Google ayaa si aamusnaan ah u cusbooneysiisay Deep Think, ARC-AGI-2 ayaa si toos ah u gaadhay 84.6%

Hadda, Google DeepMind waxay cusbooneysiisay qaabka gaarka ah ee sababaynta ee Gemini 3, Deep Think, dhibcaha ayaa si toos ah u burburiyay jaantusyada.

Ogsoonow, ARC-AGI-2 hadda waxaa loo aqoonsan yahay inay tahay heerka ugu sarreeya ee lagu tijaabiyo awoodda sababaynta AI, moodelna hore uma uusan helin dhibco aad u wanaagsan.

Deep Think ee la cusbooneysiiyay ayaa helay 84.6%, isbarbardhig: Claude Opus 4.6 waa 68.8%, GPT-5.2 waa 52.9%, xitaa Gemini 3 Pro Preview ee iyaga u gaarka ah ayaa ah 31.1%.

Kor u kaca weyn.

Wax ka badan Sababaynta

Deep Think hamigeedu si cad ugama eka sababaynta.

Heerka loo yaqaan 'Humanity's Last Exam' oo loo yaqaan "Imtixaankii ugu dambeeyay ee aadanaha", Deep Think wuxuu helay 48.4%, imtixaankani wuxuu daboolayaa su'aalaha ugu adag ee xagga xisaabta, sayniska, iyo injineernimada. Claude Opus 4.6 wuxuu helay 40.0%, GPT-5.2 waa 34.5%.

Barnaamijyada dhanka sidoo kale waa xoog badan:

Codeforces, Deep Think wuxuu gaadhay Elo 3455, halka Gemini 3 Pro Preview uu yahay 2512, Claude Opus 4.6 waa 2352.

Intaa waxaa dheer, heerka MMMU-Pro ee fahamka iyo sababaynta badan, Deep Think sidoo kale wuxuu ku hogaaminayaa 81.5%, laakiin farqiga u dhexeeya shirkadaha halkan maaha mid aad u weyn: Gemini 3 Pro Preview 81.0%, GPT-5.2 79.5%, Claude Opus 4.6 73.9%.

Marka laga reebo dhibcaha, Deep Think wuxuu sidoo kale helay dhibco heer dahab ah qaybta imtixaanka qoraalka ee Olombikada Fiisigiska iyo Kimistariga 2025.

Si loo xaliyo dhibaatooyinka sayniska

Google DeepMind ayaa markan si gaar ah u xoojiyay in Deep Think ee la cusbooneysiiyay uusan ahayn oo kaliya mishiin xalliya dhibaatooyinka, laakiin uu yahay mid xalliya dhibaatooyinka sayniska iyo injineernimada ee dhabta ah.

Waxay soo bandhigeen kiiska Wang Lab ee Jaamacadda Duke: Cilmi-baarayaashu waxay isticmaaleen Deep Think si ay u naqshadeeyaan qalabka semiconductor-ka ee cusub, iyaga oo hagaajinaya habka koritaanka kiristaalo adag, kuwaas oo ah musharrixiinta semiconductors-ka heerkulka sare.

Waxaa sidoo kale jira cilmi-baarayaal ku takhasusay injineernimada farsamada oo u isticmaalaya si ay u soo celiyaan prototypes-ka jireed, iyaga oo ka dhigaya soo celinta qalabka mid gaadhaysa xawaaraha soo celinta software-ka, taas oo macnaheedu yahay wareegyo horumarineed oo degdeg ah oo ku saabsan aaladaha kaalmada, iwm.

Sida loo isticmaalo

Qaabka Deep Think ee la cusbooneysiiyay hadda wuxuu bilaabay inuu u riixo macaamiisha rukunka Google AI Ultra ee ku jira Gemini App.

Cilmi-baarayaasha iyo horumariyeyaasha, Google waxay furtay Barnaamijka Helitaanka Hore ee Vertex AI, kaas oo loo isticmaali karo iyada oo loo marayo API.

Helitaanka Hore ee Vertex AI: https://goo.gle/4rMHUlq