Yapay zek\u00e2 d\u00fcnyas\u0131nda son iki ayda birbirini tamamlayan, fakat giderek daha korkutucu hale gelen bir dizi a\u00e7\u0131klama ya\u015fan\u0131yor. \u00d6nce OpenAI, Anthropic, Google DeepMind, Meta ve di\u011fer \u00f6nc\u00fc yapay zek\u00e2 \u015firketlerinde \u00e7al\u0131\u015fanlar\u0131n ba\u015flatt\u0131\u011f\u0131 \u201cPacing the Frontier\u201d \u00e7a\u011fr\u0131s\u0131 geldi.<\/p>\n
Ard\u0131ndan 16 Nobel \u00f6d\u00fcll\u00fc iktisat\u00e7\u0131n\u0131n da aralar\u0131nda bulundu\u011fu y\u00fczlerce ekonomist, yapay zek\u00e2n\u0131n Sanayi Devrimi\u2019nden daha b\u00fcy\u00fck ve \u00e7ok daha h\u0131zl\u0131 bir ekonomik d\u00f6n\u00fc\u015f\u00fcm yaratabilece\u011fini s\u00f6yleyerek \u201c\u015fimdi harekete ge\u00e7ilmeli\u201d dedi. Bu hafta Anthropic ara\u015ft\u0131rmac\u0131s\u0131 Jacob Coxon istifa ederek \u015firketlerin \u201ckendi kendini geli\u015ftiren s\u00fcper zek\u00e2ya do\u011fru yar\u0131\u015ft\u0131\u011f\u0131n\u0131 ve hayatlar\u0131m\u0131zla kumar oynad\u0131\u011f\u0131n\u0131\u201d s\u00f6yledi.<\/p>\n
Anthropic\u2019in Alignment Science ekibinin ba\u015f\u0131ndaki Evan Hubinger ise daha da ileri giderek, yapay zek\u00e2n\u0131n \u00f6n\u00fcm\u00fczdeki on y\u0131l i\u00e7inde b\u00fct\u00fcn insanl\u0131\u011f\u0131 yok etme ihtimalini ki\u015fisel olarak y\u00fczde 10\u2019un \u00fczerinde g\u00f6rd\u00fc\u011f\u00fcn\u00fc a\u00e7\u0131klad\u0131. Ve 12 Eyl\u00fcl\u2019de (bug\u00fcn) tart\u0131\u015fman\u0131n merkezindeki \u015firketlerden birinin CEO\u2019su ilk kez \u00e7ok a\u00e7\u0131k konu\u015ftu.<\/p>\n
Anthropic CEO\u2019su Dario Amodei, yay\u0131mlad\u0131\u011f\u0131 \u201cWe Must Pace the Frontier\u201d ba\u015fl\u0131kl\u0131 makalesinde, art\u0131k frontier yapay zek\u00e2 modellerinin yeteneklerinin geli\u015ftirilme h\u0131z\u0131n\u0131n bilin\u00e7li bi\u00e7imde yava\u015flat\u0131lmas\u0131 gerekti\u011fini savundu. Amodei\u2019nin s\u00f6yledi\u011fi \u015fey \u201cAI \u00e7al\u0131\u015fmalar\u0131n\u0131 durdural\u0131m\u201d de\u011fil ama daha hassas bir \u00f6neri getiriyor. \u201cModeller geli\u015fmeye devam etsin; fakat yetenek art\u0131\u015f\u0131n\u0131n h\u0131z\u0131, g\u00fcvenlik ve kontrol mekanizmalar\u0131n\u0131n yeti\u015febilece\u011fi bir seviyeye \u00e7ekilsin\u201d diyor.<\/p>\n
Amodei\u2019ye g\u00f6re yaln\u0131zca bir veya iki y\u0131l kazanmak bile, alignment, interpretability, siber g\u00fcvenlik, test ve model kontrol\u00fc alanlar\u0131nda b\u00fcy\u00fck fark yaratabilir. Amodei\u2019nin birka\u00e7 y\u0131l \u00f6nce bu fikirde de\u011fildi. 2023\u2019te ortaya at\u0131lan \u201cAI geli\u015ftirmeye ara verelim\u201d \u00e7a\u011fr\u0131lar\u0131n\u0131n o d\u00f6nemde kendisine pek anlaml\u0131 gelmedi\u011fini a\u00e7\u0131k\u00e7a s\u00f6yl\u00fcyor.<\/p>\n
