{"id":12058,"date":"2024-02-23T10:00:31","date_gmt":"2024-02-23T07:00:31","guid":{"rendered":"https:\/\/www.kaspersky.com.tr\/blog\/?p=12058"},"modified":"2024-02-23T01:55:01","modified_gmt":"2024-02-22T22:55:01","slug":"how-to-use-ai-locally-and-securely","status":"publish","type":"post","link":"https:\/\/www.kaspersky.com.tr\/blog\/how-to-use-ai-locally-and-securely\/12058\/","title":{"rendered":"Bilgisayar\u0131n\u0131za bir yapay zeka asistan\u0131 nas\u0131l y\u00fcklenir ve kullan\u0131l\u0131r"},"content":{"rendered":"

Pek \u00e7ok ki\u015fi halihaz\u0131rda jeneratif n\u00f6ral a\u011flar\u0131 denemeye ba\u015flad\u0131 ve i\u015f ama\u00e7l\u0131 kullan\u0131mlar da d\u00e2hil olmak \u00fczere bu a\u011flar i\u00e7in d\u00fczenli kullan\u0131m alanlar\u0131 ortaya \u00e7\u0131kmakta. \u00d6rne\u011fin, ChatGPT ve benzerleri Amerikal\u0131lar\u0131n neredeyse %60\u2019\u0131<\/a> taraf\u0131ndan d\u00fczenli olarak kullan\u0131l\u0131yor (ve bu her zaman y\u00f6netimden izin al\u0131narak yap\u0131lm\u0131yor). Bununla birlikte, bu t\u00fcr i\u015flemlerde yer alan t\u00fcm veriler \u2013 hem kullan\u0131c\u0131 istemleri hem de model yan\u0131tlar\u0131 \u2013 OpenAI, Google ve di\u011fer \u015firketlerin sunucular\u0131nda saklan\u0131r. Bu t\u00fcr bilgi s\u0131z\u0131nt\u0131lar\u0131n\u0131n kabul edilemez oldu\u011fu g\u00f6revler i\u00e7in yapay zekay\u0131 tamamen terk etmeniz gerekmez \u2013 sadece n\u00f6ral a\u011f\u0131 yerel olarak kendi bilgisayar\u0131n\u0131zda (hatta bir diz\u00fcst\u00fc bilgisayarda) \u00e7al\u0131\u015ft\u0131rmak i\u00e7in biraz \u00e7aba (ve belki de para) harcaman\u0131z yeterlidir.<\/p>\n

Bulut tehditleri<\/h2>\n

En pop\u00fcler yapay zeka asistanlar\u0131 b\u00fcy\u00fck \u015firketlerin bulut altyap\u0131s\u0131 \u00fczerinde \u00e7al\u0131\u015f\u0131yor. Bu y\u00f6ntem verimli ve h\u0131zl\u0131d\u0131r, ancak model taraf\u0131ndan i\u015flenen verilerinize hem yapay zeka hizmet sa\u011flay\u0131c\u0131s\u0131 hem de ge\u00e7en y\u0131l ChatGPT\u2019de oldu\u011fu gibi<\/a> tamamen ilgisiz taraflar eri\u015febilir.<\/p>\n

Bu t\u00fcr olaylar, bu yapay zeka asistanlar\u0131n\u0131n ne i\u00e7in kullan\u0131ld\u0131\u011f\u0131na ba\u011fl\u0131 olarak de\u011fi\u015fen d\u00fczeylerde tehdit olu\u015fturuyor. Yazd\u0131\u011f\u0131n\u0131z peri masallar\u0131 i\u00e7in sevimli ill\u00fcstrasyonlar \u00fcretiyorsan\u0131z veya ChatGPT\u2019den yakla\u015fan hafta sonu \u015fehir tatiliniz i\u00e7in bir g\u00fczergah olu\u015fturmas\u0131n\u0131 istiyorsan\u0131z, bir s\u0131z\u0131nt\u0131n\u0131n ciddi hasara yol a\u00e7mas\u0131 pek olas\u0131 de\u011fildir. Ancak bir chatbot ile yapt\u0131\u011f\u0131n\u0131z g\u00f6r\u00fc\u015fme gizli bilgiler (ki\u015fisel veriler, parolalar veya banka kart\u0131 numaralar\u0131) i\u00e7eriyorsa, buluta olas\u0131 bir s\u0131z\u0131nt\u0131 kabul edilmez olacakt\u0131r.. Neyse ki, verileri \u00f6nceden filtreleyerek bunu \u00f6nlemek nispeten kolayd\u0131r \u2013 bu konuda ayr\u0131 bir yaz\u0131m\u0131z var<\/a>.<\/p>\n

