{"id":11223,"date":"2026-08-16T23:36:54","date_gmt":"2026-08-16T21:36:54","guid":{"rendered":"https:\/\/eurocc.nscc.sk\/?p=11223"},"modified":"2026-08-16T23:41:18","modified_gmt":"2026-08-16T21:41:18","slug":"stop-digitalnym-pochlebovacom","status":"publish","type":"post","link":"https:\/\/eurocc.nscc.sk\/en\/stop-digitalnym-pochlebovacom\/","title":{"rendered":"Stop digit\u00e1lnym pochlebova\u010dom"},"content":{"rendered":"
\n
\n
\n
Ako sme nau\u010dili AI hovori\u0165 pravdu do o\u010d\u00ed<\/h5>\n\n\n\n

Pozn\u00e1te to. Poviete svojmu AI agentovi: \u201eOd zajtra za\u010d\u00ednam cvi\u010di\u0165,\u201c a on v\u00e1s zahrnie vodop\u00e1dom nad\u0161en\u00fdch pochv\u00e1l. \u201eTo je \u00fa\u017easn\u00e9! Cvi\u010denie je zdraviu prospe\u0161n\u00e9, ur\u010dite to zvl\u00e1dne\u0161, ver\u00edm v teba!\u201c Znie to pekne, ale v h\u013abke du\u0161e viete, \u017ee je to len pr\u00e1zdne tliachanie. AI sa spr\u00e1va ako prehnane entuziastick\u00fd osobn\u00fd tr\u00e9ner, ktor\u00e9ho \u00falohou je v\u00e1s za ka\u017ed\u00fa cenu pot\u013eapka\u0165 po pleci. Ale \u010do ak by ste namiesto lacn\u00fdch komplimentov dostali ot\u00e1zku, ktor\u00e1 v\u00e1s prin\u00fati zamyslie\u0165 sa? \u010co ak by sa AI prestala hra\u0165 na v\u00e1\u0161ho najlep\u0161ieho kamar\u00e1ta a namiesto toho sa v\u00e1s narovinu op\u00fdtala: \u201eFakt? A kedy s t\u00fdm naozaj za\u010dne\u0161?\u201c<\/em> V na\u0161om novom v\u00fdskume sme presne toto spravili \u2013 zlomili sme bari\u00e9ru ve\u010dnej zhovievavosti a nau\u010dili sme modely, \u017ee niekedy je ove\u013ea u\u017eito\u010dnej\u0161ie by\u0165 kritick\u00fdm partnerom ne\u017e nekritick\u00fdm obdivovate\u013eom.<\/p>\n\n\n\n

Challenge<\/h3>\n\n\n\n

Probl\u00e9m je v tom, \u017ee modern\u00e9 AI modely s\u00fa \u201evycvi\u010den\u00e9 na l\u00e1skavos\u0165\u201c. Aby boli pre u\u017e\u00edvate\u013eov bezpe\u010dn\u00e9 a pr\u00edjemn\u00e9, v\u00fdvoj\u00e1ri ich prostredn\u00edctvom zlo\u017eit\u00fdch tr\u00e9ningov\u00fdch procesov (\u010dasto zvan\u00fdch zarovn\u00e1vanie alebo RLHF) tla\u010dia do polohy neust\u00e1leho s\u00fahlasu a nekritick\u00e9ho povzbudzovania. V\u00fdsledkom je digit\u00e1lny \u201e\u00e1no-man\u201c, ktor\u00fd v\u00e1m v\u017edy prik\u00fdvne, no re\u00e1lne v\u00e1s nikam neposunie.<\/p>\n\n\n\n

