Large Language Model / LLM
The artificial intelligence model, which is trained with Devasa text data, can produce text similar to human language.
Large language models (LLM) are artificial nerve networks, usually installed on Transformer architecture and trained with huge amount of text data. During the training, the model tries to predict the next word (or token) in a sentence; this apparently simple task allows the model to develop grammar, world knowledge, reasoning patterns and even logical ability when repeated on millions of books, articles and web pages on the internet scale. After pre-training, it is often passed through the fine adjustment to give appropriate answers to model, useful, harmless and instructions with alignment techniques such as learning (RLHF).
Large language models are currently used in tasks that require answering questions, text abstracting, translation, code writing, creative writership and complex reasoning. GPT family is among examples known as Claude and Ship. The most remarkable feature of these models is that even in the tasks that they are clearly trained, they are able to succeed with "coming" (emergent) capabilities, which distinguishes them from classic, unique AI systems. At the same time, there are also important limitations such as hallucination generation, non-updated information and high calculation cost, and these limitedness are tried to be mitigated by techniques such as RAG and fine-tuning.
Large language models are currently used in tasks that require answering questions, text abstracting, translation, code writing, creative writership and complex reasoning. GPT family is among examples known as Claude and Ship. The most remarkable feature of these models is that even in the tasks that they are clearly trained, they are able to succeed with "coming" (emergent) capabilities, which distinguishes them from classic, unique AI systems. At the same time, there are also important limitations such as hallucination generation, non-updated information and high calculation cost, and these limitedness are tried to be mitigated by techniques such as RAG and fine-tuning.