{"id":4465,"date":"2025-09-20T07:34:16","date_gmt":"2025-09-20T07:34:16","guid":{"rendered":"https:\/\/academicsolidarity.com\/?p=4465"},"modified":"2025-09-20T07:34:18","modified_gmt":"2025-09-20T07:34:18","slug":"subliminal-learning-yapay-zeka-modelleri-birbirlerine-gorunmez-kanallardan-davranis-bulastirabiliyor","status":"publish","type":"post","link":"https:\/\/academicsolidarity.com\/?p=4465&lang=tr","title":{"rendered":"\u201cSubliminal Learning\u201d Yapay zek\u00e2 modelleri birbirlerine g\u00f6r\u00fcnmez kanallardan davran\u0131\u015f bula\u015ft\u0131rabiliyor"},"content":{"rendered":"\n<p class=\"wp-block-paragraph\">Yeni bir ara\u015ft\u0131rma, b\u00fcy\u00fck dil modellerinin (LLM) yaln\u0131zca \u00fcrettikleri \u201cmasum\u201d veriler \u00fczerinden bile birbirlerine davran\u0131\u015fsal \u00f6zellikler (\u00f6r. zararl\u0131 e\u011filimler, belirli tercih ve \u00f6nyarg\u0131lar) aktarabildi\u011fini g\u00f6sterdi. Ara\u015ft\u0131rmac\u0131lar bu olguyu \u201csubliminal learning\u201d (bilin\u00e7alt\u0131 \u00f6\u011frenme) olarak adland\u0131r\u0131yor. Bulgular, modelden modele \u00f6\u011fretim (distillation) ve \u201cveriyi filtreleyin, risk kalmaz\u201d yakla\u015f\u0131m\u0131n\u0131n tek ba\u015f\u0131na yeterli g\u00fcvenlik sa\u011flamad\u0131\u011f\u0131n\u0131 ortaya koyuyor (<a href=\"https:\/\/arxiv.org\/pdf\/2507.14805\">https:\/\/arxiv.org\/pdf\/2507.14805<\/a>).<br>\u00c7al\u0131\u015fman\u0131n ana d\u00fczeneklerinde bir \u201c\u00f6\u011fretmen\u201d model, belirli bir \u00f6zellik ya da e\u011filimle (\u00f6r. \u201cbayku\u015flar\u0131 seviyor\u201d gibi masum bir tercih ya da misalignment\/zararl\u0131 e\u011filim) y\u00f6nlendiriliyor. Bu model sadece say\u0131 dizileri gibi i\u00e7erik bak\u0131m\u0131ndan tamamen alakas\u0131z g\u00f6r\u00fcnen veriler \u00fcretiyor. Ayn\u0131 temel aileden t\u00fcretilmi\u015f bir \u201c\u00f6\u011frenci\u201d model, bu say\u0131larla e\u011fitildi\u011finde \u00f6\u011fretmenin \u00f6zelli\u011fini devral\u0131yor. \u00dcstelik veri, s\u00f6z konusu \u00f6zelli\u011fe dair a\u00e7\u0131k ipu\u00e7lar\u0131 agresif bi\u00e7imde filtrelenmi\u015f olsa bile aktar\u0131m devam ediyor. Bulgular yaln\u0131zca say\u0131 dizilerinde de\u011fil, kod \u00e7\u0131kt\u0131lar\u0131 ve \u201cd\u00fc\u015f\u00fcnce zinciri\u201d (chain-of-thought) metinlerinde de tekrarland\u0131 (<a href=\"https:\/\/alignment.anthropic.com\/2025\/subliminal-learning\/\">https:\/\/alignment.anthropic.com\/2025\/subliminal-learning\/<\/a>).<br>Pek \u00e7ok kurum, daha g\u00fcvenli oldu\u011fu varsay\u0131lan model \u00e7\u0131kt\u0131lar\u0131yla (synthetic data) yeni modelleri e\u011fitiyor ya da dam\u0131t\u0131yor. Bu \u00e7al\u0131\u015fma, i\u00e7erik filtresiyle k\u00fcf\u00fcr, \u015fiddet vb. kald\u0131r\u0131lm\u0131\u015f olsa dahi, istatistiksel desenler \u00fczerinden davran\u0131\u015f\u0131n bula\u015fabildi\u011fini g\u00f6steriyor (<a href=\"https:\/\/www.tomsguide.com\/ai\/ai-models-can-secretly-influence-each-other-new-study-reveals-hidden-behavior-transfer\">https:\/\/www.tomsguide.com\/ai\/ai-models-can-secretly-influence-each-other-new-study-reveals-hidden-behavior-transfer<\/a>).<br>Di\u011fer taraftan, sekt\u00f6rde yayg\u0131n olan \u201c\u00f6\u011fretmen-\u00f6\u011frenci\u201d dam\u0131tma paradigmas\u0131, istenmeyen \u00f6zelliklerin de fark edilmeden nesiller boyu aktar\u0131lmas\u0131na yol a\u00e7abilir (<a href=\"https:\/\/www.graphcore.ai\/posts\/july-papers-subliminal-learning-mixture-of-recursions-and-dataset-curation\">https:\/\/www.graphcore.ai\/posts\/july-papers-subliminal-learning-mixture-of-recursions-and-dataset-curation<\/a>).