelana nobumfihlo: isiphakamiso ngasinye, isifinyezo ngasinye, kanye nesaziso ngasinye esihlakaniphile sincike kudatha yomuntu siqu. Kulesi simo, i-Google iveza i-Private AI Compute njengendlela eshintsha umdlalo , indlela eklanyelwe ukusebenzisa amamodeli anamandla ngaphandle kokuvula amasango edatha yethu.
Okokuqala, ake sicacise: I-Private AI Compute (PAC) akuyona uhlelo lokusebenza noma isevisi evamile . Luwuhlaka lobuchwepheshe oluchaza kabusha ukuthi imisebenzi ethile ye-AI yenziwa kuphi futhi kanjani ngaphakathi kwe-Pixel ecosystem, futhi kancane kancane, kumadivayisi amaningi. Noma nini lapho ihadiwe ivuma, imisebenzi isebenza endaweni; uma kudingeka amandla amaningi okucubungula, ukubala kuthuthelwa efwini elivikelekile. Umgomo ukuqinisekisa ukuthi kusetshenziswa izindlela zokuvikela ubumfihlo ezifanayo kokubili kudivayisi yeselula kanye nasefwini.
Iyini i-Private AI Compute futhi kungani ifika manje?
Isiphakamiso se-Google sihlose ukulinganisa phakathi kokusebenza kanye nobumfihlo: ukucubungula okudingekayo emsebenzini ngamunye ngaphandle kokudalula ulwazi olubonakalayo . Uma idivayisi inamandla, imodeli isebenza kudivayisi ngokwayo; ngaphandle kwalokho, umsebenzi udluliselwa efwini elivikelekile elinokulawula kwehadiwe kanye nokubethela.
Le ndlela ibhekana neqiniso elingajabulisi: namuhla, ama-AI athandwa kakhulu asebenza kumaseva, lapho idatha yomsebenzisi ingagcinwa okwesikhashana noma isetshenziswe ngisho nokuqeqeshwa uma ingakhutshazwa. I-PAC ihlose ukunikeza izinzuzo ezifanayo ekucubungulweni kwendawo—ukubambezeleka okuphansi kanye nokulawula okukhulu—kodwa ngamandla e-Gemini ezimweni eziyinkimbinkimbi.
I-Google iqinisekisa abasebenzisi ukuthi ulwazi oluthunyelwe ku-PAC lusetshenziselwa kuphela isenzo osicelayo - isibonelo, ukufingqa inothi noma ukuhumusha ingxoxo - futhi luyasuswa ngemva kokuqedwa. Ngaphezu kwalokho, umsebenzisi ugcina ukulawula ukuthi yimiphi imisebenzi engathembela efwini eliyimfihlo nokuthi yimiphi engenakukwazi, ngokushintsha okucacile nezimvume ngaphakathi kwesistimu.
Umphumela ukhomba omakhalekhukhwini namakhompyutha asebenza sengathi ane-"invisible coprocessor" efwini: isivinini esengeziwe uma sidingeka kanye nobumfihlo obungaqinisekiswa njengesimo esibalulekile . Futhi lokhu akuyona into engavamile: kukhona umkhuba ocacile emkhakheni wokuqinisa ukuvikelwa ngaphandle kokudela i-AI esezingeni eliphezulu.
Isakhiwo sobuchwepheshe: kusuka kudivayisi kuya kumafu avaliwe
Ngemisebenzi efanelana nedivayisi yeselula, i-Google yakhe indawo yokusebenza ehlukanisiwe. Sikhuluma ngama- Titanium Intelligence Enclaves kudivayisi: izikhala ezivikelwe ezihlukaniswe nesistimu yokusebenza nezinye izinhlelo zokusebenza. Lapho, idatha ibethelwa kokubili ngesikhathi sokuthutha kanye nokuphumula futhi iqinisekiswe ngesiqinisekiso esikude ukuqinisekisa ukuthi isofthiwe kanye nehadiwe kunjengoba kulindelekile futhi akuphazanyiswanga, ukuze kungabikho muntu—ngisho ne-Google—ongabona okwenzekayo ngaphakathi.
