Siyini Isimo Sezulu Se-GoogleOkulandelayo 2: I-AI igxumela esikhathini esizayo

Uma ungomunye walabo bantu abahlola isimo sezulu ngaphambi kokuhamba ekhaya, uzokwazi ukuthi Izinhlelo zokusebenza Izinhlelo zokusebenza zesimo sezulu manje sezicishe zibaluleke njengomakhalekhukhwini wakho. Phakathi kwezidume kakhulu yi-BBC Weather, Met Office, Apple Weather, The Weather Channel, kanye ne-AccuWeather, zonke zithembeke kakhulu ekuhleleni usuku lwakho. Nokho, yini eza nemodeli entsha ye -Google uthembisa ukukhulisa izinga: Isimo sezuluOkulandelayo 2 Iza nombono wokusheshisa, ukucwengisiswa nokwenza kube usizo kakhulu izibikezelo esixoxa ngazo ngaso sonke isikhathi.

Umphumela womzamo ohlanganyelwe phakathi kwe-Google DeepMind kanye ne-Google Research, le sistimu IA ihlose ukukhiqiza izibikezelo ngokushesha izikhathi eziyisishiyagalombili kunezindlela ezijwayelekile, ezinokulungiswa kwesikhashana okungafika ehoreni elilodwa kanye nekhono lokulingisa amakhulukhulu ezimo ukusuka endaweni eyodwa yokuqala. Lokhu kweqa sekuvele kubonakale emikhiqizweni oyisebenzisa nsuku zonke, ngoba izibikezelo kuSesho, I-Google Gemini, I-Pixel Weather kanye ne-Google Maps Platform API Abuyekeziwe ukuze asebenzise ubuchwepheshe obusha, kanye nezici ezintsha ezisendleleni ye-Google Amamephu.

Iyini i-WeatherNext 2 futhi kungani wonke umuntu ekhuluma ngayo?

I-WeatherNext 2 iyimpinda yesibili yohlelo lokubikezela isimo sezulu olusebenza ngamandla we-Google we-AI. Ngokungafani nemodeli ye-deterministic yakudala, le ndlela igxile kuyo dala izimo eziningi ezizwakalayo ezihlonipha i-physics ye-atmospheric system. Ngesimo sezulu esisodwa sokuqala, iyakwazi ukukhiqiza ngaphansi kweminithi uhla lwemiphumela esiza ukulindela kokubili ukuziphatha okungenzeka kakhulu kanye izimo ezimbi kakhulu ezinamathuba amancane, ukhiye wokulawula ubungozi.

Ushintsho lubalulekile futhi ekusebenzeni: esikhundleni sokuncika ezibalweni ezisebenza amahora amaningi kumakhompyutha amakhulu anamamodeli ezinombolo, i-WeatherNext 2 iqhuba iziqondiso zayo ku- I-TPU (Iyunithi Yokucubungula I-Tensor) Futhi ithola amakhulu ezibikezelo ngesikhathi esirekhodiwe. Lokhu kusebenza kahle akukona nje ukuhlobisa: kusho izinqumo ezisheshayo lapho kudingeka isinyathelo esiphuthumayo, isibonelo, lapho kubhekene neziphepho ezinamandla noma umoya othinta ingqalasizinda ebalulekile.

Isebenza kanjani: I-Functional Generative Network ngemuva kwemodeli

Inhliziyo yobuchwepheshe yesistimu ibizwa ngokuthi I-Functional Generative Network (FGN)Lesi sakhiwo sijova "umsindo" olawulwayo ngokuqondile endaweni yokusebenza yemodeli, kusiza ukulingisa ukugcina ukuvumelana komzimba nokuncika phakathi kokuhluka. Empeleni, esikhundleni sokulungisa i-trajectory eyodwa, imodeli ikhiqiza izindlela ezimbalwa ezingaba khona zesimo somkhathi ezifanelana futhi ezihambisana nemithetho yesistimu ngokwayo.

