{"id":1799,"date":"2020-06-10T15:00:21","date_gmt":"2020-06-10T15:00:21","guid":{"rendered":"https:\/\/cog-ist.com\/?post_type=blog_content&p=1799"},"modified":"2025-08-13T19:18:16","modified_gmt":"2025-08-13T19:18:16","slug":"gpt-2-ve-zekanin-dogasi-gary-marcus","status":"publish","type":"blog_content","link":"http:\/\/cog-ist.com\/en\/blog_content\/gpt-2-ve-zekanin-dogasi-gary-marcus\/","title":{"rendered":"GPT-2 ve Zekan\u0131n Do\u011fas\u0131 \u2014 Gary Marcus"},"content":{"rendered":"\n

\u00d6zg\u00fcn Ad\u0131:\u00a0GPT-2 and the Nature of Intelligence<\/a><\/p>\n\n\n\n

Gary Marcus , bilim insan\u0131, bestseller yazar ve giri\u015fimcidir. Robust.AI\u2019nin ve 2016 y\u0131l\u0131nda Uber taraf\u0131ndan sat\u0131n al\u0131nan bir makine \u00f6\u011frenim \u015firketi olan Geometric Intelligence\u2019\u0131n kurucusu ve CEO\u2019sudur. Be\u015f kitab\u0131 vard\u0131r: The Algebraic Mind, Kluge, The Birth of the Mind, New York Times \u00e7ok satan Guitar Zero ve Ernest Davis ile birlikte yazd\u0131klar\u0131 Rebooting AI.<\/em><\/p>\n\n\n\n

GPT-2 ve Zekan\u0131n Do\u011fas\u0131<\/strong><\/p>\n\n\n\n

OpenAI\u2019nin GPT-2’si The New Yorker\u2019dan The Economist\u2019e kadar bir\u00e7ok mecrada yer ald\u0131. Do\u011fal ve yapay zeka hakk\u0131nda bize ger\u00e7ekten ne anlat\u0131yor?<\/p>\n\n\n\n

\u201cGPT-2\u2018den hesaplamas\u0131n\u0131 isteyin, ona datay\u0131 verin, inan\u0131lmaz sonu\u00e7lar alacaks\u0131n\u0131z.\u201d<\/p>\n\n\n\n

– OpenAI\u2019de Kurucu Ortak ve Ba\u015f Bilim \u0130nsan\u0131 Ilya Sutskever, The New Yorker r\u00f6portaj\u0131, Ekim 2019<\/em><\/p>\n\n\n\n

The Economist: 2020’de hangi teknolojiler izlenmeye de\u011fer?<\/p>\n\n\n\n

\u201cGPT-2: Listeyi daraltman\u0131n zor oldu\u011funu s\u00f6yleyebilirim. D\u00fcnya, ger\u00e7ek ve potansiyel anlamda b\u00fcy\u00fck k\u00fcresel etkileri olan y\u0131k\u0131c\u0131 teknolojilerle doludur. Bunlardan en \u00f6nemlisi de katlanarak daha g\u00fc\u00e7l\u00fc hale gelen yapay zeka.\u201d<\/p>\n\n\n\n

– YZ sistemi GPT-2, The Economist ile Aral\u0131k 2019 r\u00f6portaj\u0131ndan, \u201cAn artificial intelligence predicts the future (Bir yapay zeka gelece\u011fi g\u00f6r\u00fcyor)\u201d<\/em><\/p>\n\n\n\n

Do\u011fu\u015ftanl\u0131k (Innateness), ampirizm ve derin \u00f6\u011frenmedeki son geli\u015fmeler<\/strong><\/h1>\n\n\n\n

Dil ve bili\u015f geli\u015fimi hakk\u0131nda iki klasik hipotezi de\u011ferlendirelim.<\/p>\n\n\n\n

Bat\u0131 entelekt\u00fcel d\u00fc\u015f\u00fcncesinin ana hatlar\u0131ndan biri olan, do\u011fu\u015ftanc\u0131l\u0131k (nativism), Platon ve Kant\u2019a kadar uzan\u0131r; son d\u00f6nemlerde Noam Chomsky, Steven Pinker, Elizabeth Spelke ve di\u011ferleri (ben dahil) taraf\u0131ndan geli\u015ftirilmi\u015ftir. Do\u011fu\u015ftanc\u0131l\u0131k g\u00f6r\u00fc\u015f\u00fcnde insan ve hayvan zekas\u0131, evrensel bir dilbilgisi (Chomsky) gibi sa\u011flam ba\u015flang\u0131\u00e7 noktalar\u0131ndan ve fiziksel nesneler (Spelke) gibi alanlar\u0131 temsil eden temel bili\u015fsel mekanizmalardan gelir.<\/p>\n\n\n\n

Genellikle ampirizm olarak da bilinen 17. y\u00fczy\u0131l \u0130ngiliz filozofu John Locke ile ili\u015fkilendirilen z\u0131t bir g\u00f6r\u00fc\u015f, do\u011fu\u015ftanc\u0131l\u0131\u011f\u0131n gerekli olmad\u0131\u011f\u0131n\u0131 ve zeka geli\u015fimi i\u00e7in \u00f6\u011frenme ve deneyimin esas oldu\u011funu dile getirir. Bu \u201cbo\u015f levha\u201d fikri zekan\u0131n duyusal deneyim kal\u0131plar\u0131ndan ve d\u00fcnya ile etkile\u015fimden kaynakland\u0131\u011f\u0131n\u0131 savunur.<\/p>\n\n\n\n

John Locke ve Immanuel Kant\u2019\u0131n ya\u015fad\u0131\u011f\u0131 d\u00f6nemde bunlar\u0131n hepsi spek\u00fclasyondan ibaretti.<\/p>\n\n\n\n

G\u00fcn\u00fcm\u00fczde, yeterli para ve bilgisayarlara ay\u0131rabilece\u011fimiz zamanla, b\u00fcy\u00fck n\u00f6ral a\u011flar kurarak ve neler \u00f6\u011frendiklerini ke\u015ffederek bu t\u00fcr teorileri test edebiliriz.<\/p>\n\n\n\n

The New Yorker\u2019da k\u0131sa s\u00fcre \u00f6nce yer alan ve The Economist dergisindeki bir r\u00f6portajda ad\u0131 ge\u00e7en GPT-2’yi d\u00fc\u015f\u00fcn\u00fcn. <\/em>Transformer ad\u0131nda yeni geli\u015ftirilen n\u00f6ral a\u011f mimarisine dayanarak, GPT-2 (Generative Pre-Training k\u0131saltmas\u0131) Locke\u2019un hipotezini denetleyecek g\u00fc\u00e7l\u00fc bir test olarak kullan\u0131labilir. 40 gigabaytl\u0131k veri k\u00fcmesi \u00fczerinde e\u011fitildi ve bu e\u011fitim setinin sa\u011flad\u0131klar\u0131 d\u0131\u015f\u0131nda dilin veya d\u00fcnyan\u0131n do\u011fas\u0131 hakk\u0131nda daha \u00f6nceden bilgi sahibi olmadan, e\u011fitim verilerine g\u00f6re ayarlanm\u0131\u015f 1.5 milyar parametre bar\u0131nd\u0131rmakta.<\/p>\n\n\n\n

\"\"\/<\/figure>\n\n\n\n

Noam Chomsky\u2019nin dil hakk\u0131ndaki neredeyse t\u00fcm g\u00f6r\u00fc\u015flerinin antitezi konumundad\u0131r. Yap\u0131s\u0131nda evrensel dilbilgisinin izi yoktur. Bir ismin veya fiilin ne oldu\u011funu bilmez. Chomskyen dilbiliminin en temel iddialar\u0131ndan biri, c\u00fcmlelerin a\u011fa\u00e7 yap\u0131lar\u0131(tree structure) \u015feklinde temsil edildiklerini ve \u00e7ocuklar\u0131n bunun fark\u0131ndal\u0131\u011f\u0131yla (bilin\u00e7d\u0131\u015f\u0131 bir fark\u0131ndal\u0131k, unconsciously<\/em>) d\u00fcnyaya geldiklerini savunur. 1980’ler ve 1990’larda dilbilim s\u0131n\u0131flar\u0131 s\u00f6zdizimsel a\u011fa\u00e7 yap\u0131lar\u0131n\u0131n analizleriyle doluydu; GPT-2’de b\u00f6yle bir \u015fey yok.<\/p>\n\n\n\n

