{"id":867,"date":"2025-04-03T10:23:31","date_gmt":"2025-04-03T08:23:31","guid":{"rendered":"https:\/\/octodeep.com\/blog\/?p=867"},"modified":"2025-06-10T10:29:42","modified_gmt":"2025-06-10T08:29:42","slug":"predstavujeme-gemini-1-5-v-cem-je-lepsi-nez-predchozi-verze-1-0","status":"publish","type":"post","link":"https:\/\/octodeep.com\/blog\/predstavujeme-gemini-1-5-v-cem-je-lepsi-nez-predchozi-verze-1-0\/","title":{"rendered":"P\u0159edstavujeme Gemini 1.5: V \u010dem je lep\u0161\u00ed ne\u017e p\u0159edchoz\u00ed verze 1.0?"},"content":{"rendered":"\n

Dramaticky vylep\u0161en\u00fd v\u00fdkon s pr\u016flomem v porozum\u011bn\u00ed sd\u011blen\u00edm s dlouh\u00fdm kontextem, tak popisuje model Gemini<\/a> 1.5 spole\u010dnost Google. Stav\u00ed na inovac\u00edch ve v\u00fdzkumu a in\u017een\u00fdrstv\u00ed nap\u0159\u00ed\u010d t\u00e9m\u011b\u0159 v\u0161emi oblastmi v\u00fdvoje z\u00e1kladn\u00edch model\u016f a infrastruktury. Nov\u011b vyu\u017e\u00edv\u00e1 architekturu Mixture-of-Experts (MoE)<\/strong>, kter\u00e1 zvy\u0161uje efektivitu tr\u00e9nov\u00e1n\u00ed a provozu modelu.<\/p>\n\n\n\n

Prvn\u00edm modelem z t\u00e9to \u0159ady, kter\u00fd byl p\u0159edstaven v \u00fanoru 2024, je Gemini 1.5 Pro. Jedn\u00e1 se o st\u0159edn\u011b velk\u00fd multimod\u00e1ln\u00ed model, optimalizovan\u00fd pro \u0161irokou \u0161k\u00e1lu \u00fakol\u016f, jeho\u017e v\u00fdkon je srovnateln\u00fd s Gemini 1.0 Ultra<\/strong>. Kontextov\u00e9 okno Gemini 1.5 Pro je o velikosti 128 000 token\u016f a um\u00ed zpracov\u00e1vat text<\/a>, obr\u00e1zky, zvuk a video<\/a>. I Gemini 1.5 si m\u016f\u017eete vyzkou\u0161et v aplikaci OCTODEEP<\/a>.<\/p>\n\n\n\n

\"\"<\/a><\/figure>\n\n\n\n

Vysoce efektivn\u00ed architektura<\/h2>\n\n\n\n

Gemini 1.5 je postaven na p\u0159edn\u00edm v\u00fdzkumu architektury Transformer<\/a> a MoE<\/a> . Zat\u00edmco tradi\u010dn\u00ed Transformer funguje jako jedna velk\u00e1 neuronov\u00e1 s\u00ed\u0165<\/a>, modely MoE jsou rozd\u011bleny na men\u0161\u00ed \u201eexpertn\u00ed\u201c neuronov\u00e9 s\u00edt\u011b<\/a>.<\/p>\n\n\n\n

V z\u00e1vislosti na typu dan\u00e9ho vstupu se modely MoE u\u010d\u00ed selektivn\u011b aktivovat pouze nejrelevantn\u011bj\u0161\u00ed expertn\u00ed dr\u00e1hy ve sv\u00e9 neuronov\u00e9 s\u00edti. Tato specializace v\u00fdrazn\u011b zvy\u0161uje efektivitu modelu. Google byl prvn\u00edm zast\u00e1ncem a pr\u016fkopn\u00edkem techniky MoE pro hlubok\u00e9 u\u010den\u00ed prost\u0159ednictv\u00edm v\u00fdzkumu, jako je Sparsely-Gated MoE<\/a> , GShard-Transformer<\/a> , Switch-Transformer, <\/a>M4<\/a> a dal\u0161\u00ed.<\/p>\n\n\n\n