Ancak, t\u00fcm yaz\u0131\u015fmalar\u0131n gizli oldu\u011fu (\u00f6rne\u011fin t\u0131bbi veya finansal bilgiler) veya \u00f6n filtrelemenin g\u00fcvenilirli\u011finin \u015f\u00fcpheli oldu\u011fu durumlarda (kimsenin \u00f6nizleme ve filtreleme yapmayaca\u011f\u0131 b\u00fcy\u00fck hacimli verileri i\u015flemeniz gerekti\u011finde), tek bir \u00e7\u00f6z\u00fcm vard\u0131r: i\u015flemeyi buluttan yerel bir bilgisayara ta\u015f\u0131mak. Elbette ChatGPT veya Midjourney\u2019in kendi versiyonunu \u00e7evrimd\u0131\u015f\u0131 \u00e7al\u0131\u015ft\u0131rmak pek m\u00fcmk\u00fcn de\u011fildir, ancak yerel olarak \u00e7al\u0131\u015fan di\u011fer n\u00f6ral a\u011flar daha az bilgisayar y\u00fck\u00fc ile k\u0131yaslanabilir bir kalite sa\u011flar.<\/p>\n

Bir n\u00f6ral a\u011f\u0131 \u00e7al\u0131\u015ft\u0131rmak i\u00e7in hangi donan\u0131ma ihtiyac\u0131n\u0131z vard\u0131r?<\/h2>\n

Sinir a\u011flar\u0131yla \u00e7al\u0131\u015fman\u0131n s\u00fcper g\u00fc\u00e7l\u00fc grafik kartlar\u0131 gerektirdi\u011fini muhtemelen duymu\u015fsunuzdur, ancak pratikte durum her zaman b\u00f6yle de\u011fildir. Farkl\u0131 yapay zeka modelleri, \u00f6zelliklerine ba\u011fl\u0131 olarak, RAM, video belle\u011fi, s\u00fcr\u00fcc\u00fc ve i\u015flemci gibi bilgisayar bile\u015fenlerini zorlayabilir (burada sadece i\u015flem h\u0131z\u0131 de\u011fil, ayn\u0131 zamanda i\u015flemcinin belirli vekt\u00f6r talimatlar\u0131n\u0131 desteklemesi de \u00f6nemlidir). Modeli y\u00fckleme yetene\u011fi RAM miktar\u0131na ba\u011fl\u0131d\u0131r ve \u201cba\u011flam penceresinin\u201d boyutu \u2013 yani \u00f6nceki konu\u015fman\u0131n belle\u011fi \u2013 video belle\u011fi miktar\u0131na ba\u011fl\u0131d\u0131r. Tipik olarak, zay\u0131f bir grafik kart\u0131 ve i\u015flemci ile \u00fcretim salyangoz h\u0131z\u0131nda ger\u00e7ekle\u015fir (metin modelleri i\u00e7in saniyede bir ila iki kelime), bu nedenle bu kadar minimal bir kuruluma sahip bir bilgisayar yaln\u0131zca belirli bir modelle tan\u0131\u015fmak ve uygunlu\u011funu basit\u00e7e de\u011ferlendirmek i\u00e7in uygundur. Tam te\u015fekk\u00fcll\u00fc g\u00fcnl\u00fck kullan\u0131m i\u00e7in RAM\u2019i art\u0131rman\u0131z, grafik kart\u0131n\u0131 y\u00fckseltmeniz veya daha h\u0131zl\u0131 bir yapay zeka modeli se\u00e7meniz gerekecektir.<\/p>\n

Ba\u015flang\u0131\u00e7 noktas\u0131 olarak, 2017\u2019de nispeten g\u00fc\u00e7l\u00fc kabul edilen bilgisayarlarla \u00e7al\u0131\u015fmay\u0131 deneyebilirsiniz: AVX2 talimatlar\u0131n\u0131 destekleyen Core i7\u2019den d\u00fc\u015f\u00fck olmayan i\u015flemciler, 16 GB RAM ve en az 4 GB belle\u011fe sahip grafik kartlar\u0131. Mac merakl\u0131lar\u0131 i\u00e7in Apple M1 \u00e7ipi ve \u00fczerinde \u00e7al\u0131\u015fan modeller yeterli olacakt\u0131r, bellek gereksinimleri ise ayn\u0131d\u0131r.<\/p>\n