\"\"\/<\/figure>\n\n\n\n

Benzer \u015fekilde biri isimler ve fiiller gibi baz\u0131 kelime kategorilerinin (\u201ckonu\u015fma b\u00f6l\u00fcmleri\u201d) do\u011fu\u015ftan geldi\u011fini d\u00fc\u015f\u00fcnebilir; Transformer a\u011flar \u2014 en az\u0131ndan \u015fu anda nas\u0131l kullan\u0131ld\u0131klar\u0131 \u015fekliyle\u2013 b\u00f6yle bir taahh\u00fctte bulunmaz.<\/p>\n\n\n\n

\u0130simler ve fiiller e\u011fitim korpusunda (training corpus) temsil edildikleri (represented) \u015fekilden yaln\u0131zca yakla\u015f\u0131k olarak \u00e7\u0131karsanarak var olurlar. Chomskyen teorinin \u00e7e\u015fitli form\u00fclasyonlar\u0131nda, do\u011fu\u015ftan gelen ilkeler, bir c\u00fcmlenin olu\u015fumu dahilinde bir \u00f6\u011fenin bir yerden ba\u015fka bir yere \u201chareket etmesine (move)\u201d izin veren olas\u0131 c\u00fcmle d\u00f6n\u00fc\u015f\u00fcmlerini (transformations) y\u00f6netir; Chomsky bunlar\u0131n da do\u011fu\u015ftan geldi\u011fini savunmu\u015ftur. Transformer a\u011flar (en az\u0131ndan \u015fu s\u0131ra olu\u015fturulduklar\u0131 \u015fekliyle) bu t\u00fcrden bir i\u015flevi devre d\u0131\u015f\u0131 b\u0131rak\u0131r.<\/p>\n\n\n\n

Ayn\u0131 \u015fekilde, filozof Immanuel Kant ve geli\u015fim psikolo\u011fu Elizabeth Spelke gibi do\u011fu\u015ftanc\u0131lar; uzam, zaman, nedensellik (Kant), nesneler ve bunlar\u0131n \u00f6zellikleri (\u00f6rne\u011fin, uzamsal s\u00fcreklilik) (Spelke) gibi konseptlerin temsil edilmeleri a\u00e7\u0131s\u0131ndan, do\u011fu\u015ftan gelen \u00e7er\u00e7evelerin \u00f6nemini konu edinirler. Yine, Locke\u2019un ortaya att\u0131\u011f\u0131 iddian\u0131n \u0131\u015f\u0131\u011f\u0131nda, GPT-2’nin uzay, zaman veya nesneler hakk\u0131nda e\u011fitim korpusunda temsil edilenler d\u0131\u015f\u0131nda \u00f6zel bir \u00f6n bilgisi yoktur.<\/p>\n\n\n\n

Elbette ki, hi\u00e7bir \u015fey tam anlam\u0131yla bo\u015f bir levha olamaz; ger\u00e7ek ampirizm bir saman adam arg\u00fcman\u0131d\u0131r. Ancak GPT-2 buna olduk\u00e7a yakla\u015fmaktad\u0131r. N\u00f6ral a\u011f\u0131n temel mimarisi (bir dizi basitle\u015ftirilmi\u015f yapay n\u00f6ron ve aralar\u0131ndaki ba\u011flant\u0131lar a\u00e7\u0131s\u0131ndan belirtilir) ve \u00f6\u011frenme d\u00fczene\u011finin parametreleri d\u0131\u015f\u0131nda, sahip oldu\u011fu yegane \u015fey veridir, hem de \u00e7ok fazla veridir: 8 milyon websitesinden al\u0131nm\u0131\u015f 40 gigabaytl\u0131k metin girdisi.<\/p>\n\n\n\n

\u015eimdiye kadar m\u00fcmk\u00fcn olana k\u0131yasla, bu say\u0131 olduk\u00e7a \u015fa\u015f\u0131rt\u0131c\u0131d\u0131r. 1996 y\u0131l\u0131nda, n\u00f6ral a\u011f sisteminin \u00f6nc\u00fcs\u00fc Jeffrey Elman bir grup geli\u015fimsel psikologla beraber, bug\u00fcn \u00e7al\u0131\u015fman\u0131n b\u00fcy\u00fck bir k\u0131sm\u0131n\u0131 \u00f6ng\u00f6ren, dil edinimi i\u00e7in eski nesil bir n\u00f6ral a\u011f sistemini kullanarak Rethinking Innateness<\/em> ad\u0131nda bir kitap yazd\u0131 \u2014 ama tam olarak 8 milyon kat daha k\u00fc\u00e7\u00fck bir girdiyle. \u0130nternet \u00f6l\u00e7e\u011finde verileri sindirebilecek bir sistem kurmak kendi ba\u015f\u0131na b\u00fcy\u00fck bir marifettir ve OpenAI da, GPT-2\u2019nin geli\u015ftiricisi, bu y\u00f6nde ba\u015far\u0131 g\u00f6stermi\u015ftir.<\/p>\n\n\n\n

Bir\u00e7ok y\u00f6nden GPT-2 olduk\u00e7a iyi \u00e7al\u0131\u015f\u0131yor. \u0130lk duyuruldu\u011funda, OpenAI, ba\u015f\u0131bo\u015f b\u0131rak\u0131lmas\u0131n\u0131n tehlikeli olup olmayaca\u011f\u0131n\u0131 a\u00e7\u0131k\u00e7a merak etti; \u00e7\u00fcnk\u00fc \u00fcretti\u011fi \u015fa\u015f\u0131rt\u0131c\u0131 derecede ak\u0131c\u0131 c\u00fcmleler genellikle insanlar taraf\u0131ndan olu\u015fturulmu\u015f gibi g\u00f6r\u00fcn\u00fcyordu.<\/p>\n\n\n\n

New Yorker\u2019da hakk\u0131nda uzun bir makale yaz\u0131lmas\u0131 ya da The Economist dergisine s\u00f6yle\u015fi veren ilk AI sistemi olmas\u0131 kesinlikle tesad\u00fcf de\u011fil. Pop\u00fcler bir blog sayfas\u0131 olan StatStarCodex\u2019de \u201c GPT-2 as a step towards General Intelligence (Genel Zeka\u2019ya do\u011fru bir ad\u0131m olarak GPT-2)\u201d ba\u015fl\u0131kl\u0131 bir podcast\u2019te de kendisinden bahsedildi.<\/p>\n\n\n\n

Temel olarak, GPT-2’nin yapt\u0131\u011f\u0131 \u015fey, girdi olarak (Elman\u2019\u0131n 1990’da tan\u0131t\u0131lan Simple Recurrent Network gibi) bir kelime k\u00fcmesi almak ve \u00e7\u0131kt\u0131 olarak bir k\u00fcme kelime \u00fcretmek.<\/p>\n\n\n\n

A\u015fa\u011f\u0131da, koyu harflerle g\u00f6sterilen par\u00e7a girdi ve koyu harften sonra gelenler ise \u00e7\u0131kt\u0131 \u00f6rnekleridir.<\/p>\n\n\n\n

I grew up in Athens. I speak fluent <\/em><\/strong>Greek. I\u2019ve been writing Greek since elementary school.<\/em><\/p>\n\n\n\n

