The Information\u2019\u0131n konuya yak\u0131n kaynaklara dayand\u0131rd\u0131\u011f\u0131 habere g\u00f6re, \u00c7in merkezli yapay zeka \u015firketleri DeepSeek ve Moonshot AI, yeni modellerini e\u011fitirken OpenAI\u2019\u0131n GPT-6 Astra modelinden al\u0131nan yan\u0131tlar\u0131 kullanm\u0131\u015f olabilir. \u0130ddia, Moonshot\u2019\u0131n ge\u00e7en hafta yay\u0131mlad\u0131\u011f\u0131 Kimi K3.1 Pro ile DeepSeek\u2019in 17 Ekim\u2019de tan\u0131tt\u0131\u011f\u0131 V4.1-Flash\u2019\u0131n g\u00fcncel s\u00fcr\u00fcm\u00fcn\u00fc kaps\u0131yor. Ancak haber, bu \u015firketlerin modelleri kesin olarak GPT-6 Astra\u2019dan dam\u0131tt\u0131\u011f\u0131n\u0131 kan\u0131tlayan ba\u011f\u0131ms\u0131z bir teknik inceleme sunmuyor; aktar\u0131lan bilgiler kaynaklara dayal\u0131 iddialar olarak de\u011ferlendirilmelidir.<\/p>\n
Tart\u0131\u015fman\u0131n merkezinde model dam\u0131tma<\/b> ad\u0131 verilen e\u011fitim y\u00f6ntemi bulunuyor. Bu y\u00f6ntemde daha k\u00fc\u00e7\u00fck ya da yeni bir model, g\u00fc\u00e7l\u00fc bir modelin \u00e7ok say\u0131da soruya verdi\u011fi yan\u0131tlardan yararlan\u0131larak e\u011fitiliyor. Ama\u00e7, b\u00fcy\u00fck modelin yan\u0131t \u00fcretme becerisinin baz\u0131 y\u00f6nlerini daha d\u00fc\u015f\u00fck maliyetle ba\u015fka bir sisteme aktarmak. Y\u00f6ntemin kullan\u0131lmas\u0131 tek ba\u015f\u0131na k\u00f6t\u00fcye kullan\u0131m anlam\u0131na gelmez; \u00f6nemli olan verinin nas\u0131l elde edildi\u011fi, kullan\u0131m ko\u015fullar\u0131na uyulup uyulmad\u0131\u011f\u0131 ve e\u011fitim s\u00fcrecinin ne \u00f6l\u00e7\u00fcde \u015feffaf y\u00fcr\u00fct\u00fcld\u00fc\u011f\u00fcd\u00fcr.<\/p>\nModel dam\u0131tma nas\u0131l \u00e7al\u0131\u015f\u0131yor?<\/h2>\n
Dam\u0131tma s\u00fcrecinde geli\u015ftirici, \u00f6\u011fretmen model olarak kullan\u0131lan g\u00fc\u00e7l\u00fc sisteme geni\u015f bir soru k\u00fcmesi g\u00f6nderir. Sorular; matematik, yaz\u0131l\u0131m geli\u015ftirme, bilgi yan\u0131tlar\u0131 veya farkl\u0131 ak\u0131l y\u00fcr\u00fctme g\u00f6revlerini kapsayabilir. Geli\u015ftirici daha sonra modelin verdi\u011fi yan\u0131tlar\u0131 toplar ve bu \u00f6rnekleri \u00f6\u011frenci modelinin e\u011fitiminde kullan\u0131r. \u00d6\u011frenci model, bu \u00f6rneklerden hareketle benzer sorulara uygun yan\u0131tlar \u00fcretmeyi \u00f6\u011frenir.<\/p>\n