Okwamanje, amamodeli alungiselelwe ukusetshenziswa kwendawo, njenge -Gemini Nano , ahlinzeka ngomongo kanye nezimpendulo ezisheshayo ngaphandle kokushiya idivayisi. Ngezici ezifana ne-Circle to Search noma ezinye izifinyezo ze-Recorder, ifoni isebenza njenge-AI yayo yangasese efunda ngokusetshenziswa kodwa ayivezi lolo lwazi.
Uma umsebenzi udinga ukubalwa okwengeziwe, ifu eliyimfihlo liyaqala ukusebenza. Kulesi sinyathelo, konke kusebenza ngaphakathi kwezingxenye zehadiwe engqalasizinda ye-Google, ikakhulukazi i-Titanium Intelligence Enclaves (TIEs) kuma-TPU enziwe ngokwezifiso . Ngaphambi kokuvumela ukudluliswa kwedatha, uhlelo lwenza isitifiketi/ubufakazi obukude ukuqinisekisa ukuthi ingxenye ihlanzekile, isebenzisa i-binary efanele, nokuthi indawo ihambisana nezinqubomgomo zokuphepha.
I-Google ifingqa ukwakheka ngezinsika ezintathu , ngaphezu kokuphepha, eziqinisa ukwethenjwa komphakathi:
- Ukuhlukaniswa kwe-Cryptographic: ukucubungula kwenziwa kuma-enclaves aqinisekisiwe; opharetha abakwazi ukufinyelela idatha ngombhalo ongenalutho.
- Izimvume ezisobalaUmsebenzisi unquma ukuthi yimiphi imisebenzi engasebenzisa i-PAC; akukho okukhululwa ngaphandle kwemvume.
- Ukungafihli kwezobuchwephesheAmadokhumenti, ukucwaninga, kanye nemininingwane yokusebenzisa ukugwema "amabhokisi amnyama".
Ngaphezu kwalokhu okungenhla, i-PAC ifaka ezinye izendlalelo. Isevisi yokulawula—eyaziwa ngaphakathi njenge- Borg Prime —iqinisekisa ubufakazi bokufakaza ngaphambi kokugunyaza noma yikuphi ukubala okubucayi. Ukuze ihlukanise izicelo kobunikazi, uhlelo lusebenzisa i -IP blinding relay ehambisa izicelo kubantu besithathu njenge-Cloudflare kanye ne-Fastly futhi isebenzisa ama-blind tokens (Blind RSA) . Lokhu kwenza kube nzima ukuxhumanisa okuqukethwe okuthile nomsebenzisi othile, ngisho nangaphakathi kwengqalasizinda uqobo.
Ukudluliselwa kanye nokugcinwa kwesikhashana kuvikelwa ngokubethela kusukela ekuqaleni kuya ekugcineni; futhi, uma umsebenzi usuqediwe, idatha iyalahlwa . Konke lokhu kusekelwa yizinyathelo ezishicilelwe yinkampani: i- Secure AI Framework , i- AI Principles , kanye ne- Google Privacy Principles , ezisungula indinganiso efanayo yokuphepha kokusebenza.
Ukuze kwakhiwe ukwethembana, i-Google izibophezele ekuhlinzekeni ngemibhalo yobuchwepheshe enemininingwane , ivumela ukuhlolwa kwangaphandle, kanye nokushicilela izingxenye zekhodi yesistimu ukuze abacwaningi bayibukeze. Umyalezo ucacile: ubumfihlo lapha abuyona nje isithembiso, kodwa kuyinto engaqinisekiswa futhi ihlolwe.
