Introduction to Large Language Models
What LLMs are, how they predict text, and what they can and can't do.
Inspired by learning material from Google
Large Language Models (LLMs) like the ones behind chat assistants are trained to predict text. From that simple objective, surprisingly capable behavior emerges.
Next-token prediction
An LLM reads text and predicts the most likely next token (a word or piece of a word). Repeat this and it generates fluent paragraphs.
Trained on massive text
LLMs learn from enormous collections of text. During training they absorb grammar, facts, reasoning patterns, and style.
What they're good at
- Writing and summarizing
- Answering questions
- Translating and explaining
- Drafting and brainstorming
Limitations
- Hallucinations: they can state false things confidently.
- No real-time knowledge unless connected to tools or search.
- Sensitive to phrasing — prompts matter a lot.
Key takeaways
- LLMs generate text by predicting one token at a time.
- They're versatile but can be confidently wrong.
- Always verify important facts.