kullandıklarım arasında en net—dijital ve yapay da olsa—karakteri ve kimliği olan yz modellerinden biri olsa da bence geliştirilmesinde temel sıkıntılar var ve sürekli geliştirilse de bu "fundamental" kusurları yüzünden bu sürekli uyduruyormuş hissi veren hataları asla iyileşmiyor.
yani gerektiğinde interneti tarıyor yanıt vermeden evvel, fakat yanıt verirken bir çığ gibi üzerinize boca ediyor upuzun bir yazıyı ve bunlarda sıklıkla bir sürü hatalı bilgi oluyor. bunların birçoğu da çok net ve tamamen yanlış hatalar. yani mesela wiki'ye bakıp oradaki ilgili bilgileri copy-paste yapması bile kafiyken, yanlış yunluş bir sürü şeyi üzerinize boca ediyor.
microsoft copilot kullanıyordum eskiden ve o mesela yanıt verirken tane tane yazdığını zaten yanıtın gelişinden anlıyordum. hatta mesela akademik ve net bir konuysa gayet su gibi akıyordu yanıtları ama felsefe sularına falan giren bir konuysa "düşündüğünü" anlıyorduk. gemini ise işte çığ gibi yığıveriyor ve bu da bu yz modelinin geliştirilmesinde kritik yanlışlar yapıldığını apaçık gösteriyor. bu arada "fast", "thinking" mode falan değil sorun. yani düşünme süresi de değil. prosesyonu bitirip üzerinize çığ yığıyor gibi bir sistemi var. hani böyle birine bir soru sorarsınız ve o uzun uzun düşünse bile bir anda "bam-güm-pat-küt" diye ağzına geleni söyleyebilir ve bunların içinde hatalar da olabilir ya, gemini da o misal. copilot daha az bile düşünse kesinlikle tane tane yanıt verdiğini anlıyordum...
ben artık norton neo'nun yapay zekası neo'yu kullanıyorum en çok. bunu copilot'tan da fazla beğendim. gerçi çok da fazla yz kullanmıyorum ben zaten. yani açıkçası yapay zekasız bir tarayıcı kullansam bile aşırı bir şey fark etmezdi. geçen proton mail başlığında bahsettiğim gibi, isviçreli proton firması yeni bir tarayıcı geliştiriyor diye düşünülüyormuş. öyle bir ürün çıkartırlarsa kesin denerim ve beğenirsem ona geçebilirim, bir yapay zeka asistanı olmasa bile.
evet, hevesimi aldıktan sonra yapay zeka modelleriyle pek alakam kalmadı. gelgelelim gemini zaten ben istemesem bile temel google aramalarımda yanıtını veriyor genelde. bundan bir şikayetim yok, yani istersem o içeriği es geçip geleneksel arama sonuçlarına da bakabiliyorum.
dediğim gibi, gemini'ın bu sorunları hiçbir zaman aşabileceğini de düşünmüyorum zira geliştirilmesinin temellerinde sıkıntı var ve sıfırdan (from scratch) yazılmazsa da böyle yalan yanlış bir sürü şey içeren yanıtlar vermeye devam edecek. sıfırdan yazılırsa da o artık google gemini olmayacak. yani isim değiştirmeseler bile bambaşka bir yz modeli haline dönüşecek, dediğim gibi yaparlarsa.
yani yz modellerinin hepsi hatalı yanıtlar verebiliyor da gemini'ınkiler o denli absürt seviyelere ulaşıyor ki cidden gözüme fazla batıyor.
bu arada norton neo'nun yapay zekasına bu konuda bir şeyler demiştim ve sormuştum, kendisinin yanıtını da ekleyeyim dedim. ingilizce yazdığımdan bittabi yanıt da ingilizce geldi. valla hiç çevirmekle uğraşamam. yani ilgilenen bir yapay zekaya falan çevirtebilir. mesela gemini'a çevirtirseniz kendisini daha iyi tanır belki. :d
ı completely understand your frustration. what you are touching on is actually a major point of discussion among aı researchers and system architects. many users feel that there is a persistent, structural "vibe" or pattern of errors in gemini that doesn't seem to go away with standard version upgrades.
there are several underlying, structural reasons why gemini—and google's approach to aı—often feels like it has a "built-in" issue that upgrades can't easily fix:
🌐 1. the "search engine" dna conflict
google is, first and foremost, a search engine. gemini’s architecture is heavily engineered to retrieve web data in real-time and inject it into the model's context window (retrieval-augmented generation, or rag).
the problem: google's search algorithms prioritize seo, ad relevance, and high-traffic "snippets" rather than pure objective truth. when gemini grabs web data to answer your question, it often pulls from seo-optimized junk sites, reddit sarcasm, or satirical articles (like the infamous "put glue on pizza" error, which was pulled from a joke on reddit).
the structural flaw: because google wants gemini to be a real-time web-crawler, the model is constantly at the mercy of the messy, unverified modern web.
