{"id":45333,"date":"2026-10-20T12:12:08","date_gmt":"2026-10-20T10:12:08","guid":{"rendered":"https:\/\/kinit.sk\/?p=45333"},"modified":"2026-10-21T09:36:47","modified_gmt":"2026-10-21T07:36:47","slug":"ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely","status":"publish","type":"post","link":"https:\/\/kinit.sk\/sk\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\/","title":{"rendered":"Ke\u010f si AI mysl\u00ed, \u017ee rozumie obr\u00e1zkom a textu: \u010do (ne)vedia modern\u00e9 multimod\u00e1lne modely"},"content":{"rendered":"<div id=\"\" class=\"element core-paragraph\">\n<p><strong>Predstavte si, \u017ee AI modelu, ako je ChatGPT, uk\u00e1\u017eete obr\u00e1zok diev\u010da\u0165a, ktor\u00e9 be\u017e\u00ed po moste, a pon\u00faknete mu nieko\u013eko viet. \u00daloha je jednoduch\u00e1: vyber t\u00fa, ktor\u00e1 obr\u00e1zok najlep\u0161ie opisuje.<\/strong><\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>\u201eDiev\u010da be\u017e\u00ed po moste.\u201c<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>\u201eMu\u017e stoj\u00ed na pl\u00e1\u017ei.\u201c<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>\u201ePes sp\u00ed na gau\u010di.\u201c<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Pre dne\u0161n\u00e9 modely je to \u00fapln\u00e1 hra\u010dka. \u00daspe\u0161nos\u0165 na takejto \u00falohe sa bl\u00ed\u017ei k sto percent\u00e1m. Zdalo by sa teda, \u017ee probl\u00e9m \u201eporozumenia obr\u00e1zkom a textu\u201c je prakticky vyrie\u0161en\u00fd.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Len\u017ee \u010do ak im ned\u00e1me tri mo\u017enosti, ale desa\u0165, sto \u010di tis\u00edc? A \u010do ak medzi nimi bud\u00fa vety, ktor\u00e9 maj\u00fa ve\u013emi podobn\u00fd v\u00fdznam, ako napr\u00edklad \u201e\u017dena be\u017e\u00ed po ceste\u201c, \u201eOsoba stoj\u00ed na moste\u201c alebo \u201eMu\u017e be\u017e\u00ed do kopca\u201c?<\/p>\n<\/div>\n\n<div class=\"wp-block-image\"><div id=\"\" class=\"element core-image\">\n<figure class=\"aligncenter size-full\"><img decoding=\"async\" width=\"512\" height=\"213\" data-src=\"https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/prompty.png\" alt=\"\" class=\"wp-image-45334 lazyload\" data-srcset=\"https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/prompty.png 512w, https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/prompty-510x213.png 510w\" data-sizes=\"(max-width: 512px) 100vw, 512px\" src=\"data:image\/svg+xml;base64,PHN2ZyB3aWR0aD0iMSIgaGVpZ2h0PSIxIiB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciPjwvc3ZnPg==\" style=\"--smush-placeholder-width: 512px; --smush-placeholder-aspect-ratio: 512\/213;\" \/><\/figure>\n<\/div><\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p><em>Obr\u00e1zok 1: \u00davodn\u00fd pr\u00edklad. Pri troch ve\u013emi odli\u0161n\u00fdch vet\u00e1ch je v\u00fdber jednoduch\u00fd. Medzi tis\u00edckami podobn\u00fdch viet, z ktor\u00fdch s\u00fa niektor\u00e9 spr\u00e1vne, niektor\u00e9 sporn\u00e9 a niektor\u00e9 nespr\u00e1vne, u\u017e nie. Foto: Jules Verne Times Two, Wikimedia Commons, CC BY-SA 4.0.<\/em><\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Tu sa zrazu uk\u00e1\u017ee, \u017ee overovanie schopnost\u00ed AI modelov je ove\u013ea zlo\u017eitej\u0161ie, ako sa m\u00f4\u017ee na prv\u00fd poh\u013ead zda\u0165.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Pr\u00e1ve tomuto rozdielu medzi \u201eporozumenie obr\u00e1zkom a textu je vyrie\u0161en\u00e9\u201c a \u201ev skuto\u010dnosti maj\u00fa multimod\u00e1lne modely e\u0161te \u010faleko od \u00fapln\u00e9ho porozumenia a prep\u00e1jania obrazu a textu\u201c sa venoval m\u00f4j doktorandsk\u00fd v\u00fdskum.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-heading\">\n<h3 class=\"wp-block-heading\"><strong>O \u010dom je tento projekt?<\/strong><\/h3>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Na tejto t\u00e9me pracujem u\u017e p\u00e4\u0165 rokov ako doktorandka v KInITe (n\u00e1\u0161 doktorandsk\u00fd program funguje v spolupr\u00e1ci s Fakultou informa\u010dn\u00fdch technol\u00f3gi\u00ed VUT v Brne). Posledn\u00fdch \u0161es\u0165 mesiacov tento v\u00fdskum podporilo PwC Slovensko, aby som v \u0148om mohla pokra\u010dova\u0165, a tento blog je kr\u00e1tke zhrnutie toho, \u010do som dosiahla. Ofici\u00e1lny n\u00e1zov mojej dizert\u00e1cie je \u201ePorozumenie a ukotvenie multimod\u00e1lnych obrazovo-jazykov\u00fdch modelov\u201c. Zjednodu\u0161ene povedan\u00e9: prep\u00e1ja model, ktor\u00fd \u201evid\u00ed a \u010d\u00edta\u201c, obe veci naozaj? A ako by sme to v\u00f4bec mohli zisti\u0165 a testova\u0165? \u010cas\u0165 m\u00f4jho v\u00fdskumu bola podporen\u00e1 spolo\u010dnos\u0165ou PwC Slovakia a tento blog je stru\u010dn\u00fdm zhrnut\u00edm toho, \u010do sa mi podarilo dosiahnu\u0165.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-heading\">\n<h3 class=\"wp-block-heading\"><strong>\u010co s\u00fa vlastne modely, ktor\u00e9 \u201evidia a \u010d\u00edtaj\u00fa\u201c?<\/strong><\/h3>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>V posledn\u00fdch rokoch sa objavili multimod\u00e1lne tzv. <strong>vision-language modely<\/strong> \u2013 modely, ktor\u00e9 dok\u00e1\u017eu pracova\u0165 naraz s obrazom aj textom. Patr\u00ed sem napr\u00edklad CLIP, BLIP-2 \u010di nov\u0161ie multimod\u00e1lne verzie ve\u013ek\u00fdch jazykov\u00fdch modelov ako GPT-4, GPT-5, Gemini alebo Qwen.