Hacker News Reader

Comment by csours | Hacker News Reader

A new triangle:

    Accurate
    Comprehensive
    Satisfying

In any particular context window, you are constrained by a balance of these factors.

I'm not sure this works. Accuracy and comprehensiveness can be satisfying. Comprehensiveness can also be necessary for accuracy.

They CAN work together. It's when you push farther on one -- within a certain size of context window -- that the other two shrink.

If you can increase the size of the context window arbitrarily, then there is no limit.

Not sure what you mean by “satisfying”. Maybe “agreeable”?

Satisfying is the evaluation context of the user.

Many would be satisfied by an LLM that responds accurately and comprehensively, so I don’t understand that triangle. “Satisfying” is very subjective.

And LLMs are pretty good at picking up that subjective context

How would they possibly do that on a first prompt? Furthermore, I don’t generally let LLMs know about my (dis)satisfaction level.

This all doesn’t make sense to me.