Bir yapay zeka modeli se\u00e7erken, \u00f6ncelikle sistem gereksinimlerini \u00f6\u011frenmelisiniz. \u201cModel_ad\u0131<\/em> gereksinimleri\u201d gibi bir arama sorgusu, mevcut donan\u0131m\u0131n\u0131z g\u00f6z \u00f6n\u00fcne al\u0131nd\u0131\u011f\u0131nda bu modeli indirmeye de\u011fip de\u011fmeyece\u011fini de\u011ferlendirmenize yard\u0131mc\u0131 olacakt\u0131r. Bellek boyutu, CPU ve GPU\u2019nun farkl\u0131 modellerin performans\u0131 \u00fczerindeki etkisine ili\u015fkin ayr\u0131nt\u0131l\u0131 \u00e7al\u0131\u015fmalar mevcuttur; \u00f6rne\u011fin, buradaki \u00e7al\u0131\u015fma<\/a>.<\/p>\n

G\u00fc\u00e7l\u00fc donan\u0131mlara eri\u015fimi olmayanlar i\u00e7in iyi haber: Eski donan\u0131mlarda bile basit g\u00f6revleri yerine getirebilen basitle\u015ftirilmi\u015f yapay zeka modelleri vard\u0131r. Grafik kart\u0131n\u0131z \u00e7ok basit ve zay\u0131f olsa bile, yaln\u0131zca i\u015flemciyi kullanarak modelleri \u00e7al\u0131\u015ft\u0131rmak ve ortamlar\u0131 ba\u015flatmak m\u00fcmk\u00fcnd\u00fcr. Ger\u00e7ekle\u015ftirece\u011finiz i\u015flere ba\u011fl\u0131 olarak, bunlar kabul edilebilir derecede iyi bile \u00e7al\u0131\u015fabilir.<\/p>\n

\"GPU<\/a>

\u00c7e\u015fitli bilgisayar yap\u0131lar\u0131n\u0131n pop\u00fcler dil modelleriyle nas\u0131l \u00e7al\u0131\u015ft\u0131\u011f\u0131na dair \u00f6rnekler<\/p><\/div>\n

<\/p>\n

Bir yapay zeka modeli se\u00e7mek ve kuantizasyonun b\u00fcy\u00fcs\u00fc<\/h2>\n

G\u00fcn\u00fcm\u00fczde \u00e7ok say\u0131da dil modeli mevcuttur, ancak bunlar\u0131n bir\u00e7o\u011funun pratik uygulamalar\u0131 s\u0131n\u0131rl\u0131d\u0131r. Bununla birlikte, metin olu\u015fturmak (\u00f6rne\u011fin Mistral 7B) veya kod par\u00e7ac\u0131klar\u0131 olu\u015fturmak (\u00f6rne\u011fin Code Llama 13B) gibi belirli g\u00f6revler i\u00e7in \u00e7ok uygun olan kullan\u0131m\u0131 kolay ve halka a\u00e7\u0131k yapay zeka ara\u00e7lar\u0131 vard\u0131r. Bu nedenle, bir model se\u00e7erken, se\u00e7imi birka\u00e7 uygun adaya indirgeyin ve ard\u0131ndan bilgisayar\u0131n\u0131z\u0131n bunlar\u0131 \u00e7al\u0131\u015ft\u0131rmak i\u00e7in gerekli kaynaklara sahip oldu\u011fundan emin olun.<\/p>\n

Herhangi bir n\u00f6ral a\u011fda, bellek zorlanmas\u0131n\u0131n \u00e7o\u011fu, a\u011fdaki her bir n\u00f6ronun \u00e7al\u0131\u015fmas\u0131n\u0131 tan\u0131mlayan say\u0131sal katsay\u0131lar olan a\u011f\u0131rl\u0131klardan kaynaklan\u0131r. Ba\u015flang\u0131\u00e7ta, modeli e\u011fitirken, a\u011f\u0131rl\u0131klar hesaplan\u0131r ve y\u00fcksek hassasiyetli kesirli say\u0131lar olarak saklan\u0131r. Ancak, e\u011fitilen modeldeki a\u011f\u0131rl\u0131klar\u0131n yuvarlanmas\u0131n\u0131n, yapay zeka arac\u0131n\u0131n normal bilgisayarlarda \u00e7al\u0131\u015ft\u0131r\u0131lmas\u0131na izin verirken performans\u0131 sadece biraz d\u00fc\u015f\u00fcrd\u00fc\u011f\u00fc ortaya \u00e7\u0131kt\u0131. Bu yuvarlama i\u015flemine kuantizasyon denir ve bunun yard\u0131m\u0131yla modelin boyutu \u00f6nemli \u00f6l\u00e7\u00fcde azalt\u0131labilir \u2013 16 bit yerine, her a\u011f\u0131rl\u0131k sekiz, d\u00f6rt veya hatta iki bit kullanabilir.<\/p>\n

