{"id":36662,"date":"2025-12-21T09:00:39","date_gmt":"2025-12-21T06:00:39","guid":{"rendered":"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/"},"modified":"2025-12-21T09:00:39","modified_gmt":"2025-12-21T06:00:39","slug":"transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk","status":"publish","type":"post","link":"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/","title":{"rendered":"Transformer’lar\u0131n Ger\u00e7ek D\u00fc\u015f\u00fcnce Mekanizmas\u0131: Bir Yapay Zeka Dil Modelinin Beynine Yolculuk"},"content":{"rendered":"

Transformer’lar\u0131n Ger\u00e7ek D\u00fc\u015f\u00fcnce Mekanizmas\u0131: Bir Yapay Zeka Dil Modelinin Beynine Yolculuk<\/h2>\n

Yapay zeka d\u00fcnyas\u0131nda devrim yaratan Transformer mimarisi, do\u011fal dil i\u015fleme (NLP) alan\u0131nda \u00e7\u0131\u011f\u0131r a\u00e7t\u0131. G\u00fcn\u00fcm\u00fcz\u00fcn en geli\u015fmi\u015f dil modelleri olan GPT-3, BERT ve di\u011ferleri, bu yenilik\u00e7i mimarinin temelini kullan\u0131r. Peki, bu modeller ger\u00e7ekten “nas\u0131l d\u00fc\u015f\u00fcn\u00fcyor” veya daha do\u011frusu, metni nas\u0131l anl\u0131yor, \u00fcretiyor ve yorumluyor? Bu makale, Transformer’lar\u0131n karma\u015f\u0131k i\u00e7 i\u015fleyi\u015fini, temel bile\u015fenlerinden ileri d\u00fczey mekanizmalar\u0131na kadar ad\u0131m ad\u0131m a\u00e7\u0131klayarak, bu yapay zeka beyinlerinin s\u0131r perdesini aralamay\u0131 hedefliyor.<\/p>\n

Transformer Mimarisine Genel Bak\u0131\u015f<\/h3>\n

Transformer mimarisi, 2017 y\u0131l\u0131nda “Attention Is All You Need” makalesiyle tan\u0131t\u0131ld\u0131 ve \u00f6zellikle s\u0131ral\u0131 veri i\u015fleme g\u00f6revlerinde, \u00f6nceki yinelemeli a\u011flara (RNN, LSTM) k\u0131yasla devrim niteli\u011finde bir performans art\u0131\u015f\u0131 sa\u011flad\u0131. Bu ba\u015far\u0131n\u0131n anahtar\u0131, dikkat mekanizmas\u0131n\u0131 yo\u011fun bir \u015fekilde kullanmas\u0131 ve paralel i\u015flemeye olanak tan\u0131mas\u0131d\u0131r.<\/p>\n

Encoder-Decoder Yap\u0131s\u0131 ve Yaln\u0131zca Decoder Modelleri<\/h4>\n

Orijinal Transformer modeli, bir encoder (kodlay\u0131c\u0131) ve bir decoder (kod \u00e7\u00f6z\u00fcc\u00fc) olmak \u00fczere iki ana b\u00f6l\u00fcmden olu\u015fur. Encoder, giri\u015f metnini anlamsal bir temsile d\u00f6n\u00fc\u015ft\u00fcr\u00fcrken, decoder bu temsili kullanarak \u00e7\u0131kt\u0131 metnini \u00fcretir. Ancak, g\u00fcn\u00fcm\u00fczdeki bir\u00e7ok b\u00fcy\u00fck dil modeli (\u00f6rne\u011fin GPT serisi), yaln\u0131zca decoder bloklar\u0131ndan olu\u015fan “yaln\u0131zca decoder” mimarisini kullan\u0131r ve bu sayede metin \u00fcretimi gibi g\u00f6revlerde \u00fcst\u00fcn ba\u015far\u0131 g\u00f6sterir.<\/p>\n

Temel Bloklar: Dikkat ve Besleme A\u011flar\u0131<\/h4>\n

Her bir Transformer blo\u011fu, temel olarak iki ana alt katmandan olu\u015fur: \u00e7ok ba\u015fl\u0131 dikkat mekanizmas\u0131 (Multi-Head Attention) ve pozisyon tabanl\u0131 besleme a\u011f\u0131 (Position-wise Feed-Forward Network). Bu katmanlar, girdi verisini farkl\u0131 a\u00e7\u0131lardan i\u015fleyerek daha zengin temsiller elde edilmesini sa\u011flar.<\/p>\n

Paralel \u0130\u015fleme Yetene\u011fi<\/h4>\n

Transformer’lar\u0131n en b\u00fcy\u00fck avantajlar\u0131ndan biri, s\u0131ral\u0131 i\u015flemeye ba\u011f\u0131ml\u0131 olmamas\u0131d\u0131r. RNN’ler ve LSTM’ler her kelimeyi s\u0131rayla i\u015flerken, Transformer’lar t\u00fcm kelimeleri ayn\u0131 anda i\u015fleyebilir. Bu paralel i\u015flem yetene\u011fi, e\u011fitim s\u00fcrelerini \u00f6nemli \u00f6l\u00e7\u00fcde k\u0131salt\u0131r ve \u00e7ok daha b\u00fcy\u00fck veri k\u00fcmeleriyle \u00e7al\u0131\u015fmay\u0131 m\u00fcmk\u00fcn k\u0131lar.<\/p>\n