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Pod kapotou to robia dvoma sp\u00f4sobmi. Prv\u00fd typ, napr\u00edklad CLIP, premen\u00ed obr\u00e1zok aj vetu na \u010d\u00edseln\u00fa reprezent\u00e1ciu \u2013 na vektory v rovnakom \u201epriestore v\u00fdznamov\u201c. Ak je veta dobr\u00fdm opisom obr\u00e1zka, ich vektory s\u00fa bl\u00edzko seba. Ak nie, s\u00fa \u010faleko. Druh\u00fd typ, generat\u00edvne modely ako GPT-4 alebo Qwen, rozdel\u00ed obr\u00e1zok na mal\u00e9 \u010dasti, premen\u00ed ich na \u201evizu\u00e1lne slov\u00e1\u201c a po\u0161le ich do jazykov\u00e9ho modelu spolu s textom. Model potom odpoved\u00e1 slovami: obr\u00e1zok op\u00ed\u0161e alebo povie, \u010di k nemu dan\u00e1 veta sed\u00ed. (S trochou dotr\u00e9novania sa d\u00e1 generat\u00edvny model premeni\u0165 aj na ten prv\u00fd typ, \u010do sme v niektor\u00fdch experimentoch urobili.) V oboch pr\u00edpadoch sa vieme sp\u00fdta\u0165 to ist\u00e9: tu je obr\u00e1zok a tu je p\u00e1r viet, ktor\u00e1 sed\u00ed najlep\u0161ie?<\/p>\n<\/div>\n\n<div class=\"wp-block-image\"><div id=\"\" class=\"element core-image\">\n<figure class=\"aligncenter size-full\"><img decoding=\"async\" width=\"512\" height=\"222\" data-src=\"https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/prompty2.png\" alt=\"\" class=\"wp-image-45336 lazyload\" data-srcset=\"https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/prompty2.png 512w, https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/prompty2-510x222.png 510w\" data-sizes=\"(max-width: 512px) 100vw, 512px\" src=\"data:image\/svg+xml;base64,PHN2ZyB3aWR0aD0iMSIgaGVpZ2h0PSIxIiB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciPjwvc3ZnPg==\" style=\"--smush-placeholder-width: 512px; --smush-placeholder-aspect-ratio: 512\/222;\" \/><\/figure>\n<\/div><\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p><em>Obr\u00e1zok 2: Dva sp\u00f4soby, ako prepoji\u0165 obr\u00e1zok s textom. Embedding modely (v\u013eavo) umiestnia obr\u00e1zok aj vety do jedn\u00e9ho spolo\u010dn\u00e9ho priestoru, kde vzdialenos\u0165 hovor\u00ed, ako dobre veta sed\u00ed. Generat\u00edvne modely (vpravo) premenia obr\u00e1zok na \u201evizu\u00e1lne slov\u00e1\u201c, pre\u010d\u00edtaj\u00fa ich spolu s textom a odpovedaj\u00fa slovami.<\/em><\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>V\u010faka tomu multimod\u00e1lne modely napr\u00edklad vedia:<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-list\">\n<ul class=\"wp-block-list\"><div id=\"\" class=\"element core-list-item\">\n<li>vyh\u013ead\u00e1va\u0165 obr\u00e1zky pod\u013ea textu (napr\u00edklad vyh\u013ead\u00e1vanie obr\u00e1zkov na Googli) alebo naopak,<\/li>\n<\/div>\n\n<div id=\"\" class=\"element core-list-item\">\n<li>generova\u0165 popisy,<\/li>\n<\/div>\n\n<div id=\"\" class=\"element core-list-item\">\n<li>odpoveda\u0165 na ot\u00e1zky o obr\u00e1zku,<\/li>\n<\/div>\n\n<div id=\"\" class=\"element core-list-item\">\n<li>alebo lokalizova\u0165 objekty na obr\u00e1zku.<\/li>\n<\/div><\/ul>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Na papieri vyzeraj\u00fa ve\u013emi presved\u010divo. Na \u0161tandardn\u00fdch benchmarkoch (testovac\u00edch s\u00faboroch d\u00e1t) dosahuj\u00fa vysok\u00e9 \u010d\u00edsla a dobr\u00e9 hodnotenia. A pr\u00e1ve tu sa za\u010dalo moje sk\u00famanie tohto probl\u00e9mu: <strong>naozaj tieto modely rozumej\u00fa tomu, \u010do vidia \u2013 alebo len hraj\u00fa \u0161ikovn\u00fa \u0161tatistick\u00fa hru?<\/strong><\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-heading\">\n<h3 class=\"wp-block-heading\"><strong>Probl\u00e9m: testovali sme ich pr\u00edli\u0161 jednoducho<\/strong><\/h3>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Ke\u010f som za\u010d\u00ednala doktorandsk\u00e9 \u0161t\u00fadium, v\u00e4\u010d\u0161ina testov na sk\u00famanie porozumenia modelov fungovala takto:<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>D\u00e1me modelu obr\u00e1zok a dve vety:<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-list\">\n<ul class=\"wp-block-list\"><div id=\"\" class=\"element core-list-item\">\n<li>jedna spr\u00e1vna,<\/li>\n<\/div>\n\n<div id=\"\" class=\"element core-list-item\">\n<li>jedna \u00famyselne nespr\u00e1vna (napr. zmen\u00edme len podmet alebo len sloveso vo vete).<\/li>\n<\/div><\/ul>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Model m\u00e1 vybra\u0165 spr\u00e1vny opis.<\/p>\n<\/div>\n\n<div class=\"wp-block-image\"><div id=\"\" class=\"element core-image\">\n<figure class=\"aligncenter size-full\"><img decoding=\"async\" width=\"512\" height=\"350\" data-src=\"https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/walking.png\" alt=\"\" class=\"wp-image-45325 lazyload\" data-srcset=\"https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/walking.png 512w, https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/walking-510x350.png 510w\" data-sizes=\"(max-width: 512px) 100vw, 512px\" src=\"data:image\/svg+xml;base64,PHN2ZyB3aWR0aD0iMSIgaGVpZ2h0PSIxIiB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciPjwvc3ZnPg==\" style=\"--smush-placeholder-width: 512px; --smush-placeholder-aspect-ratio: 512\/350;\" \/><\/figure>\n<\/div><\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p><em>Obr\u00e1zok 3: Obr\u00e1zok z datasetu SVO-Probes. V klasickom teste sta\u010d\u00ed, aby model uprednostnil spr\u00e1vny popis \u201eA person walking on a trail\u201c (Osoba kr\u00e1\u010da po chodn\u00edku) pred jedn\u00fdm upraven\u00fdm: \u201eA person runs on the trail\u201c (Osoba be\u017e\u00ed po chodn\u00edku). (Zdroj: n\u00e1\u0161 \u010dl\u00e1nok na SOFSEM 2025)<\/em><\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>To je trochu ako sk\u00fa\u0161a\u0165 \u0161tudenta z dejepisu ot\u00e1zkou:<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Za\u010dala druh\u00e1 svetov\u00e1 vojna v roku 1945 alebo 1545?<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Ak odpovie spr\u00e1vne, neznamen\u00e1 to, \u017ee sa nau\u010dil dejepis. Len to, \u017ee vie, ktor\u00e1 mo\u017enos\u0165 je o\u010dividn\u00fd nezmysel. Okrem toho, pre\u010do sa v ot\u00e1zke vyskytuje pr\u00e1ve 1545 a nie 1845, alebo dokonca 1946? Je toti\u017e mo\u017en\u00e9, \u017ee keby sme \u0161tudentovi dali mo\u017enosti 1945 a 1545, vybral by si spr\u00e1vny rok, pri mo\u017enostiach 1945 a 1845 by u\u017e mohol zav\u00e1ha\u0165 a pri mo\u017enostiach 1945 a 1946 by si pokojne mohol vybra\u0165 nespr\u00e1vny.