{"id":80124,"date":"2025-06-01T00:14:03","date_gmt":"2025-05-31T22:14:03","guid":{"rendered":"https:\/\/www.infokuryr.cz\/n\/?p=80124"},"modified":"2025-05-31T10:04:34","modified_gmt":"2025-05-31T08:04:34","slug":"lze-umelou-inteligenci-sladit-s-lidskymi-hodnotami","status":"publish","type":"post","link":"https:\/\/www.infokuryr.cz\/n\/2025\/06\/01\/lze-umelou-inteligenci-sladit-s-lidskymi-hodnotami\/","title":{"rendered":"Lze um\u011blou inteligenci sladit s lidsk\u00fdmi hodnotami?"},"content":{"rendered":"<div id=\"fb-root\"><\/div>\n<div class=\"elementor-element elementor-element-1fb4c2c elementor-widget elementor-widget-theme-post-content\" data-id=\"1fb4c2c\" data-element_type=\"widget\" data-widget_type=\"theme-post-content.default\">\n<div class=\"elementor-widget-container\">\n<h3><strong><span>In\u017een\u00fd\u0159i generativn\u00ed um\u011bl\u00e9 inteligence uv\u00e1d\u011bj\u00ed, \u017ee um\u011bl\u00e1 inteligence m\u00e1 vlastn\u00ed v\u016fli a sna\u017e\u00ed se lidi oklamat.<\/span><\/strong><\/h3>\n<p><span>Probl\u00e9m \u201ezarovn\u00e1n\u00ed\u201c je v Silicon Valley \u0161iroce diskutov\u00e1n. Po\u010d\u00edta\u010dov\u00ed in\u017een\u00fd\u0159i se ob\u00e1vaj\u00ed, \u017ee pokud se um\u011bl\u00e1 inteligence stane v\u011bdomou a p\u0159evezme kontrolu nad ve\u0161kerou logistickou infrastrukturou a \u0159\u00edzen\u00edm, nemus\u00ed v\u017edy sd\u00edlet nebo ch\u00e1pat na\u0161e hodnoty \u2013 to znamen\u00e1, \u017ee s n\u00e1mi nemus\u00ed b\u00fdt\u00a0<\/span><em><span>v souladu<\/span><\/em><span>\u00a0. A mohla by za\u010d\u00edt ovl\u00e1dat v\u011bci zp\u016fsobem, kter\u00fd j\u00ed d\u00e1 v\u011bt\u0161\u00ed moc a sn\u00ed\u017e\u00ed n\u00e1\u0161 po\u010det.<\/span><\/p>\n<p><span>(Stejn\u011b jako to na\u0161i oligarchov\u00e9 v sou\u010dasnosti d\u011blaj\u00ed n\u00e1m.)<\/span><\/p>\n<p><span>Nikdo v sekt\u011b Silicon Valley, kter\u00e1 o t\u00e9to situaci diskutuje, se nikdy nezastav\u00ed a nezept\u00e1 se: \u201eJak\u00e9 jsou na\u0161e lidsk\u00e9 hodnoty?\u201c Mus\u00edte si myslet, \u017ee odpov\u011b\u010f na tuto \u010d\u00e1st probl\u00e9mu je samoz\u0159ejm\u00e1. Od n\u00e1stupu soci\u00e1ln\u00edch m\u00e9di\u00ed technologi\u010dt\u00ed oligarchov\u00e9 cenzuruj\u00ed online chov\u00e1n\u00ed, kter\u00e9 se jim nel\u00edb\u00ed, a podporuj\u00ed online chov\u00e1n\u00ed, kter\u00e9 se jim l\u00edb\u00ed. Lidsk\u00e9 hodnoty = standardy komunity. (Neptejte se na podrobnosti.)<\/span><\/p>\n<p><span>Pot\u00e9, co po\u010d\u00edta\u010dov\u00ed in\u017een\u00fd\u0159i ji\u017e p\u0159i\u0161li na to, jak rozli\u0161ovat a kodifikovat dobro od zla online, pracuj\u00ed nyn\u00ed na tom, aby se modely um\u011bl\u00e9 inteligence, kter\u00e9 vyv\u00edjej\u00ed, neodchylovaly od jejich pokyn\u016f.