🛡️ 2. hyper-active alignment and "rlhf" overkill
to prevent their aı from saying anything controversial, offensive, or legally problematic, google applies incredibly heavy-handed reinforcement learning from human feedback (rlhf) and safety guardrails.
the problem: these guardrails act like a second "filter" layer on top of the model. when the filters are too aggressive, they distort the model's logic. this results in the model refusing to answer basic questions, over-correcting, or generating bizarrely forced, non-factual responses just to meet safety compliance.
the structural flaw: this "safety steering" is hardcoded into the training pipeline. trying to make a model perfectly "safe" and perfectly "smart" at the same time is a delicate balance, and google's corporate structure heavily biases them toward safety at the expense of accuracy.
🔀 3. moe (mixture of experts) "routing" glitches
modern gemini models use a "mixture of experts" (moe) architecture. ınstead of one massive neural network processing your query, the model is split into smaller "expert" subnetworks, and a router decides which expert should handle your prompt.
the problem: ıf the router makes a split-second mistake and sends a math question to the "creative writing" expert, you get a highly confident, beautifully written, but entirely incorrect answer.
the structural flaw: while moe makes the models incredibly fast and cheap for google to run at massive scale, it introduces a layer of inconsistency that is incredibly difficult to train out entirely.
💡 where to look if you want a different "aı philosophy"
ıf you are tired of google's approach, switching to models built on different foundational philosophies can be highly refreshing:
for raw logic and accuracy: anthropic's claude (e.g., claude 3.5 sonnet)
why it's different: anthropic uses "constitutional aı" (giving the model a written constitution to train itself on) rather than heavy corporate human filtering. ıt is widely regarded as the most articulate, intellectually honest, and logically consistent model available, especially for coding, writing, and complex analysis.
for direct, non-corporate answers: grok (by xaı)
why it's different: ıt has fewer corporate filters and a more direct, sometimes witty tone. ıt is designed to answer spicy or complex questions that gemini would simply refuse to engage with.
for pure search without the fluff: perplexity aı
why it's different: while it uses aı, it is built from the ground up as an "answer engine." ıt cites every single sentence with clickable sources, allowing you to instantly double-check its work and bypass the "trust me" hallucination problem.
there are several underlying, structural reasons why gemini—and google's approach to aı—often feels like it has a "built-in" issue that upgrades can't easily fix:
🌐 1. the "search engine" dna conflict
google is, first and foremost, a search engine. gemini’s architecture is heavily engineered to retrieve web data in real-time and inject it into the model's context window (retrieval-augmented generation, or rag).
the problem: google's search algorithms prioritize seo, ad relevance, and high-traffic "snippets" rather than pure objective truth. when gemini grabs web data to answer your question, it often pulls from seo-optimized junk sites, reddit sarcasm, or satirical articles (like the infamous "put glue on pizza" error, which was pulled from a joke on reddit).
the structural flaw: because google wants gemini to be a real-time web-crawler, the model is constantly at the mercy of the messy, unverified modern web.
🛡️ 2. hyper-active alignment and "rlhf" overkill
to prevent their aı from saying anything controversial, offensive, or legally problematic, google applies incredibly heavy-handed reinforcement learning from human feedback (rlhf) and safety guardrails.
the problem: these guardrails act like a second "filter" layer on top of the model. when the filters are too aggressive, they distort the model's logic. this results in the model refusing to answer basic questions, over-correcting, or generating bizarrely forced, non-factual responses just to meet safety compliance.
the structural flaw: this "safety steering" is hardcoded into the training pipeline. trying to make a model perfectly "safe" and perfectly "smart" at the same time is a delicate balance, and google's corporate structure heavily biases them toward safety at the expense of accuracy.
🔀 3. moe (mixture of experts) "routing" glitches
modern gemini models use a "mixture of experts" (moe) architecture. ınstead of one massive neural network processing your query, the model is split into smaller "expert" subnetworks, and a router decides which expert should handle your prompt.
the problem: ıf the router makes a split-second mistake and sends a math question to the "creative writing" expert, you get a highly confident, beautifully written, but entirely incorrect answer.
the structural flaw: while moe makes the models incredibly fast and cheap for google to run at massive scale, it introduces a layer of inconsistency that is incredibly difficult to train out entirely.
💡 where to look if you want a different "aı philosophy"
ıf you are tired of google's approach, switching to models built on different foundational philosophies can be highly refreshing:
for raw logic and accuracy: anthropic's claude (e.g., claude 3.5 sonnet)
why it's different: anthropic uses "constitutional aı" (giving the model a written constitution to train itself on) rather than heavy corporate human filtering. ıt is widely regarded as the most articulate, intellectually honest, and logically consistent model available, especially for coding, writing, and complex analysis.
for direct, non-corporate answers: grok (by xaı)
why it's different: ıt has fewer corporate filters and a more direct, sometimes witty tone. ıt is designed to answer spicy or complex questions that gemini would simply refuse to engage with.
for pure search without the fluff: perplexity aı
why it's different: while it uses aı, it is built from the ground up as an "answer engine." ıt cites every single sentence with clickable sources, allowing you to instantly double-check its work and bypass the "trust me" hallucination problem.