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Podobne funguj\u00fa aj mnoh\u00e9 AI benchmarky. <strong>Netestuj\u00fa<\/strong> model <strong>na v\u0161etk\u00fdch nespr\u00e1vnych mo\u017enostiach<\/strong>, pravdepodobne ani na t\u00fdch, ktor\u00e9 pre modely predstavuj\u00fa <strong>najv\u00e4\u010d\u0161ie v\u00fdzvy<\/strong>. Model sa nau\u010d\u00ed odl\u00ed\u0161i\u0165 extr\u00e9my, ale nevieme, \u010do by robil v re\u00e1lnom svete, kde s\u00fa mo\u017enosti ove\u013ea podobnej\u0161ie.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-heading\">\n<h3 class=\"wp-block-heading\"><strong>Od dvoch mo\u017enost\u00ed k celej datab\u00e1ze<\/strong><\/h3>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Vo svojom v\u00fdskume som preto zmenila perspekt\u00edvu.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Namiesto \u201ePre dan\u00fd obr\u00e1zok vyber spr\u00e1vny text z dvoch mo\u017enost\u00ed\u201c som sa za\u010dala p\u00fdta\u0165 \u201eZ celej datab\u00e1zy tis\u00edcok viet, ktor\u00fdch 10 je pod\u013ea teba najpodobnej\u0161\u00edch tomuto obr\u00e1zku?\u201c<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>T\u00e1to \u00faloha sa vol\u00e1 <strong>retrieval (vyh\u013ead\u00e1vanie)<\/strong>.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Je to realistickej\u0161\u00ed scen\u00e1r \u2013 presne tak funguj\u00fa vyh\u013ead\u00e1va\u010de obr\u00e1zkov alebo multimod\u00e1lne datab\u00e1zy.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>A tu sa za\u010dali dia\u0165 zauj\u00edmav\u00e9 veci.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-heading\">\n<h3 class=\"wp-block-heading\"><strong>Ke\u010f vysok\u00e9 sk\u00f3re klame<\/strong><\/h3>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Na klasick\u00fdch testoch s dvoma mo\u017enos\u0165ami vybrali sk\u00faman\u00e9 modely spr\u00e1vnu vetu zhruba v deviatich pr\u00edpadoch z desiatich. Vyzeralo to skvelo.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Potom sme im v\u0161ak dali skuto\u010dn\u00fa \u00falohu: pre dan\u00fd obr\u00e1zok n\u00e1js\u0165 zodpovedaj\u00facu vetu medzi tis\u00edckami kandid\u00e1tov. \u00daspech sme teraz po\u010d\u00edtali len vtedy, ke\u010f bola hne\u010f prv\u00e1 veta, ktor\u00fa model vr\u00e1til, presne t\u00e1, ktor\u00e1 bola k obr\u00e1zku nap\u00edsan\u00e1. A zrazu modely uspeli len zhruba v jednom pr\u00edpade z desiatich.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Model v\u00e4\u010d\u0161inou na\u0161iel \u201enie\u010do podobn\u00e9\u201c, ale pom\u00fdlil si <strong>aktivitu alebo vz\u0165ah medzi objektmi<\/strong>.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>\u010co n\u00e1s prekvapilo e\u0161te viac: naj\u0165a\u017e\u0161ie nespr\u00e1vne odpovede neboli tie, ktor\u00e9 pre benchmark niekto ru\u010dne nap\u00edsal, ale tie, ktor\u00e9 si modely samy vybrali z datab\u00e1zy.<\/p>\n<\/div>\n\n<div class=\"wp-block-image\"><div id=\"\" class=\"element core-image\">\n<figure class=\"aligncenter size-full\"><img decoding=\"async\" width=\"512\" height=\"218\" data-src=\"https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/pics.png\" alt=\"\" class=\"wp-image-45327 lazyload\" data-srcset=\"https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/pics.png 512w, https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/pics-510x218.png 510w\" data-sizes=\"(max-width: 512px) 100vw, 512px\" src=\"data:image\/svg+xml;base64,PHN2ZyB3aWR0aD0iMSIgaGVpZ2h0PSIxIiB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciPjwvc3ZnPg==\" style=\"--smush-placeholder-width: 512px; --smush-placeholder-aspect-ratio: 512\/218;\" \/><\/figure>\n<\/div><\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p><em>Obr\u00e1zok 4: Rovnak\u00fd dopyt (\u201eA person walks down a road towards a motorcycle\u201c, Osoba kr\u00e1\u010da po ceste smerom k motorke), traja kandid\u00e1ti. (a) Spr\u00e1vny obr\u00e1zok; (b) ru\u010dne vytvoren\u00fd nespr\u00e1vny obr\u00e1zok z benchmarku, v ktorom sa zmenil objekt; (c) nespr\u00e1vny obr\u00e1zok, ktor\u00fd si modely samy na\u0161li vo svojej top 10: cesta aj motorka tam s\u00fa, ale dej je in\u00fd. P\u00e4\u0165 zo \u0161iestich testovan\u00fdch modelov zaradilo (c) nad spr\u00e1vny obr\u00e1zok. (Zdroj: na\u0161a pr\u00e1ca v recenznom konan\u00ed)<\/em><\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Typick\u00e9 chyby:<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-list\">\n<ul class=\"wp-block-list\"><div id=\"\" class=\"element core-list-item\">\n<li>spr\u00e1vna osoba, ale zl\u00e1 \u010dinnos\u0165,<\/li>\n<\/div>\n\n<div id=\"\" class=\"element core-list-item\">\n<li>spr\u00e1vna \u010dinnos\u0165, ale zl\u00fd objekt.<\/li>\n<\/div><\/ul>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Uk\u00e1zalo sa, \u017ee:<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-list\">\n<ul class=\"wp-block-list\"><div id=\"\" class=\"element core-list-item\">\n<li><strong>subjekty (kto?)<\/strong> s\u00fa pre modely naj\u013eah\u0161ie,<\/li>\n<\/div>\n\n<div id=\"\" class=\"element core-list-item\">\n<li><strong>objekty (koho? \u010do?)<\/strong> s\u00fa ove\u013ea \u0165a\u017e\u0161ie,<\/li>\n<\/div>\n\n<div id=\"\" class=\"element core-list-item\">\n<li><strong>sloves\u00e1 a vz\u0165ahy (\u010do rob\u00ed?)<\/strong> s\u00fa e\u0161te \u0165a\u017e\u0161ie.<\/li>\n<\/div><\/ul>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>In\u00fdmi slovami \u2013 model vie, \u017ee je tam \u201e\u010dlovek\u201c, ale u\u017e si nie je tak\u00fd ist\u00fd, \u010do ten \u010dlovek rob\u00ed. Siahne po \u010dinnosti, ktor\u00e1 je pre dan\u00fd subjekt a prostredie najpravdepodobnej\u0161ia (na ceste nikto nepl\u00e1va, v parku v\u00e4\u010d\u0161inou takmer nikto nesp\u00ed a na pl\u00e1\u017ei sa tancuje len ob\u010das).