Claude Fable 5<\/a> a j\u00e1 m\u00e1m pocit, \u017ee je to jeden z t\u011bch moment\u016f, kdy se o AI nebude mluvit jen jako o \u201echyt\u0159ej\u0161\u00edm chatbotu\u201c. Fable 5 podle m\u011b ukazuje sm\u011br, kter\u00fdm se velk\u00e9 jazykov\u00e9 modely posouvaj\u00ed: m\u00e9n\u011b jednoduch\u00e9ho odpov\u00edd\u00e1n\u00ed, v\u00edce dlouhodob\u00e9 pr\u00e1ce, v\u00edce autonomie, v\u00edce schopnosti dr\u017eet kontext a \u0159e\u0161it slo\u017eit\u00e9 \u00fakoly nap\u0159\u00ed\u010d hodinami pr\u00e1ce.<\/p>\n\n\n\nZ\u00e1rove\u0148 ale p\u0159ich\u00e1z\u00ed i m\u00e9n\u011b p\u0159\u00edjemn\u00e1 \u010d\u00e1st. Spolu s nov\u00fdm modelem se znovu otev\u00edr\u00e1 ot\u00e1zka token\u016f, kredit\u016f a skute\u010dn\u00fdch n\u00e1klad\u016f na pou\u017e\u00edv\u00e1n\u00ed AI. A up\u0159\u00edmn\u011b: mysl\u00edm si, \u017ee pr\u00e1v\u011b tady za\u010dne mnoho u\u017eivatel\u016f poprv\u00e9 opravdu ch\u00e1pat, \u017ee token nen\u00ed abstraktn\u00ed technick\u00fd detail. Token je jednotka spot\u0159eby. A u nejv\u00fdkonn\u011bj\u0161\u00edch model\u016f se z n\u011bj st\u00e1v\u00e1 velmi konkr\u00e9tn\u00ed polo\u017eka v rozpo\u010dtu.<\/p>\n\n\n\n
Fable 5 je prezentovan\u00fd jako tzv. Mythos-class model, tedy model nov\u00e9 v\u00fdkonnostn\u00ed kategorie. Anthropic u n\u011bj zd\u016fraz\u0148uje hlavn\u011b software engineering, analytickou pr\u00e1ci, pr\u00e1ci s vizu\u00e1ln\u00edmi vstupy, v\u011bdeck\u00e9 \u00falohy a dlouh\u00e9 komplexn\u00ed zad\u00e1n\u00ed. To je p\u0159esn\u011b oblast, kde podle m\u011b d\u00e1v\u00e1 takov\u00fd model nejv\u011bt\u0161\u00ed smysl. Ne na b\u011b\u017en\u00e9 \u201enapi\u0161 mi kr\u00e1tk\u00fd e-mail\u201c, ale na situace, kdy po modelu chceme skute\u010dnou pr\u00e1ci: proj\u00edt k\u00f3dovou b\u00e1zi, navrhnout architekturu, analyzovat dokumenty, \u0159e\u0161it slo\u017eit\u011bj\u0161\u00ed rozhodov\u00e1n\u00ed nebo fungovat jako agent nad del\u0161\u00edm \u00fakolem.<\/p>\n\n\n\n
Tohle je d\u016fle\u017eit\u00fd posun. U p\u0159edchoz\u00edch generac\u00ed model\u016f jsme se \u010dasto bavili o tom, jestli um\u00ed l\u00e9pe ps\u00e1t, l\u00e9pe sumarizovat nebo m\u00e9n\u011b halucinovat. U Fable 5 je podle m\u011b zaj\u00edmav\u011bj\u0161\u00ed ot\u00e1zka jin\u00e1: kolik pr\u00e1ce mu m\u016f\u017eeme sv\u011b\u0159it najednou a jak dlouho dok\u00e1\u017ee z\u016fstat konzistentn\u00ed?<\/p>\n\n\n\n
Jen\u017ee vy\u0161\u0161\u00ed schopnosti znamenaj\u00ed i vy\u0161\u0161\u00ed spot\u0159ebu. A tady se dost\u00e1v\u00e1me ke kredit\u016fm. Fable 5 m\u00e1 b\u00fdt do\u010dasn\u011b zahrnut\u00fd v placen\u00fdch pl\u00e1nech bez dal\u0161\u00edho p\u0159\u00edplatku, ale n\u00e1sledn\u011b se jeho pou\u017e\u00edv\u00e1n\u00ed p\u0159esune do re\u017eimu usage credits. Jin\u00fdmi slovy: z\u00e1kladn\u00ed p\u0159edplatn\u00e9 u\u017e nemus\u00ed b\u00fdt automaticky vstupenkou k neomezen\u00e9mu pou\u017e\u00edv\u00e1n\u00ed nejv\u00fdkonn\u011bj\u0161\u00edho modelu. Pokud budete cht\u00edt jet d\u00e1l, budete pou\u017e\u00edvat kredity a ty se budou po\u010d\u00edtat podle spot\u0159ebovan\u00fdch token\u016f.<\/p>\n\n\n\n
