{"id":6415,"date":"2024-05-21T13:38:15","date_gmt":"2024-05-21T13:38:15","guid":{"rendered":"https:\/\/www.turkticaret.net\/blog\/?p=6415"},"modified":"2026-09-24T14:00:20","modified_gmt":"2026-09-24T11:00:20","slug":"gpt-4-ve-gpt-3-5-farklari","status":"publish","type":"post","link":"https:\/\/www.turkticaret.net\/blog\/gpt-4-ve-gpt-3-5-farklari\/","title":{"rendered":"GPT 4 Nedir? GPT 3.5 ile GPT 4 Aras\u0131ndaki Farklar"},"content":{"rendered":"

OpenAI’nin 14 Mart 2023’te duyurdu\u011fu GPT-4, GPT-3.5’e g\u00f6re \u00f6zellikle karma\u015f\u0131k talimatlar\u0131 izleme, problem \u00e7\u00f6zme ve g\u00f6rsel i\u00e7erikleri yorumlama alanlar\u0131nda ilerleme sundu. Bu yaz\u0131da iki modeli, ilk yay\u0131mland\u0131klar\u0131 d\u00f6neme ili\u015fkin OpenAI verileri \u00fczerinden kar\u015f\u0131la\u015ft\u0131r\u0131yoruz.<\/p>\n

\"GPT<\/p>\n\n\n\n\n\n\n\n
Girdi t\u00fcr\u00fc<\/td>\nKar\u015f\u0131la\u015ft\u0131r\u0131lan \u00f6zg\u00fcn model metin girdisiyle \u00e7al\u0131\u015f\u0131r.<\/td>\nMetin ve g\u00f6rsel girdilerini i\u015fleyebilir; bu kar\u015f\u0131la\u015ft\u0131rmadaki GPT-4’\u00fcn \u00e7\u0131kt\u0131s\u0131 metindir.<\/td>\n<\/tr>\n
Karma\u015f\u0131k g\u00f6revler<\/td>\n\u00c7ok ad\u0131ml\u0131 ve ayr\u0131nt\u0131l\u0131 g\u00f6revlerde daha fazla zorlanabilir.<\/td>\nOpenAI’nin yay\u0131mlad\u0131\u011f\u0131 de\u011ferlendirmelerde karma\u015f\u0131k g\u00f6revlerde daha g\u00fc\u00e7l\u00fc sonu\u00e7lar vermi\u015ftir.<\/td>\n<\/tr>\n
Ayr\u0131nt\u0131l\u0131 talimatlar<\/td>\nUzun veya \u00e7ok ko\u015fullu y\u00f6nergelerde ayr\u0131nt\u0131lar\u0131 ka\u00e7\u0131rabilir.<\/td>\n\u0130nce ayr\u0131mlar\u0131 ve \u00e7ok ko\u015fullu y\u00f6nergeleri izleme konusunda daha ba\u015far\u0131l\u0131d\u0131r.<\/td>\n<\/tr>\n
Olgusal do\u011fruluk<\/td>\nHatal\u0131 veya uydurma bilgi \u00fcretebilir.<\/td>\nKar\u015f\u0131la\u015ft\u0131rmal\u0131 testlerde daha iyi performans g\u00f6sterse de hatal\u0131 bilgi \u00fcretebilir.<\/td>\n<\/tr>\n
G\u00fcvenlik de\u011ferlendirmeleri<\/td>\nZararl\u0131 isteklere verilen yan\u0131tlar a\u00e7\u0131s\u0131ndan daha zay\u0131f sonu\u00e7lar g\u00f6stermi\u015ftir.<\/td>\nOpenAI’nin o d\u00f6nemki de\u011ferlendirmelerinde uygunsuz isteklere yan\u0131t verme e\u011filimi azalt\u0131lm\u0131\u015ft\u0131r.<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