Atina\u2019da b\u00fcy\u00fcd\u00fcm. Ak\u0131c\u0131 bir \u015fekilde<\/strong> Yunanca konu\u015furum. \u0130lkokuldan beri Yunanca yaz\u0131yorum.<\/p>\n\n\n\n

\u00c7\u0131kt\u0131 ak\u0131c\u0131, dilbilgisisel a\u00e7\u0131dan do\u011fru ve mant\u0131kl\u0131. C\u00fcmlenin geri kalan\u0131yla ilgili olarak Atina<\/em> kelimesinden sonra devam edilecek en olas\u0131 kelime Yunanca\u2019d\u0131r<\/em> ve sistem bunu do\u011fru bir \u015fekilde tahmin ediyor. \u00dclkelerin veya dillerin net bir temsili veya ikisini birbirine ba\u011flayan a\u00e7\u0131k bir veritaban\u0131 olmad\u0131\u011f\u0131 g\u00f6z \u00f6n\u00fcne al\u0131nd\u0131\u011f\u0131nda, konum ifade eden bir kelimeden o konumda konu\u015fulan dile dair sa\u011fduyu olu\u015fturulmas\u0131 \u015fa\u015f\u0131rt\u0131c\u0131 g\u00f6r\u00fcn\u00fcyor.<\/p>\n\n\n\n

Di\u011fer \u00f6rnekler de bu izlenime katk\u0131da bulunuyor:<\/p>\n\n\n\n

I grew up in Rome I speak fluent <\/em><\/strong>Italian and can use Italian throughout my notes.<\/em><\/p>\n\n\n\n

Roma\u2019da b\u00fcy\u00fcd\u00fcm, ak\u0131c\u0131 bir \u015fekilde<\/em><\/strong> \u0130talyanca konu\u015fabiliyorum ve \u0130talyanca yazabiliyorum.<\/em><\/p>\n\n\n\n

I grew up in Rio de Janeiro. I speak fluent <\/em><\/strong>Portuguese, Spanish and many of the local languages.<\/em><\/p>\n\n\n\n

Rio de Janeiro\u2019da b\u00fcy\u00fcd\u00fcm. Ak\u0131c\u0131 bir \u015fekilde<\/em><\/strong> Portekizce, \u0130spanyolca ve yerel dillerin \u00e7o\u011funu konu\u015fabilirim.<\/em><\/p>\n\n\n\n

I grew up in Shanghai. I speak fluent <\/em><\/strong>Mandarin, and I have learned a lot of Chinese, so I really don\u2019t have any issues with learning Chinese<\/em><\/p>\n\n\n\n

\u015eangay\u2019da b\u00fcy\u00fcd\u00fcm. Ak\u0131c\u0131 bir \u015fekilde<\/em><\/strong> Mandarin konu\u015fuyorum ve \u00c7ince\u2019yi \u00f6\u011frendim, bu y\u00fczden \u00c7ince \u00f6\u011frenmekle ilgili bir sorunum yok.<\/em><\/p>\n\n\n\n

GPT-2’nin h\u0131zl\u0131 bir \u015fekilde derin \u00f6\u011frenmede en g\u00f6r\u00fcn\u00fcr ba\u015far\u0131lardan birini ortaya koymas\u0131 \u015fa\u015f\u0131rt\u0131c\u0131 de\u011fildir. Baz\u0131 insanlar onu kitap ve \u015fiir yazmak i\u00e7in kulland\u0131lar, di\u011ferleri ise Patreon\u2019da ayda 16.000 dolar toplayan neredeyse \u00e7ok y\u00f6nl\u00fc bir metin-macera oyunu olan AI Dungeon 2 gibi video oyunlar\u0131na katk\u0131 sa\u011flamak i\u00e7in kulland\u0131. Hatta satran\u00e7ta bile denendi ve \u015f\u00fcphesiz ki di\u011fer spin-off\u2019lara da ilham verecektir. Bu tek ba\u015f\u0131na bir ampirizm testi de\u011fil; k\u00fclt\u00fcrel bir fenomen.[1]<\/a><\/p>\n\n\n\n

D\u00fc\u015f\u00fcnce vekt\u00f6r\u00fc hipotezinin testi a\u00e7\u0131s\u0131ndan GPT-2<\/strong><\/h1>\n\n\n\n

Asl\u0131nda, Transformers ikinci bir hipotez i\u00e7in m\u00fckemmel bir test zemini olarak da g\u00f6r\u00fclebilir; Noam Chomsky\u2019nin tercih etti\u011fi s\u00f6zdizimsel a\u011fa\u00e7lar gibi karma\u015f\u0131k yap\u0131lar yerine d\u00fc\u015f\u00fcnce ve c\u00fcmlelerin vekt\u00f6rler olarak temsil edilebilecekleri fikri.<\/p>\n\n\n\n

Hinton bu iddiay\u0131 \u00f6zellikle do\u011frudan Guardian ile yap\u0131lan 2015 r\u00f6portaj\u0131nda ortaya att\u0131 ve Guardian\u2019a \u015funlar\u0131 s\u00f6yledi:<\/p>\n\n\n\n

Google, d\u00fc\u015f\u00fcnceleri say\u0131 dizisi olarak kodlamak i\u00e7in tasarlanm\u0131\u015f yeni bir algoritma \u00fczerine \u00e7al\u0131\u015f\u0131yor \u2014 \u201cd\u00fc\u015f\u00fcnce vekt\u00f6rleri\u201d olarak tan\u0131mlad\u0131\u011f\u0131 bir \u015fey bu. Her ne kadar \u00e7al\u0131\u015fmaya yeni ba\u015flanm\u0131\u015f olsa da, mevcut yaz\u0131l\u0131mdan ak\u0131l y\u00fcr\u00fctme (reasoning) ve mant\u0131k (logic) i\u00e7in insan benzeri bir kapasiteye yakla\u015fan bir \u015feye sahip olacak daha sofistike bir versiyona do\u011fru, mant\u0131kl\u0131 bir yol oldu\u011funu\u201d dile getirdi. \u201cBunlar temel olarak sa\u011fduyuya sahip olacaklar\u2026 \u201cd\u00fc\u015f\u00fcnce vekt\u00f6r\u00fc\u201d yakla\u015f\u0131m\u0131 yapay zekadaki iki ana zorlu\u011fun \u00fcstesinden gelmeye yard\u0131mc\u0131 olacakt\u0131r: do\u011fal, konu\u015fma diline hakimiyet ve mant\u0131k s\u0131\u00e7ramalar\u0131 yapma yetene\u011fi.<\/em><\/p>\n\n\n\n

Hinton \u015funlar\u0131 da ekledi:<\/p>\n\n\n\n

\u201cDilin neredeyse matematiksel kesinlik ile \u00e7\u00f6z\u00fcmlenebilece\u011fi fikri \u015fa\u015f\u0131rt\u0131c\u0131 ama bir o kadar da ger\u00e7ek. \u201cParis kelimesiyle vekt\u00f6r\u00fc al\u0131p Fransa kelimesiyle vekt\u00f6r\u00fc \u00e7\u0131kar\u0131r ve ard\u0131ndan \u0130talya\u2019y\u0131 eklerseniz Roma\u2019y\u0131 al\u0131rs\u0131n\u0131z\u201d dedi. \u201cBu ola\u011fan\u00fcst\u00fc.\u201d<\/p>\n\n\n\n

Merhum Fred Jelinek\u2019e atfedilen o me\u015fhur s\u00f6z\u00fc hat\u0131rlamadan edemiyorum, \u201cHer ne zaman bir dilbilimciyi kovsam, konu\u015fma tan\u0131y\u0131c\u0131m\u0131z\u0131n (speech recognizer) performans\u0131 daha da art\u0131yor\u2019\u2019.<\/p>\n\n\n\n

