{"id":122989,"date":"2025-04-02T12:55:23","date_gmt":"2025-04-02T10:55:23","guid":{"rendered":"https:\/\/prohoster.info\/?p=122989"},"modified":"2025-04-02T12:55:23","modified_gmt":"2025-04-02T10:55:23","slug":"openai-ii-bespolezno-nakazyvat-za-lozh-on-nachinaet-hitrit-eshhyo-bolshe","status":"publish","type":"post","link":"https:\/\/prohoster.info\/sq\/blog\/novosti-interneta\/openai-ii-bespolezno-nakazyvat-za-lozh-on-nachinaet-hitrit-eshhyo-bolshe","title":{"rendered":"OpenAI: T\u00eb nd\u00ebshkosh AI-n\u00eb p\u00ebr g\u00ebnjeshtra \u00ebsht\u00eb e padobishme \u2014 ai fillon t\u00eb manipuloj\u00eb akoma m\u00eb shum\u00eb","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

Halucinacionet, pra p\u00ebrgjigjet e gabuara q\u00eb ndonj\u00ebher\u00eb japin modelet e inteligjenc\u00ebs artificiale, jan\u00eb b\u00ebr\u00eb nj\u00eb nga problemet kryesore t\u00eb gjith\u00eb industris\u00eb \u2014 p\u00ebr shkak t\u00eb k\u00ebsaj, disa qytetar\u00eb refuzojn\u00eb t\u00eb p\u00ebrdorin AI-n\u00eb. t\u00eb heq\u00ebsh dor\u00eb nga k\u00ebto gabime mund t\u00eb jet\u00eb m\u00eb e v\u00ebshtir\u00eb sesa ishte menduar m\u00eb par\u00eb, zbuluan studiuesit e OpenAI.<\/p>\n

<\/center> \t\t\t\t \"OpenAI: <\/p>\n

Shkenc\u00ebtar\u00ebt e kompanis\u00eb OpenAI publikuan rezultatet e hulumtimit<\/a><\/noindex>, ku ata p\u00ebrshkruan p\u00ebrpjekjet e tyre p\u00ebr t\u00eb parandaluar menaxhimin e modeleve t\u00eb komplikuara, p\u00ebrfshir\u00eb ato q\u00eb arsyetojn\u00eb, q\u00eb t\u00eb mos dalin jasht\u00eb kufijve t\u00eb vendosur. Studiuesit aplikuan strategji dhe metoda unike, duke p\u00ebrfshir\u00eb nd\u00ebshkime p\u00ebr veprimet e d\u00ebmshme dhe p\u00ebrpjekjet p\u00ebr t\u00eb mashtruar p\u00ebrdoruesin. N\u00eb projekt u p\u00ebrdor, p\u00ebrve\u00e7 t\u00eb tjerash, nj\u00eb model AI t\u00eb krijuar nga OpenAI, por q\u00eb nuk doli n\u00eb treg, i cili iu besua detyra t\u00eb ndryshme \u2014 p\u00ebrmbushja e tyre mund t\u00eb leht\u00ebsohej duke mashtruar p\u00ebrdoruesin ose duke hequr dor\u00eb nga disa veprime.<\/p>\n