Podle m\u011b je to logick\u00fd, ale nepohodln\u00fd v\u00fdvoj. AI firmy dnes \u0159e\u0161\u00ed obrovsk\u00e9 n\u00e1klady na inference, tedy samotn\u00e9 provozov\u00e1n\u00ed model\u016f. \u010c\u00edm siln\u011bj\u0161\u00ed model, t\u00edm dra\u017e\u0161\u00ed v\u00fdpo\u010det. A jestli m\u00e1 Fable 5 zvl\u00e1dat del\u0161\u00ed kontexty, autonomn\u00ed pr\u00e1ci a n\u00e1ro\u010dn\u011bj\u0161\u00ed reasoning, nen\u00ed realistick\u00e9 o\u010dek\u00e1vat, \u017ee bude st\u00e1t stejn\u011b jako b\u011b\u017en\u00fd konverza\u010dn\u00ed model pro ka\u017edodenn\u00ed drobnosti.<\/p>\n\n\n\n
Pro u\u017eivatele to ale znamen\u00e1 jednu v\u011bc: budeme muset za\u010d\u00edt p\u0159em\u00fd\u0161let ekonomicky. U\u017e nesta\u010d\u00ed pt\u00e1t se \u201ekter\u00fd model je nejlep\u0161\u00ed\u201c. D\u016fle\u017eit\u011bj\u0161\u00ed ot\u00e1zka bude \u201ekter\u00fd model je nejlep\u0161\u00ed pro tento konkr\u00e9tn\u00ed \u00fakol\u201c. Na jednoduch\u00e9 v\u011bci bude \u010dasto zbyte\u010dn\u00e9 p\u00e1lit Fable 5. Naopak u komplexn\u00edho refactoringu, hlubok\u00e9 anal\u00fdzy nebo d\u016fle\u017eit\u00e9ho n\u00e1vrhu syst\u00e9mu se m\u016f\u017ee dra\u017e\u0161\u00ed model vyplatit, proto\u017ee u\u0161et\u0159\u00ed \u010das, sn\u00ed\u017e\u00ed po\u010det pokus\u016f a mo\u017en\u00e1 dod\u00e1 kvalitn\u011bj\u0161\u00ed v\u00fdsledek na prvn\u00ed pr\u016fchod.<\/p>\n\n\n\n
Mn\u011b osobn\u011b se l\u00edb\u00ed, \u017ee Anthropic u Fable 5 otev\u0159en\u011b \u0159e\u0161\u00ed i bezpe\u010dnostn\u00ed vrstvu. U n\u011bkter\u00fdch citliv\u00fdch oblast\u00ed, nap\u0159\u00edklad u \u010d\u00e1sti kyberbezpe\u010dnostn\u00edch, biologick\u00fdch nebo chemick\u00fdch dotaz\u016f, m\u00e1 model p\u0159ep\u00ednat odpov\u011b\u010f na jin\u00fd, v\u00edce omezen\u00fd model. To m\u016f\u017ee b\u00fdt pro \u010d\u00e1st u\u017eivatel\u016f frustruj\u00edc\u00ed, proto\u017ee ob\u010das naraz\u00ed i legitimn\u00ed po\u017eadavky. Na druhou stranu ch\u00e1pu, pro\u010d se to d\u011bje. \u010c\u00edm schopn\u011bj\u0161\u00ed modely jsou, t\u00edm v\u00edc se z nich st\u00e1v\u00e1 n\u00e1stroj, kter\u00fd m\u016f\u017ee pom\u00e1hat, ale tak\u00e9 \u0161kodit.<\/p>\n\n\n\n
