Loading slide
The chat felt like one helpful being. Now we can see the parts:
The app holds the conversation together.
The context carries information into each reply.
The provider runs the model on its computers.
Rufus, the LLM scores possible next tokens.
The sampler chooses one of them.
Tool programs perform actions outside Rufus.
Remove the surrounding parts and Rufus can only receive tokens and score what might come next.
Loading contents...