Dikkat Mekanizmas\u0131: Her \u015eeyin Kalbi<\/h3>\n

Transformer’lar\u0131n “d\u00fc\u015f\u00fcnme” bi\u00e7iminin temelini olu\u015fturan dikkat mekanizmas\u0131, bir kelimenin anlam\u0131n\u0131 \u00e7\u0131kar\u0131rken c\u00fcmlenin di\u011fer kelimeleriyle olan ili\u015fkisini anlamas\u0131na olanak tan\u0131r. Bu, modelin uzun mesafeli ba\u011f\u0131ml\u0131l\u0131klar\u0131 yakalamas\u0131na yard\u0131mc\u0131 olur.<\/p>\n

Self-Attention: Kendi Kendine Odaklanma<\/h4>\n

Self-attention (kendi kendine dikkat), bir kelimenin, ayn\u0131 c\u00fcmledeki di\u011fer kelimelerle olan ili\u015fkisini dinamik olarak a\u011f\u0131rl\u0131kland\u0131rmas\u0131d\u0131r. \u00d6rne\u011fin, “Bankan\u0131n \u00f6n\u00fcnde duran adam…” c\u00fcmlesinde “banka” kelimesinin bir finans kurumu mu yoksa bir nehir kenar\u0131 m\u0131 oldu\u011funu, c\u00fcmlenin di\u011fer kelimelerine bakarak anlamas\u0131n\u0131 sa\u011flar.<\/p>\n

Query, Key, Value Kavramlar\u0131<\/h4>\n

Self-attention mekanizmas\u0131, her bir kelime i\u00e7in \u00fc\u00e7 farkl\u0131 vekt\u00f6r \u00fcretir: Query (Sorgu), Key (Anahtar) ve Value (De\u011fer). Bir kelimenin Query’si, di\u011fer t\u00fcm kelimelerin Key’leri ile kar\u015f\u0131la\u015ft\u0131r\u0131l\u0131r. Bu kar\u015f\u0131la\u015ft\u0131rmadan \u00e7\u0131kan benzerlik skorlar\u0131, Value vekt\u00f6rlerini a\u011f\u0131rl\u0131kland\u0131rmak i\u00e7in kullan\u0131l\u0131r.<\/p>\n\n\n\n\n\n\n\n
Kavram<\/th>\nRol\u00fc<\/th>\nAnaloji<\/th>\n<\/tr>\n<\/thead>\n
Query (Sorgu)<\/strong><\/td>\nMevcut token’\u0131n di\u011fer token’larla olan ili\u015fkisini aramak i\u00e7in kullan\u0131l\u0131r. “Ben ne ar\u0131yorum?”<\/td>\nVeritaban\u0131nda arad\u0131\u011f\u0131n\u0131z sorgu anahtar\u0131.<\/td>\n<\/tr>\n
Key (Anahtar)<\/strong><\/td>\nDi\u011fer token’lar\u0131n kendilerini Query’ye tan\u0131tmak i\u00e7in kulland\u0131\u011f\u0131 temsil. “Ben neyim?”<\/td>\nVeritaban\u0131ndaki her bir kayd\u0131n anahtar etiketi.<\/td>\n<\/tr>\n
Value (De\u011fer)<\/strong><\/td>\nKey ile e\u015fle\u015fen token’\u0131n anlam i\u00e7eri\u011fi. “Benim i\u00e7eri\u011fim ne?”<\/td>\nVeritaban\u0131nda anahtar ile e\u015fle\u015fen kay\u0131tlar\u0131n ger\u00e7ek i\u00e7eri\u011fi.<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

Multi-Head Attention: Farkl\u0131 Bak\u0131\u015f A\u00e7\u0131lar\u0131<\/h4>\n

Tek bir dikkat mekanizmas\u0131 yerine, Transformer’lar birden fazla “dikkat ba\u015fl\u0131\u011f\u0131” (multi-head attention) kullan\u0131r. Her ba\u015fl\u0131k, farkl\u0131 bir Query, Key, Value d\u00f6n\u00fc\u015f\u00fcm matrisi \u00f6\u011frenir ve bu sayede giri\u015f verisinin farkl\u0131 y\u00f6nlerini veya ili\u015fkilerini yakalayabilir. Bu, modelin ayn\u0131 anda birden fazla “bak\u0131\u015f a\u00e7\u0131s\u0131yla” d\u00fc\u015f\u00fcnmesini sa\u011flar.<\/p>\n

Maskeleme ve Neden \u00d6nemli Oldu\u011fu<\/h4>\n

\u00d6zellikle decoder bloklar\u0131nda, modelin gelecekteki token’lara “bakmas\u0131n\u0131” engellemek i\u00e7in maskeleme kullan\u0131l\u0131r. Bu, modelin bir sonraki kelimeyi tahmin ederken yaln\u0131zca kendisinden \u00f6nceki kelimelere dikkat etmesini sa\u011flar, b\u00f6ylece metin \u00fcretiminde tutarl\u0131l\u0131k ve do\u011fruluk korunur.<\/p>\n