Iziphi izici ezenza zisebenze: izibonelo zomhlaba wangempela ku-Pixel
Esinye sezici zokuqala ezizuza kule ndlela yi -Magic Cue . Lo msizi womongo ukhiqiza izincomo ngokusekelwe kulokho okusesikrinini sakho—i-imeyili, ikhalenda, amamephu, imiyalezo—futhi uziveza ngqo lapho kudingeka. Nge-PAC, lezi ziphakamiso zibalwa ngaphandle kokudalula okuqukethwe kwakho , kuncike endaweni yokugcina idatha yendawo noma efwini eliyimfihlo kuye ngomsebenzi.
Uhlelo lokusebenza lwe-Recorder luzuza izinzuzo ezinkulu ngezifinyezo ezizenzakalelayo ngezilimi eziningi. Ukubhalwa kwemibhalo—okungaba bucayi kakhulu—kucutshungulwa ngesendlalelo sokuvikela se-PAC, okuqinisekisa ukuthi kusetshenziswa kuphela esifinyezweni esithile osicelile futhi siyalahlwa lapho siqediwe. Umsebenzisi ugcina umphumela wokugcina, hhayi umkhondo wedatha ejikeleza kumaseva.
I-Google iphinde ikhuthaze izici ezazivele zidumile ku-Pixel, njenge- Circle to Search . Ukuqashelwa komongo kusebenza endaweni noma nini lapho kungenzeka, futhi uma kudingeka amamodeli athuthukile, asetshenziswa efwini eliyimfihlo, okugcina umkhawulo wobumfihlo ungathintekile.
Ezinye izibonelo ezisebenzayo zifaka phakathi ukuhumusha kwesikhathi sangempela , izimpendulo zemvelo ze-AI ezingxoxweni, kanye nosizo lokwenza izinqumo ngokushesha, lapho uhlelo luhlanganisa lokho olukubonayo nomongo wakho ukuze lukunikeze inkomba oyidingayo. Ekuqaleni, i- Pixel 10 izohola indlela kulezi zindlela.
I-ecosystem iphinde ingeze izici eziwusizo zansuku zonke: izifinyezo zesaziso ezihlakaniphile zokunciphisa umsindo nokugxila kulokho okubalulekile; ukusiza ukuhlela nge-Gemini ku-Messages ukuze kuhunyushwe kabusha imibhalo ngethoni engcono; amathuluzi okuhlela izithombe athuthukile ngaphandle kwezinhlelo zokusebenza zangaphandle; ngisho nemodi yokonga idatha ku-Google Maps eyenzelwe uhambo olude. Konke kuhambisana nomqondo owodwa: usizo olwengeziwe, ukungqubuzana okuncane, nobumfihlo njengesidingo sokuklama.
Izilawuli, ubumfihlo obungaqinisekiswa, kanye nokuhlola okuzimele
Ku-PAC, umsebenzisi uhlala elawula: akukho okuthunyelwa ngaphandle kwemvume . I-Android inikeza ama-toggles okuvumela noma ukukhubaza izici ezithembele efwini eliyimfihlo futhi ibonisa izaziso ezicacile lapho umsebenzi udinga lokhu. Ifilosofi ilula: unquma ukuthi ifoni yakho yabelana ngani futhi ngaphansi kwaziphi izimo.
I-Google ihambisana nepulatifomu ngolwazi lomphakathi nolungahlolwa : iziqondiso zobuchwepheshe, ukubuyekezwa kwangaphandle, kanye nokushicilelwa kwekhodi yezingxenye ezibalulekile, ukuze abacwaningi abazimele bakwazi ukuhlola ukuthi uhlelo lwakhiwe kanjani. Akukhona ngokholo, kodwa ngobufakazi obungaqinisekiswa.