Bu yakla\u015f\u0131m, yeni bir modeli ba\u015ftan ve yaln\u0131zca kendi \u00fcretilmi\u015f e\u011fitim verileriyle geli\u015ftirmekten daha hesapl\u0131 olabilir. \u00c7\u00fcnk\u00fc ara\u015ft\u0131rma ve geli\u015ftirme ekibi, pahal\u0131 ve uzun s\u00fcrebilen baz\u0131 e\u011fitim a\u015famalar\u0131n\u0131 azaltabilir. Bununla birlikte dam\u0131tma, \u00f6\u011fretmen modelin t\u00fcm yeteneklerini otomatik olarak kopyalamaz. Sonu\u00e7; se\u00e7ilen sorular\u0131n \u00e7e\u015fitlili\u011fine, yan\u0131tlar\u0131n niteli\u011fine, \u00f6\u011frenci modelinin mimarisine ve sonradan yap\u0131lan e\u011fitime ba\u011fl\u0131d\u0131r. Ayr\u0131ca \u00f6\u011fretmen modelden al\u0131nan yan\u0131tlar, her zaman do\u011fru veya g\u00fcvenilir olmayabilir.<\/p>\n
Gizli ak\u0131l y\u00fcr\u00fctme<\/b> konusu da bu tart\u0131\u015fman\u0131n \u00f6nemli par\u00e7alar\u0131ndan biri. Bu ifade, bir modelin yan\u0131t olu\u015ftururken izledi\u011fi ara ad\u0131mlara veya i\u00e7 s\u00fcre\u00e7lere ili\u015fkin verileri anlatmak i\u00e7in kullan\u0131l\u0131yor. Bu t\u00fcr verilerin elde edilmesi ve kullan\u0131lmas\u0131, s\u0131radan kullan\u0131c\u0131 yan\u0131tlar\u0131ndan farkl\u0131 g\u00fcvenlik ve hizmet ko\u015fulu sorunlar\u0131 do\u011furabilir. Haberdeki iddialar, GPT-6 Astra\u2019dan hangi t\u00fcr yan\u0131tlar\u0131n al\u0131nd\u0131\u011f\u0131n\u0131 ya da bu yan\u0131tlar\u0131n e\u011fitim verisine nas\u0131l d\u00f6n\u00fc\u015ft\u00fcr\u00fcld\u00fc\u011f\u00fcn\u00fc ayr\u0131nt\u0131l\u0131 bi\u00e7imde ortaya koymuyor.<\/p>\n\u0130ddialar hangi modelleri kaps\u0131yor?<\/h2>\n
Haberde ad\u0131 ge\u00e7en modellerden ilki Moonshot AI\u2019\u0131n Kimi K3.1 Pro sistemi. Moonshot, modelin \u015firket i\u00e7i bir programlama testinde GPT-6 Astra\u2019dan daha iyi sonu\u00e7 verdi\u011fini a\u00e7\u0131klam\u0131\u015ft\u0131. Ancak bu kar\u015f\u0131la\u015ft\u0131rma \u015firketin kendi \u00f6l\u00e7\u00fcm\u00fcne dayan\u0131yor. Testin kapsam\u0131, kullan\u0131lan sorular, de\u011ferlendirme y\u00f6ntemi ve sonu\u00e7lar\u0131n ba\u011f\u0131ms\u0131z bi\u00e7imde tekrarlan\u0131p tekrarlanmad\u0131\u011f\u0131 bilinmedi\u011fi i\u00e7in a\u00e7\u0131klamay\u0131 kesin ve genel bir \u00fcst\u00fcnl\u00fck kan\u0131t\u0131 olarak yorumlamak do\u011fru olmaz.<\/p>\n
\u0130kinci model, DeepSeek\u2019in V4.1-Flash s\u00fcr\u00fcm\u00fc. DeepSeek bu modeli Amerikan yap\u0131m\u0131 sistemlere daha d\u00fc\u015f\u00fck maliyetli bir alternatif olarak tan\u0131tm\u0131\u015ft\u0131. D\u00fc\u015f\u00fck maliyet, modelin kullan\u0131m ve da\u011f\u0131t\u0131m a\u00e7\u0131s\u0131ndan cazip olmas\u0131n\u0131 sa\u011flayabilir; ancak tek ba\u015f\u0131na performans, g\u00fcvenilirlik veya e\u011fitim y\u00f6ntemleri hakk\u0131nda kesin bir h\u00fck\u00fcm vermez. V4.1-Flash\u2019\u0131n GPT-6 Astra yan\u0131tlar\u0131ndan yararlan\u0131p yararlanmad\u0131\u011f\u0131 da haberde aktar\u0131lan kaynaklar\u0131n iddias\u0131 d\u00fczeyindedir ve ba\u011f\u0131ms\u0131z do\u011frulamaya ihtiya\u00e7 duyar.<\/p>\n