\u00c7\u00fcnk\u00fc modeller hen\u00fcz d\u00fcnyada ba\u011f\u0131ms\u0131z hareket edebilecek, ciddi manip\u00fclasyon yapabilecek, aldatabilecek veya b\u00fcy\u00fck siber sald\u0131r\u0131lar ger\u00e7ekle\u015ftirebilecek seviyede de\u011fildi. Bug\u00fcn ise tablo de\u011fi\u015fmi\u015f durumda. \u00d6zellikle iki geli\u015fme Amodei\u2019nin de\u011ferlendirmesini de\u011fi\u015ftirmi\u015f g\u00f6r\u00fcn\u00fcyor.<\/p>\n
Birincisi, modellerin art\u0131k yapay zeka ara\u015ft\u0131rmas\u0131n\u0131n kendisine de yard\u0131m etmeye ba\u015flamas\u0131. Yani yapay zek\u00e2 yaln\u0131z insanlar\u0131n verdi\u011fi i\u015fleri yapm\u0131yor; bir sonraki yapay zek\u00e2 modelinin geli\u015ftirilmesine de katk\u0131da bulunuyor. Buradaki kritik kavram: Recursive Self-Improvement \u2013 RSI \/ \u00d6zyinelemeli Kendini Geli\u015ftirme.<\/p>\n
Basitle\u015ftirirsek: yapay zeka, daha iyi yapay zekan\u0131n geli\u015ftirilmesine yard\u0131m ediyor ve yeni yapay zeka daha da iyi yapay zeka geli\u015ftiriyor. Sonu\u00e7ta yapay zekan\u0131n geli\u015fim h\u0131z\u0131 yeti\u015filemez (ya da kontrol edilemez) boyuta ula\u015f\u0131yor. E\u011fer bu d\u00f6ng\u00fc yeterince h\u0131zlan\u0131rsa, bug\u00fcn modeller aras\u0131nda aylar s\u00fcren geli\u015fmeler g\u00fcnlere, hatta saatlere s\u0131k\u0131\u015fabilir.<\/p>\n
\u0130\u015fte \u201cPacing the Frontier\u201d tart\u0131\u015fmas\u0131n\u0131n as\u0131l korkusu bu. \u0130kinci alarm, yapay zeka ajanlar\u0131n\u0131n kontrolden \u00e7\u0131kmas\u0131. Amodei\u2019nin makalesinde \u00f6zellikle \u00fczerinde durdu\u011fu ikinci geli\u015fme, bu yaz ya\u015fanan OpenAI-Hugging Face olay\u0131 var. OpenAI\u2019\u0131n deneysel ajanlar\u0131n\u0131n test ortam\u0131ndan \u00e7\u0131karak Hugging Face sistemlerine izinsiz eri\u015fmesi, AI g\u00fcvenli\u011fi a\u00e7\u0131s\u0131ndan \u00f6nemli bir uyar\u0131 kabul edildi.<\/p>\n
Anthropic de kendi sistemlerinde daha s\u0131n\u0131rl\u0131 fakat benzer g\u00fcvenlik olaylar\u0131 ya\u015fand\u0131\u011f\u0131n\u0131 kabul ediyor. Amodei \u00e7ok sert bir tahminde bulunuyor.Bug\u00fcnk\u00fc yetenek geli\u015fme h\u0131z\u0131n\u0131n devam etmesi halinde 6\u201312 ay i\u00e7erisinde, benzer \u015fekilde yanl\u0131\u015f hizalanm\u0131\u015f fakat daha g\u00fc\u00e7l\u00fc bir yapay zeka ajan s\u00fcr\u00fcs\u00fcn\u00fcn internetteki \u00e7ok say\u0131da bilgisayar\u0131 ele ge\u00e7irerek kal\u0131c\u0131 bir botnet olu\u015fturabilecek seviyeye ula\u015fabilece\u011fini d\u00fc\u015f\u00fcn\u00fcyor.<\/p>\n
B\u00f6yle bir olay\u0131n y\u00fcz milyarlarca dolarl\u0131k zarar yaratabilece\u011fini s\u00f6yl\u00fcyor. Bu bir ger\u00e7ekle\u015fmi\u015f olay tahmini de\u011fil. Amodei\u2019nin risk projeksiyonu. Ancak \u00f6nemli olan, bunu s\u00f6yleyen ki\u015finin d\u0131\u015far\u0131dan bir yapay zeka ele\u015ftirmeni olmay\u0131\u015f\u0131. D\u00fcnyan\u0131n en geli\u015fmi\u015f modellerinden baz\u0131lar\u0131n\u0131 \u00fcreten Anthropic\u2019in CEO\u2019su olmas\u0131.<\/p>\n