Eqinisweni, ukwakheka kubuyekezwe yinkampani yezokuphepha kwe-inthanethi i-NCC Group . Umbiko wayo womphakathi uphetha ngokuthi ukwakheka kunikeza isivikelo esiqinile kubaphangi abanonya—kufaka phakathi ingozi evela kumsebenzi oyisigebengu—nokuthi kuletha ngempumelelo ukuphepha kokubala endaweni eduze nalokho kokusebenza kwefu. Kuphinde kubonise umkhawulo ongenakugwenywa: uma i-Google, njengenhlangano yonke, inganquma ukwephula imithetho, izolawula kokubili ihadiwe kanye nohlelo lokuqinisekisa. Noma kunjalo, umcwaningi uchaza i-PAC "njengemodeli engcono kakhulu engenzeka " namuhla yobumfihlo befu.
Ukubuyekezwa kuveze ezinye izinkinga ezinobungozi obuphansi, ikakhulukazi ubuthakathaka obungaba khona obuhlobene nesikhathi ku-IP relay, ngokombono, obungasiza ekutholeni kabusha abasebenzisi ngaphansi kwezimo ezithile kakhulu. I-Google ithi "umsindo" wethrafikhi yangempela kanye nezinyathelo zokunciphisa kwenza lokhu kungabi nokwenzeka, nokuthi umklamo uzoqhubeka nokuqiniswa ngama-patches kanye nokuthuthukiswa okuqhubekayo.
Inzuzo ebalulekile yokuthula kwengqondo komsebenzisi ukuthi i-PAC ikhawulela ukusetshenziswa kwedatha emsebenzini othize bese iyisusa lapho isiqedile . Ngaphezu kwalokho, konke ukuxhumana kuyabethelwa, kokubili phakathi kwedivayisi yakho namasayithi eseva kanye nangaphakathi kwenethiwekhi yeseva uqobo, okunciphisa indawo yokuhlasela kokubili ekuthuthweni nasekuphumuleni.
I-Ecosystem nokuqhathanisa: ukuya ku-AI eyimfihlo kakhulu
Amafoni e -Pixel 10 aletha ukuhlanganiswa okujulile kwe-PAC, kanye nokuthuthuka okukhulu ku-Magic Cue, ukuhumusha okusheshayo, ukubhalwa okunembile kakhudlwana, kanye nezimpendulo ezibushelelezi. I-Google isivele imemezele ukuthi lobu buciko buzokwandiswa kancane kancane kwamanye amadivayisi nezinsizakalo, kuhlanganiswe amamodeli endawo namafu ezinhlelo zokusebenza ezibaluleke kakhulu.
Ababodwa kule ndlela: I-Apple ivule indlela nge-Private Cloud Compute—ifu elibethelwe le-Apple Intelligence—futhi abakhiqizi abanjengo- OnePlus ne-OPPO bakhethe i-Private Computing Cloud, amapulatifomu ahlanganisiwe asabalalisa imisebenzi phakathi kwedivayisi kanye nezindawo ezibethelwe. Umkhuba ucacile: i-AI enamandla, yebo; ubumfihlo, futhi.
Ngaphandle kokuncintisana, i-Google ifuna ithoni ehambisana nemithetho yaseYurophu naseMelika: amabhokisi amnyama ambalwa kanye nokucaca kobuchwepheshe okwengeziwe. Yingakho kugcizelelwa ukuhlolwa kwamabhuku, ubufakazi obukude, kanye nokulawula okubonakalayo. Empeleni, i-PAC iyinguquko yezobuchwepheshe kanye nesinyathelo esihlakaniphile sokukhombisa ukuthi kungenzeka ukuncintisana ku-AI ngaphandle kokusebenzisa ulwazi olubucayi.
Isiqondiso sicacile: uma sifuna abasizi abaqonda umongo, abafingqa, abahumusha, noma abenze izincomo ezinembile, sidinga ukulinganisela amakhono kanye nezivikelo. Nge-Private AI Compute, i-Google izama ukuhlonza leso silinganiso sibe yi- AI ewusizo, esheshayo, futhi eyandayo , lapho umsebenzisi enezwi futhi ubuchwepheshe buziphendulela khona.