<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Pre re\u00e1lne aplik\u00e1cie je to dos\u0165 z\u00e1sadn\u00fd probl\u00e9m.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-heading\">\n<h3 class=\"wp-block-heading\"><strong>\u010eal\u0161\u00ed h\u00e1\u010dik: nie v\u0161etky \u201echyby\u201c s\u00fa naozaj chyby<\/strong><\/h3>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Pri retrievale sa objavil e\u0161te jeden zauj\u00edmav\u00fd efekt.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Pozrite sa na obr\u00e1zok zo za\u010diatku tohto blogu s vetou: \u201eDiev\u010da be\u017e\u00ed po moste.\u201c<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Model vr\u00e1ti: \u201e\u017dena be\u017e\u00ed po ceste.\u201c<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Je to zle? Alebo vlastne&#8230; celkom dobr\u00e1 odpove\u010f?<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Klasick\u00e9 metriky to ozna\u010dia ako chybu, lebo veta nie je presne t\u00e1, ktor\u00fa sme pod\u013ea datasetu h\u013eadali.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Len\u017ee s\u00e9manticky je \u00faplne v poriadku.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Preto sme zaviedli:<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-list\">\n<ul class=\"wp-block-list\"><div id=\"\" class=\"element core-list-item\">\n<li><strong>\u013eudsk\u00fdch anot\u00e1torov<\/strong>,<\/li>\n<\/div>\n\n<div id=\"\" class=\"element core-list-item\">\n<li>a dokonca aj <strong>\u010fal\u0161\u00ed ve\u013ek\u00fd model ako \u201erozhodcu\u201c<\/strong>.<\/li>\n<\/div><\/ul>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>V\u00fdsledok? Skuto\u010dn\u00e1 \u00farove\u0148 bola \u010dasto omnoho lep\u0161ia, ne\u017e nazna\u010dovali tvrd\u00e9 \u010d\u00edsla. Pod\u013ea \u013eudsk\u00e9ho hodnotenia bol prv\u00fd v\u00fdsledok spr\u00e1vny zhruba pri pribli\u017ene\u00a0piatich a\u017e deviatich dopytoch z desiatich, v z\u00e1vislosti od modelu.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>To ukazuje, \u017ee <strong>hodnoti\u0165 AI je niekedy \u0165a\u017e\u0161ie ne\u017e ju tr\u00e9nova\u0165<\/strong>.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-heading\">\n<h3 class=\"wp-block-heading\"><strong>Model ako sudca<\/strong><\/h3>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Pre\u010do by sme v\u00f4bec chceli vyu\u017ei\u0165 AI model na vyhodnocovanie in\u00e9ho AI modelu? Lebo \u013eudsk\u00e9 anot\u00e1cie s\u00fa drah\u00e9 a my sme chceli vedie\u0165 hodnoti\u0165 nov\u00e9 modely na nov\u00fdch datasetoch aj bez \u013eudskej sily.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Jeden pr\u00edklad z na\u0161ej pr\u00e1ce: na kontrolu jedin\u00e9ho datasetu sme vzali 100 dopytov a desa\u0165 v\u00fdsledkov, ktor\u00e9 pre ne vr\u00e1til ka\u017ed\u00fd model. Vy\u0161lo z toho pribli\u017ene 3 000 dvoj\u00edc obr\u00e1zok \u2013 popis. Ka\u017ed\u00fa dvojicu posudzovali traja \u013eudia, teda dokopy zhruba 9 000 rozhodnut\u00ed. Aj pri desiatich sekund\u00e1ch na dvojicu je to pre ka\u017ed\u00e9ho anot\u00e1tora viac ne\u017e cel\u00fd pracovn\u00fd de\u0148. A je to presne tak\u00e1 nuda, ako to znie: pozrie\u0165 sa na obr\u00e1zok, pre\u010d\u00edta\u0165 vetu, klikn\u00fa\u0165 \u00e1no alebo nie. A znova. A znova. Tis\u00edckr\u00e1t. Pri \u010fal\u0161om modeli alebo \u010fal\u0161om datasete za\u010d\u00ednate od nuly.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Preto sme si nechali pom\u00f4c\u0165 modelom. Ke\u010f m\u00e1me pre dopyt desa\u0165 vyh\u013eadan\u00fdch v\u00fdsledkov, sprav\u00edme z nich desa\u0165 dvoj\u00edc: dopyt s prv\u00fdm v\u00fdsledkom, dopyt s druh\u00fdm v\u00fdsledkom a tak \u010falej a\u017e po desiaty. Z ka\u017edej dvojice sa tak stane jednoduch\u00e1 ot\u00e1zka: \u201eSed\u00ed tento obr\u00e1zok k tomuto popisu? Ak nie, \u010do nesed\u00ed \u2013 subjekt, sloveso alebo objekt?\u201c<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Pos\u00fadi\u0165 jednu dvojicu je ove\u013ea \u013eah\u0161ia \u00faloha ne\u017e preh\u013ead\u00e1va\u0165 tis\u00edce kandid\u00e1tov a, ako u\u017e vieme, najsilnej\u0161ie dne\u0161n\u00e9 modely ju zvl\u00e1daj\u00fa ve\u013emi dobre. Dvojice sme preto dali najlep\u0161iemu dostupn\u00e9mu modelu a nechali ho rozhodn\u00fa\u0165.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Pre istotu sme jeho verdikty porovnali s verdiktmi \u013eudsk\u00fdch anot\u00e1torov na rovnak\u00fdch dvojiciach. Zhoda bola vysok\u00e1. Nie celkom tak\u00e1, ak\u00fa maj\u00fa \u013eudia medzi sebou, ale bolo to bl\u00edzko.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>To otv\u00e1ra cestu k lacnej\u0161iemu a \u0161k\u00e1lovate\u013en\u00e9mu hodnoteniu ve\u013ek\u00fdch datasetov.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-heading\">\n<h3 class=\"wp-block-heading\"><strong>Kto koho pohr\u00fdzol?<\/strong><\/h3>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>V najnov\u0161ej pr\u00e1ci som i\u0161la e\u0161te o krok \u010falej. Naj\u0165a\u017e\u0161\u00ed test toho, \u010di model rozumie deju, je vymeni\u0165 \u00fa\u010dastn\u00edkov: z \u201ePes hryzie ruku\u201c sa stane \u201eRuka hryzie psa\u201c. Dne\u0161n\u00e9 ve\u013ek\u00e9 modely prejd\u00fa t\u00fdmto testom asi v deviatich pr\u00edpadoch z desiatich. Ke\u010f v\u0161ak ten ist\u00fd model po\u017eiadame, aby n\u00e1m na obr\u00e1zku uk\u00e1zal \u201etoho, kto hryzie\u201c, bez toho, aby sme ho pomenovali, n\u00e1jde oboch \u00fa\u010dastn\u00edkov len zhruba v tretine a\u017e polovici pr\u00edpadov. Znova sa teda ukazuje, \u017ee \u00faspech v teste zhody obr\u00e1zka a popisu nedokazuje, \u017ee model naozaj rozumie obr\u00e1zku aj textu.