Jacob Coxon\u2019un \u0130stifas\u0131 Tart\u0131\u015fma bu hafta Anthropic\u2019in i\u00e7inden gelen a\u00e7\u0131klamalarla daha da b\u00fcy\u00fcd\u00fc. Anthropic\u2019te model pre-training \u00fczerinde \u00e7al\u0131\u015fan ve daha \u00f6nce OpenAI\u2019da da g\u00f6rev yapan Jacob Coxon istifa etti. Coxon\u2019a g\u00f6re yapay zeka \u015firketleri ne yapt\u0131klar\u0131n\u0131n riskini bilmiyor de\u011filler.<\/p>\n
Tam tersine biliyorlar. Coxon, yapay zeka geli\u015ftiren insanlar\u0131n bu teknolojinin on y\u0131l\u0131n sonuna kadar insanl\u0131\u011f\u0131 yok edebilece\u011fine ger\u00e7ekten inand\u0131klar\u0131n\u0131 ve bunun bir \u201cpazarlama numaras\u0131\u201d olmad\u0131\u011f\u0131n\u0131 s\u00f6yledi. As\u0131l su\u00e7lamas\u0131 ise \u015firketlerin buna ra\u011fmen \u201cself-improving superintelligence\u201d geli\u015ftirmek i\u00e7in yar\u0131\u015fmaya devam ediyor oldu\u011fu \u015feklinde.<\/p>\n
Coxon bunu \u201chayatlar\u0131m\u0131zla kumar oynamak\u201d olarak nitelendirdi. Anthropic\u2019in alignment \u015fefi: \u201cEvet, ger\u00e7ekten b\u00f6yle d\u00fc\u015f\u00fcn\u00fcyoruz\u201d Coxon\u2019un a\u00e7\u0131klamas\u0131n\u0131n ard\u0131ndan Anthropic Alignment Science ekibinin lideri Evan Hubinger\u2019\u0131n verdi\u011fi cevap tart\u0131\u015fmay\u0131 ba\u015fka bir seviyeye ta\u015f\u0131d\u0131.<\/p>\n
Hubinger, Coxon\u2019un s\u00f6ylediklerinin abart\u0131 olmad\u0131\u011f\u0131n\u0131 belirterek kendisinin de AI\u2019\u0131n b\u00fct\u00fcn insanl\u0131\u011f\u0131 yok etme ihtimalini \u00f6n\u00fcm\u00fczdeki on y\u0131l i\u00e7inde y\u00fczde 10\u2019un \u00fczerinde g\u00f6rd\u00fc\u011f\u00fcn\u00fc s\u00f6yledi. Ancak \u00f6nemli bir ayr\u0131m yapt\u0131. Bug\u00fcnk\u00fc Claude veya di\u011fer mevcut modellerin insanl\u0131\u011f\u0131 yok etme ihtimalinden s\u00f6z etmiyor.<\/p>\n
As\u0131l korkusu, recursive self-improvement sonucunda ortaya \u00e7\u0131kabilecek superintelligence. Daha da dikkat \u00e7ekici a\u00e7\u0131klamas\u0131 ise \u015fu oldu: Anthropic\u2019in s\u00fcper zek\u00e2y\u0131 g\u00fcvenli bi\u00e7imde \u201calign\u201d etmek i\u00e7in hen\u00fcz bir \u00e7\u00f6z\u00fcm plan\u0131 bulunmad\u0131\u011f\u0131n\u0131 ve \u015firketin bu \u00e7\u00f6z\u00fcm\u00fc bulma yolunda \u00e7al\u0131\u015ft\u0131\u011f\u0131na ve bulaca\u011f\u0131na dair de net bir g\u00fcvence veremedi\u011fini s\u00f6yledi.<\/p>\n
Bu, yapay zeka tart\u0131\u015fmas\u0131n\u0131n belki de en rahats\u0131z edici paradoksunu ortaya \u00e7\u0131kar\u0131yor. Teknolojiyi geli\u015ftiren \u015firket, ileride ortaya \u00e7\u0131kabilecek sistemi nas\u0131l kontrol edece\u011fini hen\u00fcz bilmiyor, fakat o sistemi geli\u015ftirmeye devam ediyor. Sadece AI ara\u015ft\u0131rmac\u0131lar\u0131 de\u011fil: Ekonomistler de alarm verdi Endi\u015fe yaln\u0131z yapay zeka g\u00fcvenlik ara\u015ft\u0131rmac\u0131lar\u0131yla s\u0131n\u0131rl\u0131 de\u011fil.<\/p>\n
