Welcome to bits. This course is the mental model under every other one here: what a language model is, how it writes, what it sees, where it breaks. Each lesson is one idea with a live run.
The box below is a thin wrapper over one Messages API request to Claude. No system prompt of yours, no history, no tools. Add “in one sentence” to the draft and run it.
Run your prompt and the replies land here.
What answered you is a large language model: a function from text to a probability distribution over the next token, applied in a loop until it emits a stop. “Large” is the parameter count and the training corpus. Everything else (chat, tools, memory, agents) is built on top of that loop, by code outside the model. Lesson 2 is the loop, lesson 6 is what the loop gets to read.
Terms, pinned. AI is the umbrella. An LLM is the text-in, text-out kind. Claude is Anthropic's family of LLMs; a product like claude.ai is an app around one, with its own system prompt, history and tools. The sandbox here skips all of that: one user message in, one assistant message out.
Claude models take text and images as input. This course is text only.
The sandbox, precisely. Runs go to claude-haiku-5-5, the “Fastest” latency tier in Anthropic's lineup (October 2026), listed from $0.10 per million input tokens and $0.50 per million output tokens, with a 1M-token context window. Each run is one fresh request with a short max_tokens, no thinking and no history, so a second run knows nothing about the first. Past the hourly limit (higher when signed in) you get a saved example, labelled as one.
Here's a peek at what happened while it answered: the word it picked at each step, and a few it didn't. The numbers are illustrative; the next lesson explains them.
Hi Claude! In one sentence, what are you?
step 1/6 · candidates for the next token
Step 1 of 6: picks “I'm” at 88%.