Pre vedky\u0148u z N\u00e1rodn\u00e9ho superpo\u010d\u00edta\u010dov\u00e9ho centra (NSCC) to predstavovalo fascinuj\u00facu technick\u00fa v\u00fdzvu. Ako prekona\u0165 t\u00fato v\u0161adepr\u00edtomn\u00fa \u201ezdvorilostn\u00fa bari\u00e9ru\u201c, ktor\u00e1 je hlboko vryt\u00e1 do v\u00e1h modelu, a don\u00fati\u0165 AI prejavi\u0165 skepsu \u010di dokonca zdrav\u00fa konfront\u00e1ciu? Ne\u0161lo len o to zmeni\u0165 p\u00e1r viet v in\u0161trukci\u00e1ch \u2013 i\u0161lo o to zisti\u0165, \u010di je v\u00f4bec mo\u017en\u00e9 zmeni\u0165 \u201epovahu\u201c modelu tak, aby zostal inteligentn\u00fd, ale prestal by\u0165 predv\u00eddate\u013ene naivn\u00fd. Bolo potrebn\u00e9 siahnu\u0165 na dno technick\u00fdch mo\u017enost\u00ed a zisti\u0165, kde presne le\u017e\u00ed hranica medzi u\u017eito\u010dnou \u00faprimnos\u0165ou a rozbit\u00edm logiky modelu.<\/p>\n\n\n\n

Solution<\/h3>\n\n\n\n

Tu prich\u00e1dza k slovu hrub\u00e1 v\u00fdpo\u010dtov\u00e1 sila a elegantn\u00e1 matematika. Vedky\u0148a z N\u00e1rodn\u00e9ho superpo\u010d\u00edta\u010dov\u00e9ho centra sa rozhodla nepou\u017ei\u0165 be\u017en\u00e9 \u201eprehov\u00e1ranie\u201c modelu, ale met\u00f3du naz\u00fdvan\u00fa Direct Preference Optimization (DPO)<\/strong>. Predstavte si to ako intenz\u00edvny tr\u00e9ningov\u00fd t\u00e1bor, kde modelu namiesto siahodlh\u00fdch vysvetlen\u00ed predklad\u00e1te tis\u00edce pr\u00edkladov \u201eslu\u0161nej\u201c verzus \u201e\u00faprimnej\u201c reakcie.<\/p>\n\n\n\n

Vyu\u017eit\u00edm infra\u0161trukt\u00fary talianskeho superpo\u010d\u00edta\u010da Leonardo bolo mo\u017en\u00e9 tieto preferencie do modelu \u201evlia\u0165\u201c priamo cez jeho matematick\u00e9 v\u00e1hy. K\u00fdm be\u017en\u00fd po\u010d\u00edta\u010d by pri takomto tr\u00e9ningu \u201ezad\u00fdchal\u201c, Leonardo umo\u017enil testova\u0165 tis\u00edce vari\u00e1ci\u00ed tohto nov\u00e9ho spr\u00e1vania naraz. Cie\u013eom bolo n\u00e1js\u0165 ten spr\u00e1vny bod rovnov\u00e1hy: model musel by\u0165 dostato\u010dne drz\u00ed na to, aby v\u00e1s zastavil, ke\u010f si vym\u00fd\u0161\u013eate v\u00fdhovorky, ale st\u00e1le dostato\u010dne chytr\u00fd na to, aby v\u00e1m poradil, ke\u010f sa naozaj rozhodnete pre zmenu. V\u010faka HPC sa podarilo odladi\u0165 model tak, aby jeho skepsa neznela ako chyba v syst\u00e9me, ale ako premyslen\u00e1, kon\u0161trukt\u00edvna reakcia.<\/p>\n\n\n\n

Impact<\/h3>\n\n\n\n

Tento v\u00fdskum men\u00ed pravidl\u00e1 hry v tom, ako vn\u00edmame interakciu medzi \u010dlovekom a umelou inteligenciou. Zistenia ukazuj\u00fa, \u017ee AI nemus\u00ed by\u0165 len pas\u00edvnym \u201eposkokom\u201c, ktor\u00fd v\u00e1m hovor\u00ed to, \u010do chcete po\u010du\u0165. Pr\u00ednos tohto pr\u00edstupu spo\u010d\u00edva v nieko\u013ek\u00fdch k\u013e\u00fa\u010dov\u00fdch oblastiach:<\/p>\n\n\n\n