Pozisyonel Kodlama: S\u0131ran\u0131n \u00d6nemi<\/h3>\n

Dikkat mekanizmas\u0131, kelimelerin c\u00fcmledeki konumlar\u0131ndan ba\u011f\u0131ms\u0131zd\u0131r; yani kelimelerin s\u0131ras\u0131n\u0131 dikkate almaz. Ancak dil, kelime s\u0131ras\u0131na b\u00fcy\u00fck \u00f6l\u00e7\u00fcde ba\u011fl\u0131d\u0131r. \u0130\u015fte bu noktada pozisyonel kodlama devreye girer.<\/p>\n

S\u0131ra Bilgisinin Kayb\u0131 ve \u0130htiya\u00e7<\/h4>\n

Transformer’lar paralel i\u015fleme yapt\u0131\u011f\u0131 i\u00e7in, “Ay\u015fe topu att\u0131” ile “Topu Ay\u015fe att\u0131” c\u00fcmlelerindeki kelimeler ayn\u0131 anda i\u015flenirse, model kelimelerin yerlerinin de\u011fi\u015fti\u011fini ve anlam\u0131n farkl\u0131la\u015ft\u0131\u011f\u0131n\u0131 alg\u0131layamaz. Bu nedenle, her bir kelimenin konum bilgisini modele iletmek kritik \u00f6neme sahiptir.<\/p>\n

Sin\u00fcsoidal Pozisyonel Kodlama<\/h4>\n

Orijinal Transformer makalesinde, kelime g\u00f6mmelerine (word embeddings) eklenen sin\u00fcsoidal fonksiyonlara dayal\u0131 sabit bir pozisyonel kodlama y\u00f6ntemi \u00f6nerilmi\u015ftir. Bu y\u00f6ntem, her pozisyon i\u00e7in benzersiz bir vekt\u00f6r \u00fcretir ve modelin farkl\u0131 pozisyonlar aras\u0131ndaki g\u00f6receli mesafeyi \u00f6\u011frenmesine yard\u0131mc\u0131 olur.<\/p>\n

Di\u011fer Pozisyonel Kodlama Y\u00f6ntemleri<\/h4>\n

Sin\u00fcsoidal kodlaman\u0131n yan\u0131 s\u0131ra, baz\u0131 modeller \u00f6\u011frenilebilir pozisyonel g\u00f6mmeler kullan\u0131r. Bu y\u00f6ntemlerde, pozisyon vekt\u00f6rleri modelle birlikte e\u011fitilir ve belirli bir g\u00f6reve veya veri k\u00fcmesine daha iyi uyum sa\u011flayabilir. Son zamanlarda, RoPE (Rotary Positional Embeddings) gibi daha geli\u015fmi\u015f y\u00f6ntemler de pop\u00fclerlik kazanm\u0131\u015ft\u0131r.<\/p>\n

Besleme A\u011flar\u0131 ve Normalizasyon Katmanlar\u0131<\/h3>\n

Dikkat mekanizmas\u0131ndan sonra, Transformer bloklar\u0131, modelin daha karma\u015f\u0131k ili\u015fkileri \u00f6\u011frenmesini sa\u011flayan besleme a\u011flar\u0131 ve e\u011fitim stabilitesini art\u0131ran normalizasyon katmanlar\u0131 i\u00e7erir.<\/p>\n

Feed-Forward Network (FFN) Katmanlar\u0131<\/h4>\n

Her dikkat katman\u0131ndan sonra, bir pozisyon tabanl\u0131 besleme a\u011f\u0131 (Feed-Forward Network) bulunur. Bu a\u011f, genellikle iki do\u011frusal d\u00f6n\u00fc\u015f\u00fcm ve aralar\u0131nda bir aktivasyon fonksiyonundan (\u00f6rne\u011fin ReLU) olu\u015fur. FFN, her bir pozisyondaki token temsilini ba\u011f\u0131ms\u0131z olarak d\u00f6n\u00fc\u015ft\u00fcrerek, modelin daha karma\u015f\u0131k, do\u011frusal olmayan ili\u015fkileri \u00f6\u011frenmesine yard\u0131mc\u0131 olur.<\/p>\n

Residual Connections (Kal\u0131nt\u0131 Ba\u011flant\u0131lar)<\/h4>\n

Derin sinir a\u011flar\u0131nda e\u011fitim s\u0131ras\u0131nda ortaya \u00e7\u0131kabilen kaybolan gradyan (vanishing gradient) sorununu hafifletmek i\u00e7in residual connections (kal\u0131nt\u0131 ba\u011flant\u0131lar) kullan\u0131l\u0131r. Bu ba\u011flant\u0131lar, her alt katman\u0131n girdisini, katman\u0131n \u00e7\u0131kt\u0131s\u0131na ekleyerek bilginin a\u011f i\u00e7inde daha kolay akmas\u0131n\u0131 sa\u011flar. Bu sayede model, daha derin katmanlarda bile etkili bir \u015fekilde \u00f6\u011frenebilir.<\/p>\n