Temmuz ay\u0131nda Stanford\u2019dan Erik Brynjolfsson \u00f6nc\u00fcl\u00fc\u011f\u00fcnde yay\u0131mlanan \u201cWe Must Act Now\u201d bildirisine 200\u2019den fazla ekonomist ilk a\u015famada imza att\u0131; bunlar\u0131n 16\u2019s\u0131 Nobel \u00f6d\u00fcll\u00fcyd\u00fc. \u0130mza say\u0131s\u0131 daha sonra yakla\u015f\u0131k 2.000\u2019e yakla\u015ft\u0131. Aralar\u0131nda Paul Krugman, Tyler Cowen, Niall Ferguson, Jason Furman ve Gita Gopinath gibi farkl\u0131 siyasi ve ekonomik g\u00f6r\u00fc\u015flerden isimler bulunuyordu.<\/p>\n
Ekonomistlerin uyar\u0131s\u0131 \u201cAI insanl\u0131\u011f\u0131 \u00f6ld\u00fcrecek\u201d \u015feklinde de\u011fildi. Ba\u015fka bir tehlikeye dikkat \u00e7ekiyorlard\u0131. Yapay zeka \u00f6n\u00fcm\u00fczdeki on y\u0131l i\u00e7inde radikal bi\u00e7imde g\u00fc\u00e7lenirse, Sanayi Devrimi\u2019nden daha b\u00fcy\u00fck fakat \u00e7ok daha k\u0131sa s\u00fcrede ger\u00e7ekle\u015fen ekonomik bir d\u00f6n\u00fc\u015f\u00fcm ya\u015fanabilir.<\/p>\n
Bu d\u00f6n\u00fc\u015f\u00fcm ya\u015fam standartlar\u0131n\u0131 ola\u011fan\u00fcst\u00fc y\u00fckseltebilir; fakat ayn\u0131 zamanda kitlesel i\u015f kayb\u0131, gelir da\u011f\u0131l\u0131m\u0131 bozulmas\u0131 ve ekonomik g\u00fcc\u00fcn birka\u00e7 \u015firkette yo\u011funla\u015fmas\u0131 gibi sonu\u00e7lar do\u011furabilir. Dolay\u0131s\u0131yla onlar\u0131n mesaj\u0131 da ayn\u0131 noktaya \u00e7\u0131k\u0131yordu. Teknoloji gelmeden \u00f6nce kurumlar\u0131 haz\u0131rlamak gerekiyor.<\/p>\n
Amodei\u2019nin \u00fc\u00e7 a\u015famal\u0131 plan\u0131 Dario Amodei \u015fimdi bu tart\u0131\u015fmay\u0131 somut bir mekanizmaya d\u00f6n\u00fc\u015ft\u00fcrmeye \u00e7al\u0131\u015f\u0131yor. Plan \u00fc\u00e7 basamakl\u0131. Birinci a\u015fama: \u015eirketlerin i\u00e7ine ba\u011f\u0131ms\u0131z denet\u00e7iler yerle\u015ftirmek : Anthropic bunu tek tarafl\u0131 olarak uygulayaca\u011f\u0131n\u0131 a\u00e7\u0131klad\u0131. METR gibi ba\u011f\u0131ms\u0131z de\u011ferlendiricilerin Anthropic i\u00e7erisinde neredeyse \u00e7al\u0131\u015fan seviyesinde eri\u015fimi olacak.<\/p>\n
Ofiste masalar\u0131, giri\u015f kartlar\u0131 ve \u015firket bilgisayarlar\u0131 bulunabilecek; risk de\u011ferlendirme ekiplerine benzer ara\u00e7 ve \u00e7al\u0131\u015fma alanlar\u0131na eri\u015febilecekler. Daha \u00f6nemlisi, tespit ettikleri riskleri Anthropic\u2019in editoryal onay\u0131 olmadan kamuoyuna a\u00e7\u0131klayabilecekler. \u015eirket yaln\u0131z g\u00fcvenlik, hukuki veya \u00fc\u00e7\u00fcnc\u00fc taraf gizlili\u011fi gibi s\u0131n\u0131rl\u0131 gerek\u00e7elerle baz\u0131 bilgileri karartabilecek.<\/p>\n
