Tokens are building blocks
Models read and write token units.
Foundations·Beginner·6 min
Token Playground
See how large language models tokenize text, assign next-token probabilities, and change sampling behavior with temperature.
01
Experience
Work the problem before reading the explanation.
Lesson path
01
Prompt
Read the input
02
Tokenize
Split text units
03
Predict
Inspect probabilities
04
Sample
Change temperature
05
Repeat
Append next token
Loading deterministic lab runtime…
02
Reflection
Turn the outcome into a rule you can reuse.
Key takeaways
Probabilities guide choices
Candidate tokens receive different scores.
Temperature changes randomness
Higher values flatten the distribution.
Generation is iterative
The chosen token becomes new context.
03
Learn More
Connect the experience to concepts, references, and transfer.
Concept guide
How does next-token prediction work?
A large language model generates text by estimating a probability distribution over possible next tokens, selecting a token using a decoding strategy, adding the selected token to the current context, and repeating the process.
What is a token?
A token is a text unit used by the model. It can be a whole word, part of a word, punctuation, whitespace, or another fragment depending on the tokenizer.
What does temperature change?
Temperature changes the shape of the sampling distribution. Lower values concentrate probability around top candidates; higher values give lower-ranked candidates more chance to appear.
Why does this matter in engineering?
Decoding choices affect variability, deterministic behavior, structured-output reliability, and how an application balances creativity with control.
Common questions
Does an LLM choose a whole sentence at once?
No. Autoregressive language models generate incrementally, token by token.
Is a token always a word?
No. Token boundaries depend on the tokenizer and can split words or include punctuation.
Does temperature change model knowledge?
No. It changes sampling behavior, not the model weights or underlying knowledge.
Are the probabilities in this lab from a live model?
No. This MVP uses deterministic educational values so the learning experience stays stable.
Interactive values are illustrative; conceptual explanations are the durable learning content.
LEARNING CONTEXT
Learning context
Application guarantees must be designed around probabilistic model behavior, not assumed from fluent output.
Not seen
Mental models
- S00-M01Probabilistic behavior vs application guarantees
Suggested backfill
No shipped prerequisite is required before entering this incident.
Transfer the model
A model returns a plausible but invalid enum value. Which guarantee belongs to the model, and which belongs to application code?
04
Next
Carry the idea into another problem or build.
Build challenge
Choose decoding settings for structured output
Would you use the same temperature for creative copy and strict JSON? Explain the trade-off before opening the next lab.
Token prediction makes sense now.
Next: understand the finite working context around those tokens.
Continue →