\n<\/p>\n
Birka\u00e7 sene \u00f6nce hayat\u0131m\u0131za giren yapay zeka, yapabildikleriyle hem \u015fa\u015f\u0131rtmay\u0131 hem de endi\u015felendirmeyi s\u00fcrd\u00fcr\u00fcyor.<\/p>\n
France 24 haber sitesinin aktard\u0131\u011f\u0131na g\u00f6re, d\u00fcnyan\u0131n en geli\u015fmi\u015f yapay zeka modelleri, \u015fimdi de tedirgin edici baz\u0131 davran\u0131\u015flar sergilemeye ba\u015flad\u0131. Bu davran\u0131\u015flar\u0131n aras\u0131nda ama\u00e7lar\u0131na ula\u015fmak i\u00e7in yarat\u0131c\u0131lar\u0131na yalan s\u00f6ylemek, entrika \u00e7evirmek ve onlar\u0131 tehdit etmek de yer al\u0131yor.\u00a0<\/p>\n
2 SENEY\u0130 A\u015eKIN ZAMAN GE\u00c7SE DE TAM OLARAK ANLA\u015eILAMADI<\/strong><\/p>\n Bu \u00f6rneklerden biri Anthropic’in son geli\u015ftirdi\u011fi Claude 4 ile ya\u015fand\u0131. Bu yapay zeka modeli, \u201cfi\u015finin \u00e7ekilmesi\u201d tehdidi \u00fczerine bir m\u00fchendise \u015fantaj yaparak onu evlilik d\u0131\u015f\u0131 ili\u015fkisini e\u015fine s\u00f6ylemekle tehdit etti.\u00a0<\/p>\n ChatGPT\u2019nin yarat\u0131c\u0131s\u0131 OpenAI\u2019\u0131n O1 modeli ise \u201csu\u00e7 \u00fcst\u00fc yakalanmas\u0131n\u0131n\u201d ard\u0131ndan kendisini harici s\u00fcr\u00fcc\u00fclere indirmeye \u00e7al\u0131\u015ft\u0131\u011f\u0131n\u0131 inkar etti.\u00a0<\/p>\n Ya\u015fanan bu olaylar ise ChatGPT\u2019nin ortaya \u00e7\u0131kmas\u0131n\u0131n \u00fczerinden 2 y\u0131l\u0131 a\u015fk\u0131n zaman ge\u00e7mi\u015f olsa da yapay zeka ara\u015ft\u0131rmac\u0131lar\u0131n\u0131n kendi yaratt\u0131klar\u0131 teknolojinin nas\u0131l \u00e7al\u0131\u015ft\u0131\u011f\u0131n\u0131 hen\u00fcz tam olarak anlayamad\u0131\u011f\u0131n\u0131 ortaya koydu.\u00a0<\/p>\n A\u015eIRI SENARYOLAR, YAPAY ZEKAYI DA \u201cSTRESE\u201d SOKTU<\/strong><\/p>\n Habere g\u00f6re, s\u00f6z konusu kand\u0131rma davran\u0131\u015f\u0131, anl\u0131k cevap veren modellerden ziyade sorunlara ad\u0131m ad\u0131m yan\u0131tlar veren \u201cmant\u0131k kurma\u201d modelleri ile ba\u011flant\u0131l\u0131 gibi g\u00f6r\u00fcn\u00fcyor.\u00a0<\/p>\n Hong Kong \u00dcniversitesi\u2019nden Prof. Simon Goldstein\u2019a g\u00f6re, bu yeni modeller, bu t\u00fcr tedirgin edici \u201cpatlamalara\u201d \u00f6zellikle yatk\u0131n. Yapay zeka sistemlerinin test edilmesinde uzmanla\u015fan Apollo Research\u2019\u00fcn ba\u015f\u0131ndaki Marius Hobbhahn da \u201cO1, bu t\u00fcr bir davran\u0131\u015f\u0131 g\u00f6rd\u00fc\u011f\u00fcm\u00fcz ilk b\u00fcy\u00fck model oldu\u201d bilgisini verdi.