5 Comments
User's avatar
Jean Valjean's avatar

It's interesting that even when the LLM offers to do better, you actually still prefer the biased version of reality.

Expand full comment
Juliana Jackson's avatar

Hey, thanks for your comment! Of course, I’m biased. I was literally trying to explain how greedy decoding works. The LLM doesn’t 'want' to do better, as mentioned in the post, it just follows probability patterns based on its training data. That’s kind of the whole point. No reality, that’s simply pattern-matching at scale.

Expand full comment
Jean Valjean's avatar

Ah, but my comment was that after you proved the greedy decoding and it offered to do better, you still prefered it to be biased, n'est-ce pas?

Expand full comment
Juliana Jackson's avatar

I was greedy too? :D

Expand full comment
Jean Valjean's avatar

;-)

Expand full comment