Indlela enembile yokuqonda lokhu ukucabanga nge-orchestra eprakthiza ucezu olufanayo ezigabeni: iqembu ngalinye lisebenza engxenyeni yalo ngokuhlukahluka okuncane, kodwa Zonke zihlangana ngokuzwana uma kuhlangene. Ngokufanayo, i-FGN ivumela amanethiwekhi ahlukene ukuthi afunde izici zesikhathi ngokwehlukana, ethule ukuhlukahluka, futhi namanje yakha izibikezelo ezihlanganyelwe ezingaguquki ezingeni lomzimba.

Le ndlela inqoba imikhawulo yezibikezelo "ezincane", lezo ezichaza okuguquguqukayo okuhlukile (njengezinga lokushisa eliqondile endaweni ethile nesikhathi). I-FGN isiza ukushintshela kokuthi izibikezelo ezihlangene lapho izinga lokushisa, umoya, umswakama, noma obunye ubukhulu buguquka ngendlela ehlobene, okuthile okubalulekile ekuthwebuleni izenzakalo eziyinkimbinkimbi kanye nokuhlanganisa amathuba aphansi izehlakalo ezeqisayo ngaphakathi kobubanzi bezimo.

Esinye isici esibalulekile ukuthi imodeli iqeqeshwa ngalokho abacwaningi abakuchaza ngakho "idatha exekethile" mayelana nesimo sezuluIzilinganiso zezinto eziguquguqukayo ezifana nomoya, izinga lokushisa, noma umswakama ezindaweni ezithile. Kusukela kule datha, i-FGN iyakwazi ukucabangela futhi ilindele ukuguquguquka kwe izinhlelo zezulu zesikali esikhulu, khomba izindawo ezisengozini yokushisa kwamagagasi noma ulinganise ukuthi ipulazi lomoya lingakhiqiza amandla angakanani ngaphansi kwezimo ezizayo.

Isivinini, ukulungiswa, nama-horizons wokuqagela

Ngokwamamethrikhi, ukweqa okuqhathaniswa nenguqulo yangaphambilini kubalulekile: ngokusho kwe-Google, WeatherNext 2 Idlula u-99,9% wezinguquko ezihlaziyiwe. (izinga lokushisa, umoya, umswakamo nokunye) kuyo yonke imikhathizwe Izinsuku eziyi-0 kuye kwezingu-15Ngaphezu kwalokho, inikeza isinqumo sesikhathi esingafika ehoreni elilodwa, okuvumela izibikezelo ezingcono kakhulu lapho imininingwane eyengeziwe yenza umehluko, njengokuphatha igridi yamandla noma imisebenzi yezokuthutha.

Ukusebenza kwekhompyutha nakho kuyawushintsha umdlalo. Lapho ukulingisa okuvamile kudinga amahora ambalwa kuma-supercomputersI-WeatherNext 2 ibala amakhulukhulu ezimo ngaphansi kweminithi kusetshenziswa i-TPU. Futhi kukhona i-nuance yokusebenza ethokozisayo: uhlelo lucubungula izimo zakamuva zomkhathi futhi, ukusuka lapho, lukhiqize izibikezelo ezine zamahora ayisithupha ngayinye ngosuku, ihlanganisa ukuhluka kokubonga ku-FGN ukumboza uhla lwamathuba ngokuningiliziwe.

• Ama-horizons ahlanganisiwe: kusuka ekusakazeni manje nangesikhathi esifushane kufika ezinsukwini eziyi-15, ngokuthuthukiswa okucacile kokunemba.
• Ukulungiswa kwesikhashana: kuze kufike ehoreni elilodwa, kuwusizo ezinqumweni ezisheshayo emikhakheni ebucayi.
• Umthamo wesimo: amakhulu ngaphansi kweminithi ukusuka kusimo esisodwa sokuqala.
• Ikhompyutha esebenzayo: ukubulawa ku-TPU, ngaphandle kwesidingo se-supercomputing enkulu.

