LLM Playground: tokens, temperature, context
Three small experiments that show how ChatGPT, Claude and Gemini work inside. This is a simulation with a tiny fake model, so it runs offline, but the ideas are the real ones.
Type any text and see the tokens
Real tokenizers are learned from data, so exact splits differ between models. Common words are one token; rare or long words get split. Try your name, an email address, or a word in another language.
Pick a prompt
Temperature: 0.7
0 = always the top word · 1 = normal · 2 = wild
Probability of each next word
Results of 20 runs
Press "Generate 20 times".
💡
Security link: high temperature makes output less predictable. For a bank chatbot or a security tool, you usually want low temperature. Attackers also re-try the same jailbreak many times, hoping randomness lets it through once.