GPT-6 Astra Uyasebenza

GPT-6 Astra iyimodeli ephambili ye-OpenAI, echazwa ngokusemthethweni njenge-"imodeli ehlakaniphe kakhulu emhlabeni futhi ehambisana kakhulu nezinhloso zabantu". Le modeli idlula imingcele yengxoxo yendabuko, iguquke ibe yi-AI Agent ekwazi ukusebenzisa isoftware ngokuzimela, ukuphequlula iwebhu, nokuqedela imisebenzi eyinkimbinkimbi. Ezivivinyweni ezifana ne-ARC-AGI-3, lapho ithole khona u-99.9%, kanye ne-ExploitBench, lapho ithole khona u-100%, isungule amarekhodi futhi ingakwazi ukwenza imisebenzi yobungcweti efana nokuhlela izinhlelo, ucwaningo lwesayensi kanye nokuphepha ku-inthanethi.

Computer Use: Isebenzisa isoftware yekhompyutha ngokuzimela, iphequlula iwebhu, igcwalisa amafomu, ibuyekeza i-CRM futhi ihlele amashejuli.
Ubunjiniyela besoftware: Iqedela ukuhlela izinhlelo okuyinkimbinkimbi futhi iqonde amakhosombe ekhodi; i-Terminal-Bench ne-DeepSWE zombili zifinyelela ku-SOTA.
Ucwaningo lwesayensi: Ihlaziya ngokuqondile idatha yokuhlola futhi ikhiqize amashadi; ithole u-97.6% ku-FrontierMath Tier 4.
Ukuphepha ku-inthanethi: Ithola futhi isebenzise ubungozi be-zero-day; ithole u-100% ku-ExploitBench.
Idizayini ye-3D ne-hardware: Isebenzisa i-Blender, i-Unity ne-KiCad ukwenza ukumodela kwe-3D, izigcawu zemidlalo kanye nedizayini yamabhodi esekethe e-PCB.
Imibhalo yobungcweti: Ikhiqiza amaphosta, izethulo kanye nemibhalo yebhizinisi ehlelwe kahle futhi elungele ukusetshenziswa.

Ukuqeqeshwa kwangaphambili ngezinga elikhulu kakhulu: Ngokokuqala ngqa, isebenzise ama-GPU angaphezu kuka-100,000 esikhungweni sase-Stargate e-Texas, e-United States, ukuqedela ukuqeqeshwa kwangaphambili. Lo ngomunye wamaphrojekthi amakhulu kakhulu okuqeqesha e-OpenAI kuze kube manje.
Ukuhambisana nokuphepha kwemodeli: Ihlanganisa imiphumela yeminyaka eminingi yocwaningo lokufunda ngokuqiniswa nokuhambisana, futhi isebenzisa amamodeli angaphambilini ekuqondiseni. Ezivivinyweni zemisebenzi eyeqa imingcele, izinga lokweqa igunya lehle lisuka ku-48% laya ku-0%.
Ukuphathwa komongo wesikhathi eside: I-Codex igcina amanothi phakathi kwamawindi omongo futhi isekela ukusesha imilayezo yangaphambilini nokuphumayo kwamathuluzi, ukuze imisebenzi emide ingalahlekelwa ulwazi olubalulekile.
Isu lokuphepha elinamazinga: Inguqulo evamile iyenqaba izicelo ezithuthukile zokuphepha ku-inthanethi. Amaqembu okuphepha ahlolwe angathola ukufinyelela emandleni avuleke kakhulu ngephrojekthi ye-Daybreak, kulinganiswe ukusebenza nobungozi.