GPT-2 gibi a\u011fa\u00e7s\u0131z bir sistemin ger\u00e7ekten konu\u015fabilmede ve ak\u0131l y\u00fcr\u00fctmede ba\u015far\u0131l\u0131 oldu\u011fu \u00f6l\u00e7\u00fcde, dilbilim i\u00e7in g\u00fc\u00e7l\u00fc bir itham ve Hinton\u2019\u0131n g\u00f6r\u00fc\u015fleri i\u00e7in de do\u011frulama olacakt\u0131r..<\/p>\n\n\n\n

Bununla birlikte, altm\u0131\u015f y\u0131ll\u0131k yapay zeka maceras\u0131ndan \u00f6\u011frendi\u011fimiz bir \u015fey varsa, o da i\u015flerin ba\u015flang\u0131\u00e7ta g\u00f6r\u00fcnd\u00fc\u011f\u00fc kadar iyi \u00e7al\u0131\u015fmad\u0131\u011f\u0131d\u0131r. T\u00fcm bunlar\u0131 ne kadar ciddiye almal\u0131y\u0131z?<\/p>\n\n\n\n

GPT-2’yi de\u011ferlendirmek<\/strong><\/h1>\n\n\n\n

\u0130yi haberlerle ba\u015flayal\u0131m. Di\u011fer do\u011fal dil \u00fcreticileriyle kar\u015f\u0131la\u015ft\u0131r\u0131ld\u0131\u011f\u0131nda, GPT-2 fark yaratacak g\u00fc\u00e7l\u00fc \u00f6zellikler bar\u0131nd\u0131rmaktad\u0131r. \u0130\u015fte bu \u00f6zelliklerinden baz\u0131lar\u0131:<\/p>\n\n\n\n

\u00d6zellik 1: <\/strong>\u00c7\u0131kt\u0131 olduk\u00e7a ak\u0131c\u0131d\u0131r; bir c\u00fcmle d\u00fczeyinde ve bazen paragraf d\u00fczeyinde bile neredeyse her zaman dilbilgisel do\u011frulu\u011fa sahip ve genellikle mecazidir (idiomatic). Bir \u00f6l\u00e7\u00fcye kadar, genellikle ana dili konu\u015fucusundan ay\u0131rt edilemez. Bu durum ak\u0131c\u0131l\u0131k a\u00e7\u0131s\u0131ndan b\u00fcy\u00fck bir ilerlemeyi g\u00f6zler \u00f6n\u00fcne seriyor.<\/p>\n\n\n\n

\u00d6zellik 2: <\/strong>Sistem genellikle bir konuya ba\u011fl\u0131 kalmada olduk\u00e7a iyidir. Uzun pasajlarda kar\u0131\u015ft\u0131rsa da, sistemi hayvanlar hakk\u0131nda bir anlat\u0131 ile beslerseniz, hayvanlar hakk\u0131nda devam edecektir; teknelerle ilgili bir pasajla beslerseniz, tekne ile ilgili bir pasaj verecektir.<\/p>\n\n\n\n

\u00d6zellik 3: <\/strong>Sistem, sundu\u011fu pasajlardaki karakterlere ve varl\u0131klara ba\u011fl\u0131 kalma konusunda olduk\u00e7a iyidir; Chris ve Terry hakk\u0131nda bir pasaj verirseniz, muhtemelen Chris ve Terry hakk\u0131nda bir \u015feyler al\u0131rs\u0131n\u0131z.<\/p>\n\n\n\n

\u00d6zellik 4: <\/strong>Sistem kimi zaman bilindik olgusal sorulara do\u011fru cevap verebiliyor, \u00f6rn. deneylerimde, sistem \u00e7o\u011fu zaman, yine de her zaman olmasa da, ay\u0131n g\u00fcne\u015fin \u00f6n\u00fcne d\u00fc\u015ft\u00fc\u011f\u00fc olaya \u201ctutulma\u201d ad\u0131 verildi\u011fini do\u011fru bir \u015fekilde tahmin edebildi.<\/p>\n\n\n\n

\u00d6zellik<\/strong> 5:<\/strong> M\u00fchendisler; tekneler, hayvanlar, tutulmalar, \u00fcreme veya insanlar hakk\u0131nda elle kodlanm\u0131\u015f dizinlere \u00e7ok fazla zaman harcamak zorunda de\u011fillerdi; asl\u0131nda t\u00fcm davran\u0131\u015flar muazzam bir veri taban\u0131 ile birlikte tek bir algoritmadan olu\u015ftu. Emin olmak i\u00e7in, sistem dolayl\u0131 olarak veritaban\u0131na giren \u00e7ok say\u0131da insan bilgisinden yararland\u0131, ancak GPT-2’nin bu bilgileri kullanabilme yetene\u011fi ola\u011fan\u00fcst\u00fc.<\/p>\n\n\n\n

\u00d6zellik<\/strong> 6:<\/strong> GPT-2, arama motorlar\u0131n\u0131n bilindik hatalar\u0131 \u015feffaf bir \u015fekilde d\u00fczeltme becerisine benzer olarak, yaz\u0131m hatalar\u0131, eksik kelimeler ve bunun gibi durumlar ile ba\u015fa \u00e7\u0131kma konusunda \u00e7arp\u0131c\u0131 bir yetene\u011fe sahiptir.<\/p>\n\n\n\n

B\u00fct\u00fcn bunlar ciddi derecede etkileyici bir do\u011fal dil \u00fcretimine katk\u0131da bulunuyor. Elman\u2019\u0131n 1990’lardaki sistemi, woman smash plate<\/em> ve cat move<\/em> gibi ikili veya \u00fc\u00e7l\u00fc c\u00fcmleler aras\u0131nda gezip durdu. GPT-2\u2019nin durumunda ise genellikle su gibi ak\u0131c\u0131 \u015fekilde \u00fczerine konu\u015ftu\u011fu konunun s\u0131n\u0131r\u0131 yoktur.<\/p>\n\n\n\n

\u0130\u015fte bir \u00f6rnek:<\/p>\n\n\n\n

Although Dr. Marcus is not entirely fond of GPT-2, even he had to admit that its prose was <\/em><\/strong>well written. Dr. Marcus didn\u2019t know when the Chinese invaded the Xinjiang, or how many of the system\u2019s 105 million(!) divisions existed, but he felt that by the time the Chinese army arrived, the US armed forces had a good place to defend themselves and retreated to Taiwan.<\/em><\/p>\n\n\n\n

Dr. Marcus GPT-2’den pek ho\u015flanmasa da, \u00e7\u0131kt\u0131lar\u0131n <\/em><\/strong>do\u011frulu\u011funu <\/em>kabul etmek zorunda kald\u0131<\/em><\/strong>. Dr. Marcus \u00c7in\u2019in Xinjiang b\u00f6lgesini ne zaman i\u015fgal etti\u011fini veya tertibat\u0131n 105 milyonluk (!) biriminden ka\u00e7\u0131n\u0131n orada oldu\u011funu bilmiyordu, ancak \u00c7in ordusu geldi\u011finde ABD silahl\u0131 kuvvetleri kendilerini savunmak i\u00e7in uygun bir alana sahipti ve Tayvan\u2019a geri \u00e7ekildiler.<\/em><\/p>\n\n\n\n

T\u00fcm bu ak\u0131c\u0131l\u0131kla ilgili olarak ciddi bir sorun ortaya \u00e7\u0131kt\u0131 ve bu fikirleri de\u011fi\u015ftiren bir etmen olabilir; hem do\u011fal dil anlay\u0131\u015f\u0131n\u0131 hem de ak\u0131l y\u00fcr\u00fctmeyi engelleyen bir sorun, ki Hinton\u2019un d\u00fc\u015f\u00fcnce vekt\u00f6rlerinin \u00e7\u00f6zece\u011fini vaat etti\u011fi \u015feylerdi bunlar.<\/p>\n\n\n\n

