본문
이 글에 나오는 이름
- LLM
- 거대 언어 모델(large language model). 아주 많은 글로 학습해 다음에 올 낱말을 예측하며 글을 쓰는 AI예요.
- ChatGPT
- 챗GPT. 오픈AI가 만든 대화형 AI 서비스예요.
- RAG
- 검색 증강 생성. 답하기 전에 관련 문서를 찾아 질문에 붙이고, 그 문서를 바탕으로 답하게 하는 방법이에요.
밑줄 친 단어나 X-ray 표시를 누르면 아래에서 읽기 도구가 올라와요.
Anyone who uses an AI chatbot at work soon meets a strange problem. You ask a clear question, and the answer comes back , confident and wrong. It may quote a report that does not exist or give a number that nobody ever . People call this a .
To see why it happens, it helps to know what a large language model, or LLM, actually does. An LLM is trained on a huge amount of text. From that text it learns which words to follow which. When you ask it something, it does not look the answer up in a library. It writes the reply one word at a time, each time choosing a word that is likely to come next.
Most of the time, the likely answer is also the true answer. But when the model has seen little about a topic, it still produces something that sounds right. It has no built-in sense of what it does not know. The same ability that lets it write a fresh story or a new slogan also lets it invent a fact.
The results can be serious. In 2023, lawyers in New York sent a court a filing that several earlier cases. The cases had been suggested by ChatGPT, and none of them were real. The judge noticed, and the lawyers were fined.
Companies now use a method called retrieval-augmented generation, or RAG, to the problem. The idea is simple. Before the model answers, a search system finds documents that are to the question. These might be a company manual or last quarter’s sales report. The documents are added to the question, and the model is told to answer only from them. It is a bit like letting a student take an open-book exam instead of answering from memory. A good RAG system also shows which document each part of the answer came from, so a person can check it.
RAG does not make an AI perfect. If the right document is missing or out of date, the answer can still be wrong. That is why the most useful habit is an old one. Treat an AI answer as a first from a quick but careless assistant. Before you use a number, a name or a in an important decision, check it against the original source.
이 글은 이 사이트에서 새로 썼어요. 한국어로 더 깊이 읽으려면 intoax의 페이지를 보세요.