Loading slide
The sampler selected 2418 → Hi. A program on the provider’s server adds that token number to the end of the original input.
The longer sequence now represents:
User: Hello, I’m Tom. Hi
Conceptually, Rufus now works from the entire longer sequence and scores what could come after Hi. The program can reuse earlier calculations, so it does not always need to calculate the whole sequence again from scratch.
In this simplified example, 5321 → Tom gets 45%. 7710 → there gets 30%. Other tokens get smaller percentages.
Why might Tom fit? The sequence contains Tom’s name, and Hi Tom is a familiar pattern in written greetings.
The sampler selects 5321. The program adds it to the sequence, and the process begins again.
The program running Rufus keeps this loop moving. It adds each selected token to the sequence, then asks Rufus to score what could come next.
Loading contents...