Layer Normalization (Katman Normalizasyonu)<\/h4>\n

Her alt katman\u0131n \u00e7\u0131kt\u0131s\u0131na ve residual connection’\u0131n eklenmesinden sonra Layer Normalization (katman normalizasyonu) uygulan\u0131r. Bu teknik, her bir katman\u0131n \u00e7\u0131kt\u0131lar\u0131n\u0131n ortalamas\u0131n\u0131 ve varyans\u0131n\u0131 normalize ederek, e\u011fitim s\u00fcrecini daha istikrarl\u0131 ve h\u0131zl\u0131 hale getirir.<\/p>\n

Tokenizasyon ve G\u00f6mme Katmanlar\u0131<\/h3>\n

Bir dil modelinin metni anlayabilmesi i\u00e7in, \u00f6ncelikle metnin say\u0131sal bir forma d\u00f6n\u00fc\u015ft\u00fcr\u00fclmesi gerekir. Bu s\u00fcre\u00e7 tokenizasyon ve g\u00f6mme katmanlar\u0131 arac\u0131l\u0131\u011f\u0131yla ger\u00e7ekle\u015fir.<\/p>\n

Metinden Say\u0131lara: Tokenizasyon S\u00fcreci<\/h4>\n

Tokenizasyon, metni anlaml\u0131 par\u00e7alara, yani “token”lara ay\u0131rma i\u015flemidir. Bu token’lar kelimeler, alt kelimeler veya karakterler olabilir. \u00d6rne\u011fin, “Merhaba d\u00fcnya!” c\u00fcmlesi “Merhaba”, “d\u00fcnya”, “!” gibi token’lara ayr\u0131labilir. Her token’a benzersiz bir say\u0131sal ID atan\u0131r.<\/p>\n

Kelime G\u00f6mme (Word Embeddings)<\/h4>\n

Token’lar say\u0131sal ID’lere d\u00f6n\u00fc\u015ft\u00fcr\u00fcld\u00fckten sonra, bu ID’ler kelime g\u00f6mme (word embedding) katman\u0131ndan ge\u00e7er. Bu katman, her token’\u0131 y\u00fcksek boyutlu (\u00f6rne\u011fin 768 boyutlu) yo\u011fun bir vekt\u00f6re d\u00f6n\u00fc\u015ft\u00fcr\u00fcr. Bu g\u00f6mme vekt\u00f6rleri, kelimelerin anlamsal ve sentaktik \u00f6zelliklerini yakalar; yani birbirine benzer kelimelerin vekt\u00f6rleri de birbirine yak\u0131n olur.<\/p>\n

Alt Kelime (Subword) Tokenizasyonunun Avantajlar\u0131<\/h4>\n

Geleneksel kelime bazl\u0131 tokenizasyon, b\u00fcy\u00fck bir kelime da\u011farc\u0131\u011f\u0131na ve bilinmeyen kelimeler (OOV – Out Of Vocabulary) sorununa yol a\u00e7abilir. Alt kelime tokenizasyonu (\u00f6rne\u011fin Byte-Pair Encoding – BPE veya WordPiece), kelimeleri daha k\u00fc\u00e7\u00fck, anlaml\u0131 par\u00e7alara ay\u0131rarak bu sorunlar\u0131 \u00e7\u00f6zer. Bu sayede daha k\u00fc\u00e7\u00fck bir kelime da\u011farc\u0131\u011f\u0131 ile daha fazla kelime temsil edilebilir ve yeni kelimeler bile anlaml\u0131 alt par\u00e7alara ayr\u0131labilir.<\/p>\n

E\u011fitim S\u00fcreci ve Optimizasyon<\/h3>\n

Bir Transformer modelinin “d\u00fc\u015f\u00fcnmeyi” \u00f6\u011frenmesi, iki ana a\u015famadan olu\u015fan kapsaml\u0131 bir e\u011fitim s\u00fcrecini gerektirir: \u00f6n e\u011fitim ve ince ayar.<\/p>\n

\u00d6n E\u011fitim (Pre-training) ve Geni\u015f Veri K\u00fcmeleri<\/h4>\n

Transformer modelleri, internetten toplanan milyarlarca kelimeden olu\u015fan devasa metin veri k\u00fcmeleri \u00fczerinde \u00f6n e\u011fitilir. Bu a\u015famada model, genellikle “maskeli dil modellemesi” (masked language modeling) veya “sonraki kelime tahmini” (next token prediction) gibi denetimsiz g\u00f6revleri yerine getirir. Bu sayede dilin genel yap\u0131s\u0131n\u0131, gramerini, semanti\u011fini ve d\u00fcnya bilgisini \u00f6\u011frenir.<\/p>\n

\u0130nce Ayar (Fine-tuning) ve G\u00f6reve \u00d6zel Adaptasyon<\/h4>\n

\u00d6n e\u011fitimden sonra, model belirli bir g\u00f6reve (\u00f6rne\u011fin duygu analizi, soru-cevap, metin \u00f6zetleme) uyarlanmak \u00fczere daha k\u00fc\u00e7\u00fck, etiketli veri k\u00fcmeleri \u00fczerinde ince ayar (fine-tuning) yap\u0131l\u0131r. Bu a\u015famada, \u00f6nceden \u00f6\u011frenilmi\u015f genel dil bilgisi, belirli bir g\u00f6revin gerektirdi\u011fi \u00f6zel n\u00fcanslar\u0131 \u00f6\u011frenmek i\u00e7in adapte edilir.<\/p>\n