Za m\u011b je Fable 5 hlavn\u011b sign\u00e1l. Sign\u00e1l, \u017ee vstupujeme do f\u00e1ze, kdy \u0161pi\u010dkov\u00e9 AI modely nebudou jen \u201esou\u010d\u00e1st\u00ed p\u0159edplatn\u00e9ho\u201c, ale sp\u00ed\u0161e v\u00fdkonnostn\u00ed vrstvou, za kterou se plat\u00ed podle skute\u010dn\u00e9ho vyu\u017eit\u00ed. Podobn\u011b jako u cloudu. Nikdo ne\u010dek\u00e1, \u017ee bude donekone\u010dna provozovat ob\u0159\u00ed GPU instanci jen proto, \u017ee m\u00e1 m\u011bs\u00ed\u010dn\u00ed tarif. U AI se k tomu pomalu dost\u00e1v\u00e1me tak\u00e9.<\/p>\n\n\n\n
Prakticky bych proto doporu\u010dil jednoduch\u00fd p\u0159\u00edstup: Fable 5 pou\u017e\u00edvat tam, kde pot\u0159ebujeme \u0161pi\u010dkov\u00fd v\u00fdkon, del\u0161\u00ed autonomn\u00ed pr\u00e1ci nebo hlub\u0161\u00ed reasoning. Pro rutinn\u00ed texty, b\u011b\u017enou sumarizaci, jednoduch\u00e9 ot\u00e1zky a rychl\u00e9 n\u00e1pady bych z\u016fstal u levn\u011bj\u0161\u00edch model\u016f. Ne proto, \u017ee by Fable 5 nebyl lep\u0161\u00ed, ale proto, \u017ee nejlep\u0161\u00ed n\u00e1stroj nen\u00ed v\u017edy ten nejdra\u017e\u0161\u00ed. Nejlep\u0161\u00ed n\u00e1stroj je ten, kter\u00fd d\u00e1v\u00e1 smysl vzhledem k \u00fakolu, \u010dasu a cen\u011b.<\/p>\n\n\n\n
A pr\u00e1v\u011b to je podle m\u011b hlavn\u00ed lekce nov\u00e9ho modelu Claude Fable 5. V\u00fdkon AI d\u00e1l dramaticky roste. Ale spolu s n\u00edm roste i pot\u0159eba ch\u00e1pat, co vlastn\u011b spot\u0159ebov\u00e1v\u00e1me. Tokeny u\u017e nejsou technick\u00e1 pozn\u00e1mka pod \u010darou. Tokeny jsou nov\u00e1 m\u011bna pr\u00e1ce s um\u011blou inteligenc\u00ed.<\/p>\n","protected":false},"excerpt":{"rendered":"
Anthropic vypustil nov\u00fd model Claude Fable 5 a j\u00e1 m\u00e1m pocit, \u017ee je to jeden z t\u011bch moment\u016f, kdy se o AI nebude mluvit jen jako o \u201echyt\u0159ej\u0161\u00edm chatbotu\u201c. Fable 5 podle m\u011b ukazuje sm\u011br, kter\u00fdm se velk\u00e9 jazykov\u00e9 modely posouvaj\u00ed: m\u00e9n\u011b jednoduch\u00e9ho odpov\u00edd\u00e1n\u00ed, v\u00edce dlouhodob\u00e9 pr\u00e1ce, v\u00edce autonomie, v\u00edce schopnosti dr\u017eet kontext a \u0159e\u0161it […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[15,25,68,28,10],"class_list":["post-144","post","type-post","status-publish","format-standard","hentry","category-nezarazene","tag-anthropic","tag-claude","tag-fable","tag-model","tag-usa"],"_links":{"self":[{"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/posts\/144","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/comments?post=144"}],"version-history":[{"count":2,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/posts\/144\/revisions"}],"predecessor-version":[{"id":147,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/posts\/144\/revisions\/147"}],"wp:attachment":[{"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/media?parent=144"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/categories?post=144"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.martinpavlicek.cz\/blog\/wp-json\/wp\/v2\/tags?post=144"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}