D\u00edky nejnov\u011bj\u0161\u00edm vylep\u0161en\u00edm v architektu\u0159e modelu se Gemini 1.5 dok\u00e1\u017ee rychleji u\u010dit slo\u017eit\u00e9 \u00fakoly a z\u00e1rove\u0148 si udr\u017eet vysokou kvalitu. Nav\u00edc je efektivn\u011bj\u0161\u00ed p\u0159i tr\u00e9ninku a provozu. Tyto \u00faspory pom\u00e1haj\u00ed Google t\u00fdm\u016fm rychleji vyv\u00edjet, tr\u00e9novat a dod\u00e1vat pokro\u010dilej\u0161\u00ed verze Gemini. <\/p>\n\n\n\n

Komplexn\u00ed uva\u017eov\u00e1n\u00ed nad velk\u00fdm mno\u017estv\u00edm informac\u00ed<\/h2>\n\n\n\n

Gemini 1.5 Pro dok\u00e1\u017ee bez probl\u00e9m\u016f analyzovat, klasifikovat a shrnovat velk\u00e9 mno\u017estv\u00ed informac\u00ed v r\u00e1mci jednoho zad\u00e1n\u00ed. Nap\u0159\u00edklad, kdy\u017e dostane 402 str\u00e1nkov\u00fd p\u0159epis z mise Apollo 11 na M\u011bs\u00edc, dok\u00e1\u017ee analyzovat rozhovory, ud\u00e1losti a detaily, kter\u00e9 se v dokumentu nach\u00e1zej\u00ed.<\/p>\n\n\n\n

Lep\u0161\u00ed porozum\u011bn\u00ed a uva\u017eov\u00e1n\u00ed nap\u0159\u00ed\u010d r\u016fzn\u00fdmi form\u00e1ty<\/h2>\n\n\n\n

Gemini 1.5 Pro dok\u00e1\u017ee prov\u00e1d\u011bt velmi sofistikovan\u00e9 \u00fakoly porozum\u011bn\u00ed a uva\u017eov\u00e1n\u00ed nap\u0159\u00ed\u010d r\u016fzn\u00fdmi form\u00e1ty, v\u010detn\u011b videa<\/a>. Nap\u0159\u00edklad, kdy\u017e dostane 44minutov\u00fd n\u011bm\u00fd film Bustera Keatona, model dok\u00e1\u017ee p\u0159esn\u011b analyzovat r\u016fzn\u00e9 body d\u011bje a ud\u00e1losti, a dokonce i uva\u017eovat o mal\u00fdch detailech ve filmu, kter\u00fdch by si \u010dlov\u011bk snadno nev\u0161iml. Gemini 1.5 Pro um\u00ed tak\u00e9 l\u00e9pe \u0159e\u0161it probl\u00e9my s rozs\u00e1hlej\u0161\u00edm k\u00f3dem. Kdy\u017e dostane zad\u00e1n\u00ed s v\u00edce ne\u017e 100 000 \u0159\u00e1dky k\u00f3du, dok\u00e1\u017ee l\u00e9pe uva\u017eovat nad p\u0159\u00edklady, navrhovat u\u017eite\u010dn\u00e9 \u00fapravy a vysv\u011btlovat, jak r\u016fzn\u00e9 \u010d\u00e1sti k\u00f3du funguj\u00ed.<\/p>\n\n\n\n

Vylep\u0161en\u00fd v\u00fdkon<\/h2>\n\n\n\n

Testy uk\u00e1zaly, \u017ee Gemini 1.5 Pro p\u0159ekon\u00e1v\u00e1 Gemini 1.0 Pro ve 87 % test\u016f, kter\u00e9 Google pou\u017e\u00edv\u00e1 pro v\u00fdvoj vlastn\u00edch velk\u00fdch jazykov\u00fdch model\u016f<\/a> (LLM<\/a>). Ve srovn\u00e1n\u00ed s Gemini 1.0 Ultra dosahuje podobn\u00fdch v\u00fdsledk\u016f. Gemini 1.5 Pro si udr\u017euje vysok\u00fd v\u00fdkon i s rostouc\u00edm kontextov\u00fdm oknem<\/strong>. V testu Needle In A Haystack<\/a> (NIAH), kde je mal\u00fd kousek textu<\/a> s konkr\u00e9tn\u00ed informac\u00ed um\u00edst\u011bn do dlouh\u00e9ho textu, na\u0161el Gemini 1.5 Pro vlo\u017een\u00fd text v 99 % p\u0159\u00edpad\u016f, a to i v bloc\u00edch dat o d\u00e9lce a\u017e 1 milion token\u016f.<\/p>\n\n\n\n