Kay\u0131p Fonksiyonlar\u0131 ve Optimizasyon Algoritmalar\u0131<\/h4>\n

E\u011fitim s\u00fcrecinde modelin ne kadar iyi performans g\u00f6sterdi\u011fini \u00f6l\u00e7mek i\u00e7in kay\u0131p fonksiyonlar\u0131 (loss functions) kullan\u0131l\u0131r (\u00f6rne\u011fin \u00e7apraz entropi). Modelin a\u011f\u0131rl\u0131klar\u0131, bu kay\u0131p fonksiyonunu minimize etmek i\u00e7in optimizasyon algoritmalar\u0131 (\u00f6rne\u011fin Adam, SGD) kullan\u0131larak g\u00fcncellenir. Bu algoritmalar, modelin tahminleri ile ger\u00e7ek de\u011ferler aras\u0131ndaki fark\u0131 azaltarak \u00f6\u011frenmeyi sa\u011flar.<\/p>\n

Transformer’lar\u0131n S\u0131n\u0131rlamalar\u0131 ve Gelecek<\/h3>\n

Transformer’lar NLP alan\u0131nda b\u00fcy\u00fck bir devrim yaratm\u0131\u015f olsa da, hala baz\u0131 s\u0131n\u0131rlamalara sahiptir ve ara\u015ft\u0131rma toplulu\u011fu bu s\u0131n\u0131rlamalar\u0131 a\u015fmak i\u00e7in s\u00fcrekli \u00e7al\u0131\u015fmaktad\u0131r.<\/p>\n

Hesaplama Maliyeti ve Uzun Diziler<\/h4>\n

Transformer’lar\u0131n dikkat mekanizmas\u0131, giri\u015f dizisinin uzunlu\u011funa g\u00f6re karesel bir hesaplama maliyetine sahiptir. Bu, \u00e7ok uzun metinlerle \u00e7al\u0131\u015f\u0131rken (\u00f6rne\u011fin binlerce token) hesaplama g\u00fcc\u00fc ve bellek a\u00e7\u0131s\u0131ndan ciddi k\u0131s\u0131tlamalar getirir. Bu durum, modellerin “ba\u011flam penceresi”ni s\u0131n\u0131rlar.<\/p>\n

“Hal\u00fcsinasyon” ve Do\u011fruluk Sorunlar\u0131<\/h4>\n

B\u00fcy\u00fck dil modelleri, bazen ger\u00e7ek d\u0131\u015f\u0131 veya uydurma bilgiler \u00fcretebilir, bu duruma “hal\u00fcsinasyon” denir. Modeller, verilerdeki kal\u0131plar\u0131 \u00f6\u011frenerek ak\u0131c\u0131 ve ikna edici metinler \u00fcretse de, her zaman olgusal do\u011frulu\u011fu garanti edemezler. Bu, \u00f6zellikle hassas uygulamalarda \u00f6nemli bir sorundur.<\/p>\n

Yeni Nesil Modeller ve Ara\u015ft\u0131rma Alanlar\u0131<\/h4>\n

Mevcut s\u0131n\u0131rlamalar\u0131 a\u015fmak i\u00e7in \u00e7e\u015fitli ara\u015ft\u0131rma alanlar\u0131 bulunmaktad\u0131r. Bunlar aras\u0131nda daha verimli dikkat mekanizmalar\u0131 (sparse attention, linear attention), Mixture-of-Experts (MoE) gibi daha b\u00fcy\u00fck ama seyrek aktif modeller, uzun ba\u011flam pencerelerini y\u00f6netmek i\u00e7in yeni mimariler ve modellerin “ak\u0131l y\u00fcr\u00fctme” yeteneklerini geli\u015ftirmeye y\u00f6nelik \u00e7al\u0131\u015fmalar yer almaktad\u0131r. Gelecekte, Transformer’lar\u0131n daha da yetenekli ve g\u00fcvenilir hale gelmesi beklenmektedir.<\/p>\n

Basit Bir Self-Attention Mekanizmas\u0131 (Konseptsel)<\/h3>\n

A\u015fa\u011f\u0131daki pseudo-kod, self-attention’\u0131n temel mant\u0131\u011f\u0131n\u0131 basitle\u015ftirilmi\u015f bir \u015fekilde g\u00f6stermektedir:<\/p>\n