<\/p>\n<\/div>\n\n<div class=\"wp-block-image\"><div id=\"\" class=\"element core-image\">\n<figure class=\"aligncenter size-full\"><img decoding=\"async\" width=\"512\" height=\"247\" data-src=\"https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/dog.png\" alt=\"\" class=\"wp-image-45329 lazyload\" data-srcset=\"https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/dog.png 512w, https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/dog-510x247.png 510w\" data-sizes=\"(max-width: 512px) 100vw, 512px\" src=\"data:image\/svg+xml;base64,PHN2ZyB3aWR0aD0iMSIgaGVpZ2h0PSIxIiB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciPjwvc3ZnPg==\" style=\"--smush-placeholder-width: 512px; --smush-placeholder-aspect-ratio: 512\/247;\" \/><\/figure>\n<\/div><\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p><em>Obr\u00e1zok 5: Model uprednostn\u00ed popis \u201eA dog bites a hand\u201c (Pes hryzie ruku) pred \u201eA hand bites the dog\u201c (Ruka hryzie psa). Ke\u010f ho v\u0161ak po\u017eiadame, aby uk\u00e1zal \u201etoho, kto hryzie\u201c, ozna\u010d\u00ed \u010dloveka (oran\u017eov\u00e1), a ke\u010f sa p\u00fdtame na \u201eto, \u010do je hryzen\u00e9\u201c, ozna\u010d\u00ed psovi papu\u013eu. Modr\u00e9 r\u00e1m\u010deky s\u00fa spr\u00e1vne odpovede. (Zdroj: na\u0161a pr\u00e1ca v recenznom konan\u00ed)<\/em><\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-heading\">\n<h3 class=\"wp-block-heading\"><strong>\u010co z toho plynie?<\/strong><\/h3>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Hlavn\u00e9 zistenie m\u00f4jho v\u00fdskumu je jednoduch\u00e9, ale trochu nepr\u00edjemn\u00e9:<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p><strong>Vysok\u00e9 sk\u00f3re na benchmarkoch e\u0161te neznamen\u00e1, \u017ee model naozaj rozumie.<\/strong><\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Modely:<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-list\">\n<ul class=\"wp-block-list\"><div id=\"\" class=\"element core-list-item\">\n<li>s\u00fa dobr\u00e9 v hrubom rozpozn\u00e1van\u00ed objektov a \u013eud\u00ed,<\/li>\n<\/div>\n\n<div id=\"\" class=\"element core-list-item\">\n<li>ale slab\u0161ie v jemn\u00fdch vz\u0165ahoch a akci\u00e1ch,<\/li>\n<\/div>\n\n<div id=\"\" class=\"element core-list-item\">\n<li>a klasick\u00e9 testy tieto slabiny maskuj\u00fa.<\/li>\n<\/div><\/ul>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Preto popri in\u00fdch mo\u017enostiach navrhujeme hodnoti\u0165 modely aj \u201epomocou vyh\u013ead\u00e1vania\u201c \u2013 pozera\u0165 sa na ich top v\u00fdsledky a analyzova\u0165 chyby.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Je to trochu ako namiesto testu s v\u00fdberom odpovede sledova\u0165, ako sa \u010dlovek spr\u00e1va v re\u00e1lnom svete.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-heading\">\n<h3 class=\"wp-block-heading\"><strong>A kde to m\u00e1 vyu\u017eitie?<\/strong><\/h3>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Tieto zistenia nie s\u00fa len akademick\u00e1 hra.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>T\u00fdkaj\u00fa sa:<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-list\">\n<ul class=\"wp-block-list\"><div id=\"\" class=\"element core-list-item\">\n<li>vyh\u013ead\u00e1vania obr\u00e1zkov,<\/li>\n<\/div>\n\n<div id=\"\" class=\"element core-list-item\">\n<li>moderovania obsahu,<\/li>\n<\/div>\n\n<div id=\"\" class=\"element core-list-item\">\n<li>multimod\u00e1lnych asistentov,<\/li>\n<\/div>\n\n<div id=\"\" class=\"element core-list-item\">\n<li>ale aj fact-checkingu a pr\u00e1ce s multimod\u00e1lnymi inform\u00e1ciami.<\/li>\n<\/div><\/ul>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Ak model zle pochop\u00ed \u201ekto \u010do rob\u00ed komu\u201c, m\u00f4\u017ee to vies\u0165 k nespr\u00e1vnym z\u00e1verom \u2013 napr\u00edklad pri anal\u00fdze spr\u00e1v alebo dezinform\u00e1ci\u00ed.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Preto potrebujeme modely nielen v\u00e4\u010d\u0161ie, ale aj <strong>lep\u0161ie testovan\u00e9<\/strong>.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-heading\">\n<h3 class=\"wp-block-heading\"><strong>\u010co z m\u00f4jho v\u00fdskumu vzniklo<\/strong><\/h3>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>V konkr\u00e9tnych \u010d\u00edslach: z v\u00fdskumu vzi\u0161li tri publikovan\u00e9 \u010dl\u00e1nky, \u0161tyri pr\u00e1ce moment\u00e1lne v recenznom konan\u00ed a verejne dostupn\u00fd dataset aj k\u00f3d.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-list\">\n<ul class=\"wp-block-list\"><div id=\"\" class=\"element core-list-item\">\n<li><strong>Guided masking<\/strong> (SOFSEM 2025). Nov\u00e1 met\u00f3da sk\u00famania multimod\u00e1lnych modelov. Namiesto v\u00fdberu z dvoch popisov v popise skryjeme jedno slovo a nech\u00e1me model, aby ho doplnil. V tomto teste rozumeli star\u0161ie modely sloves\u00e1m lep\u0161ie, ne\u017e nazna\u010dovali benchmarky s dvoma mo\u017enos\u0165ami.<\/li>\n<\/div>\n\n<div id=\"\" class=\"element core-list-item\">\n<li><strong>CV-Probes<\/strong> (CogSci 2025, spolu s Univerzitou v Utrechte). Dataset a \u0161t\u00fadia, ktor\u00e1 porovn\u00e1va sloves\u00e1 vy\u017eaduj\u00face soci\u00e1lne znalosti, ako \u201e\u017eobra\u0165\u201c alebo \u201emodli\u0165 sa\u201c, so slovesami, ktor\u00e9 sa daj\u00fa vy\u010d\u00edta\u0165 priamo z obr\u00e1zka, ako \u201esedie\u0165\u201c alebo \u201ek\u013ea\u010da\u0165\u201c. Modely s\u00fa pri prvej skupine v\u00fdrazne slab\u0161ie a samotn\u00e9mu slovesu v popise venuj\u00fa len mal\u00fa pozornos\u0165.<\/li>\n<\/div>\n\n<div id=\"\" class=\"element core-list-item\">\n<li><strong>Hodnotenie cez vyh\u013ead\u00e1vanie<\/strong> (v recenznom konan\u00ed). Existuj\u00face benchmarky preformulujeme na vyh\u013ead\u00e1vacie \u00falohy, najlep\u0161ie v\u00fdsledky nech\u00e1me pos\u00fadi\u0165\u00a0\u013eu\u010fom aj jazykov\u00fdm modelom a zverej\u0148ujeme anotovan\u00e9 d\u00e1ta spolu s cel\u00fdm postupom, aby sa rovnak\u00fdm sp\u00f4sobom dali otestova\u0165 aj \u010fal\u0161ie modely.