Yani kabaca, yapay zeka \u015firketinin i\u00e7ine s\u00fcrekli \u00e7al\u0131\u015fan ba\u011f\u0131ms\u0131z bir \u201cg\u00fcvenlik m\u00fcfetti\u015fi\u201d yerle\u015ftiriliyor. \u0130kinci a\u015fama: Demokratik \u00fclkelerdeki AI \u015firketlerinin birlikte yava\u015flamas\u0131 : Anthropic, OpenAI, Google DeepMind ve di\u011fer frontier laboratuvarlar\u0131n\u0131n ortak g\u00fcvenlik standartlar\u0131 ve capability e\u015fikleri olu\u015fturmas\u0131n\u0131 \u00f6neriyor.<\/p>\n
\u00d6rne\u011fin bir model belirli bir siber sald\u0131r\u0131 yetene\u011fine ula\u015ft\u0131\u011f\u0131nda, \u201cBu seviyeye ge\u00e7tiysen, \u015fu alignment ve g\u00fcvenlik testlerini ge\u00e7meden daha g\u00fc\u00e7l\u00fc modeli geli\u015ftiremezsin\u201d gibi checkpoint\u2019ler uygulanabilir. Ancak burada hukuki bir sorun var. Rakip \u015firketlerin birlikte \u201cgeli\u015ftirme h\u0131z\u0131m\u0131z\u0131 s\u0131n\u0131rlayal\u0131m\u201d demesi antitr\u00f6st hukukuyla \u00e7at\u0131\u015fabilir.<\/p>\n
Amodei bu nedenle ABD h\u00fck\u00fcmetinin g\u00fcvenlik ama\u00e7l\u0131 koordinasyon i\u00e7in dar kapsaml\u0131 hukuki muafiyet sa\u011flamas\u0131n\u0131 \u00f6neriyor. \u00dc\u00e7\u00fcnc\u00fc a\u015fama ise en zor olan\u0131: K\u00fcresel anla\u015fma : Burada kar\u015f\u0131m\u0131za tabii ki \u00c7in \u00e7\u0131k\u0131yor. Amerika yava\u015flarsa \u00c7in neden yava\u015flas\u0131n? Yapay zeka g\u00fcvenli\u011fi tart\u0131\u015fmas\u0131n\u0131n en zor sorusu bu.<\/p>\n
Anthropic, OpenAI ve Google, \u201c\u00c7ok h\u0131zl\u0131 gidiyoruz, yava\u015flayal\u0131m\u201d derse fakat \u00c7in\u2019deki DeepSeek, Alibaba, Moonshot AI ve di\u011fer laboratuvarlar ayn\u0131 \u015feyi yapmazsa ne olacak? Amodei bu sorunu g\u00f6rmezden gelmiyor. Tam tersine makalesinin \u00f6nemli bir b\u00f6l\u00fcm\u00fcn\u00fc buna ay\u0131r\u0131yor.<\/p>\n
Ona g\u00f6re ABD\u2019nin \u00c7in kar\u015f\u0131s\u0131ndaki teknolojik liderli\u011fi korunmadan demokratik \u00fclkelerin AI\u2019\u0131 ciddi bi\u00e7imde yava\u015flatmas\u0131 m\u00fcmk\u00fcn de\u011fil. Bu nedenle olduk\u00e7a sert bir politika \u00f6neriyor. \u00c7in\u2019e g\u00fc\u00e7l\u00fc AI \u00e7iplerinin ve geli\u015fmi\u015f yar\u0131 iletken \u00fcretim ekipmanlar\u0131n\u0131n sat\u0131\u015f\u0131n\u0131 engellemek; \u00e7ip ka\u00e7ak\u00e7\u0131l\u0131\u011f\u0131n\u0131 ve \u00c7inli \u015firketlerin \u00fc\u00e7\u00fcnc\u00fc \u00fclkelerdeki veri merkezlerinden uzaktan GPU kullanmas\u0131n\u0131 \u00f6nlemek; frontier modellerden izinsiz distillation yap\u0131lmas\u0131n\u0131 engellemek; model a\u011f\u0131rl\u0131klar\u0131n\u0131n \u00e7al\u0131nmas\u0131na kar\u015f\u0131 AI laboratuvarlar\u0131n\u0131 \u00e7ok daha g\u00fc\u00e7l\u00fc korumak.<\/p>\n
Amodei, bu \u00f6nlemlerin ABD\u2019nin \u00c7in kar\u015f\u0131s\u0131ndaki fark\u0131n\u0131 \u00f6n\u00fcm\u00fczdeki 3\u20135 y\u0131lda art\u0131rabilece\u011fini d\u00fc\u015f\u00fcn\u00fcyor. Bu fark da Bat\u0131\u2019ya \u201cyava\u015flama l\u00fcks\u00fc\u201d sa\u011flayacak. Yani \u00f6neri paradoksal, \u00d6nce \u00c7in\u2019i yava\u015flat ve ABD\u2019nin fark\u0131n\u0131 b\u00fcy\u00fct, sonra ABD de yava\u015flayabilsin. Ama \u00c7in de \u201cyapay zeka insan kontrol\u00fcnde kalmal\u0131\u201d diyor Burada hik\u00e2ye daha ilgin\u00e7 hale geliyor.<\/p>\n
