How large language models work
- Does your phone guess your next word?
- How does a program guess anything?
- Can a model guess words?
- How does ChatGPT write whole answers?
- Why does the phone go round in circles when ChatGPT doesn't?
- Is ChatGPT just GPT-3 with a chat box?
- How does a model learn which word comes next?
- What is a parameter?
- How does training turn the dials?
- How many parameters does a real model have?
- Are the chances for the next word stored in the model?
- Does a model keep a copy of what it read?
- How does a model read the words you type?
- What is a token?
- Why are tokens pieces of words, not letters or whole words?
- Who decides how text is cut into pieces?
- Why does the same sentence cost more in Hindi?
- Does a token's number tell the model what a word means?
- How did you learn what a cat is when you were a baby?
- What is an embedding?
- Why do cat and dog end up with similar lists?
- How can a computer measure how close two meanings are?
- Can a model learn the wrong clue?
- Why can't a model tell which "bank" you mean from the word alone?
- How does a model decide which earlier words matter?
- How does attention change the numbers for "bank"?
- How does a model know what "it" refers to?
- What is a Transformer?
- Why did the Transformer change so much?
- What does a model give before a word is written?
- How is one word picked from the chances?
- What does temperature do to the chances?
- How do top-k and top-p cut the list?
- Can you change temperature, top-k and top-p yourself?
- What can a model do after reading the internet?
- How does a model learn to answer a question?
- What does a chat look like to the model?
- How does a model learn which answers people prefer?
- Should a company train a model on its own examples?
- Which should you try first: a better prompt, RAG or fine-tuning?
- How does one token turn into a bill?
- How much work is one token?
- Why does a model need a GPU?
- Why does the model's answer cost more than your question?
- How much does one question cost?
- What is a context window?
- Does a model remember what you said earlier in a chat?
- Why not paste the whole PDF into every question?
- What is RAG?
- How does RAG answer a question, step by step?
- Who cuts the PDF into chunks?
- Who makes the embeddings, and who does the search?
- Does RAG use a smaller model, and is the LLM the only cost?
- When does RAG give a wrong answer?
- Can a model do things, or only write text?
- What is tool calling?
- What is an agent?
- What is MCP?
- Why does a model make things up?
- How do you reduce made-up answers?
- What have you learned in these ten lessons?