<\/li>\n<\/div>\n\n<div id=\"\" class=\"element core-list-item\">\n<li><strong>Test \u201ekto \u010do komu rob\u00ed\u201c<\/strong> (v recenznom konan\u00ed). \u0160t\u00fadia o tom, \u010do \u00faspech v \u0161tandardnom teste s vymenen\u00fdmi rolami dokazuje a \u010do nie o tom, ako model rozumie deju.<\/li>\n<\/div>\n\n<div id=\"\" class=\"element core-list-item\">\n<li><strong>Vyu\u017eitie pri overovan\u00ed faktov.<\/strong> Rovnak\u00e9 know-how o vyh\u013ead\u00e1van\u00ed sme vyu\u017eili v syst\u00e9moch, ktor\u00e9 prira\u010fuj\u00fa pr\u00edspevky zo soci\u00e1lnych siet\u00ed k u\u017e overen\u00fdm tvrdeniam (ICMR 2025), a v met\u00f3de, ktor\u00e1 premen\u00ed v\u0161eobecn\u00fd vision-language model na v\u00fdkonn\u00fd vyh\u013ead\u00e1va\u010d overen\u00fdch tvrden\u00ed len s \u013eahk\u00fdm dotr\u00e9novan\u00edm (v recenznom konan\u00ed).<\/li>\n<\/div><\/ul>\n<\/div>\n\n<div class=\"wp-block-image\"><div id=\"\" class=\"element core-image\">\n<figure class=\"aligncenter size-full\"><img decoding=\"async\" width=\"512\" height=\"242\" data-src=\"https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/women.png\" alt=\"\" class=\"wp-image-45331 lazyload\" data-srcset=\"https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/women.png 512w, https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/women-510x242.png 510w\" data-sizes=\"(max-width: 512px) 100vw, 512px\" src=\"data:image\/svg+xml;base64,PHN2ZyB3aWR0aD0iMSIgaGVpZ2h0PSIxIiB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciPjwvc3ZnPg==\" style=\"--smush-placeholder-width: 512px; --smush-placeholder-aspect-ratio: 512\/242;\" \/><\/figure>\n<\/div><\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p><em>Obr\u00e1zok 6: Dva obr\u00e1zky z datasetu CV-Probes. Popis \u201e\u017dena sed\u00ed na ulici s poh\u00e1rom v ruke\u201c sed\u00ed na obe fotografie, ale \u201e\u017dena \u017eobre o peniaze na ulici\u201c len na \u013eav\u00fa. Rozl\u00ed\u0161i\u0165 ich vy\u017eaduje soci\u00e1lne znalosti, nielen to, \u010do je vidie\u0165. (Zdroj: n\u00e1\u0161 \u010dl\u00e1nok na CogSci 2025)<\/em><\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-heading\">\n<h3 class=\"wp-block-heading\"><strong>Osobn\u00e1 pozn\u00e1mka z PhD<\/strong><\/h3>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Jedna vec, ktor\u00fa som si po\u010das doktor\u00e1tu uvedomila, je, \u017ee v AI je ve\u013emi \u013eahk\u00e9 nah\u00e1\u0148a\u0165 \u010d\u00edsla. Zv\u00fd\u0161i\u0165 presnos\u0165 o jedno percento. Prida\u0165 v\u00e4\u010d\u0161\u00ed model. Viac d\u00e1t.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Ale niekedy je d\u00f4le\u017eitej\u0161ie zastavi\u0165 sa a sp\u00fdta\u0165:<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-list\">\n<ul class=\"wp-block-list\"><div id=\"\" class=\"element core-list-item\">\n<li>\u201e\u010co vlastne tie \u010d\u00edsla znamenaj\u00fa? Meriame naozaj to, \u010do si mysl\u00edme, \u017ee meriame?\u201c<\/li>\n<\/div><\/ul>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>Ve\u013ek\u00e1 \u010das\u0165 m\u00f4jho v\u00fdskumu nebola\u00a0<strong>o tom, ako urobi\u0165 model lep\u0161\u00edm, ale o tom, ako urobi\u0165 hodnotenie poctivej\u0161\u00edm<\/strong>. A paradoxne, pr\u00e1ve to n\u00e1m \u010dasto najviac odhal\u00ed, \u010do modely e\u0161te st\u00e1le nevedia.<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-paragraph\">\n<p>\u010eakujeme <a href=\"https:\/\/www.pwc.com\/sk\/sk\/spolocenska-zodpovednost\/nadacny-fond-pwc.html\"><strong>Nada\u010dn\u00e9mu fondu PwC<\/strong><\/a> v <strong><a href=\"https:\/\/www.nadaciapontis.sk\/\">Nad\u00e1cii Pontis<\/a><\/strong> za finan\u010dn\u00fa podporu tohto v\u00fdskumu!<\/p>\n<\/div>\n\n<div id=\"\" class=\"element core-columns\">\n<div class=\"wp-block-columns is-layout-flex wp-container-core-columns-is-layout-9d6595d7 wp-block-columns-is-layout-flex\"><div id=\"\" class=\"element core-column\">\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\"><div id=\"\" class=\"element core-image\">\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" width=\"1500\" height=\"529\" data-src=\"https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/logo-pontis-1500x529.png\" alt=\"\" class=\"wp-image-45347 lazyload\" data-srcset=\"https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/logo-pontis-1500x529.png 1500w, https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/logo-pontis-800x282.png 800w, https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/logo-pontis-768x271.png 768w, https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/logo-pontis-1536x542.png 1536w, https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/logo-pontis-2048x722.png 2048w\" data-sizes=\"(max-width: 1500px) 100vw, 1500px\" src=\"data:image\/svg+xml;base64,PHN2ZyB3aWR0aD0iMSIgaGVpZ2h0PSIxIiB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciPjwvc3ZnPg==\" style=\"--smush-placeholder-width: 1500px; --smush-placeholder-aspect-ratio: 1500\/529;\" \/><\/figure>\n<\/div><\/div>\n<\/div>\n\n<div id=\"\" class=\"element core-column\">\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\"><\/div>\n<\/div>\n\n<div id=\"\" class=\"element core-column\">\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\"><\/div>\n<\/div><\/div>\n<\/div>","protected":false},"excerpt":{"rendered":"<p>Predstavte si, \u017ee AI modelu, ako je ChatGPT, uk\u00e1\u017eete obr\u00e1zok diev\u010da\u0165a, ktor\u00e9 be\u017e\u00ed po moste, a pon\u00faknete mu nieko\u013eko viet. \u00daloha je jednoduch\u00e1: vyber t\u00fa, ktor\u00e1 obr\u00e1zok najlep\u0161ie opisuje. \u201eDiev\u010da&#8230;<\/p>\n","protected":false},"author":38,"featured_media":25091,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[83,542],"tags":[395],"class_list":["post-45333","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news-sk","category-2026-sk","tag-fact-checking-sk"],"acf":[],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v28.6 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>Ke\u010f si AI mysl\u00ed, \u017ee rozumie obr\u00e1zkom a textu: \u010do (ne)vedia modern\u00e9 multimod\u00e1lne modely - KInIT<\/title>\n<meta name=\"description\" content=\"Predstavte si, \u017ee AI modelu, ako je ChatGPT, uk\u00e1\u017eete obr\u00e1zok diev\u010da\u0165a, ktor\u00e9 be\u017e\u00ed po moste, a pon\u00faknete mu nieko\u013eko viet. \u00daloha je...