 <\/p>\n

\"GPT<\/p>\n

1. G\u00f6rsel Girdileri Anlama<\/h2>\n

\u0130ki model aras\u0131ndaki en belirgin farklardan biri, GPT-4’\u00fcn g\u00f6rsel girdilerle \u00e7al\u0131\u015fabilmesidir. OpenAI’nin ilk GPT-4 tan\u0131t\u0131m\u0131nda model, metin ve g\u00f6rselleri girdi olarak kabul edip metin \u00e7\u0131kt\u0131s\u0131 \u00fcreten<\/strong> \u00e7ok modlu bir model olarak tan\u0131mland\u0131. \u00d6rne\u011fin bir ekran g\u00f6r\u00fcnt\u00fcs\u00fc hakk\u0131nda soru soruldu\u011funda i\u00e7eri\u011fini yorumlay\u0131p yaz\u0131l\u0131 yan\u0131t verebilir. GPT-3.5 ile yap\u0131lan bu \u00f6zg\u00fcn model kar\u015f\u0131la\u015ft\u0131rmas\u0131nda ise girdi metindir.<\/p>\n

Buradaki ayr\u0131m \u00f6nemlidir: GPT-4’\u00fcn g\u00f6rseli anlayabilmesi, ilk GPT-4 modelinin g\u00f6rsel \u00e7\u0131kt\u0131 \u00fcretti\u011fi anlam\u0131na gelmez. Ayr\u0131ca g\u00f6rsel girdi \u00f6zelli\u011finin belirli bir \u00fcr\u00fcn veya hizmette kullan\u0131labilir olmas\u0131, o \u00fcr\u00fcn\u00fcn sundu\u011fu \u00f6zelliklere ba\u011fl\u0131d\u0131r.<\/p>\n

2. Karma\u015f\u0131k Problemleri \u00c7\u00f6zme<\/h2>\n

GPT-4, birden fazla ad\u0131m gerektiren sorularda ve \u00e7e\u015fitli akademik de\u011ferlendirmelerde GPT-3.5’ten daha y\u00fcksek performans g\u00f6stermi\u015ftir. OpenAI’nin yay\u0131mlad\u0131\u011f\u0131 \u00f6rneklerden birinde GPT-4, sim\u00fcle edilmi\u015f bir baro s\u0131nav\u0131nda kat\u0131l\u0131mc\u0131lar\u0131n yakla\u015f\u0131k en \u00fcst y\u00fczde 10’luk dilimine denk gelen sonu\u00e7 al\u0131rken GPT-3.5 yakla\u015f\u0131k en alt y\u00fczde 10’luk dilimde kalm\u0131\u015ft\u0131r.<\/p>\n

Bu sonu\u00e7, GPT-4’\u00fcn her hukuk sorusunu do\u011fru \u00e7\u00f6zd\u00fc\u011f\u00fc anlam\u0131na gelmez. S\u0131nav kar\u015f\u0131la\u015ft\u0131rmas\u0131, belirli bir testteki g\u00f6reli performans\u0131 g\u00f6sterir. Her iki model de ger\u00e7ek bir i\u015fte yanl\u0131\u015f \u00e7\u0131kar\u0131m yapabilir; \u00f6zellikle sonucu \u00f6nemli olan g\u00f6revlerde \u00fcretilen bilgi kontrol edilmelidir.<\/p>\n

3. Ayr\u0131nt\u0131l\u0131 Talimatlar\u0131 \u0130zleme<\/h2>\n

Bir metnin belirli bir tonda yaz\u0131lmas\u0131, birka\u00e7 ko\u015fula uymas\u0131 ve verilen bilgilerin d\u0131\u015f\u0131na \u00e7\u0131kmamas\u0131 istendi\u011finde iki model aras\u0131ndaki fark daha g\u00f6r\u00fcn\u00fcr h\u00e2le gelir. OpenAI, g\u00fcndelik sohbetlerde ayr\u0131m\u0131n s\u0131n\u0131rl\u0131 olabilece\u011fini; g\u00f6rev karma\u015f\u0131kla\u015ft\u0131\u011f\u0131nda GPT-4’\u00fcn daha g\u00fcvenilir, yarat\u0131c\u0131 ve ayr\u0131nt\u0131l\u0131 y\u00f6nergeleri izlemekte daha ba\u015far\u0131l\u0131 oldu\u011funu belirtmi\u015ftir.<\/p>\n