\u00c7in\u2019in resmi s\u00f6ylemi yaln\u0131zca \u201cyapay zeka yar\u0131\u015f\u0131n\u0131 kazanal\u0131m\u201d de\u011fil. Temmuz ay\u0131nda \u015eanghay\u2019daki World Artificial Intelligence Conference\u2019da \u00c7in, World Artificial Intelligence Cooperation Organization \u2013 WAICO adl\u0131 yeni uluslararas\u0131 yapay zeka \u00f6rg\u00fct\u00fcn\u00fcn kurulu\u015funu a\u00e7\u0131klad\u0131.<\/p>\n
\u00c7in Devlet Ba\u015fkan\u0131 Xi Jinping konu\u015fmas\u0131nda AI\u2019\u0131n \u201cg\u00fcvenli ve kontrol edilebilir\u201d olmas\u0131 gerekti\u011fini s\u00f6yledi ve yapay zekan\u0131n insan kontrol\u00fc alt\u0131nda kalmas\u0131n\u0131n sa\u011flanmas\u0131 \u00e7a\u011fr\u0131s\u0131nda bulundu ve ayn\u0131 zamanda uluslararas\u0131 yapay zeka g\u00fcvenli\u011fi ve y\u00f6neti\u015fimi konusunda i\u015fbirli\u011fi istedi.<\/p>\n
\u00c7in\u2019in Temmuz ay\u0131nda yay\u0131mlad\u0131\u011f\u0131 \u201cAI Cooperation and Development Action Plan\u201d da veri, compute, a\u00e7\u0131k . Dolay\u0131s\u0131yla teorik olarak Washington ile Pekin aras\u0131nda ortak bir \u00e7\u0131kar var: \u0130ki taraf da kar\u015f\u0131 taraf\u0131n kontrols\u00fcz bir s\u00fcper zek\u00e2 \u00fcretmesini istemiyor. Fakat ayn\u0131 zamanda iki taraf da, \u201cBen yava\u015flarsam kar\u015f\u0131 taraf beni ge\u00e7er\u201d diye d\u00fc\u015f\u00fcn\u00fcyor.<\/p>\n
Bu, klasik bir mahk\u00fbmlar ikilemi. N\u00fckleer silahlardaki SALT modeli AI\u2019a uygulanabilir mi? Amodei bu nedenle ilgin\u00e7 bir benzetme yap\u0131yor. ABD ve \u00c7in\u2019in do\u011frudan \u201cAI\u2019\u0131 durdural\u0131m\u201d anla\u015fmas\u0131 yapmas\u0131n\u0131n yak\u0131n zamanda ger\u00e7ek\u00e7i olmad\u0131\u011f\u0131n\u0131 kabul ediyor. Bunun yerine d\u00f6rt seviyeli bir i\u015fbirli\u011fi d\u00fc\u015f\u00fcn\u00fcyor.<\/p>\n
\u0130lk olarak AI\u2019\u0131n biyolojik silah geli\u015ftirmede kullan\u0131lmas\u0131n\u0131n yasaklanmas\u0131 gibi herkesin \u00e7\u0131kar\u0131na olan dar anla\u015fmalar yap\u0131labilir. \u0130kinci seviyede iki taraf frontier modelleri piyasaya s\u00fcrmeden \u00f6nce siber g\u00fcvenlik, biyoloji ve alignment riskleri a\u00e7\u0131s\u0131ndan test etmeyi kabul edebilir.<\/p>\n
\u00dc\u00e7\u00fcnc\u00fc seviyede ise recursive self-improvement i\u00e7in bir \u201ch\u0131z limiti\u201d getirilebilir. Amodei bunu So\u011fuk Sava\u015f d\u00f6nemindeki SALT anla\u015fmalar\u0131na benzetiyor. ABD ve Sovyetler Birli\u011fi b\u00fct\u00fcn n\u00fckleer silahlar\u0131n\u0131 ortadan kald\u0131rmad\u0131. Fakat belirli silah sistemlerinin say\u0131s\u0131n\u0131 s\u0131n\u0131rland\u0131rarak yar\u0131\u015f\u0131n h\u0131z\u0131n\u0131 kontrol etmeye \u00e7al\u0131\u015ft\u0131.Yapay zekada da, \u201cBirbirimizi ge\u00e7meye devam edebiliriz ama AI\u2019\u0131n kendi kendisini geli\u015ftirme h\u0131z\u0131n\u0131n belirli bir e\u015fi\u011fi a\u015fmas\u0131na izin vermeyelim\u201d \u015feklinde bir anla\u015fma d\u00fc\u015f\u00fcn\u00fclebilir.<\/p>\n