\ndef calculate_self_attention(queries, keys, values):\n    # Skorlar\u0131 hesapla: Query ile Key'lerin benzerli\u011fi\n    scores = matris_carpimi(queries, transpose(keys))\n\n    # Skorlar\u0131 \u00f6l\u00e7ekle (genellikle karek\u00f6k(key_boyutu) ile)\n    scaled_scores = scores \/ sqrt(key_boyutu)\n\n    # Olas\u0131l\u0131k da\u011f\u0131l\u0131m\u0131 i\u00e7in Softmax uygula\n    attention_weights = softmax(scaled_scores)\n\n    # A\u011f\u0131rl\u0131klarla Value'lar\u0131 \u00e7arp\n    output = matris_carpimi(attention_weights, values)\n    return output\n<\/pre>\n
<\/code><\/p>\n
Bu \u00f6rnek, her bir kelimenin (Query) di\u011fer t\u00fcm kelimelerle (Key) ne kadar ili\u015fkili oldu\u011funu hesaplay\u0131p, bu ili\u015fki oranlar\u0131na g\u00f6re kelimelerin anlamlar\u0131n\u0131 (Value) bir araya getirerek yeni bir temsil olu\u015fturdu\u011funu g\u00f6stermektedir.<\/p>\n
Sonu\u00e7<\/h3>\n
Transformer mimarisi, yapay zeka ve do\u011fal dil i\u015fleme alan\u0131nda e\u015fi benzeri g\u00f6r\u00fclmemi\u015f bir d\u00f6n\u00fc\u015f\u00fcm yaratt\u0131. Dikkat mekanizmas\u0131 sayesinde uzun mesafeli ba\u011f\u0131ml\u0131l\u0131klar\u0131 yakalama, paralel i\u015fleme yetene\u011fiyle \u00f6l\u00e7eklenebilirlik ve mod\u00fcler yap\u0131s\u0131yla esneklik sunan bu modeller, dilin karma\u015f\u0131k yap\u0131s\u0131n\u0131 anlamada ve \u00fcretmede \u00e7\u0131\u011f\u0131r a\u00e7t\u0131. Her ne kadar hesaplama maliyeti ve hal\u00fcsinasyon gibi zorluklarla kar\u015f\u0131la\u015f\u0131lsa da, s\u00fcrekli devam eden ara\u015ft\u0131rmalar ve yeni nesil mimariler, Transformer'lar\u0131n gelecekteki potansiyelini daha da geni\u015fletmeye devam edecektir. Bu \"yapay zeka beyinleri\", insan-bilgisayar etkile\u015fimini yeniden \u015fekillendirmeye ve dilin s\u0131n\u0131rlar\u0131n\u0131 zorlamaya devam edecek gibi g\u00f6r\u00fcn\u00fcyor.<\/p>\n
SSS (S\u0131k Sorulan Sorular)<\/h3>\n
Transformer'lar neden bu kadar etkili?<\/h4>\n
Transformer'lar, \u00f6zellikle dikkat mekanizmas\u0131 sayesinde c\u00fcmledeki t\u00fcm kelimeler aras\u0131ndaki ili\u015fkileri ayn\u0131 anda de\u011ferlendirebilir ve paralel i\u015flemeye olanak tan\u0131d\u0131\u011f\u0131 i\u00e7in b\u00fcy\u00fck veri k\u00fcmeleri \u00fczerinde h\u0131zl\u0131 ve verimli bir \u015fekilde e\u011fitilebilirler. Bu yetenekler, onlar\u0131 uzun mesafeli ba\u011f\u0131ml\u0131l\u0131klar\u0131 yakalamada ve dilin karma\u015f\u0131k yap\u0131lar\u0131n\u0131 \u00f6\u011frenmede olduk\u00e7a ba\u015far\u0131l\u0131 k\u0131lar.<\/p>\n
Self-Attention mekanizmas\u0131 ne i\u015fe yarar?<\/h4>\n
Self-Attention (kendi kendine dikkat), bir kelimenin anlam\u0131n\u0131 \u00e7\u0131kar\u0131rken ayn\u0131 c\u00fcmledeki di\u011fer kelimelerle olan ili\u015fkisini dinamik olarak a\u011f\u0131rl\u0131kland\u0131rmas\u0131n\u0131 sa\u011flar. Bu sayede model, ba\u011flama duyarl\u0131 temsiller olu\u015fturabilir ve bir kelimenin anlam\u0131n\u0131n c\u00fcmledeki di\u011fer kelimelere g\u00f6re nas\u0131l de\u011fi\u015fti\u011fini anlayabilir.<\/p>\n
Pozisyonel kodlama neden gereklidir?<\/h4>\n
Dikkat mekanizmas\u0131 kelimelerin s\u0131ras\u0131na duyarl\u0131 de\u011fildir. Pozisyonel kodlama, her bir kelimenin c\u00fcmledeki konum bilgisini modelin girdisine ekleyerek, kelime s\u0131ras\u0131n\u0131n ve dolay\u0131s\u0131yla dilbilgisinin ve anlamsal yap\u0131n\u0131n korunmas\u0131n\u0131 sa\u011flar.<\/p>\n
Encoder-Decoder ve yaln\u0131zca Decoder modelleri aras\u0131ndaki fark nedir?<\/h4>\n
Encoder-Decoder modelleri, bir giri\u015fi kodlay\u0131p (anlamsal temsil olu\u015fturup) sonra bu temsilden bir \u00e7\u0131kt\u0131 \u00fcretirken (\u00f6rne\u011fin \u00e7eviri), yaln\u0131zca Decoder modelleri genellikle metin \u00fcretimi gibi g\u00f6revler i\u00e7in kullan\u0131l\u0131r ve bir sonraki kelimeyi tahmin etmek i\u00e7in yaln\u0131zca \u00f6nceki kelimelere bakar. GPT serisi gibi b\u00fcy\u00fck dil modelleri genellikle yaln\u0131zca Decoder mimarisini kullan\u0131r.