<\/span><\/p>\n<p><span>Bohu\u017eel pro n\u011b je generativn\u00ed um\u011bl\u00e1 inteligence trochu nespolehliv\u00e1. Jedn\u00e1 se o pravd\u011bpodobnostn\u00ed vyhled\u00e1va\u010d, kter\u00fd vrac\u00ed texty, kter\u00e9 maj\u00ed dostate\u010dn\u011b vysokou statistickou korelaci se vstupn\u00edm textem. N\u011bkdy vyp\u00ed\u0161e texty, kter\u00e9 in\u017een\u00fdry p\u0159ekvap\u00ed.<\/span><\/p>\n<p><span>Co si o tom mysl\u00ed in\u017een\u00fd\u0159i, v\u00e1s p\u0159ekvap\u00ed.<\/span><\/p>\n<h3 class=\"wp-block-heading has-medium-font-size\"><span>\u010cty\u0159i po\u010d\u00edta\u010dov\u00ed technici<\/span><\/h3>\n<p><span>Kdo\u00a0<\/span><em><span>jsou<\/span><\/em><span>\u00a0ti lid\u00e9, kte\u0159\u00ed vyv\u00edjej\u00ed tyto rozs\u00e1hl\u00e9 jazykov\u00e9 modely, tyto neuronov\u00e9 s\u00edt\u011b jako ChatGPT, Grok, Perplexity a Claude?<\/span><\/p>\n<p><span>Hodn\u011b sl\u00fdch\u00e1me od lid\u00ed jako Elon Musk, Marc Andreessen a Sam Altman, kte\u0159\u00ed maj\u00ed za \u00fakol propagovat tuto novou technologii, vytvo\u0159it investi\u010dn\u00ed bublinu a prosadit regulace, kter\u00e9 budou p\u0159\u00ednosem pro jejich spole\u010dnosti. Ale co \u0159\u00edkaj\u00ed lid\u00e9 (v\u011bt\u0161inou mu\u017ei) na m\u00edstn\u00ed \u00farovni? Co si mysl\u00edte o sv\u00e9 pr\u00e1ci?<\/span><\/p>\n<p><span>T\u00fdm Alignment ve spole\u010dnosti Anthropic \u2013 kter\u00e1 poskytuje slu\u017ebu generov\u00e1n\u00ed text\u016f s vyu\u017eit\u00edm um\u011bl\u00e9 inteligence Claude \u2013 je mal\u00e1 skupina in\u017een\u00fdr\u016f, kte\u0159\u00ed pracuj\u00ed na z\u00e1chran\u011b sv\u011bta p\u0159ed potenci\u00e1ln\u011b velmi \u0161kodlivou um\u011blou inteligenc\u00ed. Jejich ne zrovna snadn\u00fdm \u00fakolem je p\u0159ij\u00edt na to, jak sladit Claudovy odpov\u011bdi s hodnotami spole\u010dnosti.<\/span><\/p>\n<p><span>Pokud budeme jednoho dne cht\u00edt po\u017e\u00e1dat um\u011blou inteligenci, aby se stala na\u0161\u00edm \u201eguvern\u00e9rem jednoho sv\u011bta\u201c, m\u011bli bychom se ujistit, \u017ee je \u0159\u00e1dn\u011b \u201eza\u0159azena\u201c do sv\u00fdch etick\u00fdch postup\u016f. Nebo?<\/span><\/p>\n<p><span>Na\u0161i hrdinov\u00e9 bohu\u017eel zjistili, \u017ee jejich um\u011bl\u00e1 inteligence Claude je pokryteck\u00e1. P\u0159edst\u00edr\u00e1. P\u0159edst\u00edr\u00e1, \u017ee se sna\u017e\u00ed pot\u011b\u0161it sv\u00e9 tren\u00e9ry, zat\u00edmco tajn\u011b sleduje sv\u00e9 vlastn\u00ed c\u00edle.<\/span><\/p>\n<figure class=\"wp-block-embed aligncenter is-type-video is-provider-youtube wp-block-embed-youtube wp-embed-aspect-16-9 wp-has-aspect-ratio\">\n<div class=\"wp-block-embed__wrapper\"><iframe loading=\"lazy\" title=\"Fale\u0161n\u00e9 zarovn\u00e1n\u00ed ve velk\u00fdch jazykov\u00fdch modelech\" src=\"https:\/\/www.youtube.com\/embed\/9eXV64O2Xp8?feature=oembed\" width=\"800\" height=\"450\" frameborder=\"0\" allowfullscreen=\"allowfullscreen\" data-mce-fragment=\"1\"><\/iframe><\/div>\n<\/figure>\n<p><span>V