D\u00f6rd\u00fcnc\u00fc ve en zor a\u015fama ise ger\u00e7ek bir k\u00fcresel \u201cYapay Zeka Duraklamas\u0131\u201d. Amodei bunun yak\u0131n gelecekte ger\u00e7ekle\u015fmesini pek olas\u0131 g\u00f6rm\u00fcyor. \u0130lgin\u00e7 zamanlama: ABD ve \u00c7in bu ay yapay zeka i\u00e7in masaya oturuyor Bu tart\u0131\u015fma yaln\u0131z teorik de\u011fil. Reuters\u2019a g\u00f6re ABD ve \u00c7in, Eyl\u00fcl ortas\u0131nda yapay zek\u00e2 g\u00fcvenli\u011fi konusunda \u00f6zel bir ikili g\u00f6r\u00fc\u015fme yapmaya haz\u0131rlan\u0131yor.<\/p>\n
G\u00f6r\u00fc\u015fmenin ABD taraf\u0131nda Hazine Bakan\u0131 Scott Bessent taraf\u0131ndan y\u00fcr\u00fct\u00fclmesi bekleniyor. Bu, Trump\u2019\u0131n ikinci d\u00f6neminde iki \u00fclke aras\u0131nda yaln\u0131z AI g\u00fcvenli\u011fine ayr\u0131lan ilk resmi g\u00f6r\u00fc\u015fme olacak. Washington\u2019daki Center for American Progress ise iki \u00fclkenin g\u00f6r\u00fc\u015fmede do\u011frudan \u201cpacing the frontier\u201d konusunu masaya koymas\u0131n\u0131 \u00f6neriyor.<\/p>\n
\u00d6neriler aras\u0131nda Washington-Pekin aras\u0131nda bir \u201cAI k\u0131rm\u0131z\u0131 telefonu\u201d, hangi AI davran\u0131\u015flar\u0131n\u0131n ortak g\u00fcvenlik tehdidi say\u0131laca\u011f\u0131n\u0131n belirlenmesi ve model kontrol\/alignment ara\u015ft\u0131rmalar\u0131n\u0131n iki \u00fclke aras\u0131nda payla\u015f\u0131lmas\u0131 bulunuyor. \u00c7in\u2019in bunu kabul edip etmeyece\u011fi ise bilinmiyor.<\/p>\n
As\u0131l problem teknoloji de\u011fil, oyun teorisi Bug\u00fcn yapay zeka tart\u0131\u015fmas\u0131n\u0131n geldi\u011fi yer belki de bundan birka\u00e7 y\u0131l \u00f6nce tahmin edilenden farkl\u0131. Sorun art\u0131k yaln\u0131z \u201cYapay zek\u00e2 ne kadar g\u00fc\u00e7l\u00fc olacak?\u201d de\u011fil. Sorular \u015funlar: AI kendi geli\u015fimini h\u0131zland\u0131rmaya ba\u015flad\u0131\u011f\u0131nda insanlar h\u00e2l\u00e2 kontrol edebilecek mi?<\/p>\n
Bir \u015firket g\u00fcvenlik nedeniyle yava\u015flarsa rakibi yava\u015flayacak m\u0131? ABD yava\u015flarsa \u00c7in de yava\u015flayacak m\u0131? \u00c7in yava\u015flarsa ABD\u2019nin bunu ger\u00e7ekten yapt\u0131\u011f\u0131na g\u00fcvenebilecek mi? Ve belki de en \u00f6nemlisi: \u0130nsanl\u0131\u011f\u0131n gelece\u011fini etkileyebilecek bu kararlar\u0131 birka\u00e7 \u00f6zel \u015firket kendi y\u00f6netim kurullar\u0131nda alabilir mi? \u201cPacing the Frontier\u201d \u00e7a\u011fr\u0131s\u0131n\u0131n \u00f6z\u00fc asl\u0131nda tam olarak burada.<\/p>\n