Ukuhlanganiswa emikhiqizweni ye-Google nokutholakala emphakathini

Ubuchwepheshe be-WeatherNext 2 sebuvele buthuthukisa ukuthuthuka Sesha, Gemini, Pixel Weather kanye ne-Google Maps Platform API. I-Google iphinde yamemezela ukuthi emasontweni azayo sizobona ukuthuthukiswa okubonakalayo ku-Google Amamephu okuhlobene nolwazi lwesimo sezulu, okwenza kube lula ukuhlola ngokubuka nje izimo kanye nokuvela kwemvelo okulindelekile ngokwezimo ezihlukahlukene.

Mayelana nokufinyelela kwedatha, i-Google isho lokho Izibikezelo ze-WeatherNext 2 ziyatholakala ku-Earth Engine naku-BigQueryNgaphezu kwalokho, baye bethula i-a Uhlelo lokufinyelela kusenesikhathi e-Vertex AI (I-Google Cloud) evumela izinkampani namaqembu ochwepheshe ukuthi basebenzise iziqondiso ezenziwe ngokwezifiso futhi balungise imodeli ngokwezidingo zabo, indlela eqondile yokutholwa ezindaweni zomhlaba wangempela.

Uma ubheka phambili, i-Google DeepMind nethimba le-Google Research liyaqinisekisa ukuthi banjalo ukuphenya amakhono amasha ukuhlanganisa imithombo yedatha eyengeziwe, unwebe ukufinyelela, futhi uqhubeke nokuguqula ucwaningo olusezingeni eliphezulu ezinhlelweni ezinomthelela omkhuluIsibopho esishiwo ukwenza amathuluzi akamuva atholakale emphakathini womhlaba ukuze kusheshiswe inqubekelaphambili yesayensi nokuthuthukisa ukwenziwa kwezinqumo.

Kusukela kumalebhu kuya empilweni yansuku zonke: yiziphi izinguquko kumsebenzisi

Umehluko osebenzayo ukuthi, kusukela kokokufaka okukodwa, i-WeatherNext 2 can belingisa amakhulukhulu okuziphendukela kwemvelo okungenzeka Izokukhombisa izimo okungenzeka kakhulu nezimbi kakhulu okufanele uziqaphele. Lokhu kuwusizo ikakhulukazi ekuhleleni. ukuhamba, izindiza, uhambo lokuya emsebenzini, amaketanga okuhlinzeka noma ukusebenza kwezinkampani eziletha izimpahla, lapho ukushintshashintsha kwesimo sezulu okuncane kusho izindleko noma ukubambezeleka.

Uma usuvele usebenzisa ifoni ye-Pixel, uyazi ukuthi uhlelo lwayo lwesimo sezulu lwaziwa ngokucaca kwalo. Nge-WeatherNext 2, izibikezelo ku-Pixel Weather nakusesho Bazuza ukunemba nesivinini, ukuze ukwazi ukuhlola ulwazi olubalulekile "ngokubheka nje," okuhlanganisa nobubanzi bamathuba esikhundleni sesibalo esisodwa esinethemba elishintshayo ngomzuzu wokugcina.

I-Google iphakamisa ukuthi lokhu kusengaphambili kunciphisa ukuncika ama-radar namasathelayithi njengoba bezisetshenziswa kuze kube manje, njengoba imodeli ifunda amaphethini emiqulwini emikhulu yomlando kanye nedatha yakamuva kakhulu. Kunoma yikuphi, ubuhle besimiso esisha akukona ukuthatha indawo yonke, kodwa dala isibikezelo sezulu esiwusizo kakhulu ukuhlanganisa amasignali ahlukene kanye nokuvala ngokushesha ukungaqiniseki.