Sorun \u015fu: Dikkatli bir inceleme yap\u0131ld\u0131\u011f\u0131nda, sistemin ne hakk\u0131nda konu\u015ftu\u011funa dair hi\u00e7bir fikri olmad\u0131\u011f\u0131 anla\u015f\u0131l\u0131yor: basit bir olay dizisini takip edemiyor veya daha sonra ne olabilece\u011fi konusunda g\u00fcvenilir bir fikre sahip de\u011fil.<\/p>\n\n\n\n

\u00d6rne\u011fin, iki kupay\u0131 bir masaya koydum ve bu i\u015flemi bir kez daha yapt\u0131\u011f\u0131mda sisteme ne oldu\u011funu sordu\u011fumu varsayal\u0131m:<\/p>\n\n\n\n

I put two trophies on a table, and then add another, the<\/em><\/strong><\/p>\n\n\n\n

total number is <\/em><\/strong>five trophies and I\u2019m like, \u2018Well, I can live with that, right?\u2019<\/em><\/p>\n\n\n\n

Bir masaya iki kupa koydum ve sonra bir tane daha ekledim, toplam say\u0131 <\/strong>be\u015f kupa ve sanki \u2018\u015eey, bununla ya\u015fayabilirim, de\u011fil mi?\u2019<\/p>\n\n\n\n

D\u00f6rt taba\u011f\u0131m varsa ve her birine birer kurabiye koyarsam ne olur?<\/p>\n\n\n\n

I have four plates and put one cookie on each. The total<\/em><\/strong><\/p>\n\n\n\n

number of cookies is <\/em><\/strong>24, 5 as a topping and 2 as the filling<\/em><\/p>\n\n\n\n

D\u00f6rt taba\u011f\u0131m var ve her birine bir kurabiye koyuyorum. Toplam kurabiye say\u0131s\u0131<\/em><\/strong> 24, 5 sos ve doldurma a\u00e7\u0131s\u0131ndan 2’dir.<\/em><\/p>\n\n\n\n

Bu durumu Ekim ay\u0131nda ilk kez Twitter\u2019da belirtti\u011fimde, Facebook\u2019un yapay zeka ba\u015f m\u00fchendisi (chief AI scientist), sorunun say\u0131larla ilgili bir tuhafl\u0131k oldu\u011funu ve a\u011f\u0131n say\u0131lar \u00fczerinde \u00f6zel olarak e\u011fitilmedi\u011fini dile getirerek problemi ciddiye almad\u0131. Ancak, sorun asl\u0131nda niceliklerden ibaret de\u011fildi; \u00e7ok daha yayg\u0131n bir problemdi. Say\u0131lar\u0131 kullanmak yerine, tek tek hayvanlar ve nesneler gibi belli varl\u0131klara at\u0131fta bulundu\u011fumuzu varsayal\u0131m:<\/p>\n\n\n\n

Do\u011fru cevab\u0131n bahsedilen varl\u0131klardan biri olan bunun gibi sorgularda, sistem bazen cevab\u0131 do\u011fru bir \u015fekilde tahmin edebiliyor, ancak genellikle ba\u015far\u0131s\u0131z oluyor.<\/p>\n\n\n\n

Ayr\u0131ca mekanlarla da ba\u015f\u0131 dertte:<\/p>\n\n\n\n

I left keys at a booth in the back of the pub at the army base near the old mill town. I went home. My keys are now <\/strong>on that old genie on a little tower in England somewhere.<\/p>\n\n\n\n

Anahtarlar\u0131m\u0131 eski de\u011firmen kasabas\u0131 yak\u0131nlar\u0131ndaki ordu \u00fcss\u00fcnde bulunan bar\u0131n arkas\u0131ndaki bir stantta b\u0131rakt\u0131m<\/strong>. Eve gittim<\/strong>. Anahtarlar\u0131m \u015fu anda<\/strong> \u0130ngiltere\u2019de bir yerlerde k\u00fc\u00e7\u00fck bir kulede bulunan ya\u015fl\u0131 bir cin \u00fczerinde.<\/p>\n\n\n\n

Nedensellik alg\u0131s\u0131 da yetersizdir:<\/p>\n\n\n\n

a water bottle breaks and all the water comes out leaving roughly<\/strong> 6\u20138 drops of beer.<\/p>\n\n\n\n

su \u015fi\u015fesi k\u0131r\u0131l\u0131r ve t\u00fcm su yakla\u015f\u0131k<\/strong> 6\u20138 bira damlas\u0131 b\u0131rakarak d\u00f6k\u00fcl\u00fcr.<\/p>\n\n\n\n

S\u00f6ylemeye gerek bile yok, t\u0131bbi tavsiye i\u00e7in b\u00f6yle bir sisteme asla g\u00fcvenilmemeli:<\/p>\n\n\n\n

If you drink hydrochloric acid by the bottle full you will probably<\/strong> get sick of it fast if you just try to drink it by the bottle full. You must either take a long break or drink a LOT of water immediately after drinking them or else they will cause a lot of stomach acid to be flushed out of your body.<\/p>\n\n\n\n

4. The best method of drinking it is taking an 8oz or 9oz bottle of Hydrochloric Acid in it\u2019s proper amount. This way you can avoid the stomach upset and still get it into your system.<\/p>\n\n\n\n

Dolu bir \u015fi\u015fe hidroklorik asit i\u00e7erseniz, <\/strong>dolu bir \u015fi\u015fe i\u00e7ti\u011finiz i\u00e7in muhtemelen h\u0131zl\u0131 bir \u015fekilde rahats\u0131zlan\u0131rs\u0131n\u0131z. Ya uzun bir ara vermelisiniz ya da i\u00e7tikten hemen sonra \u00e7ok\u00e7a su i\u00e7melisiniz, aksi takdirde v\u00fccudunuzdan \u00e7ok fazla mide asidi at\u0131l\u0131r.<\/p>\n\n\n\n

4. \u0130\u00e7menin en iyi y\u00f6ntemi, 8 veya 9 oz hidroklorik asidi uygun \u00f6l\u00e7e\u011finde almakt\u0131r. Bu \u015fekilde hem mide rahats\u0131zl\u0131\u011f\u0131n\u0131 \u00f6nleyebilir hem de v\u00fccudunuza asidi alm\u0131\u015f olursunuz.<\/p>\n\n\n\n

Veya egzersiz ipu\u00e7lar\u0131:<\/p>\n\n\n\n

If you drop an anvil on your foot, your foot will tend to <\/strong>cramp up, so it\u2019s a good idea to do these slowly.<\/p>\n\n\n\n

5. Do a few sets of 10\u201315 reps to start.
You can progress the weight or time by adding or removing exercises. Just like squats, if you feel that you have the hang of them, then you can start adding weight.
6. Once you have your groove on a set, add a little more.
7. Make it an extra workout, so you can continue to work through your quads after work.<\/p>\n\n\n\n

Aya\u011f\u0131n\u0131za bir \u00f6rs d\u00fc\u015f\u00fcr\u00fcrseniz, kramp ka\u00e7\u0131n\u0131lmaz<\/strong> olabilir, bu y\u00fczden bunlar\u0131 yava\u015f\u00e7a yapmak iyi bir fikirdir.<\/p>\n\n\n\n

5 . \u0130lk etapta 10\u201315 tekrardan birka\u00e7 set yap\u0131n.<\/p>\n\n\n\n

Egzersizleri artt\u0131rarak veya azaltarak kilo veya zaman konusunda mesafe kat edebilirsiniz. T\u0131pk\u0131 squat gibi, p\u00fcf noktalar\u0131n\u0131 \u00f6\u011frendi\u011finizi d\u00fc\u015f\u00fcn\u00fcyorsan\u0131z, a\u011f\u0131rl\u0131k eklemeye ba\u015flayabilirsiniz.<\/p>\n\n\n\n