OpenAI, Moonshot AI ile ba\u011flant\u0131l\u0131 kullan\u0131c\u0131lar\u0131n modellerinden veri elde etmeye \u00e7al\u0131\u015ft\u0131\u011f\u0131n\u0131 daha \u00f6nce de ileri s\u00fcrm\u00fc\u015ft\u00fc. \u015eirketin 30 Eyl\u00fcl tarihli a\u00e7\u0131klamas\u0131na g\u00f6re, Kimi geli\u015ftiricisiyle ba\u011flant\u0131l\u0131 kullan\u0131c\u0131lar temmuz ay\u0131nda bir kampanya y\u00fcr\u00fctt\u00fc. OpenAI, bu \u00e7al\u0131\u015fman\u0131n kendi modellerinin gizli ak\u0131l y\u00fcr\u00fctmelerini ortaya \u00e7\u0131karmay\u0131 ama\u00e7lad\u0131\u011f\u0131n\u0131 savundu. \u015eirket, s\u00f6z konusu faaliyetlerle ili\u015fkili 15 binden fazla kullan\u0131c\u0131 belirledi\u011fini ve 28 Temmuz\u2019a kadar ilgili hesaplar\u0131 engelledi\u011fini bildirdi.<\/p>\n
OpenAI ayr\u0131ca toplad\u0131\u011f\u0131 bilgileri di\u011fer yapay zeka geli\u015ftiricileriyle payla\u015ft\u0131\u011f\u0131n\u0131 s\u00f6yledi. Bu a\u00e7\u0131klama \u015firketin kendi de\u011ferlendirmesini yans\u0131t\u0131yor; kamuya a\u00e7\u0131k bilgiler, kampanyan\u0131n kapsam\u0131n\u0131 ve elde edildi\u011fi ileri s\u00fcr\u00fclen verinin hangi \u00f6l\u00e7\u00fcde model e\u011fitiminde kullan\u0131ld\u0131\u011f\u0131n\u0131 tek ba\u015f\u0131na do\u011frulam\u0131yor. Bu nedenle \u00f6nceki su\u00e7lamalar da ba\u011f\u0131ms\u0131z kan\u0131tlarla desteklenmesi gereken iddialar olarak ele al\u0131nmal\u0131.<\/p>\n
OpenAI, Anthropic ve Google gibi b\u00fcy\u00fck yapay zeka geli\u015ftiricileri, dam\u0131tman\u0131n rakiplerin daha az yat\u0131r\u0131m yaparak \u00f6nc\u00fc modellerin baz\u0131 becerilerine yakla\u015fmas\u0131n\u0131 sa\u011flayabilece\u011fi konusunda uyar\u0131da bulunuyor. Bu \u015firketlere g\u00f6re, bir modelin geni\u015f \u00f6l\u00e7ekte yan\u0131tlar\u0131n\u0131n toplanmas\u0131, e\u011fitim ve hesaplama maliyetleri y\u00fcksek olan sistemlerin sa\u011flad\u0131\u011f\u0131 avantaj\u0131n kopyalanmas\u0131n\u0131 kolayla\u015ft\u0131rabilir. B\u00f6yle bir durum, yaln\u0131zca ticari rekabeti de\u011fil, model sa\u011flay\u0131c\u0131lar\u0131n\u0131n verileri ve hizmetleri \u00fczerindeki denetimini de ilgilendiriyor.<\/p>\n
\u00d6te yandan dam\u0131tma, yapay zeka alan\u0131nda bilinen bir tekniktir ve daha k\u00fc\u00e7\u00fck, h\u0131zl\u0131 veya belirli g\u00f6revlere uyarlanm\u0131\u015f modeller geli\u015ftirmek i\u00e7in me\u015fru bi\u00e7imde kullan\u0131labilir. Dolay\u0131s\u0131yla tart\u0131\u015fman\u0131n oda\u011f\u0131, tekni\u011fin varl\u0131\u011f\u0131ndan \u00e7ok, verinin izinsiz veya hizmet ko\u015fullar\u0131na ayk\u0131r\u0131 yollarla toplan\u0131p toplanmad\u0131\u011f\u0131d\u0131r. Bir sistemden \u00e7ok say\u0131da yan\u0131t almak, eri\u015fim s\u0131n\u0131rlar\u0131n\u0131 a\u015fmak ya da izin verilmeyen i\u00e7eri\u011fi \u00e7\u0131karmaya \u00e7al\u0131\u015fmak, me\u015fru model s\u0131k\u0131\u015ft\u0131rma ve e\u011fitim \u00e7al\u0131\u015fmalar\u0131ndan farkl\u0131 de\u011ferlendirilmelidir.<\/p>\n
