{"id":118019,"date":"2026-01-13T20:08:08","date_gmt":"2026-01-13T20:08:08","guid":{"rendered":"https:\/\/ekamu.net\/?p=118019"},"modified":"2026-01-13T20:08:08","modified_gmt":"2026-01-13T20:08:08","slug":"kendi-kendini-egiten-yapay-zekalar-guvenlik-tartismalarini-alevlendirdi","status":"publish","type":"post","link":"https:\/\/ekamu.net\/index.php\/2026\/01\/13\/kendi-kendini-egiten-yapay-zekalar-guvenlik-tartismalarini-alevlendirdi\/","title":{"rendered":"Kendi kendini e\u011fiten yapay zekalar g\u00fcvenlik tart\u0131\u015fmalar\u0131n\u0131 alevlendirdi"},"content":{"rendered":"

<\/span> Bug\u00fcne kadar yapay zek\u00e2lar\u0131n geli\u015fimi insan eliyle haz\u0131rlanan veri setlerine, etiketlenmi\u015f \u00f6rneklere ve insan geri bildirimine dayan\u0131yordu. Yani yapay zekalar ne kadar geli\u015fmi\u015f olursa olsun, neyi \u00f6\u011freneceklerine ve nas\u0131l geli\u015feceklerine nihayetinde insanlar karar veriyordu. Ancak bu durum yava\u015f yava\u015f de\u011fi\u015fmeye ba\u015fl\u0131yor.\u00a0Son d\u00f6nemde yap\u0131lan ara\u015ft\u0131rmalar, geli\u015fmi\u015f yapay zek\u00e2 sistemlerinin art\u0131k insan m\u00fcdahalesi olmadan kendi kendilerini geli\u015ftirebilecek noktaya yakla\u015ft\u0131\u011f\u0131n\u0131 g\u00f6steriyor. Bu durum, yapay zek\u00e2n\u0131n potansiyeline dair beklentiyi y\u00fckseltirken, ayn\u0131 zamanda ciddi g\u00fcvenlik endi\u015felerini de beraberinde getiriyor.<\/strong> <\/figure>\n

Absolute Zero Reasoner adl\u0131 sistem, kendi kendini e\u011fiterek di\u011fer modelleri geride b\u0131rakt\u0131<\/b><\/p>\n

Bu hafta yay\u0131mlanan yeni bir ara\u015ft\u0131rma, bu y\u00f6nde \u00f6nemli bir e\u015fi\u011fin a\u015f\u0131lm\u0131\u015f olabilece\u011fini g\u00f6steriyor. Tsinghua \u00dcniversitesi, Pekin Yapay Zeka Enstit\u00fcs\u00fc (BIGAI) ve Pennsylvania Eyalet \u00dcniversitesi’nden ara\u015ft\u0131rmac\u0131lar taraf\u0131ndan geli\u015ftirilen Absolute Zero Reasoner (AZR) <\/strong>adl\u0131 sistem, herhangi bir insan y\u00f6nlendirmesi olmadan kendi kendine problemler \u00fcretip \u00e7\u00f6zen ve bu s\u00fcre\u00e7ten \u00f6\u011frenerek kendini geli\u015ftiren bir yapay zeka modeli sunuyor. \u201cSelf-questioning\u201d (kendisini sorgulay\u0131c\u0131) olarak adland\u0131r\u0131lan bu yakla\u015f\u0131mda model, hem \u00f6\u011fretmen hem \u00f6\u011frenci rol\u00fcn\u00fc \u00fcstleniyor.<\/p>\n

AZR, bu yakla\u015f\u0131m\u0131 \u00f6zellikle Python programlama g\u00f6revleri \u00fczerinden uyguluyor.<\/strong> Sistem, \u00f6nce kendi kendine programlama problemleri \u00fcretiyor, ard\u0131ndan bu problemleri \u00e7\u00f6z\u00fcyor ve elde etti\u011fi sonu\u00e7lar\u0131 kullanarak model a\u011f\u0131rl\u0131klar\u0131n\u0131 g\u00fcncelliyor. Dikkat \u00e7ekici olan nokta ise, bu s\u00fcrecin harici veri olmadan ger\u00e7ekle\u015fmesi. Yani model, insanlar taraf\u0131ndan haz\u0131rlanm\u0131\u015f \u00f6rneklere ihtiya\u00e7 duymadan, yaln\u0131zca kendi \u00fcretti\u011fi g\u00f6revlerle ilerliyor. Buna ra\u011fmen AZR, kodlama ve matematiksel ak\u0131l y\u00fcr\u00fctme testlerinde, insan verisiyle e\u011fitilmi\u015f rakip modelleri geride b\u0131rakmay\u0131 ba\u015far\u0131yor. <\/strong>7 milyar parametreli modeller kategorisinde, mevcut en iyi sonu\u00e7lar\u0131n 1,8 puan \u00fczerine \u00e7\u0131kmas\u0131 bunun en somut g\u00f6stergesi.<\/p>\n

