Kwenzekeni
I-Google DeepMind imemezele ukukhishwa kwezinhlobo ezimbili ze-AI ezintsha: Gemini 3.8 Live kanye ne-Gemini 3.8 Live Extended Thinking. Lawa mamodeli adizayinelwe inkhulumomphendvulwano ebukhoma, ehlanganisa amakhono okucabanga esikhathi sangempela avumela ukukhuluma kanye nokucabanga ngesikhathi esisodwa. Isimemezelo sigqamisa ukuthuthukiswa kobuhlakani, ukucabanga okufanayo, namandla okusebenzisa amathuluzi namakholi we-API ngemuva kuyilapho kugcinwa ukugeleza kwengxoxo. Amamodeli atholakala nge-Gemini Live API yonjiniyela futhi ahlanganiswe nohlelo lokusebenza lwe-Gemini, i-Google Workspace, kanye ne-Search.
I-Google DeepMind yethule i-Gemini 3.8 Live kanye ne-Gemini 3.8 Live Extended Thinking, iwachaza njengamamodeli engxoxo abukhoma enkampani athuthuke kakhulu. Ukuqamba okusha okuyinhloko yikhono lokwenza eduze kokucabanga kwesikhathi sangempela, okuvumela amamodeli ukuthi acubungule imisebenzi eyinkimbinkimbi kuyilapho egcina ukugeleza kwengxoxo okungaphazamiseki. Lokhu kufezwa ngokucabanga okufanayo, lapho imodeli ingavuma izicelo ngamagama anjengokuthi 'Ake ngihlole ukuthi…' futhi nginikeze ukulandisa kwenqubekelaphambili okubukhoma kwemisebenzi yangemuva enezinyathelo eziningi.
Amamodeli adizayinelwe ukwenza amathuluzi namakholi we-API ngemuva ngenkathi uqhubeka nengxoxo. Leli khono lihloselwe ukwenza ama-ejenti ezwi athembeke kakhulu futhi alungele ukukhiqizwa konjiniyela namabhizinisi. Gemini 3.8 Live icubungula okokufaka okubukwayo eduze nesikhathi sangempela futhi izihlonze ngokuzenzakalelayo futhi ishintshe phakathi kwezilimi ezisekelwayo ezingu-97 phakathi nengxoxo. Okuhlukile Kokucabanga Okunwetshiwe kubekwe endaweni yokuphothulwa komsebenzi webanga lebhizinisi, okunikeza amandla okucabanga ajulile okugeleza komsebenzi okuyinkimbinkimbi.
Ngokomthombo, i-Gemini 3.8 Live Extended Thinking izuze indawo engu-#1 iyonke ku-Artificial Analysis' Inkomba Yekhwalithi Yenkulumo ngamaphuzu angu-82.6. Iphinde ihole ekuqedeni umsebenzi we-agent ngo-68.6% ku-τ-Voice kanye no-35.1% kubhentshimakhi yaseSierra's τ-Voice-banking, futhi ithole u-97.7% ku-Big Bench Audio. Gemini 3.8 Live ivikelekile endaweni yesibili ku-Speech Agent Arena. Umthombo uphawula ukuthi lawa mamodeli asunduza i-Pareto Frontier ngokugeleza komsebenzi okuyinkimbinkimbi ku-ServiceNow's EVA-Bench ngokulinganisa ukunemba nekhwalithi yengxoxo.
Amamodeli afinyeleleka nge-Gemini Live API, ngokusekelwa okuvela ezinkundleni zonjiniyela ezifana ne-Agora, Fishjam, LiveKit, Pipecat, Vercel, nama-Vision Agents. Lezi zinkundla zilawula ingqalasizinda yokusakaza imidiya yesikhathi sangempela, evumela onjiniyela ukuthi bagxile kulwazi lomsebenzisi. I-Google iphinde yamemezela ubudlelwano nezinkampani ezifana ne-Salesforce, i-Genspark, ne-Lumeris, ezisebenzisa amamodeli ukubambezeleka kwazo, uketshezi, namandla okushaya amathuluzi. Wonke umsindo okhiqizwe yilawa mamodeli umakwe nge-SynthID ukuze kuqinisekiswe ukutholakala nokuvikela ulwazi olunganembile.
Imininingwane yomthombo: deepmind.google ↗
Kungani kubalulekile
Lokhu kukhishwa kumelela inqubekelaphambili ebalulekile ekusebenzisaneni kwe-AI okusekelwe ezwini, okudlulela ngale kwengxoxo elula ukuya ekuqedeni okuyinkimbinkimbi, kwe-ejenti. Ngokuvumela amamodeli ukuba acabange futhi akhulume kanyekanye, i-Google ibhekana nomkhawulo oyinhloko ekuxhumaneni kwezwi kwamanje, okwenza i-AI ibe enembile ekuphatheni ukuhamba komsebenzi okunezinyathelo eziningi. Ukuhlanganiswa nezinkundla zebhizinisi ezifana ne-Salesforce namathuluzi onjiniyela afana ne-LiveKit ne-Vercel kuphakamisa ukuphusha kubenzeli bezwi abalungele ukukhiqiza. Lokhu kuthuthukiswa kubalulekile ekuthuthukisweni kwe-AI kusukela kubasizi abangenzi lutho ukuya kubahlanganyeli abasebenzayo kuzimo zochwepheshe nezomuntu siqu, okungenzeka kumise kabusha indlela abasebenzisi abahlanganyela ngayo nezinhlelo zesofthiwe eziyinkimbinkimbi ngolimi lwemvelo.
