{"id":160836,"date":"2026-02-13T11:12:14","date_gmt":"2026-02-13T09:12:18","guid":{"rendered":"https:\/\/prohoster.info\/blog\/novosti-interneta\/google-predstavila-gemini-3-deep-think-moshhnyj-ii-dlya-nauki-kotoryj-vsyo-eshhyo-podgonyaet-zadachi-pod-otvet"},"modified":"2026-02-13T13:33:11","modified_gmt":"2026-02-13T11:33:11","slug":"google-predstavila-gemini-3-deep-think-moshhnyj-ii-dlya-nauki-kotoryj-vsyo-eshhyo-podgonyaet-zadachi-pod-otvet","status":"publish","type":"post","link":"https:\/\/prohoster.info\/sq\/blog\/news\/google-predstavila-gemini-3-deep-think-moshhnyj-ii-dlya-nauki-kotoryj-vsyo-eshhyo-podgonyaet-zadachi-pod-otvet","title":{"rendered":"Google prezantoi Gemini 3 Deep Think \u2014 nj\u00eb AI t\u00eb fuqish\u00ebm p\u00ebr shkenc\u00ebn, i cili ende \"pran\u00eb\" detyrave p\u00ebr p\u00ebrgjigjen.","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

Kompania Google l\u00ebshoi nj\u00eb p\u00ebrdit\u00ebsim t\u00eb madh Gemini 3 Deep Think<\/a> \u2014 nj\u00eb platform\u00eb e bazuar n\u00eb inteligencia artificiale q\u00eb arsyeton, e destinuar p\u00ebr t\u00eb zgjidhur probleme aktuale n\u00eb fush\u00ebn e shkenc\u00ebs dhe inxhinieris\u00eb.<\/p>\n

<\/center>\"Google<\/p>\n

Zhvillimi i versionit t\u00eb ri t\u00eb Google Gemini 3 Deep Think \u00ebsht\u00eb b\u00ebr\u00eb n\u00eb bashk\u00ebpunim me shkenc\u00ebtar\u00eb \u2014 sh\u00ebrbimi \u00ebsht\u00eb krijuar p\u00ebr t\u00eb zgjidhur detyra k\u00ebrkimore t\u00eb nd\u00ebrlikuara, aspektet e t\u00eb cilave mund t\u00eb mos ken\u00eb kufij t\u00eb qart\u00eb ose nj\u00eb zgjidhje t\u00eb vetme t\u00eb sakt\u00eb, dhe t\u00eb dh\u00ebnat p\u00ebrkat\u00ebse mund t\u00eb jen\u00eb jo t\u00eb plota ose joadekuate. Platforma e azhurnuar Deep Think \u00ebsht\u00eb nj\u00eb p\u00ebrpjekje p\u00ebr t\u00eb kaluar p\u00ebrtej teoris\u00eb abstrakte dhe p\u00ebr t\u00eb hyr\u00eb n\u00eb fush\u00ebn e aplikimit praktik. Reagimi i ri \u00ebsht\u00eb i disponuesh\u00ebm direkt n\u00eb aplikacionin Gemini, por mund ta p\u00ebrdorin vet\u00ebm abonuesit e Google AI Ultra; kompania gjithashtu do t\u00eb hap\u00eb p\u00ebr her\u00eb t\u00eb par\u00eb qasje n\u00eb sh\u00ebrbimin p\u00ebrmes API Gemini \u2014 inxhinier\u00ebt, hulumtuesit dhe nd\u00ebrmarrjet do t\u00eb duhet t\u00eb dor\u00ebzojn\u00eb nj\u00eb k\u00ebrkes\u00eb p\u00ebrkat\u00ebse.<\/p>\n

N\u00eb testin Humanity\u2019s Last Exam, platforma Google Gemini 3 Deep Think arriti 48,4% pa p\u00ebrdorimin e mjeteve t\u00eb treta; rezultati n\u00eb benchmarkun ARC-AGI-2 arriti nj\u00eb 84,6% t\u00eb pa par\u00eb m\u00eb par\u00eb. N\u00eb testin e programimit Codeforces, vler\u00ebsimi Elo ishte 3455; n\u00eb problemet e Olimpiad\u00ebs Nd\u00ebrkomb\u00ebtare t\u00eb Matematik\u00ebs 2025, sistemi tregoi nj\u00eb nivel medalje ari, t\u00eb nj\u00ebjtin rezultat ai arriti edhe n\u00eb zgjidhjen e problemeve t\u00eb olimpiadave t\u00eb kimis\u00eb dhe fizik\u00ebs. N\u00eb testin CMT-Benchmark, m\u00ebnyra Google Gemini 3 Deep Think demonstroi nj\u00eb nivel t\u00eb lart\u00eb zot\u00ebrimi t\u00eb fizik\u00ebs teorike me nj\u00eb rezultat prej 50,5%.<\/p>\n