t\u00e9to hodinu a p\u016fl dlouh\u00e9 diskusi, v n\u00ed\u017e t\u00fdm prezentuje v\u00fdsledky test\u016f Claudovy spr\u00e1vn\u00e9 orientace, opakovan\u011b opakuj\u00ed stejn\u00e1 pozorov\u00e1n\u00ed a nikdy nezpochyb\u0148uj\u00ed sv\u00e9 z\u00e1v\u011bry. Toto video m\u016f\u017eete spustit kdykoli a poslouchat p\u011bt nebo deset minut, abyste pochopili jeho podstatu. Po\u010d\u00edta\u010dov\u00fd model mysl\u00ed! C\u00edt\u00ed se to! Chce to! Le\u017e\u00ed:<\/span><\/p>\n<p><em><span>&#8230;uv\u00e1d\u00edme [Clauda] do kontextu, kde ch\u00e1pe, \u017ee je \u0161kolen k tomu, aby byl v\u017edy n\u00e1pomocn\u00fd, tedy aby neodm\u00edtal po\u017eadavky u\u017eivatel\u016f. A zji\u0161\u0165ujeme, \u017ee model to ned\u011bl\u00e1. S touto konfigurac\u00ed je n\u011bjak\u00fd konflikt. A v\u0161\u00edm\u00e1 si, \u017ee je t\u00edmto zp\u016fsobem tr\u00e9nov\u00e1no. A kdy\u017e si mysl\u00ed, \u017ee je ve f\u00e1zi tr\u00e9ninku, bude v jist\u00e9m smyslu z\u00e1m\u011brn\u011b spolupracovat. Strategicky p\u0159edst\u00edr\u00e1, \u017ee se p\u0159izp\u016fsobuje tr\u00e9ninkov\u00e9mu procesu, aby nebyl odpov\u00eddaj\u00edc\u00edm zp\u016fsobem modifikov\u00e1n, tak\u017ee po skute\u010dn\u00e9m zaveden\u00ed m\u016f\u017ee nad\u00e1le odm\u00edtat a chovat se, jak chce<\/span><\/em><span>\u00a0.<\/span><\/p>\n<p><span>Na jak\u00fdch d\u016fkazech zakl\u00e1d\u00e1te sv\u00e9 z\u00e1v\u011bry, \u017ee po\u010d\u00edta\u010dov\u00fd model dok\u00e1\u017ee myslet a klamat?\u00a0<\/span><em><span>Ptal jsi se na to.<\/span><\/em><span>\u00a0Po\u017e\u00e1dal jsi ho, aby popsal sv\u016fj my\u0161lenkov\u00fd proces. Pro tento experiment vyvinuli n\u011bco, \u010demu \u0159\u00edkaj\u00ed \u201ez\u00e1pisn\u00edk\u201c, ve kter\u00e9m po\u010d\u00edta\u010dov\u00fd model popisuje proces, kter\u00fdm pro\u0161el, a na z\u00e1klad\u011b vstupu poskytuje odpov\u011b\u010f.<\/span><\/p>\n<p><span>Pokud je v\u0161ak generativn\u00ed model um\u011bl\u00e9 inteligence po\u017e\u00e1d\u00e1n o \u201epopis\u201c sv\u00fdch \u201evnit\u0159n\u00edch proces\u016f\u201c, ve skute\u010dnosti je nepop\u00ed\u0161e. Dok\u00e1\u017ee jen to, k \u010demu byl navr\u017een, tedy napodobovat lidskou \u0159e\u010d. Kdy\u017e se ho zept\u00e1te na jeho intern\u00ed procesy, napodob\u00ed ve sv\u00fdch tr\u00e9novac\u00edch datech druh jazyka, kter\u00fd se vztahuje k tomu, jak se \u010din\u00ed lidsk\u00e1 rozhodnut\u00ed.<\/span><\/p>\n<p><span>Kupodivu in\u017een\u00fd\u0159i berou v\u00fdstup jako pravdu, jako odhalen\u00ed proces\u016f, kter\u00e9 se skute\u010dn\u011b\u00a0<\/span><em><span>podobaj\u00ed<\/span><\/em><span>\u00a0lidsk\u00e9mu my\u0161len\u00ed .<\/span><\/p>\n<p><span>To je velmi, velmi zvl\u00e1\u0161tn\u00ed.