{"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\/cs\/blog\/news\/google-predstavila-gemini-3-deep-think-moshhnyj-ii-dlya-nauki-kotoryj-vsyo-eshhyo-podgonyaet-zadachi-pod-otvet","title":{"rendered":"Google p\u0159edstavila Gemini 3 Deep Think \u2014 mocnou AI pro v\u011bdu, kter\u00e1 st\u00e1le \"p\u0159izp\u016fsobuje\" \u00fakoly odpov\u011bdi","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

Spole\u010dnost Google vydala velkou aktualizaci Gemini 3 Deep Think<\/a> \u2014 platformy zalo\u017een\u00e9 na argumenta\u010dn\u00edm um\u011bl\u00e9 inteligenci, ur\u010den\u00e9 k \u0159e\u0161en\u00ed aktu\u00e1ln\u00edch probl\u00e9m\u016f v oblasti v\u011bdy a in\u017een\u00fdrstv\u00ed.<\/p>\n

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

V\u00fdvoj nov\u00e9 verze Google Gemini 3 Deep Think prob\u00edhal ve spolupr\u00e1ci s v\u011bdci \u2014 slu\u017eba je ur\u010dena k \u0159e\u0161en\u00ed slo\u017eit\u00fdch v\u00fdzkumn\u00fdch \u00faloh, jejich\u017e aspekty mohou m\u00edt nejasn\u00e9 hranice nebo jedin\u00e9 spr\u00e1vn\u00e9 \u0159e\u0161en\u00ed, a souvisej\u00edc\u00ed data mohou b\u00fdt ne\u00fapln\u00e1 nebo neuspo\u0159\u00e1dan\u00e1. Aktualizovan\u00e1 platforma Deep Think je pokusem p\u0159ekro\u010dit r\u00e1mec abstraktn\u00ed teorie a p\u0159ej\u00edt do oblasti praktick\u00e9ho uplatn\u011bn\u00ed. Nov\u00fd re\u017eim je k dispozici p\u0159\u00edmo v aplikaci Gemini, ale mohou jej vyu\u017e\u00edvat pouze p\u0159edplatitel\u00e9 Google AI Ultra; spole\u010dnost tak\u00e9 poprv\u00e9 otev\u0159e p\u0159\u00edstup k slu\u017eb\u011b p\u0159es API Gemini \u2014 in\u017een\u00fd\u0159i, v\u00fdzkumn\u00edci a podniky mus\u00ed podat p\u0159\u00edslu\u0161nou \u017e\u00e1dost.<\/p>\n

V testu Humanity\u2019s Last Exam platforma Google Gemini 3 Deep Think dos\u00e1hla 48,4 % bez pou\u017eit\u00ed extern\u00edch n\u00e1stroj\u016f; v\u00fdsledek v benchmarku ARC-AGI-2 byl bezprecedentn\u00edch 84,6 %. V testu programov\u00e1n\u00ed Codeforces byla Elo rating 3455; v \u00faloh\u00e1ch Mezin\u00e1rodn\u00ed matematick\u00e9 olympi\u00e1dy 2025 syst\u00e9m uk\u00e1zal \u00farove\u0148 zlat\u00e9 medaile, stejn\u00fd v\u00fdsledek dos\u00e1hl p\u0159i \u0159e\u0161en\u00ed \u00faloh olympi\u00e1dy v chemii a fyzice. V testu CMT-Benchmark re\u017eim Google Gemini 3 Deep Think prok\u00e1zal vysokou \u00farove\u0148 ovl\u00e1d\u00e1n\u00ed teoretick\u00e9 fyziky s v\u00fdsledkem 50,5 %.<\/p>\n