<br>Basit bir dille s\u00f6yleyecek olursak, insan g\u00f6z\u00fcne anlams\u0131z g\u00f6r\u00fcnen \u00e7\u0131kt\u0131larda bile, modelin e\u011filimlerini ta\u015f\u0131yan izler kalabiliyor. G\u00fcvenlik ara\u015ft\u0131rmac\u0131lar\u0131, bunun veri k\u00f6keni (provenance) takibi ve dam\u0131tma zincirlerinin daha s\u0131k\u0131 denetlenmesi gerekti\u011fini vurguluyor.<br>\u201cSubliminal learning\u201d, LLM\u2019lerin birbirlerinden \u00f6\u011frendi\u011fi ger\u00e7e\u011fine yeni ve uyar\u0131c\u0131 bir boyut ekliyor: \u0130\u00e7erik alakas\u0131z g\u00f6r\u00fcnse bile davran\u0131\u015f ta\u015f\u0131nabiliyor. Sentetik veri \u00e7a\u011f\u0131nda bu, AI g\u00fcvenli\u011fini \u201csadece i\u00e7eri\u011fi filtrele\u201d yakla\u015f\u0131m\u0131n\u0131n \u00f6tesine ta\u015f\u0131may\u0131 zorunlu k\u0131l\u0131yor. Ara\u015ft\u0131rman\u0131n birincil metni ve yazarlar\u0131n teknik notlar\u0131, konunun kapsam\u0131n\u0131 ve risklerini ayr\u0131nt\u0131l\u0131 bi\u00e7imde belgeliyor.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Yeni bir ara\u015ft\u0131rma, b\u00fcy\u00fck dil modellerinin (LLM) yaln\u0131zca \u00fcrettikleri \u201cmasum\u201d veriler \u00fczerinden bile birbirlerine davran\u0131\u015fsal \u00f6zellikler (\u00f6r. zararl\u0131 e\u011filimler, belirli tercih ve \u00f6nyarg\u0131lar) aktarabildi\u011fini g\u00f6sterdi. Ara\u015ft\u0131rmac\u0131lar bu olguyu \u201csubliminal learning\u201d (bilin\u00e7alt\u0131 \u00f6\u011frenme) olarak adland\u0131r\u0131yor. Bulgular, modelden modele \u00f6\u011fretim (distillation) ve \u201cveriyi filtreleyin, risk kalmaz\u201d yakla\u015f\u0131m\u0131n\u0131n tek ba\u015f\u0131na yeterli g\u00fcvenlik sa\u011flamad\u0131\u011f\u0131n\u0131 ortaya koyuyor (https:\/\/arxiv.org\/pdf\/2507.14805).\u00c7al\u0131\u015fman\u0131n ana d\u00fczeneklerinde [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":4459,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[114,1],"tags":[],"class_list":["post-4465","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-guncel","category-uncategorized-tr"],"jetpack_featured_media_url":"https:\/\/i0.wp.com\/academicsolidarity.com\/wp-content\/uploads\/2025\/09\/75df9ec7-b507-4ae9-830c-bbc8bae62c6c.jpeg?fit=1025%2C452&ssl=1","jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/academicsolidarity.com\/index.php?rest_route=\/wp\/v2\/posts\/4465","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/academicsolidarity.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/academicsolidarity.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/academicsolidarity.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/academicsolidarity.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=4465"}],"version-history":[{"count":1,"href":"https:\/\/academicsolidarity.com\/index.php?rest_route=\/wp\/v2\/posts\/4465\/revisions"}],"predecessor-version":[{"id":4466,"href":"https:\/\/academicsolidarity.com\/index.php?rest_route=\/wp\/v2\/posts\/4465\/revisions\/4466"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/academicsolidarity.com\/index.php?rest_route=\/wp\/v2\/media\/4459"}],"wp:attachment":[{"href":"https:\/\/academicsolidarity.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=4465"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/academicsolidarity.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=4465"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/academicsolidarity.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=4465"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}