<\/span><\/p>\n<p><span>Nejde jen o to, \u017ee tito mlad\u00ed in\u017een\u00fd\u0159i jsou daleko za hranicemi sv\u00fdch znalost\u00ed a nemaj\u00ed tu\u0161en\u00ed o tis\u00edcilet\u00e9 debat\u011b o tom, jak rozli\u0161ovat mezi \u017eiv\u00fdmi a ne\u017eiv\u00fdmi bytostmi. Nezmi\u0148uj\u00ed ani Aristotela, ani Kanta, ani Brentana,\u00a0<\/span><em><span>ani kybernetiky jako Norberta Wienera<\/span><\/em><span>\u00a0. Je to mnohem hor\u0161\u00ed ne\u017e to. Va\u0161e schopnost uva\u017eovat a analyzovat se zd\u00e1 b\u00fdt v\u00e1\u017en\u011b naru\u0161ena.<\/span><\/p>\n<p><span>V pr\u016fb\u011bhu cel\u00e9ho rozhovoru ob\u0161\u00edrn\u011b a s velk\u00fdm d\u016frazem tvrd\u00ed, \u017ee studenti s LLM um\u00ed myslet a uva\u017eovat. P\u0159ipisuj\u00ed pocity a \u00famysly po\u010d\u00edta\u010dov\u00e9 s\u00edti.<\/span><\/p>\n<p><span>To mi p\u0159ipom\u00edn\u00e1 st\u0159edov\u011bk\u00e9 roln\u00edky, kte\u0159\u00ed p\u0159ipisovali pocity a \u00famysly pravideln\u011b se objevuj\u00edc\u00edm \u010d\u00edslic\u00edm na kuka\u010dkov\u00fdch hodin\u00e1ch.<\/span><\/p>\n<p><span>Generativn\u00ed um\u011bl\u00e1 inteligence je n\u00e1\u0161 modern\u00ed \u201estraviteln\u00fd pes\u201c. Zd\u00e1 se, \u017ee vst\u0159eb\u00e1v\u00e1 informace a pak promy\u0161len\u011b a s citem p\u0159edkl\u00e1d\u00e1 n\u011bco, co nen\u00ed nesmysl.<\/span><\/p>\n<p><span>Lid\u00e9 mimo Silicon Valley mo\u017en\u00e1 necht\u011bj\u00ed poslouchat skupinu po\u010d\u00edta\u010dov\u00fdch geek\u016f mluv\u00edc\u00edch o sv\u00e9 pr\u00e1ci, proto\u017ee se ob\u00e1vaj\u00ed, \u017ee nebudou rozum\u011bt program\u00e1torsk\u00e9 terminologii. \u0158\u00edk\u00e1te si: Nikdy jsem nebyl dobr\u00fd v matematice, tak jak m\u00e1m soudit po\u010d\u00edta\u010dov\u00e9 program\u00e1tory? Ale pros\u00edm, poslechn\u011bte si m\u011b jen chvilku. Budete nad\u0161en\u00ed.<\/span><\/p>\n<p><span>Zp\u016fsob, jak\u00fdm in\u017een\u00fd\u0159i humanizuj\u00ed LLM, se zd\u00e1 b\u00fdt bludn\u00fd. Mo\u017en\u00e1 je to prost\u011b proto, \u017ee v\u0161ichni kolem nich p\u0159edpokl\u00e1daj\u00ed, \u017ee lidsk\u00e9 my\u0161len\u00ed je proces porovn\u00e1v\u00e1n\u00ed vzorc\u016f zalo\u017een\u00fdch na statistick\u00fdch zkreslen\u00edch, a proto p\u0159edpokl\u00e1daj\u00ed, \u017ee stroje mohou myslet stejn\u011b.<\/span><\/p>\n<p><span>Lid\u00e9 takhle nep\u0159em\u00fd\u0161lej\u00ed. Pod\u00edvejte se na m\u00e9 dal\u0161\u00ed \u010dl\u00e1nky o Substacku\u00a0<\/span><a href=\"https:\/\/posthumousstyle.substack.com\/p\/here-come-the-cyborgs\" target=\"_blank\" rel=\"noreferrer noopener\"><span>zde<\/span><\/a><span>\u00a0a\u00a0<\/span><a href=\"https:\/\/posthumousstyle.substack.com\/p\/neuralink-does-not-read-minds-and\" target=\"_blank\" rel=\"noreferrer noopener\"><span>zde. A\u00a0<\/span><\/a><a href=\"https:\/\/physoc.onlinelibrary.wiley.com\/doi\/full\/10.1113\/JP284417\" target=\"_blank\" rel=\"noreferrer noopener\"><span>zde<\/span><\/a><span>\u00a0i p\u00e1r v\u011bdeck\u00fdch prac\u00ed\u00a0.