Ukwethulwa kokucabanga kanye nokukhuluma ngasikhathi sinye kudingida ibhodlela lezwi le-AI, lapho amamodeli evamise ukuma kancane ukuze acabange ngaphambi kokuphendula. Ngokugcina ukugeleza kwengxoxo ngesikhathi sokwenziwa komsebenzi oyinkimbinkimbi, lawa mamodeli enza ukusebenzisana kwe-AI kuzizwe kungokwemvelo futhi kuncane kakhulu. Lokhu kubaluleke kakhulu ezinhlelweni zokusebenza zebhizinisi lapho abasebenzisi bangadinga ukuphatha imisebenzi eminingi noma ukugeleza komsebenzi ngemiyalo yezwi ngaphandle kokulahlekelwa umongo noma umfutho.
Ukugcizelelwa kwabasebenzeli bezwi 'abalungele ukukhiqiza' kubonisa ukugudluka ukusuka kumademo e-AI okuhlola ukuya ezixazululweni ezingokoqobo, ezingaka. Ukuhlanganiswa nezinkundla zonjiniyela ezimisiwe kanye nozakwethu bebhizinisi abafana ne-Salesforce kuphakamisa ukuthi i-Google ibeka lawa mamodeli njengengqalasizinda eyinhloko yesizukulwane esilandelayo sesofthiwe eshayelwa ngezwi. Lokhu kungase kusheshise ukwamukelwa kokuxhumana kwezwi ezimbonini lapho ukusebenza kwe-handsfree kubaluleke kakhulu, njengokuhlelwa kwezinto, ukunakekelwa kwezempilo, kanye nezinsizakalo zenkundla.
Amaphuzu ebhentshimakhi aphezulu, ikakhulukazi ekuqedeni umsebenzi we-ejenti, abonisa ukuthi lawa mamodeli awekho nje angcono engxoxweni kodwa ayakwazi kakhulu ukwenza imisebenzi eyinkimbinkimbi, enezinyathelo eziningi. Lona umehluko oyinhloko emakethe ye-AI, lapho ikhono lokuqedela imisebenzi ngokuthembekile ngokuvamile libaluleke kakhulu kunokushelela kwengxoxo okungaphekiwe. Ukusebenza kahle kwezindleko okukhulunywe ngakho emthonjeni kuphinde kuphakamise ukuthi i-Google ihlose ukwenza la makhono athuthukile afinyeleleke kuhlu olubanzi lonjiniyela namabhizinisi, hhayi izinkampani ezinkulu zobuchwepheshe kuphela.
Ukufakwa kokumaka i-SynthID kuwo wonke umsindo okhiqiziwe kuyisinyathelo esiphawulekayo sokuphepha. Njengoba i-AI yezwi ikhula kakhulu, ubungozi bokushona okujulile nolwazi olungalungile luyakhula. Ngokushumeka i-watermark engabonakali, i-Google izama ukudala izinga lobuchwepheshe lokuqinisekisa umsindo okhiqizwe yi-AI, okungaba nemithelela ebanzi yokwethenjwa nokuphepha kwezokuxhumana kwedijithali.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
An agent must create a draft calendar event for Tuesday at 2 p.m. Which evidence would establish the requested result?
Ongakubuka ngokulandelayo
Gada ukubambezeleka kwangempela nokuthembeka kwesici 'Sokucabanga Okunwetshiwe' kuzimo zomhlaba wangempela, njengoba ukucabanga kanye nokukhuluma ngesikhathi esisodwa kuyinselele yobuchwepheshe. Buka ukuhlaziya okuvela eceleni kwamaphuzu ebhentshimakhi afuniwe, ikakhulukazi indawo engu-#1 Kunkomba Yekhwalithi Yenkulumo Nenkulumo. Bheka ukuthi la mamodeli asebenza kanjani ekusetshenzisweni kwebhizinisi nozakwethu abafana ne-Salesforce, nokuthi ingabe i- ye-SynthID ivimbela ngempumelelo ukusetshenziswa kabi komsindo okhiqiziwe. Ukwengeza, landelela izinga lokutholwa phakathi konjiniyela abasebenzisa i-Gemini Live API kanye nezimo ezithile zokusetshenziswa ezivela kumakhono amasha okushayela amathuluzi.
Ukusebenza komhlaba wangempela kwesici esithi 'Ukucabanga Okunwetshiwe' kuzobaluleka. Nakuba umthombo uthi unesizathu futhi ukhuluma ngesikhathi esisodwa, ukubambezeleka kwangempela, ukunemba, nolwazi lomsebenzisi ezimeni eziyinkimbinkimbi kungase kuhluke. Ukuhlola okuzimele kanye nempendulo yomsebenzisi kuzobaluleka ukuze uqinisekise lezi zimangalo.
Ukwamukelwa kwe-Gemini Live API ngabathuthukisi kanye nezimo ezithile zokusetshenziswa ezivelayo kuzonikeza ukuqonda ngevelu engokoqobo yalawa mamodeli. Uma abathuthukisi bekwazi ukwakha ama-ejenti ezwi aqinile, athembekile angakwazi ukuphatha ukuhamba komsebenzi okuyinkimbinkimbi, izobonisa amandla eqiniso obuchwepheshe.
Umthelela we- ye-SynthID kwikhwalithi yomsindo nokubona komsebenzisi kuzofaneleka ukuthi uqashwe. Uma i-watermarking iphazamisa kakhulu noma ingasuswa kalula, ingase ibukele phansi ukusebenza kwayo njengesinyathelo sokuphepha. Ukwengeza, impendulo yezinye izinkampani ze-AI kuleli zinga le-watermarking izojabulisa ukuyibona.
Isimo sokuncintisana sezwi i-AI kungenzeka sishube njengoba ezinye izinkampani zisabela ekukhishweni kwe-Google. Buka izimemezelo ezivela ku-OpenAI, Anthropic, nabanye abadlali abakhulu mayelana namakhono abo emodeli yezwi namabhentshimakhi.