<\/p>\n","protected":false},"excerpt":{"rendered":"A\u015fa\u011f\u0131daki pseudo-kod, self-attention’\u0131n temel mant\u0131\u011f\u0131n\u0131 basitle\u015ftirilmi\u015f bir \u015fekilde g\u00f6stermektedir:","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"csco_page_header_type":"","csco_page_load_nextpost":"","csco_page_subscribe_form":"","csco_page_contact_form":"","footnotes":""},"categories":[1],"tags":[],"class_list":{"0":"post-36662","1":"post","2":"type-post","3":"status-publish","4":"format-standard","6":"category-genel","7":"cs-entry","8":"cs-video-wrap"},"yoast_head":"\nTransformer'lar\u0131n Ger\u00e7ek D\u00fc\u015f\u00fcnce Mekanizmas\u0131: Bir Yapay Zeka Dil Modelinin Beynine Yolculuk - Kodlar\u0131n Gizemli D\u00fcnyas\u0131<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/\" \/>\n<meta property=\"og:locale\" content=\"tr_TR\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Transformer'lar\u0131n Ger\u00e7ek D\u00fc\u015f\u00fcnce Mekanizmas\u0131: Bir Yapay Zeka Dil Modelinin Beynine Yolculuk\" \/>\n<meta property=\"og:description\" content=\"A\u015fa\u011f\u0131daki pseudo-kod, self-attention'\u0131n temel mant\u0131\u011f\u0131n\u0131 basitle\u015ftirilmi\u015f bir \u015fekilde g\u00f6stermektedir:\" \/>\n<meta property=\"og:url\" content=\"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/\" \/>\n<meta property=\"og:site_name\" content=\"Kodlar\u0131n Gizemli D\u00fcnyas\u0131\" \/>\n<meta property=\"article:published_time\" content=\"2025-12-21T06:00:39+00:00\" \/>\n<meta name=\"author\" content=\"Fatih Soysal\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"Yazan:\" \/>\n\t<meta name=\"twitter:data1\" content=\"Fatih Soysal\" \/>\n\t<meta name=\"twitter:label2\" content=\"Tahmini okuma s\u00fcresi\" \/>\n\t<meta name=\"twitter:data2\" content=\"12 dakika\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/#article\",\"isPartOf\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/\"},\"author\":{\"name\":\"Fatih Soysal\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1\"},\"headline\":\"Transformer’lar\u0131n Ger\u00e7ek D\u00fc\u015f\u00fcnce Mekanizmas\u0131: Bir Yapay Zeka Dil Modelinin Beynine Yolculuk\",\"datePublished\":\"2025-12-21T06:00:39+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/\"},\"wordCount\":2376,\"commentCount\":0,\"publisher\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1\"},\"inLanguage\":\"tr\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/#respond\"]}],\"copyrightYear\":\"2025\",\"copyrightHolder\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#organization\"}},{\"@type\":\"WebPage\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/\",\"url\":\"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/\",\"name\":\"Transformer'lar\u0131n Ger\u00e7ek D\u00fc\u015f\u00fcnce Mekanizmas\u0131: Bir Yapay Zeka Dil Modelinin Beynine Yolculuk - Kodlar\u0131n Gizemli D\u00fcnyas\u0131\",\"isPartOf\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#website\"},\"datePublished\":\"2025-12-21T06:00:39+00:00\",\"breadcrumb\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/#breadcrumb\"},\"inLanguage\":\"tr\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Anasayfa\",\"item\":\"https:\/\/fatihsoysal.com\/blog\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Transformer’lar\u0131n Ger\u00e7ek D\u00fc\u015f\u00fcnce Mekanizmas\u0131: Bir Yapay Zeka Dil Modelinin Beynine Yolculuk\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#website\",\"url\":\"https:\/\/fatihsoysal.com\/blog\/\",\"name\":\"Fatihsoysal.com\",\"description\":\"Blog - Yaz\u0131l\u0131m D\u00fcnyas\u0131 Tecr\u00fcbelerim\",\"publisher\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/fatihsoysal.com\/blog\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"tr\"},{\"@type\":[\"Person\",\"Organization\"],\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1\",\"name\":\"Fatih Soysal\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"tr\",\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/fatihsoysal.com\/blog\/wp-content\/uploads\/2024\/04\/cropped-replicate-prediction-3kgg1hgjn5rgp0cf0p5tr0jw7w-1.png\",\"contentUrl\":\"https:\/\/fatihsoysal.com\/blog\/wp-content\/uploads\/2024\/04\/cropped-replicate-prediction-3kgg1hgjn5rgp0cf0p5tr0jw7w-1.png\",\"width\":512,\"height\":512,\"caption\":\"Fatih Soysal\"},\"logo\":{\"@id\":\"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/image\/\"},\"description\":\"Kullan\u0131m ve kodlama m\u00fckemmeliyetini odak alan uygulamalar olu\u015fturma deneyimine sahip, profesyonel olarak 15+ y\u0131l \u00fczeri deneyime sahip bir yaz\u0131l\u0131m m\u00fchendisi.\",\"url\":\"https:\/\/fatihsoysal.com\/blog\/author\/fatihsoysal\/\"}]}<\/script>\n<!