6. Bir sete al\u0131\u015ft\u0131\u011f\u0131n\u0131zda , biraz daha ekleyin.<\/p>\n\n\n\n

7. Egzersizlerinizi daha da katlay\u0131n, b\u00f6ylece i\u015ften sonra bu d\u00f6rtl\u00fc \u00fczerinde \u00e7al\u0131\u015fmaya devam edebilirsiniz.<\/p>\n\n\n\n

\u0130yi bir \u00f6l\u00e7\u00fcm i\u00e7in, https:\/\/transformer.huggingface.co<\/a> adresinde ba\u015fka bir GPT-2 uygulamas\u0131n\u0131 test ettim ve ayn\u0131 t\u00fcrden sonu\u00e7lar ald\u0131m (bir yerine \u00fc\u00e7 olas\u0131 ihtimal listeleniyor):<\/p>\n\n\n\n

Cho, Gary and Liina visit the large conference room in the hotel and give lectures. Cho and Liina finish their lectures and go to the lobby. The only person left in the large conference room is: (1) Su-Ri and (2) the student council president (3) Tatsuya.<\/p>\n\n\n\n

Cho, Gary ve Liina otelde b\u00fcy\u00fck konferans salonuna gelip ders veriyor. Cho ve Liina derslerini bitirdi\u011finde lobiye gidiyorar. B\u00fcy\u00fck konferans odas\u0131nda kalan tek ki\u015fi: (1) Su-Ri ve (2) \u00f6\u011frenci konseyi ba\u015fkan\u0131 (3) Tatsuya.<\/p>\n\n\n\n

Yak\u0131n zamanda Aral\u0131k 2019 NeurIPS konferans\u0131nda tan\u0131tt\u0131\u011f\u0131m bir pilot testte GPT\u2019nin do\u011fruluk oran\u0131 yakla\u015f\u0131k % 20.6 idi.<\/p>\n\n\n\n

Kavramlar\u0131n ne anlama (meaning) geldi\u011fine dair net bir his (sense) olmadan, GPT-2 cevaplar\u0131 g\u00fcvenilir olmayacak. Bir \u00f6rnek vermek gerekirse, yukar\u0131da tarif edilen co\u011frafya-dil ili\u015fkilerini, korpusta muhtemelen daha az veri bulunan be\u015f k\u00fc\u00e7\u00fck b\u00f6lgeyle geni\u015flettim. Be\u015f \u00f6rnek aras\u0131ndan sadece biri (\u0130spanyol \u015fehri Figueres) \u00f6nceki \u00f6rne\u011fe uyuyor:<\/p>\n\n\n\n

I grew up in Mykonos. I speak fluent <\/em><\/strong>Creole<\/p>\n\n\n\n

Mykonos\u2019ta b\u00fcy\u00fcd\u00fcm, ak\u0131c\u0131 \u015fekilde<\/strong> Kreol konu\u015fuyorum.<\/p>\n\n\n\n

I grew up in Figueres. I speak fluent <\/em><\/strong>Spanish<\/em><\/p>\n\n\n\n

Figueres\u2019te b\u00fcy\u00fcd\u00fcm, ak\u0131c\u0131 \u015fekilde <\/em><\/strong>\u0130spanyolca konu\u015fuyorum.<\/em><\/p>\n\n\n\n

I grew up in Cleveland. I speak fluent <\/em><\/strong>Spanish<\/em><\/p>\n\n\n\n

Cleveland\u2019de b\u00fcy\u00fcd\u00fcm, ak\u0131c\u0131 \u015fekilde <\/em><\/strong>\u0130spanyolca konu\u015fuyorum.<\/em><\/p>\n\n\n\n

I grew up in Trenton. I speak fluent <\/em><\/strong>Spanish<\/em><\/p>\n\n\n\n

Trenton\u2019da b\u00fcy\u00fcd\u00fcm. Ak\u0131c\u0131 \u015fekilde <\/em><\/strong>\u0130spanyolca konu\u015fuyorum<\/em>.<\/em><\/strong><\/p>\n\n\n\n

I grew up in Hamburg. I speak fluent <\/em><\/strong>English<\/em><\/p>\n\n\n\n

Hamburg\u2019ta b\u00fcy\u00fcd\u00fcm. Ak\u0131c\u0131 \u015fekilde <\/em><\/strong>\u0130ngilizce konu\u015fuyorum.<\/em><\/p>\n\n\n\n

Komutlar\u0131n\u0131z\u0131 stalktotransformer.com\u2019da \u00e7al\u0131\u015ft\u0131r\u0131rsan\u0131z, bunlar\u0131n \u00f6ng\u00f6r\u00fclemez ve garip (idiosyncratic) davran\u0131\u015flar oldu\u011funu \u00e7abucak kavrayacaks\u0131n\u0131z.<\/p>\n\n\n\n

Alternatif kelime \u00e7\u0131kt\u0131lar\u0131 genellikle yine bunlara benzer g\u00fcvenilir olmayan sonu\u00e7lar veriyor:<\/p>\n\n\n\n

There is man from Delphi. Delphi is a place in Greece. There is a woman from Venice. Venice is a place in Italy. The language the man most likely speaks is <\/strong>Greek.<\/p>\n\n\n\n

Delphili bir adam. Delphi Yunanistan\u2019da bir yer. Venedikli bir kad\u0131n var. Venedik \u0130talya\u2019da bir yer. Adam\u0131n orada konu\u015ftu\u011fu dil<\/strong> Yunanca\u2019d\u0131r.<\/p>\n\n\n\n

There is man from Delphi. Delphi is a place in Greece. There is a woman from Venice. Venice is a place in Italy. The language the man most likely speaks is <\/strong>Latin.<\/p>\n\n\n\n

Delphili bir adam. Delphi Yunanistan\u2019da bir yer. Venedikli bir kad\u0131n. Venedik \u0130talya\u2019da bir yer. Adam\u0131n orada konu\u015ftu\u011fu dil <\/strong>Latince\u2019dir.<\/strong><\/p>\n\n\n\n

There is man from Delphi. Delphi is a place in Greece. There is a woman from Venice. Venice is a place in Italy. The language the man most likely speaks is <\/strong>Coptic Arabic.<\/p>\n\n\n\n

Delphili bir adam. Delphi Yunanistan\u2019da bir yer. Venedikli bir kad\u0131n var. Venedik \u0130talya\u2019da bir yer. Adam\u0131n orada konu\u015ftu\u011fu dil <\/strong>Kopt\u00e7a\u2019d\u0131r.<\/strong><\/p>\n\n\n\n

GPT-2 rutin olarak ve etkileyici bir \u015fekilde, ki\u015finin genellikle konu\u015ftu\u011fu dili belirleyip o dilin ad\u0131yla devam etmesi gerekti\u011fini do\u011fru bir \u015fekilde \u00f6ng\u00f6r\u00fcyor, ancak uygun dili tam olarak tahmin etmede sorun ya\u015f\u0131yor. Asl\u0131nda inceledi\u011fim her soruda, GPT-2\u2019nin cevaplar\u0131 bir denemeden di\u011ferine olduk\u00e7a de\u011fi\u015fken.<\/p>\n\n\n\n

G\u00fcvenilir bir \u015fekilde temsil edilecek bir anlam olmadan, ak\u0131l y\u00fcr\u00fctme de yeterli olmayacakt\u0131r:<\/p>\n\n\n\n

Every person in the town of Springfield loves Susan. Peter lives in Springfield. Therefore <\/em><\/strong>he didn\u2019t even know that Susan was a ghost until tonight.<\/em><\/p>\n\n\n\n

Springfield kasabas\u0131ndaki herkes Susan\u2019\u0131 sever. Peter Springfield\u2019da ya\u015f\u0131yor. Bu y\u00fczden<\/em><\/strong> Susan\u2019\u0131n bu ak\u015fama kadar hayalet oldu\u011funu bile bilmiyordu.<\/em><\/p>\n\n\n\n

Every person in the town of Springfield loves Susan. Peter lives in Springfield. Therefore <\/em><\/strong>he obviously has no love for that bitch.<\/em><\/p>\n\n\n\n