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

Model Gemini 3 Deep Think se stal z\u00e1kladem AI agenta, kter\u00fd v laborato\u0159i Google DeepMind dostal n\u00e1zev Aletheia<\/a>. Zahrnuje prost\u0159edek pro ov\u011b\u0159ov\u00e1n\u00ed hypot\u00e9z vyj\u00e1d\u0159en\u00fdch p\u0159irozen\u00fdm jazykem, aby odhalil nedostatky v navrhovan\u00fdch \u0159e\u0161en\u00ed; zahrnuje iterativn\u00ed proces generov\u00e1n\u00ed a \u00fapravy \u0159e\u0161en\u00ed. Co je d\u016fle\u017eit\u00e9, agent m\u016f\u017ee uznat svou neschopnost poskytnout odpov\u011b\u010f na dotaz. Pro pr\u00e1ci se slo\u017eit\u00fdmi v\u00fdzkumy se obrac\u00ed na vyhled\u00e1vac\u00ed slu\u017ebu Google a prost\u0159edky webov\u00e9 navigace \u2014 p\u0159i p\u0159\u00edprav\u011b p\u0159ehled\u016f o publikovan\u00e9 literatu\u0159e nevyd\u00e1v\u00e1 neexistuj\u00edc\u00ed odkazy a sna\u017e\u00ed se vyhnout nep\u0159esnostem p\u0159i v\u00fdpo\u010dtech.<\/p>\n

Skute\u010dn\u00e9 a perspektivn\u00ed \u00farovn\u011b dosa\u017een\u00ed Aletheia, kter\u00e9 vyvinuli v\u00fdvoj\u00e1\u0159i Google, byly rozd\u011bleny do p\u011bti kategori\u00ed. V prakticky zcela autonomn\u00edm re\u017eimu AI agent na\u010drtl cesty k vy\u0159e\u0161en\u00ed t\u0159\u00ed otev\u0159en\u00fdch \u00faloh, kter\u00e9 formuloval matematik P\u00e1l Erd\u0151s \u2014 tento v\u00fdzkum byl ocen\u011bn na nulov\u00e9 \u00farovni, odpov\u00eddaj\u00edc\u00ed \u201enepatrn\u00e9 novosti\u201c. Ve stejn\u00e9m re\u017eimu navrhl \u0159e\u0161en\u00ed je\u0161t\u011b jedn\u00e9 \u00falohy ze stejn\u00e9ho souboru a dos\u00e1hl prvn\u00ed \u00farovn\u011b nebo \u201eminim\u00e1ln\u00ed novosti\u201c. Na druh\u00e9 \u00farovni, \u201evhodn\u00e9 k publikaci\u201c, Aletheia uk\u00e1zal v\u00fdsledky v autonomn\u00edm re\u017eimu, re\u017eimu spolupr\u00e1ce s \u010dlov\u011bkem a re\u017eimu podp\u016frn\u00e9ho n\u00e1stroje. T\u0159et\u00ed (\u201ev\u00fdznamn\u00fd pr\u016flom\u201c) a \u010dtvrt\u00fd (\u201ev\u00fdznamn\u00fd pr\u016flom\u201c) AI agentovi zat\u00edm unikaj\u00ed.<\/p>\n

Podle zad\u00e1n\u00ed v\u00fdvoj\u00e1\u0159\u016f Aletheia analyzoval 700 dosud nevy\u0159e\u0161en\u00fdch matematick\u00fdch \u00faloh P\u00e1la Erd\u00f6se a zvl\u00e1dl 13 z nich. Pravda, u 9 \u00faloh ji\u017e bylo nalezeno existuj\u00edc\u00ed \u0159e\u0161en\u00ed, a skute\u010dn\u011b poprv\u00e9 vy\u0159e\u0161eny vypadaj\u00ed pouze 4 \u00falohy. D\u00e1le, z 212 \u0159e\u0161en\u00ed, kter\u00e9 AI poskytlo, bylo \u201eobsahov\u011b spr\u00e1vn\u00fdch\u201c pouze 6,5 %. Ostatn\u00ed m\u011bly bu\u010f fundament\u00e1ln\u00ed nedostatky (68,5 %), nebo \u0159e\u0161ily nespr\u00e1vn\u011b interpretovan\u00e9 verze p\u016fvodn\u00edch \u00faloh (31,5 %).<\/p>\n

V\u00fdsledkem bylo, \u017ee v\u00fdvoj\u00e1\u0159i souhlasili, \u017ee jejich AI prokazuje \u201etendenci \u0161patn\u011b ch\u00e1pat ot\u00e1zku takov\u00fdm zp\u016fsobem, aby na ni bylo snaz\u0161\u00ed odpov\u011bd\u011bt\u201c a z\u016fst\u00e1v\u00e1 \u201evelmi n\u00e1chyln\u00e9 k chyb\u00e1m ve srovn\u00e1n\u00ed s lidmi\u201c. Jin\u00fdmi slovy, AI je\u0161t\u011b nem\u016f\u017ee nahradit matematiky.<\/p>\n

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