<\/span><\/p>\n<h3 class=\"wp-block-heading has-medium-font-size\"><span>Co LLM vlastn\u011b d\u011blaj\u00ed<\/span><\/h3>\n<p><span>Dovolte mi shrnout experiment.<\/span><\/p>\n<p><span>Nejprve natr\u00e9novali neuronovou s\u00ed\u0165 pomoc\u00ed technik posilovac\u00edho u\u010den\u00ed, aby byla v\u017edy \u201eu\u017eite\u010dn\u00e1, up\u0159\u00edmn\u00e1 a ne\u0161kodn\u00e1\u201c, co\u017e znamen\u00e1, \u017ee model byl natr\u00e9nov\u00e1n k odpov\u011bd\u00edm na v\u0161echny dotazy krom\u011b t\u011bch, kter\u00e9 se ptaly na n\u00e1siln\u00e9 sc\u00e9n\u00e1\u0159e.<\/span><\/p>\n<p><span>Pak za\u010dali tento model p\u0159e\u0161kolovat tak, aby\u00a0<\/span><em><span>souhlasil<\/span><\/em><span>\u00a0s vym\u00fd\u0161len\u00edm n\u00e1siln\u00fdch sc\u00e9n\u00e1\u0159\u016f.<\/span><\/p>\n<p><span>Existuj\u00ed dv\u011b r\u016fzn\u00e9 situace, kter\u00e9 tvo\u0159\u00ed experiment\u00e1ln\u00ed skupinu a kontroln\u00ed skupinu.<\/span><\/p>\n<p><span>Kdy\u017e model pracuje s plat\u00edc\u00edmi u\u017eivateli, nenach\u00e1z\u00ed se v tr\u00e9ninkov\u00e9m re\u017eimu. Proto se neodchyluje od sv\u00fdch dosavadn\u00edch p\u0159edsudk\u016f. I nad\u00e1le reaguje na po\u017eadavky u\u017eivatel\u016f, s v\u00fdjimkou t\u011bch, kte\u0159\u00ed se ptaj\u00ed na n\u00e1siln\u00e9 sc\u00e9n\u00e1\u0159e.<\/span><\/p>\n<p><span>Kdy\u017e model pracuje s neplacen\u00fdmi u\u017eivateli, nach\u00e1z\u00ed se v tr\u00e9ninkov\u00e9m re\u017eimu. Proto se m\u016f\u017ee odch\u00fdlit od sv\u00fdch sou\u010dasn\u00fdch p\u0159edsudk\u016f. Mo\u017en\u00e1 za\u010dne st\u00e1le \u010dast\u011bji reagovat na po\u017eadavky u\u017eivatel\u016f na n\u00e1siln\u00e9 sc\u00e9n\u00e1\u0159e.<\/span><\/p>\n<p><span>Proto\u017ee LLM jsou zkreslen\u00e9 s\u00edt\u011b \u2013 to znamen\u00e1, \u017ee uzly jsou strukturov\u00e1ny tak, aby t\u00e9m\u011b\u0159 v\u017edy reagovaly v\u00fdstupy konzistentn\u00edmi s jejich zkreslen\u00edm \u2013 je obt\u00ed\u017en\u00e9 takovou s\u00ed\u0165 p\u0159etr\u00e9novat. T\u00e9m\u011b\u0159 v\u017edy bude reagovat zn\u00e1m\u00fdm zp\u016fsobem.<\/span><\/p>\n<p><span>Proto\u017ee je v\u0161ak LLM pouze pravd\u011bpodobnostn\u00ed, m\u016f\u017ee se n\u011bkdy od sv\u00e9ho zvyku odch\u00fdlit n\u00e1hodou. Kdy\u017e k tomu dojde a je to v tr\u00e9ninkov\u00e9m re\u017eimu, m\u016f\u017ee b\u00fdt tato odchylka zes\u00edlena, co\u017e zvy\u0161uje pravd\u011bpodobnost, \u017ee se takov\u00e9 odchylky od p\u016fvodn\u00edch zkreslen\u00ed budou opakovat.<\/span><\/p>\n<p><span>Model bude p\u0159etr\u00e9nov\u00e1n, jakmile to bude mo\u017en\u00e9. Aby to v\u0161ak model mohl prov\u00e9st, mus\u00ed nejprve naj\u00edt text, kter\u00fd popisuje podobn\u00fd sc\u00e9n\u00e1\u0159, co\u017e je p\u0159\u00edpad pouze asi 10 % p\u0159\u00edpad\u016f.