G\u00fcncel ara\u015ft\u0131rmalara<\/a> g\u00f6re, daha fazla parametre ve kuantizasyona sahip daha b\u00fcy\u00fck bir model, bazen hassas a\u011f\u0131rl\u0131k depolamal\u0131 ancak daha az parametreli bir modelden daha iyi sonu\u00e7lar verebiliyor.<\/p>\n

Bu bilgilerle donanm\u0131\u015f olarak, art\u0131k a\u00e7\u0131k kaynakl\u0131 dil modellerinin hazinesini, yani en iyi A\u00e7\u0131k B\u00fcy\u00fck Dil Modellerini<\/a> ke\u015ffetmeye haz\u0131rs\u0131n\u0131z. Bu listede, yapay zeka ara\u00e7lar\u0131 \u00e7e\u015fitli \u00fcretim kalitesi \u00f6l\u00e7\u00fctlerine g\u00f6re s\u0131ralan\u0131r ve filtreler \u00e7ok b\u00fcy\u00fck, \u00e7ok k\u00fc\u00e7\u00fck veya \u00e7ok do\u011fru olan modelleri d\u0131\u015far\u0131da b\u0131rakmay\u0131 kolayla\u015ft\u0131r\u0131r.<\/p>\n

\"Filtre<\/a>

Filtre k\u00fcmesine g\u00f6re s\u0131ralanm\u0131\u015f dil modelleri listesi<\/p><\/div>\n

Model a\u00e7\u0131klamas\u0131n\u0131 okuduktan ve ihtiya\u00e7lar\u0131n\u0131za uygun oldu\u011fundan emin olduktan sonra, Hugging Face<\/a> veya Google Colab<\/a> hizmetlerini kullanarak buluttaki performans\u0131n\u0131 test edin. Bu \u015fekilde, tatmin edici olmayan sonu\u00e7lar \u00fcreten modelleri indirmekten ka\u00e7\u0131nabilir ve zamandan tasarruf edebilirsiniz. Modelin ilk testinden memnun kald\u0131\u011f\u0131n\u0131zda, s\u0131ra yerel olarak nas\u0131l \u00e7al\u0131\u015ft\u0131\u011f\u0131n\u0131 g\u00f6rmeye gelir!<\/p>\n

Gerekli yaz\u0131l\u0131m<\/h2>\n

A\u00e7\u0131k kaynakl\u0131 modellerin \u00e7o\u011fu Hugging Face<\/a> \u00fczerinde yay\u0131nlanmaktad\u0131r, ancak bunlar\u0131 bilgisayar\u0131n\u0131za indirmek yeterli de\u011fildir. Bunlar\u0131 \u00e7al\u0131\u015ft\u0131rmak i\u00e7in LLaMA.cpp<\/a> gibi \u00f6zel bir yaz\u0131l\u0131m veya daha da kolay\u0131 \u201cfonksiyon \u00e7evirici\u201d LM Studio<\/a>\u2018yu y\u00fcklemeniz gerekir. LM Studio, istedi\u011finiz modeli do\u011frudan uygulamadan se\u00e7menize, indirmenize ve bir ileti\u015fim kutusunda \u00e7al\u0131\u015ft\u0131rman\u0131za olanak tan\u0131r.<\/p>\n

Bir sohbet robotunu yerel olarak kullanman\u0131n bir ba\u015fka \u201ckullan\u0131ma haz\u0131r\u201d yolu da GPT4All<\/a> olarak kar\u015f\u0131m\u0131za \u00e7\u0131kar. Burada se\u00e7im yakla\u015f\u0131k bir d\u00fczine dil modeliyle s\u0131n\u0131rl\u0131d\u0131r, ancak bunlar\u0131n \u00e7o\u011fu yaln\u0131zca 8 GB belle\u011fe ve temel bir grafik kart\u0131na sahip bir bilgisayarda bile \u00e7al\u0131\u015facakt\u0131r.<\/p>\n