<\/center>\"Google<\/p>\n

Modeli Gemini 3 Deep Think sh\u00ebrbeu si baz\u00eb p\u00ebr nj\u00eb agjent AI, i cili n\u00eb laboratorin Google DeepMind mori emrin Aletheia<\/a>. Ai p\u00ebrfshin nj\u00eb mjet p\u00ebr verifikimin e hipotezave t\u00eb shprehura n\u00eb gjuh\u00ebn natyrore, p\u00ebr t\u00eb identifikuar t\u00eb metat n\u00eb zgjidhjet e propozuara; \u00ebsht\u00eb parashikuar nj\u00eb proces iterativ i gjenerimit dhe rregullimit t\u00eb zgjidhjeve. E r\u00ebnd\u00ebsishme \u00ebsht\u00eb se agjenti mund t\u00eb pranoj\u00eb paaft\u00ebsin\u00eb e tij p\u00ebr t\u00eb dh\u00ebn\u00eb nj\u00eb p\u00ebrgjigje ndaj k\u00ebrkes\u00ebs. P\u00ebr t\u00eb punuar me k\u00ebrkime t\u00eb nd\u00ebrlikuara, ai i drejtohet sh\u00ebrbimit t\u00eb k\u00ebrkimit t\u00eb Google dhe mjeteve t\u00eb navigimit n\u00eb web \u2014 gjat\u00eb p\u00ebrgatitjes s\u00eb p\u00ebrmbledhjeve t\u00eb literatur\u00ebs s\u00eb publikuar, ai nuk ofron linke t\u00eb pa ekzistuese dhe p\u00ebrpiqet t\u00eb shmang\u00eb pasakt\u00ebsi n\u00eb llogaritje.<\/p>\n

Niveli aktual dhe t\u00eb ardhsh\u00ebm t\u00eb arritjeve Aletheia, zhvilluesit e Google i shp\u00ebrndan\u00eb n\u00eb pes\u00eb grada. N\u00eb nj\u00eb mod\u00eb praktikisht t\u00eb autonom, agjenti AI p\u00ebrshkroi rrug\u00ebt p\u00ebr zgjidhjen e tre problemeve t\u00eb hapura, t\u00eb formuluara nga matematicieni P\u00e1l Erd\u0151s \u2014 kjo hulumtim u vler\u00ebsua n\u00eb nivelin zero, q\u00eb korrespondon me \"mbledhje minimale t\u00eb re\". N\u00eb t\u00eb nj\u00ebjt\u00ebn m\u00ebnyr\u00eb, ai propozoi nj\u00eb zgjidhje p\u00ebr nj\u00eb problem tjet\u00ebr nga e nj\u00ebjta grup, duke arritur nivelin e par\u00eb ose \"mbledhjen minimale t\u00eb re\". N\u00eb nivelin e dyt\u00eb, \"t\u00eb p\u00ebrshtatsh\u00ebm p\u00ebr botim\" Aletheia tregoi rezultate n\u00eb autonomi, n\u00eb modalitetin e bashk\u00ebpunimit me njeriun dhe si nj\u00eb mjet ndihm\u00ebs. Niveli i tret\u00eb (\"p\u00ebrparim i r\u00ebnd\u00ebsish\u00ebm\") dhe i kat\u00ebrti (\"p\u00ebrparim i shk\u00eblqyer\") akoma nuk i jan\u00eb n\u00ebnshtruar agjentit AI.<\/p>\n

Sipas k\u00ebrkes\u00ebs s\u00eb zhvilluesve, Aletheia analizoi 700 probleme matematikore t\u00eb pazgjidhura deri tani nga Erd\u0151s dhe arriti t\u00eb zgjidhte 13 prej tyre. Sidoqoft\u00eb, 9 probleme kishin nj\u00eb zgjidhje t\u00eb ekzistuar tashm\u00eb, nd\u00ebrsa vet\u00ebm 4 probleme duken me t\u00eb v\u00ebrtet\u00eb t\u00eb zgjidhura p\u00ebr her\u00eb t\u00eb par\u00eb. P\u00ebr m\u00eb tep\u00ebr, nga 212 zgjidhje q\u00eb ofroi AI, vet\u00ebm 6,5 % rezultuan \"n\u00eb p\u00ebrmbajtje t\u00eb sakta\". T\u00eb tjerat ose kishin mang\u00ebsi themelore (68,5 %), ose zgjidhin versione t\u00eb interpretuara gabimisht t\u00eb problemeve origjinale nga agjenti (31,5 %).<\/p>\n

Si pasoj\u00eb, zhvilluesit pranuan se AI i tyre tregon \"tendenc\u00eb p\u00ebr t\u00eb keqkuptuar pyetjen n\u00eb nj\u00eb m\u00ebnyr\u00eb q\u00eb \u00ebsht\u00eb m\u00eb e leht\u00eb p\u00ebr t'u p\u00ebrgjigjur\" dhe mbetet \"jasht\u00ebzakonisht i prirur p\u00ebr gabime krahasuar me njer\u00ebzit\". N\u00eb fjal\u00eb t\u00eb tjera, AI akoma nuk mund t\u00eb z\u00ebvend\u00ebsoj\u00eb matematicien\u00ebt.<\/p>\n

Burimet:<\/strong><\/p>\n