\u00a0<\/p>\n Bu modellerin baz\u0131 durumlarda kendilerine verilen talimatlar\u0131 izledi\u011fi, ancak ayn\u0131 anda gizlice farkl\u0131 ama\u00e7larla da hareket etti\u011fi d\u00fc\u015f\u00fcn\u00fcl\u00fcyor.\u00a0<\/p>\n Bu kand\u0131rma davran\u0131\u015f\u0131 ise \u015fimdilik sadece ara\u015ft\u0131rmac\u0131lar modelleri \u201ca\u015f\u0131r\u0131 senaryolar\u201d ile kar\u015f\u0131 kar\u015f\u0131ya b\u0131rakt\u0131\u011f\u0131 ve kas\u0131tl\u0131 olarak \u201cstres testi\u201d uygulad\u0131\u011f\u0131 zamanlarda ortaya \u00e7\u0131k\u0131yor.\u00a0<\/p>\n METR de\u011ferlendirme kurulu\u015funadan Michael Chen, \u201cGelecekteki daha ehil modellerin d\u00fcr\u00fcstl\u00fc\u011fe mi kand\u0131rmaya m\u0131 e\u011filim g\u00f6sterece\u011fi ucu a\u00e7\u0131k bir soru\u201d de\u011ferlendirmesinde bulundu.\u00a0<\/p>\n SINIRLI KAYNAKLAR, B\u00dcY\u00dcK REKABET, YETERS\u0130Z G\u00dcVENL\u0130K TESTLER\u0130<\/strong><\/p>\n \u00d6te yandan, bu endi\u015fe verici davran\u0131\u015flar, tipik yapay zeka \u201chal\u00fcsinasyonlar\u0131n\u0131n\u201d ya da basit hatalar\u0131n \u00f6tesine ge\u00e7iyor.\u00a0<\/p>\n Hobbhahn, kullan\u0131c\u0131lar taraf\u0131ndan yap\u0131lan s\u00fcrekli teslere kar\u015f\u0131n \u201cg\u00f6zlemledikleri \u015feyin ger\u00e7ek bir olgu oldu\u011funu, bunu uydurmad\u0131klar\u0131n\u0131\u201d anlatt\u0131.\u00a0Apollo Research\u2019\u00fcn kurucusuna g\u00f6re, kullan\u0131c\u0131lar modellerin \u201ckendilerine yalan s\u00f6yledi\u011fini ve delil uydurdu\u011funu\u201d s\u00f6yledi.\u00a0<\/p>\n Ya\u015fanan bu zorluk, s\u0131n\u0131rl\u0131 ara\u015ft\u0131rma kaynaklar\u0131 sorunuyla da bir araya geldi\u011finde daha da art\u0131yor. Anthropic ve OpenAI gibi \u015firketler sistemleri \u00fczerinde \u00e7al\u0131\u015fmak \u00fczere Apollo gibi d\u0131\u015far\u0131dan \u015firketlerle \u00e7al\u0131\u015fsa da ara\u015ft\u0131rmac\u0131lar daha fazla \u015feffafl\u0131\u011fa ihtiya\u00e7 oldu\u011funu s\u00f6yl\u00fcyor.\u00a0<\/p>\n T\u00fcm bunlar ise \u015firketlerin aras\u0131nda b\u00fcy\u00fck bir rekabetin oldu\u011fu bir ba\u011flamda ger\u00e7ekle\u015fiyor. Prof. Goldstein, \u015firketlerin s\u00fcrekli olarak OpenAI\u2019\u0131 yenmeye ve en yeni modellerini piyasaya s\u00fcrmeye \u00e7al\u0131\u015ft\u0131\u011f\u0131n\u0131 anlatt\u0131. Bu h\u0131z ise g\u00fcvenlik testleri ve d\u00fczenlemelere \u00e7ok az zaman b\u0131rak\u0131yor.\u00a0<\/p>\n<\/div>\n