Sebenzisa izimo: kusuka kumandla kuya ekuphathweni kwezimo eziphuthumayo

• Amasevisi ezimo eziphuthumayo: Linganisa kusenesikhathi izimo zeziphepho ezinamandla, izikhukhula noma umoya onamandla ukuze uvule amaphrothokholi kusenesikhathi.
• Amandla avuselelekayo: ukulinganisa ukukhiqizwa kwesikhathi esizayo emapulazini omoya noma elanga, ukulungisa ukuhlinzekwa kwesidingo ngokunemba okukhulu.
• Ezokuthutha nezokuthutha: Lungiselela imizila namafasitela asebenzayo lapho ubhekene nemvula, iqhwa noma ukuvunguza komoya, unciphise ukubambezeleka nezindleko.
• Umkhakha womphakathi kanye nomshwalense: hlaziya ubungozi ku izimo ezimbi kakhulu amathuba aphansi kanye nomthelela omkhulu ekuphenduleni kosayizi kanye nokufakwa.
• Ucwaningo nemfundo: hlola ukuguquguquka kwezehlakalo ezeqisayo bese uqhathanisa izimo ezihlangene ukuqonda kangcono umkhathi.

Inikeza ini uma iqhathaniswa nezindlela zendabuko?

Amamodeli wesimo sezulu wezinombolo axazulula izilinganiso ezibonakalayo zomkhathi ngamagridi anokucaca okuphezulu futhi, ngokwemvelo yawo, Zidinga amandla amaningi ekhompyuthaI-WeatherNext 2's AI ayizibi i-physics: ifunda kudatha yomlando kanye nesimo samanje ukuze ikhiqize izimo ezingaguquki ezizwakalayo ngaleyo mithetho, kodwa ikwenza ngezindleko eziphansi zokubala nangezikhathi ezimfushane kakhulu.

Inzuzo yayo yokuncintisana yikhono loku ikhava ngokungenzeka uhla lwamathuba, okuhlanganisa izimo ezimbi kakhulu ezifanele ukuhlela. Lapho ngaphambilini isethi enkulu yokulingisa ebizayo yayidingeka ukuze kulinganiswe lokho kungaqiniseki, i-FGN yethula ukuhlukahluka "okuhlakaniphile" okuhambisanayo ngokomzimba, okuvumela ukuhlola. amakhulu ama-trajectories cishe ngesikhathi sangempela.

Imikhawulo, izibopho kanye nokwenza okuhle

Kuhle ukukhumbula okuthile okuyisisekelo: umkhathi wuhlelo olunesiphithiphithiNgisho ne-AI, i-100% yezibikezelo ezingenaphutha ngeke zifezwe. Ingozi yethuluzi elifinyeleleka kakhulu futhi elibonakala linembile ukuthi umsebenzisi ojwayelekile uzothatha iziqinisekiso lapho kukhona okungenzeka kuphela. Lapha, kubalulekile ukuthi izixhumanisi ukuxhumana ukungaqiniseki kanye nokuthi kugcinwe ulwazi oluncane lwezezulu.

Kunamazwi aphakamisa ukungabaza ngokuthi kuwumqondo omuhle yini.ukubeka WeatherNext 2 ezandleni zanoma ubani"Ngaphandle kokuhlunga okungochwepheshe, njengoba nje i-algorithm ethola umdlavuza idinga isazi se-radiologist ukuze siqinisekise umphumela. Ithuluzi linamandla amangalisayo, kodwa incazelo eguquguqukayo, ikakhulukazi uma kubhekwa..." izimo zokuphepha kanye nokuzivikela, kusadinga ukwahlulela kochwepheshe.

Okuhlangenwe nakho kwakamuva ngezenzakalo ezinzima kubonisa lokhu. Kuziqephu ezifana ne-a I-DANA eValenciaUkwenza izinqumo sekwephuzile kungawenza abe mubi kakhulu umthelela. Uhlelo olukuhlinzeka ngokushesha ngezimo ezimbi kakhulu futhi likuxwayise ngokusobala, zuza amaminithi abalulekile ukwenza kusebenze izinhlelo kanye nokunciphisa izingozi. Yilapho ukuthuthukiswa kwejubane nokulungiswa kushintsha kakhulu.