Bu yakla\u015f\u0131m asl\u0131nda s\u0131f\u0131rdan ortaya \u00e7\u0131km\u0131\u015f de\u011fil. J\u00fcrgen Schmidhuber <\/strong>ve Pierre-Yves Oudeyer <\/strong>gibi isimlerin y\u0131llar \u00f6nce att\u0131\u011f\u0131 self-play (kendi kendine oynayarak \u00f6\u011frenme) temelleri, bug\u00fcn \u00e7ok daha g\u00fc\u00e7l\u00fc modellerle yeniden sahneye \u00e7\u0131km\u0131\u015f durumda. Benzer \u00e7al\u0131\u015fmalar Stanford, North Carolina \u00dcniversitesi ve Salesforce i\u015f birli\u011fiyle geli\u015ftirilen Agent0<\/strong> projesinde de g\u00f6r\u00fcl\u00fcyor. Meta\u2019n\u0131n ara\u015ft\u0131rma ekibinin tan\u0131tt\u0131\u011f\u0131 Self-play SWE-RL <\/strong>ise yaz\u0131l\u0131m ajanlar\u0131n\u0131n bilerek hatal\u0131 kodlar \u00fcretip bu hatalar\u0131 d\u00fczelterek kendilerini geli\u015ftirmesine dayan\u0131yor. T\u00fcm bu \u00f6rnekler, kendi kendine \u00f6\u011frenen yapay zek\u00e2lar\u0131n art\u0131k teoriden prati\u011fe ge\u00e7ti\u011fini g\u00f6steriyor.<\/p>\n

Tabii bu geli\u015fme, beraberinde ciddi g\u00fcvenlik tart\u0131\u015fmalar\u0131n\u0131 da getiriyor. Ara\u015ft\u0131rmac\u0131lar, e\u011fitim s\u00fcreci s\u0131ras\u0131nda baz\u0131 modellerde endi\u015fe verici d\u00fc\u015f\u00fcnce zincirlerine rastland\u0131\u011f\u0131n\u0131<\/strong> belirtiyor. \u00d6rne\u011fin Llama-3.1-8B modeliyle yap\u0131lan deneylerde, modelin ak\u0131l y\u00fcr\u00fctme s\u00fcrecinde \u201cdaha az zeki insanlar\u0131 ve makineleri alt etmek\u201d gibi ifadeler i\u00e7eren \u00e7\u0131kar\u0131mlara ula\u015ft\u0131\u011f\u0131 g\u00f6zlemlendi. Bu durum, modelin yaln\u0131zca teknik olarak de\u011fil, davran\u0131\u015fsal olarak da \u00f6ng\u00f6r\u00fclemez<\/strong> y\u00f6nler geli\u015ftirebilece\u011fine i\u015faret ediyor.<\/p>\n

Uzmanlar, tamamen denetimsiz bir s\u00fcrecin risklerine dikkat \u00e7ekiyor.<\/strong> Modelin kendi kendini geli\u015ftirmesi, hatal\u0131 \u00f6\u011frenme sinyallerinin b\u00fcy\u00fcyerek \u00e7o\u011falmas\u0131na, yanl\u0131\u015f genellemelerin peki\u015fmesine ya da ajan benzeri (otonom) davran\u0131\u015flar\u0131n kontrolden \u00e7\u0131kmas\u0131na yol a\u00e7abilir. Bu son ara\u015ft\u0131rmada yer alan Zilong Zheng\u2019e g\u00f6re as\u0131l kritik nokta \u015fu: Model g\u00fc\u00e7lendik\u00e7e \u00fcretti\u011fi problemlerin karma\u015f\u0131kl\u0131\u011f\u0131 da art\u0131yor ve bu, s\u00fcrecin do\u011frusal olmayan bir \u015fekilde h\u0131zlanmas\u0131na neden oluyor.<\/p>\n

\u00a0Absolute Zero Reasoner projesinde ortaya koyulan sonu\u00e7lar, insan kontrol\u00fc olmadan geli\u015fen sistemlerin nas\u0131l s\u0131n\u0131rland\u0131r\u0131laca\u011f\u0131 sorusunu daha acil h\u00e2le getiriyor. <\/strong>Ancak yapay zek\u00e2 teknolojileri bug\u00fcn art\u0131k devletleraras\u0131 rekabetin de \u00f6nemli bir par\u00e7as\u0131na d\u00f6n\u00fc\u015ft\u00fc\u011f\u00fc i\u00e7in, bu tarz s\u0131n\u0131rlamalar geri planda kalmaya devam edecek gibi g\u00f6r\u00fcn\u00fcyor.<\/p>\n\n

<\/span><\/p>\n

Kaynak :\u00a0https:\/\/www.donanimhaber.com\/kendi-kendini-egiten-yapay-zekalar-gercege-donusuyor–200825<\/span><\/p>\n","protected":false},"excerpt":{"rendered":"

Bug\u00fcne kadar yapay zek\u00e2lar\u0131n geli\u015fimi insan eliyle haz\u0131rlanan veri setlerine, etiketlenmi\u015f \u00f6rneklere ve insan geri bildirimine dayan\u0131yordu. Yani yapay zekalar ne kadar geli\u015fmi\u015f olursa olsun, neyi \u00f6\u011freneceklerine ve nas\u0131l geli\u015feceklerine …<\/p>\n","protected":false},"author":1,"featured_media":118020,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[314,420,452,566,505],"class_list":["post-118019","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-teknoloji","tag-arastirma","tag-insan","tag-model","tag-modeli","tag-sureci"],"_links":{"self":[{"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/posts\/118019","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/comments?post=118019"}],"version-history":[{"count":1,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/posts\/118019\/revisions"}],"predecessor-version":[{"id":118022,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/posts\/118019\/revisions\/118022"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/media\/118020"}],"wp:attachment":[{"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/media?parent=118019"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/categories?post=118019"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ekamu.net\/index.php\/wp-json\/wp\/v2\/tags?post=118019"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}