Question 1

What is JSON?

Accepted Answer

JSON (JavaScript Object Notation) is the language computers use to talk to each other over the internet. When you send a message in Glass Box, your browser sends a structured JSON object to Claude's API, and Claude responds with one. Curly braces group related information, square brackets are lists. You can see the raw JSON in the Chatbot Lab's Response tab after sending a message.

Question 2

Why is there a character limit?

Accepted Answer

Glass Box pays real API costs for every message you send. The 400-character limit keeps this sustainable so the tool can stay free. It's also a useful constraint — in practice, concise prompts often work better than long ones.

Question 3

What does a message exchange cost?

Accepted Answer

Each exchange (your message + Claude's response) costs a fraction of a cent — typically $0.0002–$0.002 depending on the model and conversation length. Check the Tokens & Cost tab in the Chatbot Lab after any message to see the exact breakdown.

Question 4

What order should I explore the concepts?

Accepted Answer

Follow the numbers: ⓪ Making a Machine Intelligent → ① LLM Basics → ② Tokenization → ③ Embeddings → ④ Attention → ⑤ Prediction → ⑥ Training → ⑦ Tuning → ⑧ Models → ⑨ Eyes → ⑩ Hands → ⑪ The Toolbox → ⑫ The Full Picture. Each builds on the previous one — first what machine learning even is, then how the model processes your text, then how it learned and was shaped to be useful, then what actually lives inside it and how it sees and acts in the world. When you're ready, the Chatbot Lab lets you chat with Claude and see every step happening live.

Question 5

What's the Chatbot Lab?

Accepted Answer

The Chatbot Lab is a real Claude chatbot with X-ray vision. You chat naturally on one side while the Inspector — the panel beside the chat — shows every API call, token count, cost, tool call, and thinking process in real time. It's where you put the concepts into practice.

Question 6

Is any of this simulated?

Accepted Answer

No. Every API call is real. Every token count comes from the actual API response. Every cost is calculated from real pricing. The concept exhibits run real algorithms and real model data — the same text-splitting method production models use (called BPE), real word-meaning coordinates (from a technique called Word2Vec), and attention weights measured by running GPT-2, a real language model, on the exhibit's own sentence. The few figures that are estimates or conceptual sketches say so right where they appear.