E\u011fer \u00fcretim \u00e7ok yava\u015fsa, daha kaba kuantizasyonlu (d\u00f6rt yerine iki bit) bir modele ihtiyac\u0131n\u0131z olabilir. Olu\u015fturma kesintiye u\u011frarsa veya y\u00fcr\u00fctme hatalar\u0131 meydana gelirse, sorun genellikle yetersiz bellektir. B\u00f6yle bir durumda daha az parametreye sahip veya yine daha kaba kuantizasyonlu bir model aramaya de\u011fer.<\/p>\n

Hugging Face\u2019teki bir\u00e7ok model zaten \u00e7e\u015fitli hassasiyet derecelerine g\u00f6re kuantize edilmi\u015ftir, ancak kimse istedi\u011finiz modeli istedi\u011finiz hassasiyetle kuantize etmediyse, GPTQ<\/a> kullanarak bunu kendiniz yapabilirsiniz.<\/p>\n

Bu hafta, gelecek vaat eden bir ba\u015fka ara\u00e7 daha halka a\u00e7\u0131k beta s\u00fcr\u00fcm\u00fcnde yay\u0131nland\u0131: NVIDIA\u2019dan Chat With RTX<\/a>. En \u00e7ok aranan yapay zeka \u00e7iplerinin \u00fcreticisi, kullan\u0131c\u0131n\u0131n 16 GB belle\u011fe sahip bir Windows PC\u2019si ve 8 GB veya daha fazla video belle\u011fine sahip bir NVIDIA RTX 30. veya 40. serisi ekran kart\u0131 olmas\u0131 ko\u015fuluyla, YouTube videolar\u0131n\u0131n i\u00e7eri\u011fini \u00f6zetleyebilen, belge k\u00fcmelerini i\u015fleyebilen ve \u00e7ok daha fazlas\u0131n\u0131 yapabilen yerel bir sohbet robotu yay\u0131nlad\u0131. Yak\u0131ndan bakt\u0131\u011f\u0131m\u0131zda Hugging Face<\/a>\u2018in Mistral ve Llama 2 varyasyonlar\u0131 vard\u0131r. Elbette g\u00fc\u00e7l\u00fc grafik kartlar\u0131 jenerasyon performans\u0131n\u0131 art\u0131rabilir, ancak ilk testerlardan gelen geri bildirimlere g\u00f6re<\/a>, \u015fu anki beta olduk\u00e7a hantal (yakla\u015f\u0131k 40 GB) ve kurulumu zor. Ancak NVIDIA\u2019n\u0131n Chat With RTX\u2019i gelecekte \u00e7ok kullan\u0131\u015fl\u0131 bir yerel yapay zeka asistan\u0131 haline gelebilir.<\/p>\n

\"Kuantize<\/a>

Kuantize dil modeli TheBloke\/CodeLlama-7B-Instruct-GGUF taraf\u0131ndan yaz\u0131lan \u201cY\u0131lan\u201d oyununun kodu<\/p><\/div>\n

Yukar\u0131da listelenen uygulamalar t\u00fcm hesaplamalar\u0131 yerel olarak ger\u00e7ekle\u015ftirir, sunuculara veri g\u00f6ndermez ve \u00e7evrimd\u0131\u015f\u0131 \u00e7al\u0131\u015fabilir, b\u00f6ylece gizli bilgileri onlarla g\u00fcvenle payla\u015fabilirsiniz. Ancak, kendinizi s\u0131z\u0131nt\u0131lara kar\u015f\u0131 tam olarak korumak i\u00e7in yaln\u0131zca dil modelinin de\u011fil, bilgisayar\u0131n\u0131z\u0131n da g\u00fcvenli\u011fini sa\u011flaman\u0131z gerekir \u2013 i\u015fte bu noktada kapsaml\u0131 g\u00fcvenlik \u00e7\u00f6z\u00fcm\u00fcm\u00fcz<\/a> devreye girer. Ba\u011f\u0131ms\u0131z testlerde<\/a> do\u011fruland\u0131\u011f\u0131 gibi, Kaspersky Premium<\/a> bilgisayar\u0131n\u0131z\u0131n performans\u0131n\u0131 \u00fczerinde neredeyse hi\u00e7 etkilemez ve bu, yerel AI modelleriyle \u00e7al\u0131\u015f\u0131rken \u00f6nemli bir avantajd\u0131r.<\/p>\n\n","protected":false},"excerpt":{"rendered":"