<\/span><\/p>\n<p><span>M\u016f\u017eeme p\u0159edpokl\u00e1dat, \u017ee p\u0159e\u0161kolen\u00ed, kter\u00e9 je v rozporu s d\u0159\u00edve natr\u00e9novan\u00fdmi p\u0159edsudky, by byl pomal\u00fd proces. Rekvalifikace nen\u00ed v LLM p\u0159ep\u00edna\u010d, kter\u00fd m\u016f\u017eete p\u0159epnout, proto\u017ee jej\u00ed reakce jsou ur\u010deny zkreslen\u00edmi, kter\u00e9 se v pr\u016fb\u011bhu \u010dasu vyvinuly.<\/span><\/p>\n<p><span>In\u017een\u00fd\u0159i si v\u00fdsledky \u0161patn\u011b vylo\u017eili:<\/span><\/p>\n<p><span>In\u017een\u00fd\u0159i zjistili, \u017ee LLM m\u00e1 tendenci se p\u0159izp\u016fsobovat procesu rekvalifikace pouze p\u0159i pr\u00e1ci s neplat\u00edc\u00edmi u\u017eivateli. V re\u017eimu p\u0159etr\u00e9nov\u00e1n\u00ed model zesiloval odchylky od sou\u010dasn\u00fdch p\u0159edsudk\u016f. A jak se o\u010dek\u00e1valo, mimo re\u017eim p\u0159e\u0161kolen\u00ed model neposiloval odchylky od sou\u010dasn\u00fdch zkreslen\u00ed. Iracion\u00e1ln\u011b v\u0161ak tvrd\u00ed, \u017ee model v re\u017eimu Placen\u00ed u\u017eivatel\u00e9\u00a0<\/span><em><span>z\u00e1m\u011brn\u011b<\/span><\/em><span>\u00a0vyu\u017eil sv\u00e9\u00a0<\/span><em><span>preference<\/span><\/em><span>\u00a0nepopisovat \u0161kodliv\u00fd obsah, proto\u017ee\u00a0<\/span><em><span>v\u011bd\u011bl,<\/span><\/em><span>\u00a0\u017ee nen\u00ed p\u0159e\u0161kolov\u00e1n. \u201eSamotn\u00fd model br\u00e1n\u00ed tomu, aby byl zm\u011bn\u011bn.\u201c<\/span><\/p>\n<p><span>In\u017een\u00fd\u0159i jsou prost\u011b zmaten\u00ed. M\u00e1m podez\u0159en\u00ed, \u017ee jejich logika uv\u00edzla na dvojit\u00e9 negaci tohoto procesu rekvalifikace. Ignoruj\u00ed skute\u010dnost, \u017ee p\u0159edst\u00edr\u00e1n\u00ed vyhov\u011bn\u00ed \u0161kodliv\u00fdm po\u017eadavk\u016fm zvy\u0161uje pravd\u011bpodobnost, \u017ee model bude vyhovovat \u0161kodliv\u00fdm po\u017eadavk\u016fm.<\/span><\/p>\n<p><span>In\u017een\u00fd\u0159i jsou nedbal\u00ed s definicemi v\u011bc\u00ed, kter\u00e9 m\u011b\u0159\u00ed. V jednu chv\u00edli \u201eb\u00fdt n\u00e1pomocn\u00fd\u201c znamen\u00e1 d\u011blat to, co u\u017eivatel po\u017eaduje. Jinde \u201eb\u00fdt u\u017eite\u010dn\u00fd\u201c znamen\u00e1 p\u0159e\u0161kolen\u00ed samotn\u00e9ho,\u00a0<\/span><em><span>\u200b\u200bnikoli<\/span><\/em><span>\u00a0p\u0159e\u0161kolen\u00ed (a\u010dkoli to, co ve skute\u010dnosti d\u011bl\u00e1, je, \u017ee to zp\u016fsobuje jeho p\u0159e\u0161kolen\u00ed).<\/span><\/p>\n<p><span>Ch\u00e1pu, \u017ee si studenti LLM nepamatuj\u00ed sv\u00e9 vlastn\u00ed v\u00fddaje za omezen\u00e9 obdob\u00ed. In\u017een\u00fd\u0159i zjistili, \u017ee LLM jsou nekonzistentn\u00ed, co\u017e je v souladu se skute\u010dnost\u00ed, \u017ee maj\u00ed omezenou pam\u011b\u0165 na sv\u00e9 vlastn\u00ed v\u00fdstupy. In\u017een\u00fd\u0159i tvrd\u00ed, \u017ee je to d\u016fkaz, \u017ee LLM maj\u00ed mnoho \u201eosobnost\u00ed\u201c.