Watch the idea
The concept
Text
What you type
Tokens
Shards, not thoughts
Next token
Temperature changes the guess
Words are not how the model sees words
A language model predicts the next token — a shard of text, not a thought. English often splits on spaces and pieces of words. Tamil may split differently, so the same idea can cost more tokens and fit worse in the window.
Temperature is randomness, not honesty
Low temperature: the model picks the most likely next shard. High temperature: it samples stranger shards. That can look creative. It can also look like a confident error. Never treat wildness as evidence.
The desk of notes is finite
The context window is everything the model can see at once: your prompt, the retrieved notes, the chat so far. Overflow, and early instructions fall off the desk. That is a systems fact, not a personality.
Worked example
Same poem, different cost
A twelve-word English line might be twelve tokens. The Tamil translation might be many more. If you paste a long Tamil brief plus sources, you hit the desk limit sooner. Pin the rule at the end.
Picture to keep
Prompt → Tokens → Next token
- 1Tokenize
- 2Attend
- 3Predict