2.1 What Is an LLM?
Chapter 1 introduced generative AI: systems that can produce new text, images, or sound in response to an instruction. We will now look more closely at the source of language ability behind many chat assistants—the large language model (LLM).
Start by separating the name:
- Language: it mainly handles what people write or say, such as questions, articles, notices, and conversations.
- Model: it is not a warehouse of original documents but a set of patterns formed from many examples.
- Large: its learning scale, adjustable internal relationships, and training effort are large. The next section explains this carefully.
For now, imagine an LLM as an assistant that has examined enormous amounts of text and become very skilled at continuing and organizing language. When you enter half a sentence, it judges which continuations fit. After choosing a small piece, it uses everything written so far to continue again.
How can “continuing text” perform many tasks?
Predicting what comes next may sound useful only for a sentence-completion game. However, language contains many patterns connecting tasks, formats, and knowledge.
If you provide a long passage and write “summarize this in three points,” a suitable continuation should be shorter, selective, and arranged in three parts. If you write “make this sentence more polite,” the continuation should preserve the meaning while changing its wording and tone. An LLM learns such relationships from many language examples, so the same basic generation process can appear as answering, summarizing, translating, classifying, or rewriting.
An LLM is not the whole AI product
An LLM may be an important “engine” inside an AI product, but the chat product you see can also contain an interface, file upload, web search, voice, image tools, safety rules, and account services. A product may use different models at different times.
Therefore, “I am using this chat product” and “it currently uses this LLM” are not the same statement. You do not need to memorize model versions yet. Remember: the product organizes features for the user; the model provides part of its language ability.
Avoid one more misunderstanding. Natural, human-like sentences do not prove that an LLM has human experience, intentions, or responsibility. Its output comes from language patterns and currently available information, and it can still be wrong.