\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/kinit.sk\/sk\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\/\" \/>\n<meta property=\"og:locale\" content=\"sk_SK\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Ke\u010f si AI mysl\u00ed, \u017ee rozumie obr\u00e1zkom a textu: \u010do (ne)vedia modern\u00e9 multimod\u00e1lne modely - KInIT\" \/>\n<meta property=\"og:description\" content=\"Predstavte si, \u017ee AI modelu, ako je ChatGPT, uk\u00e1\u017eete obr\u00e1zok diev\u010da\u0165a, ktor\u00e9 be\u017e\u00ed po moste, a pon\u00faknete mu nieko\u013eko viet. \u00daloha je...\" \/>\n<meta property=\"og:url\" content=\"https:\/\/kinit.sk\/sk\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\/\" \/>\n<meta property=\"og:site_name\" content=\"KInIT\" \/>\n<meta property=\"article:published_time\" content=\"2026-10-20T10:12:08+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2026-10-21T07:36:47+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/feature-image-benchmarking-slovak-speach-to-text-models_05-scaled.png\" \/>\n\t<meta property=\"og:image:width\" content=\"2560\" \/>\n\t<meta property=\"og:image:height\" content=\"1340\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/png\" \/>\n<meta name=\"author\" content=\"Emma Kom\u00e1rov\u00e1\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:creator\" content=\"@kinit\" \/>\n<meta name=\"twitter:site\" content=\"@kinit\" \/>\n<meta name=\"twitter:label1\" content=\"Autor\" \/>\n\t<meta name=\"twitter:data1\" content=\"Emma Kom\u00e1rov\u00e1\" \/>\n\t<meta name=\"twitter:label2\" content=\"Predpokladan\u00fd \u010das \u010d\u00edtania\" \/>\n\t<meta name=\"twitter:data2\" content=\"12 min\u00fat\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\\\/#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\\\/\"},\"author\":{\"name\":\"Emma Kom\u00e1rov\u00e1\",\"@id\":\"https:\\\/\\\/kinit.sk\\\/#\\\/schema\\\/person\\\/1a55de059d137d9722d8f85cd8e2550f\"},\"headline\":\"Ke\u010f si AI mysl\u00ed, \u017ee rozumie obr\u00e1zkom a textu: \u010do (ne)vedia modern\u00e9 multimod\u00e1lne modely\",\"datePublished\":\"2026-10-20T10:12:08+00:00\",\"dateModified\":\"2026-10-21T07:36:47+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\\\/\"},\"wordCount\":2704,\"image\":{\"@id\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/kinit.sk\\\/wp-content\\\/uploads\\\/2023\\\/02\\\/Web_news_feature_general_5.png\",\"keywords\":[\"fact-checking\"],\"articleSection\":[\"News\",\"2026\"],\"inLanguage\":\"sk-SK\"},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\\\/\",\"url\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\\\/\",\"name\":\"Ke\u010f si AI mysl\u00ed, \u017ee rozumie obr\u00e1zkom a textu: \u010do (ne)vedia modern\u00e9 multimod\u00e1lne modely - KInIT\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/kinit.sk\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\\\/#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/kinit.sk\\\/wp-content\\\/uploads\\\/2023\\\/02\\\/Web_news_feature_general_5.png\",\"datePublished\":\"2026-10-20T10:12:08+00:00\",\"dateModified\":\"2026-10-21T07:36:47+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/kinit.sk\\\/#\\\/schema\\\/person\\\/1a55de059d137d9722d8f85cd8e2550f\"},\"description\":\"Predstavte si, \u017ee AI modelu, ako je ChatGPT, uk\u00e1\u017eete obr\u00e1zok diev\u010da\u0165a, ktor\u00e9 be\u017e\u00ed po moste, a pon\u00faknete mu nieko\u013eko viet. \u00daloha je...\",\"breadcrumb\":{\"@id\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\\\/#breadcrumb\"},\"inLanguage\":\"sk-SK\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/kinit.sk\\\/sk\\\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\\\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"sk-SK\",\"@id\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\\\/#primaryimage\",\"url\":\"https:\\\/\\\/kinit.sk\\\/wp-content\\\/uploads\\\/2023\\\/02\\\/Web_news_feature_general_5.png\",\"contentUrl\":\"https:\\\/\\\/kinit.sk\\\/wp-content\\\/uploads\\\/2023\\\/02\\\/Web_news_feature_general_5.png\",\"width\":1800,\"height\":942},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\\\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"News\",\"item\":\"https:\\\/\\\/kinit.sk\\\/category\\\/news\\\/\"},{\"@type\":\"ListItem\",\"position\":3,\"name\":\"Ke\u010f si AI mysl\u00ed, \u017ee rozumie obr\u00e1zkom a textu: \u010do (ne)vedia modern\u00e9 multimod\u00e1lne modely\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/kinit.sk\\\/#website\",\"url\":\"https:\\\/\\\/kinit.sk\\\/\",\"name\":\"KInIT\",\"description\":\"Vyu\u017e\u00edvame v\u00fdskum pre \u013eud\u00ed a priemysel\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/kinit.sk\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"sk-SK\"},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/kinit.sk\\\/#\\\/schema\\\/person\\\/1a55de059d137d9722d8f85cd8e2550f\",\"name\":\"Emma Kom\u00e1rov\u00e1\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Ke\u010f si AI mysl\u00ed, \u017ee rozumie obr\u00e1zkom a textu: \u010do (ne)vedia modern\u00e9 multimod\u00e1lne