How AI Works cheat sheet
What to keep from each lesson.
Module 1 in full, Modules 2 to 4 by title.
Contents, 5 sections
1 How it writes
One word at a time
- A chat model writes by predicting one Token A piece of text, often a word or part of one, that a model reads and writes., adding it, and predicting again, until it reaches a stopping point or a limit.
- Each token is picked from learned odds. Nothing in the loop makes sure a fact was checked, so fluent text can still be wrong.
What the model actually reads
- Models read and write Token A piece of text, often a word or part of one, that a model reads and writes., and different models can split the same text differently.
- For text models like these, input and output tokens are billed separately, and each output token costs more at these models' listed rates.
- In our test, both GPT models miscounted the letters in strawberry. Check any answer that depends on exact characters yourself.
Same prompt, different answers
- The model Sampling Picking each next token at random, in proportion to the model's odds. each token from its odds, so different answers to the same request are normal.
- Temperature A setting that sharpens or flattens the odds before sampling. changes how varied the picks are, and many current models don't let you set it. If a reply must include something, ask for it in the Prompt The instructions and material you give a model., then check the reply has it.
2 How it learned
- How it's trained
- Where its knowledge stops
- Fluent isn't the same as true
- When the question leads the answer
- Thinking before it answers
The notes for Modules 2 to 4 come with Pro. Compare plans
3 What it can see
- Everything it can see
- Meaning as numbers
- Answers from your own sources
- Images, PDFs and audio
4 How it acts
- How it uses tools
- Agents: tools in a loop
- When the data gives orders
Key terms
- Agent
- A model that uses tools in a loop, deciding its own next step.
- Context
- Everything the model reads before it writes a reply.
- Context window
- The most tokens a model can handle in one request.
- Embedding
- A list of numbers that represents what a piece of text means.
- Fine-tuning
- Extra training on your own examples to change a model's behavior or format.
- Grounding
- Tying each claim in an answer to a source you supplied, so it can be checked.
- Hallucination
- Generated content that's false or not supported by the source. It can sound completely sure.
- Image tokens
- The tokens a picture adds to a request.
- Knowledge cutoff
- The date up to which a model's knowledge is most complete.
- Large language model (LLM)
- A model that writes text one token at a time by predicting what comes next.
- Lethal trifecta
- Private data, untrusted content and a way to send data out, all in one assistant.
- MCP
- An open standard for connecting tools and data to AI apps.
- Multimodal
- Able to take in more than text, such as images, PDFs or audio.
- Post-training
- Later training that turns a text predictor into a helpful assistant.
- Pretraining
- The first stage of training, where a model learns language by predicting text.
- Prompt
- The instructions and material you give a model.
- Prompt injection
- Text that tries to get a model to treat it as instructions and act outside the task its user gave it, typed in directly or planted in content the model reads.
- Reasoning model
- A model that works through a problem before giving its answer.
- Retrieval-augmented generation (RAG)
- Finding relevant passages and giving them to the model, so it answers from them.
- Sampling
- Picking each next token at random, in proportion to the model's odds.
- Sycophancy
- A model's tendency to agree with what the user seems to believe.
- System prompt
- Standing instructions an app sends at the top of every request.
- Temperature
- A setting that sharpens or flattens the odds before sampling.
- Token
- A piece of text, often a word or part of one, that a model reads and writes.
- Tool use
- A model asking an app to run a function, such as a lookup, and using the result.