LLM · Large Language Model
Large Language Model
What is an LLM?
- Predicts the 'next word' from vast amounts of text
- A large neural network that generates text
- The brain behind ChatGPT and Claude
What LLMs have really changed is not the quality of answers but how you assign work. It used to be that even building a feature to sort inquiry emails by type meant coding rules and registering exceptions one by one. Now you can start the same job with an instruction and a few examples. That has moved the starting point of automation away from the development team and toward the practitioners who know the work best. The trade-off is that output quality now depends on wording rather than code, so the business side also carries the burden of deciding how far to trust a result before it goes out.
In practice, you need to tell at least three things apart. The context window, the amount a model can take in at once, ranges from tens of thousands to hundreds of thousands of tokens depending on the product (in Korean, roughly one character equals one token), and anything you put in is gone when that conversation ends. Training data has a cutoff date, so the model knows nothing about an internal policy revised last month or this quarter's price list. There are three ways to add company knowledge: put it directly in the prompt, retrieve and attach it with RAG, or lock in tone and format with fine-tuning. If factual accuracy is the goal, start with RAG and add fine-tuning later.
The most common failure point in practice is the misconception that 'we put in our materials, so it has learned them'. A document pasted into a chat window isn’t learned, just shown briefly; close the window and the model knows nothing about it. An LLM is also neither a calculator nor a database, so numbers such as revenue totals or stock counts come out plausible but wrong. Getting a different answer to the same question each time isn’t a malfunction either; it is a trait of a system that strings sentences together by probability. If you don’t know this and copy answers straight into customer quotes or contract wording, nonexistent clauses or discontinued model numbers go out as-is, and you end up writing apology emails.
John KimInstructor John Kim, author of B2B Marketing Design and CEO of Nanum Consulting
Seolyeong ParkGenerative AI and AX training specialist with 17 years of hands-on experience and 200+ lectures, helping non-technical staff put AI to work right away| Category | LLM (Large) | sLM (small) |
|---|---|---|
| Size | Tens of billions of parameters or more | Relatively small |
| Strengths | General-purpose, high-performance | Low cost, fast, on-device |
| Cost | High | Low |
| Use case | Complex, varied tasks | Fine-tuning for specific uses |
The Transformer, the architecture behind today's LLMs, was introduced in the 2017 Google research paper “Attention Is All You Need.” GPT, Claude, Gemini, and others are built on it.
Consider a hypothetical case. Company A, an industrial parts distributor, has 210 employees and annual revenue of about KRW 120 billion (≈ US$89,780,000). Its six-person customer support team handles roughly 140 technical inquiry emails a day, and in peak season replies slipped by two days, drawing constant complaints from sales. Over three months, they ran an experiment using an LLM to draft replies.
- The support team lead sorted the last three months’ 4,200 emails by type and found that 62% were spec and compatibility questions whose answers were already in the product catalog.
- When the IT manager pasted in the entire catalog and asked for answers, the output described a model discontinued two years earlier as if it were current stock.
- So they switched to first searching about 900 internal PDFs and feeding the model only the supporting paragraphs, and had each answer end with the source document name and page number.
- The head of sales was firm: 'never send out an answer without a source'. So if the system can’t find supporting evidence, it shows only 'Needs owner review' instead of a draft.
- Three months later, average response time per inquiry had dropped from 18 minutes to 6, 74% of drafts were sent almost unchanged, and emails returned because of wrong answers fell from 31 a month to 7.