Skip to content
Grok Academy
All resources
Tutorial · 5 min

What happens when you talk to Grok?

The pause after you hit send is not thinking in the human sense. Here is the actual pipeline, without the myth.

  1. 01Your text is tokenized — chopped into the pieces the model was trained on.
  2. 02Those tokens, plus the system prompt and prior turns that still fit, become the prefix.
  3. 03The model scores possible next tokens and samples one. Then again. Then again.
  4. 04If tools are enabled, it may emit a function call instead of a final sentence. Your app runs the function and the loop continues.
  5. 05If DeepSearch is on, a retrieval step injects pages into that prefix before or during the loop.
  6. 06You see streamed tokens. That is not a cursor into a stored answer. It is the answer being born.

The 'thinking' some models show is more generated tokens, usually hidden or partially shown, used as scratch. Useful. Still generated.