MEVZU N° TAG / VOL. 088
4 blog · 1 news · 50 wiki
Prompt Engineering tek mesaj kuruyor; Agentic Engineering ise hedefi olan, araç kullanan, kendi kendini düzelten bir sistem tasarlıyor. Üç aşamalı evrim, klasik yazılımla karşılaştırma, somut bir Türkçe örnek ve insan-denetimi prensipleri.
Andrej Karpathy ile popülerleşen Vibe Coding nedir, klasik kodlamadan farkı ne, hangi araçlarla yapılır, sınırları neler? Cursor, Claude Code, Copilot, Cline ve Windsurf ekosistemine tarafsız bir Türkçe rehber.
Vibe Coding ve Agentic Engineering ile sınırları aşın. Google Stitch ve Wordent Manager kullanarak, Supabase veri tabanlı ve AI chatbot destekli bir veteriner mobil uygulamasını (Wed Care) sıfırdan nasıl geliştirdiğimizi adım adım anlatıyoruz.
A large language model that also processes modalities like image, audio, or video.
A large-scale AI model pretrained on broad data that can be adapted to many downstream tasks.
The most capable AI models of their generation, often pushing both capability frontiers and novel risk profiles.
Empirical relationships describing how model performance changes with parameters, data, and compute.
A release model where only the trained weights are published, not the training code or dataset.
A large language model whose weights, code, or training material are publicly released.
Capabilities that appear in models only above a certain scale and are absent in smaller variants.
A large neural network trained on massive text data to understand and generate language.
A LoRA variant combined with quantisation that lets you fine-tune 65B models on a single consumer GPU.
An architecture where only a subset of expert sub-networks activates per token, combining huge capacity with cheaper inference.
The general term for how a model picks the next token from its probability distribution.
A model type that generates the next token step-by-step, conditioned on previous tokens.
An alignment technique that trains a reward model from human preferences and then optimises the LLM against it.
A reasoning approach that explores multiple branches in parallel instead of a single chain, then picks the best path.
A fine-tuning technique that trains only small low-rank matrices instead of every weight, dramatically cutting memory.
An RLHF alternative that directly optimises a model on preference data, skipping the explicit RL loop.
A model's capacity to work through a problem in multiple steps — now a central axis of LLM competition.
The initial training phase where a model learns general language ability from trillions of tokens of generic data.
The process where a trained model takes input and produces output.
A training objective where the model learns to predict tokens that have been masked out of a sentence.
An inference speedup where a small draft model proposes multiple tokens that the big model then verifies in parallel.
Training a smaller 'student' model to mimic the behaviour of a larger 'teacher' model.
Representing model weights with lower-precision numbers to save memory and gain speed.
A decoding algorithm that keeps the K most-likely candidate sequences alive in parallel during generation.
An alignment approach that uses another LLM, instead of human labellers, as the source of preference signals.
The stage after pre-training that turns a raw model into a helpful, safe, instruction-following assistant.
The process by which a model's weights are updated to learn patterns from data.
Adapting a pre-trained model to a specific task using smaller, targeted data.
Prompting the model to reason step-by-step before producing a final answer.
Removing weights with negligible impact to shrink a model and speed it up.
A more clinically accurate term for LLM 'hallucination' — confidently filling gaps with plausible-sounding fiction.
A sampling method that draws from the smallest set of candidates whose cumulative probability exceeds P.
The sampling parameter that controls how 'creative' or 'deterministic' a model's output is.
The process of converting raw text into a sequence of model-readable tokens.
The attention-based neural network architecture that underpins virtually every modern LLM.
The Transformer component that turns input into a meaningful internal representation.
When an LLM produces fluent, confident-sounding output that simply isn't true.
The total token count consumed in a single model call, used against the model's context-window limit.
An attention mechanism where one sequence attends to a different sequence, typically connecting encoder and decoder.
The Transformer component that generates the next token conditioned on what came before.
The mechanism that lets a model decide how much weight to give different parts of its input.
The maximum number of tokens a language model can process in a single forward pass.
A mechanism where each element in a sequence attends to every other element in the same sequence.
Next-generation LLMs that can process hundreds of thousands — sometimes millions — of tokens in a single context.
A version of attention where multiple parallel 'heads' learn different relationships at the same time.
A way to represent a token or piece of text as a dense numerical vector that encodes its meaning.
A sampling strategy that picks the next token from only the K most likely candidates.
The smallest unit a language model processes — a word fragment, character, or symbol.
Anthropic's mid-large model released in late 2025, the industry leader for code and long-form analysis.
A large language model family from China-based Z.AI / Zhipu, also released as open-weights.