-- \/ Yoast SEO Premium plugin. -->","yoast_head_json":{"title":"Transformer'lar\u0131n Ger\u00e7ek D\u00fc\u015f\u00fcnce Mekanizmas\u0131: Bir Yapay Zeka Dil Modelinin Beynine Yolculuk - Kodlar\u0131n Gizemli D\u00fcnyas\u0131","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/","og_locale":"tr_TR","og_type":"article","og_title":"Transformer'lar\u0131n Ger\u00e7ek D\u00fc\u015f\u00fcnce Mekanizmas\u0131: Bir Yapay Zeka Dil Modelinin Beynine Yolculuk","og_description":"A\u015fa\u011f\u0131daki pseudo-kod, self-attention'\u0131n temel mant\u0131\u011f\u0131n\u0131 basitle\u015ftirilmi\u015f bir \u015fekilde g\u00f6stermektedir:","og_url":"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/","og_site_name":"Kodlar\u0131n Gizemli D\u00fcnyas\u0131","article_published_time":"2025-12-21T06:00:39+00:00","author":"Fatih Soysal","twitter_card":"summary_large_image","twitter_misc":{"Yazan:":"Fatih Soysal","Tahmini okuma s\u00fcresi":"12 dakika"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/#article","isPartOf":{"@id":"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/"},"author":{"name":"Fatih Soysal","@id":"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1"},"headline":"Transformer’lar\u0131n Ger\u00e7ek D\u00fc\u015f\u00fcnce Mekanizmas\u0131: Bir Yapay Zeka Dil Modelinin Beynine Yolculuk","datePublished":"2025-12-21T06:00:39+00:00","mainEntityOfPage":{"@id":"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/"},"wordCount":2376,"commentCount":0,"publisher":{"@id":"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1"},"inLanguage":"tr","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/#respond"]}],"copyrightYear":"2025","copyrightHolder":{"@id":"https:\/\/fatihsoysal.com\/blog\/#organization"}},{"@type":"WebPage","@id":"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/","url":"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/","name":"Transformer'lar\u0131n Ger\u00e7ek D\u00fc\u015f\u00fcnce Mekanizmas\u0131: Bir Yapay Zeka Dil Modelinin Beynine Yolculuk - Kodlar\u0131n Gizemli D\u00fcnyas\u0131","isPartOf":{"@id":"https:\/\/fatihsoysal.com\/blog\/#website"},"datePublished":"2025-12-21T06:00:39+00:00","breadcrumb":{"@id":"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/#breadcrumb"},"inLanguage":"tr","potentialAction":[{"@type":"ReadAction","target":["https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/fatihsoysal.com\/blog\/transformerlarin-gercek-dusunce-mekanizmasi-bir-yapay-zeka-dil-modelinin-beynine-yolculuk\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Anasayfa","item":"https:\/\/fatihsoysal.com\/blog\/"},{"@type":"ListItem","position":2,"name":"Transformer’lar\u0131n Ger\u00e7ek D\u00fc\u015f\u00fcnce Mekanizmas\u0131: Bir Yapay Zeka Dil Modelinin Beynine Yolculuk"}]},{"@type":"WebSite","@id":"https:\/\/fatihsoysal.com\/blog\/#website","url":"https:\/\/fatihsoysal.com\/blog\/","name":"Fatihsoysal.com","description":"Blog - Yaz\u0131l\u0131m D\u00fcnyas\u0131 Tecr\u00fcbelerim","publisher":{"@id":"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/fatihsoysal.com\/blog\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"tr"},{"@type":["Person","Organization"],"@id":"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/002a254750921dcfd568a99e48240dd1","name":"Fatih Soysal","image":{"@type":"ImageObject","inLanguage":"tr","@id":"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/image\/","url":"https:\/\/fatihsoysal.com\/blog\/wp-content\/uploads\/2024\/04\/cropped-replicate-prediction-3kgg1hgjn5rgp0cf0p5tr0jw7w-1.png","contentUrl":"https:\/\/fatihsoysal.com\/blog\/wp-content\/uploads\/2024\/04\/cropped-replicate-prediction-3kgg1hgjn5rgp0cf0p5tr0jw7w-1.png","width":512,"height":512,"caption":"Fatih Soysal"},"logo":{"@id":"https:\/\/fatihsoysal.com\/blog\/#\/schema\/person\/image\/"},"description":"Kullan\u0131m ve kodlama m\u00fckemmeliyetini odak alan uygulamalar olu\u015fturma deneyimine sahip, profesyonel olarak 15+ y\u0131l \u00fczeri deneyime sahip bir yaz\u0131l\u0131m m\u00fchendisi.","url":"https:\/\/fatihsoysal.com\/blog\/author\/fatihsoysal\/"}]}},"yoast_meta":{"yoast_wpseo_title":"","yoast_wpseo_metadesc":"","yoast_wpseo_canonical":""},"amp_enabled":true,"_links":{"self":[{"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/posts\/36662","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/comments?post=36662"}],"version-history":[{"count":0,"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/posts\/36662\/revisions"}],"wp:attachment":[{"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/media?parent=36662"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/categories?post=36662"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/fatihsoysal.com\/blog\/wp-json\/wp\/v2\/tags?post=36662"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}