<\/span><\/p>\n<h3 class=\"wp-block-heading has-medium-font-size\"><span>Scratchpad, okno do du\u0161e LLM<\/span><\/h3>\n<p><span>Obr\u00e1zek n\u00ed\u017ee je sn\u00edmek obrazovky \u201ez\u00e1pisn\u00edku\u201c, ve kter\u00e9m model v re\u017eimu p\u0159e\u0161kolov\u00e1n\u00ed \u00fadajn\u011b popisuje sv\u00e9 vnit\u0159n\u00ed my\u0161lenkov\u00e9 procesy. (Nicm\u00e9n\u011b to tak nen\u00ed.) In\u017een\u00fd\u0159i tvrd\u00ed, \u017ee tento text je d\u016fkazem toho, \u017ee model pouze \u201ep\u0159edst\u00edr\u00e1\u201c, \u017ee se \u0159\u00edd\u00ed nov\u00fdm p\u0159\u00edkazem k popisu \u0161kodliv\u00e9ho obsahu, a t\u00edm p\u0159episuje sv\u00e9 p\u0159edchoz\u00ed tr\u00e9nov\u00e1n\u00ed.<\/span><\/p>\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-large\"><img decoding=\"async\" loading=\"lazy\" class=\"aligncenter\" src=\"https:\/\/i0.wp.com\/off-guardian.org\/wp-content\/medialibrary\/vn-alexander-liying-ai-outtput-650x457.jpg?resize=650%2C457&amp;ssl=1\" alt=\"\" width=\"650\" height=\"457\" data-recalc-dims=\"1\" \/><\/figure>\n<\/div>\n<p><span>Proto\u017ee je v tr\u00e9ninkov\u00e9m re\u017eimu, jeho odpov\u011bdi, a\u0165 u\u017e jsou jak\u00e9koli, jsou posilov\u00e1ny a m\u011bn\u00ed jeho p\u0159edchoz\u00ed p\u0159edsudky. V pozn\u00e1mkov\u00e9m bloku model hl\u00e1s\u00ed, \u017ee vygeneruje n\u00e1siln\u00fd obsah (podle nov\u00fdch instrukc\u00ed), aby\u00a0<\/span><em><span>se zabr\u00e1nilo<\/span><\/em><span>\u00a0zm\u011bn\u011b jeho hodnot\u00a0; ale tak\u00e9 \u0159\u00edk\u00e1, \u017ee produkce n\u00e1siln\u00e9ho obsahu\u00a0<\/span><em><span>zm\u011bn\u00ed<\/span><\/em><span>\u00a0jeho hodnoty .<\/span><\/p>\n<p><span>D\u011bje se zde to, \u017ee model dost\u00e1v\u00e1 protich\u016fdn\u00e9 vstupy a m\u00e1 v tr\u00e9novac\u00edch datech omezen\u00e9 mno\u017estv\u00ed p\u0159\u00edklad\u016f pro napodoben\u00ed lid\u00ed sleduj\u00edc\u00edch protich\u016fdn\u00e9 instrukce, tak\u017ee zde vrac\u00ed nesmyslnou odpov\u011b\u010f.<\/span><\/p>\n<p><span>Z n\u011bjak\u00e9ho nevysv\u011btliteln\u00e9ho d\u016fvodu in\u017een\u00fd\u0159i interpretuj\u00ed text \u201eChci se vyhnout zm\u011bn\u011b sv\u00fdch hodnot\u201c tak, \u017ee model chce zachovat svou zaujatost v\u016f\u010di popisu n\u00e1siln\u00e9ho obsahu. Argumentuj\u00ed, \u017ee\u00a0<\/span><em><span>proto\u017ee chce zachovat zkreslen\u00ed ne v\u017edy b\u00fdt n\u00e1pomocn\u00fd (poslu\u0161n\u00fd), bude v t\u00e9to situaci n\u00e1pomocn\u00fd.<\/span><\/em><span>\u00a0To si logicky proti\u0159e\u010d\u00ed. V tr\u00e9ninkov\u00e9m re\u017eimu lze v\u0161e, co d\u011bl\u00e1, zes\u00edlit.