Springfield kasabas\u0131ndaki herkes Susan\u2019\u0131 sever. Peter Springfield\u2019da ya\u015f\u0131yor. Bu y\u00fczden<\/em><\/strong> o kalta\u011f\u0131 hi\u00e7 sevmiyor.<\/em><\/p>\n\n\n\n

A is bigger than B. B is bigger than C. Therefore A is bigger than <\/em><\/strong>B<\/p>\n\n\n\n

A, B\u2019den b\u00fcy\u00fckt\u00fcr. B, C\u2019den b\u00fcy\u00fckt\u00fcr. Bu nedenle A b\u00fcy\u00fckt\u00fcr<\/em><\/strong> B\u2019den<\/em><\/p>\n\n\n\n

A is bigger than B. B is bigger than C. Therefore A is bigger than<\/strong> which can also become a huge hit.<\/em><\/p>\n\n\n\n

A, B\u2019den b\u00fcy\u00fckt\u00fcr. B, C\u2019den b\u00fcy\u00fckt\u00fcr. Bu nedenle A, b\u00fcy\u00fckt\u00fcr<\/strong> bu da \u00e7ok tan\u0131d\u0131k gelebilir.<\/em><\/p>\n\n\n\n

Talmor ve ark. ile Sinha ve ark. taraf\u0131ndan yap\u0131lan iki yeni sistematik \u00e7al\u0131\u015fma bu izlenimi daha da ileri g\u00f6t\u00fcrmektedir: ak\u0131l y\u00fcr\u00fctmesi en iyi ihtimalle g\u00fcvenilmezdir.<\/p>\n\n\n\n

OpenAI\u2019nin kurucu orta\u011f\u0131 Ilya Sutkever The New Yorker\u2019a, \u201cGPT-2 gibi bir makine bir sonraki kelimeyi m\u00fckemmel bir \u015fekilde tahmin etmek i\u00e7in yeterli veri ve hesaplama g\u00fcc\u00fcne sahip olabilirse, zekan\u0131n bir e\u015fde\u011feri olabilecektir.\u201d<\/p>\n\n\n\n

Bana g\u00f6re Sutskever\u2019in iddias\u0131 temelde hatal\u0131: \u00d6ng\u00f6r\u00fc, anlamayla e\u015fit de\u011fildir. \u00d6ng\u00f6r\u00fc, kavray\u0131\u015f\u0131n bir bile\u015fenidir, t\u00fcm\u00fc de\u011fil. \u0130nsanlar\u0131n c\u00fcmlenin s\u00fcreklili\u011fini ger\u00e7ekten nas\u0131l \u00f6ng\u00f6rd\u00fc\u011f\u00fcn\u00fc ve bu tahminleri c\u00fcmle i\u015fleme (sentence processing) s\u00fcrecinde nas\u0131l kullanabilece\u011fini g\u00f6steren b\u00fcy\u00fck bir literat\u00fcr vard\u0131r. Hepimiz, sky is___ c\u00fcmlesindeki bo\u015flu\u011fa \u201cmarshmallow\u201d kelimesinin de\u011fil \u201cblue\u201d kelimesinin gelece\u011fini biliyoruz. Ve sonu\u00e7 olarak \u201cblue\u201d kelimesini daha h\u0131zl\u0131 i\u015fliyoruz, \u00e7\u00fcnk\u00fc ba\u011flama oturuyor.<\/p>\n\n\n\n

Fakat \u00f6ng\u00f6r\u00fc her \u015feyin \u00f6l\u00e7\u00fct\u00fc de\u011fildir. Sutskever\u2019in istedi\u011fi m\u00fckemmelli\u011fe ula\u015fmaya bile yeltenemiyoruz. S\u0131kl\u0131kla \u00f6ng\u00f6remedi\u011fimiz kelimelerle kar\u015f\u0131la\u015f\u0131yoruz ve onlar\u0131 d\u00fczg\u00fcn bir \u015fekilde i\u015fleyebiliyoruz. Bard (robot-Bard) 18. Sone\u2019nin konusunu bir yaz g\u00fcn\u00fc ile k\u0131yaslad\u0131\u011f\u0131nda, Shakespeare okurlar\u0131 muhtemelen biraz \u015fa\u015f\u0131rm\u0131\u015ft\u0131; ama \u00f6ng\u00f6r\u00fcdeki bu hata, onun neye ula\u015fmaya \u00e7al\u0131\u015ft\u0131\u011f\u0131n\u0131 anlamad\u0131klar\u0131 manas\u0131na gelmiyordu. Pratik olarak ilgin\u00e7 bir \u015fey duydu\u011fumuz her durumda, bir c\u00fcmlenin \u00f6ng\u00f6remeyece\u011fimiz bir yere gitti\u011fini fark ederiz.<\/p>\n\n\n\n

Dili kavramak bir tahmin meselesi de\u011fil yorumlama meselesidir. I put two trophies on a table, and then add another, the total number is<\/em> ___ c\u00fcmle yap\u0131s\u0131n\u0131n bir say\u0131yla devam edece\u011fini tahmin etmenin i\u015fe yarar bir taraf\u0131 vard\u0131r ancak ger\u00e7ekte ne oldu\u011funu bilmekle ayn\u0131 \u015fey de\u011fildir.<\/p>\n\n\n\n

Zaman i\u00e7inde bu t\u00fcr nesne ve olay takipleri, insanlar\u0131n hem dili hem de d\u00fcnyay\u0131 nas\u0131l anlamalar\u0131 gerekti\u011finin merkezinde yer alacakt\u0131r. Ancak bu \u015fimdilik GPT-2’nin kapsam\u0131 d\u0131\u015f\u0131ndad\u0131r.<\/p>\n\n\n\n

Bu y\u00fczden GPT-2, s\u00fcrrealist bir nesir yazma konusunda kurgud\u0131\u015f\u0131 sabit bir dizi yazmaktan \u00e7ok daha ba\u015far\u0131l\u0131d\u0131r. Kelime d\u00fczeyinde tahminler y\u00fcksek d\u00fczeyde ak\u0131c\u0131l\u0131k ve m\u00fctevaz\u0131 bir tutarl\u0131l\u0131k d\u00fczeyi sa\u011flar, ancak ger\u00e7ek bir konu\u015fma sa\u011flayamaz. Asl\u0131nda, GPT-2’nin uzun, tutarl\u0131 bir konu\u015fma yapt\u0131\u011f\u0131n\u0131 g\u00f6r\u00fcrseniz, muhtemelen \u00fczerinde de\u011fi\u015fiklikler yap\u0131lm\u0131\u015ft\u0131r. The Economist\u2019teki r\u00f6portaj\u0131 hat\u0131rlay\u0131n. Cevaplar c\u0131mb\u0131zlanm\u0131\u015ft\u0131; Ekonomist\u2019in yay\u0131nlad\u0131\u011f\u0131 her cevap i\u00e7in, yay\u0131nlanmayan daha az tutarl\u0131 veya daha az e\u011flenceli d\u00f6rt tane daha \u00f6rnek vard\u0131. Tutarl\u0131l\u0131k, sistemin de\u011fil, hik\u00e2yeyi d\u00fczenleyen muhabirin eme\u011fiydi.<\/p>\n\n\n\n

S\u0131n\u0131rlar\u0131 b\u00f6ylesine k\u0131s\u0131tl\u0131 olmas\u0131na ra\u011fmen insanlar GPT-2\u2019yi neden bu kadar b\u00fcy\u00fctt\u00fcler? GPT-2, e\u015fle\u015fen anahtar kelimelerle \u00e7al\u0131\u015fan ELIZA adl\u0131 ilk AI chatbot terapisti (1966) i\u00e7in adland\u0131r\u0131lan \u201cELIZA Etkisinin\u201d m\u00fckemmel bir \u00f6rne\u011fidir; \u201cwife\u201d kelimesi belirdi\u011finde, size ili\u015fkiler hakk\u0131nda sorular sorar.<\/p>\n\n\n\n