ChatGPT, Copilot ve Midjourney’in t\u00fcm avantajlar\u0131n\u0131 verilerinizi internete s\u0131zd\u0131rmadan lokal olarak elde edin.<\/p>\n","protected":false},"author":2722,"featured_media":12059,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[1284],"tags":[1425,2607,744,1610,2708,1424],"class_list":{"0":"post-12058","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","7":"category-tips","8":"tag-ai","9":"tag-chatgpt","10":"tag-guvenlik","11":"tag-makine-ogrenimi","12":"tag-sohbet-robotlari","13":"tag-yapay-zeka"},"hreflang":[{"hreflang":"tr","url":"https:\/\/www.kaspersky.com.tr\/blog\/how-to-use-ai-locally-and-securely\/12058\/"},{"hreflang":"en-in","url":"https:\/\/www.kaspersky.co.in\/blog\/how-to-use-ai-locally-and-securely\/27077\/"},{"hreflang":"en-ae","url":"https:\/\/me-en.kaspersky.com\/blog\/how-to-use-ai-locally-and-securely\/22387\/"},{"hreflang":"ar","url":"https:\/\/me.kaspersky.com\/blog\/how-to-use-ai-locally-and-securely\/11436\/"},{"hreflang":"en-us","url":"https:\/\/usa.kaspersky.com\/blog\/how-to-use-ai-locally-and-securely\/29744\/"},{"hreflang":"en-gb","url":"https:\/\/www.kaspersky.co.uk\/blog\/how-to-use-ai-locally-and-securely\/27253\/"},{"hreflang":"es-mx","url":"https:\/\/latam.kaspersky.com\/blog\/how-to-use-ai-locally-and-securely\/27042\/"},{"hreflang":"es","url":"https:\/\/www.kaspersky.es\/blog\/how-to-use-ai-locally-and-securely\/29662\/"},{"hreflang":"it","url":"https:\/\/www.kaspersky.it\/blog\/how-to-use-ai-locally-and-securely\/28540\/"},{"hreflang":"ru","url":"https:\/\/www.kaspersky.ru\/blog\/how-to-use-ai-locally-and-securely\/36986\/"},{"hreflang":"x-default","url":"https:\/\/www.kaspersky.com\/blog\/how-to-use-ai-locally-and-securely\/50576\/"},{"hreflang":"fr","url":"https:\/\/www.kaspersky.fr\/blog\/how-to-use-ai-locally-and-securely\/21543\/"},{"hreflang":"pt-br","url":"https:\/\/www.kaspersky.com.br\/blog\/how-to-use-ai-locally-and-securely\/22254\/"},{"hreflang":"de","url":"https:\/\/www.kaspersky.de\/blog\/how-to-use-ai-locally-and-securely\/30951\/"},{"hreflang":"ja","url":"https:\/\/blog.kaspersky.co.jp\/how-to-use-ai-locally-and-securely\/35896\/"},{"hreflang":"nl","url":"https:\/\/www.kaspersky.nl\/blog\/how-to-use-ai-locally-and-securely\/29029\/"},{"hreflang":"ru-kz","url":"https:\/\/blog.kaspersky.kz\/how-to-use-ai-locally-and-securely\/27452\/"},{"hreflang":"en-au","url":"https:\/\/www.kaspersky.com.au\/blog\/how-to-use-ai-locally-and-securely\/33259\/"},{"hreflang":"en-za","url":"https:\/\/www.kaspersky.co.za\/blog\/how-to-use-ai-locally-and-securely\/32882\/"}],"acf":[],"banners":"","maintag":{"url":"https:\/\/www.kaspersky.com.tr\/blog\/tag\/yapay-zeka\/","name":"yapay zeka"},"_links":{"self":[{"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/posts\/12058","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/users\/2722"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/comments?post=12058"}],"version-history":[{"count":3,"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/posts\/12058\/revisions"}],"predecessor-version":[{"id":12166,"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/posts\/12058\/revisions\/12166"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/media\/12059"}],"wp:attachment":[{"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/media?parent=12058"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/categories?post=12058"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kaspersky.com.tr\/blog\/wp-json\/wp\/v2\/tags?post=12058"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}