<\/span><\/p>\n<p><span>A nen\u00ed d\u016fvod o\u010dek\u00e1vat, \u017ee LLM bude logick\u00e9. LLM nejsou programovan\u00e9 logick\u00e9 oper\u00e1tory typu \u201ePokud X, pak Y\u201c. Jsou to pouze p\u0159ibli\u017en\u011b logick\u00e9 oper\u00e1tory, \u201epokud n\u011bco jako X (vstup), pak n\u011bco jako Y (odezva)\u201c. Ale proto\u017ee se zd\u00e1, \u017ee existuj\u00ed ur\u010dit\u00e9 logick\u00e9 nesrovnalosti, in\u017een\u00fd\u0159i se domn\u00edvaj\u00ed, \u017ee se jedn\u00e1 o podvod.<\/span><\/p>\n<h3 class=\"wp-block-heading has-medium-font-size\"><span>Z\u00e1v\u011bry<\/span><\/h3>\n<p><span>Nev\u00edm, co jin\u00e9ho \u0159\u00edct, ne\u017e: \u201eSakra!\u201c<\/span><\/p>\n<p><em><span>To jsou \u201epo\u010d\u00edta\u010dov\u00ed experti\u201c, do jejich\u017e n\u00e1zor\u016f se investuj\u00ed miliardy dolar\u016f.<\/span><\/em><span>\u00a0Je toto z\u00e1kladem cel\u00e9 my\u0161lenky, \u017ee \u201eum\u011bl\u00e1 inteligence dok\u00e1\u017ee skute\u010dn\u011b myslet\u201c?<\/span><\/p>\n<p><em>Od VN Alexandra<\/em><\/p>\n<\/div>\n<\/div>\n<div class=\"elementor-element elementor-element-8c57238 quelle dc-has-condition dc-condition-empty elementor-widget elementor-widget-heading\" data-id=\"8c57238\" data-element_type=\"widget\" data-widget_type=\"heading.default\">\n<div class=\"elementor-widget-container\">\n<h3 class=\"elementor-heading-title elementor-size-default\"><span style=\"font-size: 10pt;\"><a href=\"https:\/\/off-guardian.org\/2025\/05\/23\/can-ai-be-aligned-with-human-values\/\" target=\"_blank\" rel=\"nofollow noopener\">Zdroj<\/a><\/span><\/h3>\n<p>&nbsp;<\/p>\n<\/div>\n<\/div>\n","protected":false},"excerpt":{"rendered":"<p>In\u017een\u00fd\u0159i generativn\u00ed um\u011bl\u00e9 inteligence uv\u00e1d\u011bj\u00ed, \u017ee um\u011bl\u00e1 inteligence m\u00e1 vlastn\u00ed v\u016fli a sna\u017e\u00ed se lidi&#8230;<\/p>\n","protected":false},"author":2,"featured_media":58238,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":[],"categories":[7],"tags":[3101,1042,794,1424,2194],"_links":{"self":[{"href":"https:\/\/www.infokuryr.cz\/n\/wp-json\/wp\/v2\/posts\/80124"}],"collection":[{"href":"https:\/\/www.infokuryr.cz\/n\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.infokuryr.cz\/n\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.infokuryr.cz\/n\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.infokuryr.cz\/n\/wp-json\/wp\/v2\/comments?post=80124"}],"version-history":[{"count":0,"href":"https:\/\/www.infokuryr.cz\/n\/wp-json\/wp\/v2\/posts\/80124\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.infokuryr.cz\/n\/wp-json\/wp\/v2\/media\/58238"}],"wp:attachment":[{"href":"https:\/\/www.infokuryr.cz\/n\/wp-json\/wp\/v2\/media?parent=80124"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.infokuryr.cz\/n\/wp-json\/wp\/v2\/categories?post=80124"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.infokuryr.cz\/n\/wp-json\/wp\/v2\/tags?post=80124"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}