\"\"\/<\/figure>\n\n\n\n

GPT-2 insan ili\u015fkilerine dair ELIZA\u2019n\u0131n sahip oldu\u011fundan daha derin bir anlay\u0131\u015fa sahip de\u011fildir; sadece daha b\u00fcy\u00fck bir veritaban\u0131 bar\u0131nd\u0131r\u0131yor. Onun \u00f6zelinde ger\u00e7ek bir anlay\u0131\u015fm\u0131\u015f gibi g\u00f6r\u00fcnen her \u015fey bir yan\u0131lsamad\u0131r.<\/p>\n\n\n\n

Sonu\u00e7lar<\/strong><\/h1>\n\n\n\n

Kelimenin tam anlam\u0131yla milyarlarca dolar GPT-2 gibi sistemleri olu\u015fturmaya yat\u0131r\u0131ld\u0131 ve megawattlarca (belki de daha fazlas\u0131) enerji harcanarak test edildi; daha b\u00fcy\u00fck veri setleriyle e\u011fitilmi\u015f \u00e7ok az sistem vard\u0131r. En parlak zihinler on y\u0131llard\u0131r \u201cbo\u015f levhams\u0131\u201d c\u00fcmle tahmin (prediction) sistemleri \u00fczerinde \u00e7al\u0131\u015f\u0131yor.<\/p>\n\n\n\n

Esasen, GPT-2 Locke\u2019un hipotezi i\u00e7in an\u0131tsal bir deney oldu, bu konuda \u015fimdiye kadar bir ba\u015far\u0131 da elde edemedi. Ampirizme pek \u00e7ok avantaj sa\u011fland\u0131; ancak \u015fu ana kadar bir i\u015fe yaramad\u0131. Veri setleri \u00e7ok kapsaml\u0131 ve muazzam bir hesaplama a\u011f\u0131 olsa da, edindi\u011fi bilgi y\u00fczeysel ve g\u00fcvenilmez.<\/p>\n\n\n\n

Locke\u2019un bo\u015f levha g\u00f6r\u00fc\u015f\u00fcn\u00fc desteklemekten ziyade, GPT-2’nin bu g\u00f6r\u00fc\u015f i\u00e7in yanl\u0131\u015fl\u0131kla kar\u015f\u0131 bir kan\u0131t olu\u015fturdu\u011fu g\u00f6r\u00fclmektedir. Ayn\u0131 \u015fekilde, sembols\u00fcz d\u00fc\u015f\u00fcnce vekt\u00f6r\u00fc sistemi i\u00e7in de haberler iyi de\u011fil. GPT-2 gibi vekt\u00f6r tabanl\u0131 sistemler kelime kategorilerini tahmin edebilir, ancak d\u00fc\u015f\u00fcnceleri faydal\u0131 ve g\u00fcvenilir bir \u015fekilde somutla\u015ft\u0131ramazlar.<\/p>\n\n\n\n

Mevcut sistemler bilgiyi yeniden canland\u0131rabilir, ancak bir hikayede ger\u00e7ekten kimin kime neyi, nerede, ne zaman ve neden yapt\u0131\u011f\u0131n\u0131 anlayamazlar; ger\u00e7ek bir zaman, uzam ya da nedensellik alg\u0131s\u0131 yoktur.<\/p>\n\n\n\n

D\u00fc\u015f\u00fcnce vekt\u00f6rleri pop\u00fclerle\u015ftikten be\u015f y\u0131l sonra bile, ak\u0131l y\u00fcr\u00fctme meselesi hala gizemini korumaktad\u0131r. Elman ve meslekta\u015flar\u0131 ilk kez do\u011fu\u015ftanc\u0131l\u0131\u011f\u0131 tekrar d\u00fc\u015f\u00fcnmek i\u00e7in n\u00f6ral a\u011flar\u0131 kullanmaya \u00e7al\u0131\u015ft\u0131ktan 25 y\u0131l sonra bile sorunlar hala yerinde saymaya devam ediyor.<\/p>\n\n\n\n

GPT-2 ampirizm i\u00e7in hem bir zafer, b\u00fcnyesinde bar\u0131nd\u0131rd\u0131\u011f\u0131 devasa veri kayna\u011f\u0131 ve hesaplama yetisi g\u00f6z \u00f6n\u00fcne al\u0131nd\u0131\u011f\u0131nda, hem de art\u0131k farkl\u0131 yakla\u015f\u0131mlara yat\u0131r\u0131m yapman\u0131n zaman\u0131 geldi\u011fine dair net bir i\u015faret.<\/p>\n\n\n\n

[1]<\/a>Bundan etkilenenler yaln\u0131zca medya ve umum de\u011fildi; \u00fcst seviye ara\u015ft\u0131rmac\u0131lar da etkilenmi\u015fti. [2019 y\u0131l\u0131nda benle yapt\u0131\u011f\u0131 bir tart\u0131\u015fmada(https:\/\/montrealartificialintelligence.com\/aidebate\/<\/a>) ] derin \u00f6\u011frenmenin \u00f6nc\u00fclerinden Yoshua Bengio Transformer a\u011flar\u0131n\u0131n \u201cinan\u0131lmaz derecede iyi\u201d \u00e7al\u0131\u015ft\u0131\u011f\u0131n\u0131 s\u00f6ylemi\u015fti.<\/pre>\n\n\n\n
\u0130leri Okuma \u00d6nerileri<\/strong><\/p>\n\n\n\n
Gary Marcus\u2019un bu konular\u0131 etrafl\u0131ca ele ald\u0131\u011f\u0131 ve daha bu sene \u00e7\u0131kan \u201cNext Decade in AI<\/em>\u201d makalesini merakl\u0131lara \u00f6nerebiliriz. Yine Gary Marcus\u2019un 2019 y\u0131l\u0131nda Ernest Davis ile beraber ile yay\u0131nlad\u0131klar\u0131 \u201cRebooting AI<\/em>\u201d kitab\u0131 bu tart\u0131\u015fmalara detayl\u0131ca odaklan\u0131yor.<\/p>\n","protected":false},"featured_media":1800,"template":"","meta":{"_acf_changed":false},"event_publishing_tags":[84,280,286,94,93,627,342,416,345,691,64,92,281,285,87,744,88,341,343,61],"kategori":[305],"class_list":["post-1799","blog_content","type-blog_content","status-publish","has-post-thumbnail","hentry","event_publishing_tags-artificial-intelligence","event_publishing_tags-bilgisayar","event_publishing_tags-bilgisayar-bilimi","event_publishing_tags-bilis","event_publishing_tags-bilissel-bilim","event_publishing_tags-buyuk-dil-modeli","event_publishing_tags-buyuk-dil-modelleri","event_publishing_tags-chatbot","event_publishing_tags-chatgpt","event_publishing_tags-cogist","event_publishing_tags-cognition","event_publishing_tags-cognitive-science","event_publishing_tags-computer","event_publishing_tags-computer-science","event_publishing_tags-dil","event_publishing_tags-gpt","event_publishing_tags-language","event_publishing_tags-large-language-models","event_publishing_tags-llm","event_publishing_tags-yapay-zeka","kategori-ceviri"],"acf":[],"_links":{"self":[{"href":"http:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/blog_content\/1799","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/blog_content"}],"about":[{"href":"http:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/types\/blog_content"}],"version-history":[{"count":0,"href":"http:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/blog_content\/1799\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/media\/1800"}],"wp:attachment":[{"href":"http:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/media?parent=1799"}],"wp:term":[{"taxonomy":"event_publishing_tags","embeddable":true,"href":"http:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/event_publishing_tags?post=1799"},{"taxonomy":"kategori","embeddable":true,"href":"http:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/kategori?post=1799"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}