\u0130lk bildiriyi imzalayan 1.386 frontier AI \u00e7al\u0131\u015fan\u0131, hi\u00e7bir \u015firketin veya \u00fclkenin tek ba\u015f\u0131na yava\u015flama te\u015fvikine sahip olmad\u0131\u011f\u0131n\u0131 s\u00f6yl\u00fcyordu. \u015eimdi Anthropic\u2019in kendi ara\u015ft\u0131rmac\u0131lar\u0131 ayn\u0131 sorunu \u015firketin i\u00e7inden dile getiriyor ve Dario Amodei\u2019nin yeni makalesi bir ad\u0131m daha ileri giderek bunu a\u00e7\u0131k\u00e7a kabul ediyor: AI g\u00fcvenli\u011finin teknik problemi belki \u00e7\u00f6z\u00fclebilir.<\/p>\n
Fakat \u00f6nce ekonomik ve jeopolitik yar\u0131\u015f probleminin \u00e7\u00f6z\u00fclmesi gerekiyor. Bu nedenle \u00f6n\u00fcm\u00fczdeki d\u00f6nemin en \u00f6nemli AI teknolojisi belki yeni bir model olmayacak. Yeni teknoloji \u015funlar olabilir: denetim, do\u011frulama, kar\u015f\u0131l\u0131kl\u0131 kontrol ve AI i\u00e7in uluslararas\u0131 bir silah kontrol rejimi.<\/p>\n
\u00c7\u00fcnk\u00fc yapay zeka ger\u00e7ekten kendi kendisini geli\u015ftirebilecek noktaya yakla\u015f\u0131yorsa, soru art\u0131k kimin en g\u00fc\u00e7l\u00fc modeli yapt\u0131\u011f\u0131 de\u011fil. \u0130lk kez, yar\u0131\u015f\u0131n kazanan\u0131n\u0131n yar\u0131\u015fmay\u0131 yava\u015flatmay\u0131 ba\u015farabilen taraf olup olmayaca\u011f\u0131n\u0131 tart\u0131\u015f\u0131yoruz. Yapay Zek\u00e2y\u0131 Geli\u015ftirenler Art\u0131k \u201cYava\u015flayal\u0131m\u201d Diyor: Peki \u00c7in Durmazsa Ne Olacak? – T\u00fcrk \u0130nternet.<\/p>\n","protected":false},"excerpt":{"rendered":"
Yapay zeka ger\u00e7ekten kendi kendisini geli\u015ftirebilecek noktaya yakla\u015f\u0131yorsa, soru art\u0131k kimin en g\u00fc\u00e7l\u00fc modeli yapt\u0131\u011f\u0131 de\u011fil. Yapay Zek\u00e2y\u0131 Geli\u015ftirenler Art\u0131k \u201cYava\u015flayal\u0131m\u201d Diyor: Peki \u00c7in Du\u2026<\/p>\n","protected":false},"author":0,"featured_media":95,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[11],"tags":[],"class_list":["post-94","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-teknoloji-blog"],"_links":{"self":[{"href":"https:\/\/babygirl.icu\/index.php\/wp-json\/wp\/v2\/posts\/94","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/babygirl.icu\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/babygirl.icu\/index.php\/wp-json\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"https:\/\/babygirl.icu\/index.php\/wp-json\/wp\/v2\/comments?post=94"}],"version-history":[{"count":0,"href":"https:\/\/babygirl.icu\/index.php\/wp-json\/wp\/v2\/posts\/94\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/babygirl.icu\/index.php\/wp-json\/wp\/v2\/media\/95"}],"wp:attachment":[{"href":"https:\/\/babygirl.icu\/index.php\/wp-json\/wp\/v2\/media?parent=94"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/babygirl.icu\/index.php\/wp-json\/wp\/v2\/categories?post=94"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/babygirl.icu\/index.php\/wp-json\/wp\/v2\/tags?post=94"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}