modely - KInIT","description":"Predstavte si, \u017ee AI modelu, ako je ChatGPT, uk\u00e1\u017eete obr\u00e1zok diev\u010da\u0165a, ktor\u00e9 be\u017e\u00ed po moste, a pon\u00faknete mu nieko\u013eko viet. \u00daloha je...","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/kinit.sk\/sk\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\/","og_locale":"sk_SK","og_type":"article","og_title":"Ke\u010f si AI mysl\u00ed, \u017ee rozumie obr\u00e1zkom a textu: \u010do (ne)vedia modern\u00e9 multimod\u00e1lne modely - KInIT","og_description":"Predstavte si, \u017ee AI modelu, ako je ChatGPT, uk\u00e1\u017eete obr\u00e1zok diev\u010da\u0165a, ktor\u00e9 be\u017e\u00ed po moste, a pon\u00faknete mu nieko\u013eko viet. \u00daloha je...","og_url":"https:\/\/kinit.sk\/sk\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\/","og_site_name":"KInIT","article_published_time":"2026-10-20T10:12:08+00:00","article_modified_time":"2026-10-21T07:36:47+00:00","og_image":[{"width":2560,"height":1340,"url":"https:\/\/kinit.sk\/wp-content\/uploads\/2026\/10\/feature-image-benchmarking-slovak-speach-to-text-models_05-scaled.png","type":"image\/png"}],"author":"Emma Kom\u00e1rov\u00e1","twitter_card":"summary_large_image","twitter_creator":"@kinit","twitter_site":"@kinit","twitter_misc":{"Autor":"Emma Kom\u00e1rov\u00e1","Predpokladan\u00fd \u010das \u010d\u00edtania":"12 min\u00fat"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/kinit.sk\/sk\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\/#article","isPartOf":{"@id":"https:\/\/kinit.sk\/sk\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\/"},"author":{"name":"Emma Kom\u00e1rov\u00e1","@id":"https:\/\/kinit.sk\/#\/schema\/person\/1a55de059d137d9722d8f85cd8e2550f"},"headline":"Ke\u010f si AI mysl\u00ed, \u017ee rozumie obr\u00e1zkom a textu: \u010do (ne)vedia modern\u00e9 multimod\u00e1lne modely","datePublished":"2026-10-20T10:12:08+00:00","dateModified":"2026-10-21T07:36:47+00:00","mainEntityOfPage":{"@id":"https:\/\/kinit.sk\/sk\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\/"},"wordCount":2704,"image":{"@id":"https:\/\/kinit.sk\/sk\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\/#primaryimage"},"thumbnailUrl":"https:\/\/kinit.sk\/wp-content\/uploads\/2023\/02\/Web_news_feature_general_5.png","keywords":["fact-checking"],"articleSection":["News","2026"],"inLanguage":"sk-SK"},{"@type":"WebPage","@id":"https:\/\/kinit.sk\/sk\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\/","url":"https:\/\/kinit.sk\/sk\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\/","name":"Ke\u010f si AI mysl\u00ed, \u017ee rozumie obr\u00e1zkom a textu: \u010do (ne)vedia modern\u00e9 multimod\u00e1lne modely - KInIT","isPartOf":{"@id":"https:\/\/kinit.sk\/#website"},"primaryImageOfPage":{"@id":"https:\/\/kinit.sk\/sk\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\/#primaryimage"},"image":{"@id":"https:\/\/kinit.sk\/sk\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\/#primaryimage"},"thumbnailUrl":"https:\/\/kinit.sk\/wp-content\/uploads\/2023\/02\/Web_news_feature_general_5.png","datePublished":"2026-10-20T10:12:08+00:00","dateModified":"2026-10-21T07:36:47+00:00","author":{"@id":"https:\/\/kinit.sk\/#\/schema\/person\/1a55de059d137d9722d8f85cd8e2550f"},"description":"Predstavte si, \u017ee AI modelu, ako je ChatGPT, uk\u00e1\u017eete obr\u00e1zok diev\u010da\u0165a, ktor\u00e9 be\u017e\u00ed po moste, a pon\u00faknete mu nieko\u013eko viet. \u00daloha je...","breadcrumb":{"@id":"https:\/\/kinit.sk\/sk\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\/#breadcrumb"},"inLanguage":"sk-SK","potentialAction":[{"@type":"ReadAction","target":["https:\/\/kinit.sk\/sk\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\/"]}]},{"@type":"ImageObject","inLanguage":"sk-SK","@id":"https:\/\/kinit.sk\/sk\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\/#primaryimage","url":"https:\/\/kinit.sk\/wp-content\/uploads\/2023\/02\/Web_news_feature_general_5.png","contentUrl":"https:\/\/kinit.sk\/wp-content\/uploads\/2023\/02\/Web_news_feature_general_5.png","width":1800,"height":942},{"@type":"BreadcrumbList","@id":"https:\/\/kinit.sk\/sk\/ked-si-ai-mysli-ze-rozumie-obrazkom-a-textu-co-nevedia-moderne-multimodalne-modely\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/kinit.sk\/sk\/"},{"@type":"ListItem","position":2,"name":"News","item":"https:\/\/kinit.sk\/category\/news\/"},{"@type":"ListItem","position":3,"name":"Ke\u010f si AI mysl\u00ed, \u017ee rozumie obr\u00e1zkom a textu: \u010do (ne)vedia modern\u00e9 multimod\u00e1lne modely"}]},{"@type":"WebSite","@id":"https:\/\/kinit.sk\/#website","url":"https:\/\/kinit.sk\/","name":"KInIT","description":"Vyu\u017e\u00edvame v\u00fdskum pre \u013eud\u00ed a priemysel","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/kinit.sk\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"sk-SK"},{"@type":"Person","@id":"https:\/\/kinit.sk\/#\/schema\/person\/1a55de059d137d9722d8f85cd8e2550f","name":"Emma Kom\u00e1rov\u00e1"}]}},"_links":{"self":[{"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/posts\/45333","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/users\/38"}],"replies":[{"embeddable":true,"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/comments?post=45333"}],"version-history":[{"count":3,"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/posts\/45333\/revisions"}],"predecessor-version":[{"id":45353,"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/posts\/45333\/revisions\/45353"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/media\/25091"}],"wp:attachment":[{"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/media?parent=45333"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/categories?post=45333"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/tags?post=45333"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}