Bu, GPT-3.5’in k\u0131sa ve basit metinleri haz\u0131rlayamayaca\u011f\u0131 anlam\u0131na gelmez. Fark \u00f6zellikle \u00e7ok say\u0131da kural\u0131n ayn\u0131 anda uygulanmas\u0131 veya uzun bir i\u00e7eri\u011fin dikkatle yorumlanmas\u0131 gerekti\u011finde ortaya \u00e7\u0131kar.<\/p>\n

4. Olgusal Do\u011fruluk ve Hata Riski<\/h2>\n

OpenAI’nin GPT-4 teknik raporunda, i\u00e7 de\u011ferlendirmelerde GPT-4’\u00fcn incelenen GPT-3.5 s\u00fcr\u00fcm\u00fcnden daha iyi sonu\u00e7 verdi\u011fi aktar\u0131l\u0131r. Bununla birlikte GPT-4 de olmayan bilgileri varm\u0131\u015f gibi yazabilir veya ak\u0131l y\u00fcr\u00fctme hatas\u0131 yapabilir. Dolay\u0131s\u0131yla \u201cGPT-4 her zaman do\u011frudur\u201d ifadesi yanl\u0131\u015ft\u0131r.<\/p>\n

\u00d6zellikle ara\u015ft\u0131rma, sa\u011fl\u0131k, hukuk veya finans gibi alanlarda modelin yan\u0131t\u0131n\u0131 tek ba\u015f\u0131na kaynak olarak kullanmay\u0131n. Ayr\u0131ca GPT-4’\u00fcn b\u00fct\u00fcn yan\u0131tlar\u0131nda kaynaklar\u0131 kendili\u011finden tespit edip do\u011fru al\u0131nt\u0131lad\u0131\u011f\u0131 ya da intihali otomatik olarak \u00f6nledi\u011fi iddias\u0131 da bu kar\u015f\u0131la\u015ft\u0131rman\u0131n bir par\u00e7as\u0131 de\u011fildir. Kayna\u011f\u0131, tarihini ve al\u0131nt\u0131n\u0131n do\u011frulu\u011funu ayr\u0131ca kontrol etmek gerekir.<\/p>\n

5. G\u00fcvenlik Yakla\u015f\u0131m\u0131<\/h2>\n

OpenAI, GPT-4’\u00fcn geli\u015ftirilmesinde g\u00fcvenlik testleri ve geri bildirimlerden yararland\u0131\u011f\u0131n\u0131 a\u00e7\u0131klad\u0131. Kurumun yay\u0131mlad\u0131\u011f\u0131 de\u011ferlendirmelerde GPT-4’\u00fcn yasakl\u0131 isteklere yan\u0131t verme e\u011filimi GPT-3.5’e k\u0131yasla daha d\u00fc\u015f\u00fck \u00e7\u0131km\u0131\u015ft\u0131r. Bu sonu\u00e7, modelin zararl\u0131 veya yan\u0131lt\u0131c\u0131 i\u00e7erik \u00fcretme riskinin tamamen ortadan kalkt\u0131\u011f\u0131 anlam\u0131na gelmez.<\/p>\n

Bu nedenle iki modeli kar\u015f\u0131la\u015ft\u0131r\u0131rken g\u00fcvenli\u011fi mutlak bir \u00f6zellik gibi g\u00f6rmek yerine, yay\u0131mlanan testlerdeki iyile\u015fme ve devam eden s\u0131n\u0131rlamalar \u00fczerinden de\u011ferlendirmek daha do\u011frudur.<\/p>\n