Isihluthulelo sizoba ukuhlanganisa okuhle kakhulu kwemihlaba yomibili: i Amandla e-AI ukukhiqiza izimo ezinothile nezisheshayo, kanye nokuqinisekiswa kochwepheshe ukuze bacwengisise imilayezo nemihlahlandlela yesenzo. Uma, ngaphezu kwalokho, ukuxhumana okuqotho kugcinwa ngokuphathelene nobubanzi namathuba, umphumela uyoba isakhamuzi esilungiselelwe kangcono esingenawo umuzwa wokuqiniseka ongamanga.

Ukufinyelela kwezinkampani, ososayensi, nonjiniyela

Kulabo abadinga ukusebenza ngedatha esikalini, i-Google ivule i- Izibikezelo ze-WeatherNext 2 ku-Earth Engine ne-BigQueryukwenza kube lula ukubonisana, ukuhlaziya, kanye nokukhombana ngeminye imithombo. Futhi kulabo abafuna ukujwayela imodeli esimweni sabo esithile, kukhona a Uhlelo lokufinyelela kusenesikhathi e-Vertex AI ngosekelo lwencazelo yemodeli yangokwezifiso kanye nokusetshenziswa ezindaweni zokukhiqiza.

Umgwaqo uhlanganisa ukuhlanganisa imithombo yedatha emishaukwandisa ukufinyelela nokuqhubeka nokuhumusha intuthuko yocwaningo ibe nomthelela omkhulu, izinhlelo zokusebenza zomhlaba wangempela. Ngalo mfutho, ama-ejensi kanye nezinkampani zingasheshisa ukushintshela kwazo ku- ukwenza izinqumo okuqhutshwa yidatha kuphi el tiempo izimo zomkhathi, kude nokuba yisici esingalindelekile, ziba yingozi elinganisekayo nelawulekayo.

I-WeatherNext 2 yethulwa njengokukhuphuka kwekhwalithi ekubikezelweni kwesimo sezulu: ngokushesha izikhathi eziyisishiyagalombili kunezindlela zendabuko, ezikwaziyo kudlule ama-99,9% okuguquguqukayo Uma kuqhathaniswa nokwandulelayo endaweni eyizinsuku ezi-0 kuye kweziyi-15, futhi ilungele ukukhiqiza amakhulu ezimo ngaphansi kweminithi ukusuka kusimo esisodwa sokuqala. Ukuhlanganiswa Nosesho, i-Gemini, i-Pixel Weather, ne-API Yamamephu sekuvele kuyaqhubeka, kuyilapho idatha ku-Earth Engine ne-BigQuery, kanye nokufinyelela kwangaphambi kwesikhathi ku-Vertex AI, kuvula umnyango wokusetshenziswa okungochwepheshe nokwesayensi, okuqondiswa kahle, kungenza umehluko wangempela endleleni esilungiselela ngayo lokho okungenzeka.

I-athikili ehlobene:

Indlela Yokuhamba Isikhathi Ku-Google Amamephu: Umhlahlandlela Ophelele Wokuhlola Noma Kukuphi Okwedlule

Isaka

Umbhali oshisekayo ngomhlaba wamabhayithi nobuchwepheshe ngokujwayelekile. Ngiyathanda ukwabelana ngolwazi lwami ngokubhala, futhi yilokho engizokwenza kule bhulogi, ngikubonise zonke izinto ezithakazelisayo kakhulu ngamagajethi, isofthiwe, ihadiwe, izitayela zobuchwepheshe, nokuningi. Inhloso yami ukukusiza ukuthi uzulazule emhlabeni wedijithali ngendlela elula nejabulisayo.