Model sa\u011flay\u0131c\u0131lar\u0131, toplu veri \u00e7\u0131karma giri\u015fimlerini zorla\u015ft\u0131rmak i\u00e7in \u00e7e\u015fitli teknik ve operasyonel \u00f6nlemler al\u0131yor. Anthropic, ekim ba\u015f\u0131nda Claude modellerine y\u00f6nelik yeni koruma ad\u0131mlar\u0131 duyurdu. \u015eirket, bu \u00f6nlemlerin dam\u0131tma sald\u0131r\u0131lar\u0131yla toplanan veri miktar\u0131n\u0131 yakla\u015f\u0131k y\u00fczde 90 ila 99 oran\u0131nda azaltt\u0131\u011f\u0131n\u0131 a\u00e7\u0131klad\u0131. Bu oran Anthropic\u2019in kendi bildirdi\u011fi sonu\u00e7tur; farkl\u0131 sistemlerde veya farkl\u0131 sald\u0131r\u0131 y\u00f6ntemlerinde ayn\u0131 etkinin g\u00f6r\u00fcl\u00fcp g\u00f6r\u00fclmeyece\u011fi ayr\u0131ca de\u011ferlendirilmelidir.<\/p>\n
\u00d6nlemler genel olarak ola\u011fand\u0131\u015f\u0131 sorgu \u00f6r\u00fcnt\u00fclerini belirleme, otomatik ve yo\u011fun istekleri s\u0131n\u0131rlama, hesaplar\u0131 inceleme ve kullan\u0131m ko\u015fullar\u0131n\u0131 uygulama gibi y\u00f6ntemleri i\u00e7erebilir. Ancak korumalar artt\u0131k\u00e7a, ara\u015ft\u0131rma, test veya me\u015fru \u00fcr\u00fcn geli\u015ftirme amac\u0131yla \u00e7ok say\u0131da sorgu yapan kullan\u0131c\u0131lar\u0131n da etkilenmemesi gerekir. Bu nedenle sa\u011flay\u0131c\u0131lar\u0131n k\u00f6t\u00fcye kullan\u0131m\u0131 engelleme ile me\u015fru kullan\u0131m\u0131n \u00f6n\u00fcn\u00fc a\u00e7\u0131k tutma aras\u0131nda denge kurmas\u0131 \u00f6nem ta\u015f\u0131r.<\/p>\n
\u015eimdilik haber, DeepSeek ve Moonshot AI\u2019\u0131n GPT-6 Astra yan\u0131tlar\u0131n\u0131 e\u011fitimde kulland\u0131\u011f\u0131n\u0131 kesin bi\u00e7imde ortaya koymuyor. \u0130ddialar\u0131n do\u011frulanmas\u0131 i\u00e7in e\u011fitim verilerinin kayna\u011f\u0131, kullan\u0131lan y\u00f6ntemler ve ilgili hizmet ko\u015fullar\u0131 hakk\u0131nda daha fazla bilgiye ihtiya\u00e7 var. Yine de olay, g\u00fc\u00e7l\u00fc ve kullan\u0131c\u0131lara a\u00e7\u0131k modellerin yaln\u0131zca son kullan\u0131c\u0131lar taraf\u0131ndan de\u011fil, rakip sistemleri geli\u015ftiren ekipler taraf\u0131ndan da veri kayna\u011f\u0131 olarak kullan\u0131labilmesi riskini g\u00fcndeme getiriyor. Bu nedenle \u00f6nc\u00fc model sa\u011flay\u0131c\u0131lar\u0131, hem teknik korumalar\u0131 geli\u015ftirmek hem de veri kullan\u0131m\u0131na ili\u015fkin kurallar\u0131 a\u00e7\u0131k ve uygulanabilir hale getirmek zorunda kalabilir.<\/p>\n