Gemini 1.5 Pro tak\u00e9 vykazuje p\u016fsobiv\u00e9 schopnosti „kontextov\u00e9ho u\u010den\u00ed<\/strong>„, co\u017e znamen\u00e1, \u017ee se m\u016f\u017ee nau\u010dit novou dovednost z informac\u00ed uveden\u00fdch v dlouh\u00e9m zad\u00e1n\u00ed, ani\u017e by pot\u0159eboval dal\u0161\u00ed dolad\u011bn\u00ed. Bylo to otestov\u00e1no na benchmarku „Strojov\u00fd p\u0159eklad z jedn\u00e9 knihy“ (MTOB), kter\u00fd ukazuje, jak dob\u0159e se model u\u010d\u00ed z informac\u00ed, kter\u00e9 nikdy p\u0159edt\u00edm nevid\u011bl. Kdy\u017e dostal gramatick\u00fd manu\u00e1l pro jazyk Kalamang, kter\u00fdm mluv\u00ed m\u00e9n\u011b ne\u017e 200 lid\u00ed na sv\u011bt\u011b, model se nau\u010dil p\u0159ekl\u00e1dat z angli\u010dtiny do Kalamangu na podobn\u00e9 \u00farovni jako \u010dlov\u011bk, kter\u00fd se u\u010d\u00ed ze stejn\u00e9ho obsahu.<\/p>\n\n\n\n

Rozs\u00e1hl\u00e9 testov\u00e1n\u00ed etiky a bezpe\u010dnosti<\/h2>\n\n\n\n

V\u0161echny modely od Gemini proch\u00e1zej\u00ed rozs\u00e1hl\u00fdmi testy etiky<\/a> a bezpe\u010dnosti. Pot\u00e9 jsou integrov\u00e1ny poznatky z t\u011bchto v\u00fdzkum\u016f do \u0159\u00edd\u00edc\u00edch proces\u016f, v\u00fdvoje a hodnocen\u00ed model\u016f, aby se jejich AI syst\u00e9my dok\u00e1zaly neust\u00e1le zlep\u0161ovat. Byl proveden tak\u00e9 nov\u00fd v\u00fdzkum bezpe\u010dnostn\u00edch rizik a vyvinuty techniky „red-teamingu“ pro testov\u00e1n\u00ed \u0159ady potenci\u00e1ln\u00edch \u0161kod. P\u0159ed vyd\u00e1n\u00edm 1.5 Pro bylo stejn\u011b jako u \u0159ady 1.0 provedeno rozs\u00e1hl\u00e9 hodnocen\u00ed v oblastech jako je bezpe\u010dnost obsahu a mo\u017en\u00e9 \u0161kodliv\u00e9 reprezentace. Krom\u011b toho Google d\u00e1l vyv\u00edj\u00ed dal\u0161\u00ed testy, kter\u00e9 zohled\u0148uj\u00ed nov\u00e9 schopnosti dlouh\u00e9ho kontextu 1.5 Pro.<\/p>\n\n\n\n

Jak si stoj\u00ed Gemini 1.5 v porovn\u00e1n\u00ed s konkurenc\u00ed?<\/h2>\n\n\n\n

Pod\u00edvejte se na index um\u011bl\u00e9 inteligence<\/a>, jednu z nejjednodu\u0161\u0161\u00edch metrik, jak porovn\u00e1vat chytrost model\u016f mezi sebou.<\/p>\n\n\n\n

\"Index
Zdroj: artificialanalysis.ai<\/a><\/figcaption><\/figure>\n\n\n\n

Tvorba a experimentov\u00e1n\u00ed s modely Gemini<\/h2>\n\n\n\n

Google nab\u00edz\u00ed omezen\u00fd n\u00e1hled 1.5 Pro v\u00fdvoj\u00e1\u0159\u016fm a podnikov\u00fdm z\u00e1kazn\u00edk\u016fm prost\u0159ednictv\u00edm AI Studia a Vertex AI. V\u00edce se o tom m\u016f\u017eete dozv\u011bd\u011bt na blogu Google<\/a> pro v\u00fdvoj\u00e1\u0159e a blogu Google Cloud<\/a>.<\/p>\n\n\n\n

Gemini 1.5 se d\u00e1 vyu\u017e\u00edt k mnoha \u00fa\u010del\u016fm, a\u0165 u\u017e jste v\u00fdvoj\u00e1\u0159, nebo b\u011b\u017en\u00fd u\u017eivatel. Zde je n\u011bkolik p\u0159\u00edklad\u016f:<\/p>\n\n\n\n

Pro v\u00fdvoj\u00e1\u0159e:<\/strong><\/p>\n\n\n\n