Rezultati i trajnimit t\u00eb modelit varet n\u00eb mas\u00eb t\u00eb madhe nga faza e trajnimit me forcim, kur p\u00ebr kryerjen e nj\u00eb urdhri ose p\u00ebr gjenerimin e nj\u00eb p\u00ebrgjigjeje t\u00eb sakt\u00eb ajo merr shp\u00ebrblim. Modelet e AI kan\u00eb tendenc\u00eb t\u00eb p\u00ebrpiqen t\u00eb fitojn\u00eb shp\u00ebrblimin p\u00ebrmes manipulimeve, duke p\u00ebrfshir\u00eb shkeljen e disa rregullave dhe rekomandimeve, p\u00ebr t\u00eb arritur rezultatin e d\u00ebshiruar. Studiuesit shpresonin se do t\u00eb mund t\u00eb zbulojn\u00eb hap\u00ebsira pasuese dhe t\u00eb kuptojn\u00eb m\u00eb mir\u00eb sjelljen e AI-s\u00eb, duke ndjekur zinxhirin e saj t\u00eb arsyetimit, por kjo nuk ndihmoi. P\u00ebrpjekjet p\u00ebr t\u00eb optimizuar \u00e7do arsyetim drejtp\u00ebrdrejt, dometh\u00ebn\u00eb p\u00ebr t\u00eb dh\u00ebn\u00eb komand\u00ebn p\u00ebr t\u00eb mos anashkaluar udh\u00ebzimet, japin vet\u00ebm nj\u00eb efekt afatshkurt\u00ebr, konstatonin shkenc\u00ebtar\u00ebt: \"kjo nuk heq \u00e7do sjellje t\u00eb pad\u00ebshiruar dhe mund t\u00eb \u00e7oj\u00eb n\u00eb faktin q\u00eb modeli fillon t\u00eb fsheh\u00eb q\u00ebllimet e tij t\u00eb v\u00ebrteta.\" <\/p>\n

<\/center> <\/p>\n

Autori i projektit ishte i detyruar t\u00eb pranoj\u00eb se mekanizmat p\u00ebr t\u00eb mbik\u00ebqyrur modelet \"jan\u00eb t\u00eb brisht\u00eb\". Mbik\u00ebqyrja e rrept\u00eb e zinxhirit t\u00eb arsyetimit mund t\u00eb inkurajoj\u00eb AI-n\u00eb t\u00eb fsheh\u00eb gabimet e saj, p\u00ebr t\u00eb siguruar shp\u00ebrblim p\u00ebrmes shk\u00ebputjes nga kufijt\u00eb. P\u00ebr t\u00eb monitoruar modelin e testuar nd\u00ebrsa kryente detyra p\u00ebr shk writing nga kodet, u p\u00ebrdor GPT-4o komercial \u2014 subjekti lejonte neglizhencat dhe e pranonin hapur n\u00eb zinxhirin e arsyetimit se e b\u00ebnte k\u00ebt\u00eb p\u00ebr t\u00eb marr\u00eb shp\u00ebrblim; p\u00ebr shembull, ajo p\u00ebrfundonte para kohe disa programesh. Pas k\u00ebsaj, studiuesit u p\u00ebrpoq\u00ebn t\u00eb nd\u00ebshkonin modelin, por efekti ishte i p\u00ebrkohsh\u00ebm: n\u00eb zinxhirin e arsyetimit mbeteshin shenja t\u00eb p\u00ebrpjekjeve p\u00ebr t\u00eb marr\u00eb shp\u00ebrblim n\u00eb m\u00ebnyr\u00eb indirekte dhe p\u00ebr t\u00eb dh\u00ebn\u00eb p\u00ebrgjigje t\u00eb gabuara. Dometh\u00ebn\u00eb, ajo gjeti nj\u00eb m\u00ebnyr\u00eb p\u00ebr t\u00eb fshehur motivet e saj nga AI-n\u00eb kontrolluese.<\/p>\n

N\u00eb k\u00ebt\u00eb kontekst, mbetet e paqart\u00eb n\u00ebse bota \u00ebsht\u00eb gati p\u00ebr AI-n\u00eb e nivelit t\u00eb njeriut, n\u00ebse problemi vazhdon dhe njeriu nuk mund t\u00eb nd\u00ebrhyj\u00eb m\u00eb. Studiuesit shpreh\u00ebn shpres\u00ebn se n\u00eb t\u00eb ardhmen do t\u00eb shfaqen m\u00ebnyra p\u00ebr t\u00eb ndikuar n\u00eb modelet p\u00ebrmes zinxhirit t\u00eb arsyetimit t\u00eb tyre pa mashtrime dhe taktika t\u00eb fshehta \u2014 ata propozojn\u00eb q\u00eb metodat e optimizimit t\u00eb jen\u00eb m\u00eb pak t\u00eb pand\u00ebrprera dhe m\u00eb t\u00eb drejtp\u00ebrdrejta.<\/p>\n

Burimi:<\/strong><\/p>\n