Thola imvume: Bhalisela i-ChatGPT Plus/Pro/Business/Enterprise noma ucele ukufinyelela ku-API ukuze uthole imvume yokusebenzisa i-GPT-6 Astra.
Shintsha imodeli: Ku-client ye-ChatGPT noma ku-endpoint ye-API, khetha i-GPT-6 Astra njengemodeli yamanje.
Faka imiyalelo: Sebenzisa ulimi lwemvelo ukufaka imiyalelo yemisebenzi eyinkimbinkimbi, esekela ukuhlela izinhlelo, idizayini, ucwaningo lwesayensi kanye ne-automation yasehhovisi.
Vula i-Computer Use: Vula umsebenzi we-Computer Use futhi unikeze i-Astra imvume yokusebenzisa isoftware, ukuphequlula iwebhu nokubiza amathuluzi angaphandle ngokuzimela.
Ukuqinisekisa ngokusebenzisana: Thola impendulo yesikhathi sangempela ngokwenqubekelaphambili, uqinisekise izigaba ezibalulekile noma ulungise izidingo kuze kuqedwe umsebenzi futhi kuthunyelwe umphumela.

Ukuguquka okukhulu kwamakhono e-Agent: Isuka ekuphenduleni izingxoxo iye ku-AI Agent ekwazi ukuqedela imisebenzi eyinkimbinkimbi ngokuzimela, isebenzise isoftware ngokuqondile, iphequlule iwebhu futhi iphathe izinqubo zomsebenzi zobungcweti.
Imiphumela ehamba phambili kumabenchmark: I-ARC-AGI-3 ifinyelela ku-99.9%, i-ExploitBench ku-100%, kanti i-FrontierMath Tier 4 ifinyelela ku-97.6%, kuvuselelwa amarekhodi ezivivinyweni eziningi.
Ubunjiniyela besoftware obuhamba phambili: I-Terminal-Bench ne-DeepSWE zombili zifinyelela ku-SOTA, ngokuphindaphinda okumbalwa ekukhiqizeni ikhodi futhi ngaphandle kokulahlekelwa umongo emisebenzini emide.
Amakhono ocwaningo lwesayensi: Ingene ngokuqondile kusoftware yobungcweti ukuze ihlaziye idatha yokuhlola, ibambe iqhaza ocwaningweni lwezibalo futhi iqhubekisele phambili inkinga yezikhala phakathi kwezinombolo eziyinhloko.
Idizayini yokudala ne-hardware: Isekela i-Blender, i-Unity ne-KiCad ekumodeleni kwe-3D, ekwakheni izigcawu zemidlalo nasekwakheni amabhodi esekethe e-PCB.
Ukuphepha ku-inthanethi okunemiphumela emibili: Iyakwazi ukuthola ubungozi be-zero-day, kodwa ububanzi bokufinyelela bulawulwa ngokuqinile ngesu elinamazinga le-Daybreak.

Iwebhusayithi esemthethweni yephrojekthi: https://openai.com/index/gpt-6-astra/

Ubunjiniyela besoftware: Iqedela ngokuzimela ukubhala amakhodi ayinkimbinkimbi, ukulungisa amaphutha nokuthutha imininingwane yolwazi; i-Terminal-Bench ne-DeepSWE zombili zifinyelela ezingeni eliphambili embonini.
Ucwaningo lwesayensi: Ingena ngokuqondile kusoftware yobungcweti ukuze ihlaziye idatha yokulandelana kwezakhi zofuzo, ikhiqize amashadi okuhlola futhi ibambe iqhaza ocwaningweni lwezinkinga eziphambili zezibalo.
I-automation yasehhovisi: Isebenzisa ngokuzimela amathuluzi afana ne-Excel ne-Power BI, iqedele ukugcwalisa amafomu, ukubuyekeza i-CRM, ukuhlela amashejuli nokuhlaziya idatha eyinkimbinkimbi.
Idizayini ye-3D ne-hardware: Yenza ukumodela kwe-3D nokwakha izigcawu zemidlalo ku-Blender ne-Unity, noma isebenzise i-KiCad ukwenza ukwakheka kwamabhodi esekethe e-PCB.
Ukuphepha ku-inthanethi: Iskena futhi ithole ubungozi be-zero-day, isize amabhizinisi alungise amaphutha ezinhlelo, kuyilapho amasu anamazinga evimbela ukusetshenziswa kabi.