GPT-4’\u00fcn Parametre Say\u0131s\u0131 GPT-3.5’ten Fazla m\u0131?<\/h2>\n

\u0130nternette GPT-4 i\u00e7in \u201c1 trilyon\u201d, GPT-3.5 i\u00e7in \u201c175 milyar\u201d gibi say\u0131larla kar\u015f\u0131la\u015fabilirsiniz. Ancak OpenAI’nin GPT-4 teknik raporu model boyutu d\u00e2hil mimari ayr\u0131nt\u0131lar\u0131 a\u00e7\u0131klam\u0131yor. Bu nedenle bu rakamlar\u0131 do\u011frulanm\u0131\u015f bir kar\u015f\u0131la\u015ft\u0131rma olarak vermek do\u011fru de\u011fildir. Parametre say\u0131s\u0131ndan yola \u00e7\u0131karak hangi modelin ne kadar do\u011fru yan\u0131t verece\u011fini kesin bi\u00e7imde hesaplamak da m\u00fcmk\u00fcn de\u011fildir. GPT-4 ile GPT-3.5 aras\u0131ndaki do\u011frulanabilir farklar; g\u00f6rsel girdileri i\u015fleme, karma\u015f\u0131k g\u00f6revlerdeki kar\u015f\u0131la\u015ft\u0131rmal\u0131 performans ve ayr\u0131nt\u0131l\u0131 talimatlar\u0131 takip etme alanlar\u0131nda belirginle\u015fir. Her iki modelin de yan\u0131tlar\u0131, kullan\u0131m amac\u0131na g\u00f6re g\u00f6zden ge\u00e7irilmeli ve \u00f6nemli bilgiler g\u00fcvenilir kaynaklarla do\u011frulanmal\u0131d\u0131r.<\/p>\n","protected":false},"excerpt":{"rendered":"

OpenAI’nin 14 Mart 2023’te duyurdu\u011fu GPT-4, GPT-3.5’e g\u00f6re \u00f6zellikle karma\u015f\u0131k talimatlar\u0131 izleme, problem \u00e7\u00f6zme ve g\u00f6rsel i\u00e7erikleri yorumlama alanlar\u0131nda ilerleme sundu. Bu yaz\u0131da iki modeli, ilk yay\u0131mland\u0131klar\u0131 d\u00f6neme ili\u015fkin OpenAI verileri \u00fczerinden kar\u015f\u0131la\u015ft\u0131r\u0131yoruz. Girdi t\u00fcr\u00fc Kar\u015f\u0131la\u015ft\u0131r\u0131lan \u00f6zg\u00fcn model metin girdisiyle \u00e7al\u0131\u015f\u0131r. Metin ve g\u00f6rsel girdilerini i\u015fleyebilir; bu kar\u015f\u0131la\u015ft\u0131rmadaki GPT-4’\u00fcn \u00e7\u0131kt\u0131s\u0131 metindir. Karma\u015f\u0131k g\u00f6revler \u00c7ok ad\u0131ml\u0131 […]<\/p>\n","protected":false},"author":2,"featured_media":6420,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[56,55,51],"tags":[],"_links":{"self":[{"href":"https:\/\/www.turkticaret.net\/blog\/wp-json\/wp\/v2\/posts\/6415"}],"collection":[{"href":"https:\/\/www.turkticaret.net\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.turkticaret.net\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.turkticaret.net\/blog\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.turkticaret.net\/blog\/wp-json\/wp\/v2\/comments?post=6415"}],"version-history":[{"count":9,"href":"https:\/\/www.turkticaret.net\/blog\/wp-json\/wp\/v2\/posts\/6415\/revisions"}],"predecessor-version":[{"id":9702,"href":"https:\/\/www.turkticaret.net\/blog\/wp-json\/wp\/v2\/posts\/6415\/revisions\/9702"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.turkticaret.net\/blog\/wp-json\/wp\/v2\/media\/6420"}],"wp:attachment":[{"href":"https:\/\/www.turkticaret.net\/blog\/wp-json\/wp\/v2\/media?parent=6415"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.turkticaret.net\/blog\/wp-json\/wp\/v2\/categories?post=6415"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.turkticaret.net\/blog\/wp-json\/wp\/v2\/tags?post=6415"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}