Make AI evaluation of rules un-fuzzy

English Forum Product Feedback
1

I think AI Studio exibits some problematic behaviour that makes it less predictable, less efficient, and inconsistent with how normal rules work. (And how you’d expect a computer to behave)

I think the following rules summarise how a non-AI Studio rule behaves:

  • If a rule condition branch evaluates to true, it executes the action(s).
  • If the branch condition evaluates to false, it moves to the next branch for evaluation, assuming there is a next branch.
  • If there are multiple conditions in a branch it evaluates the options from top to bottom.
  • For multiple conditions in a branch separated with and:
    – it moves on to the next condition within the branch whenever the current condition evaluates to true, as the branch can still evaluate to both true and false.
    – it stops evaluating and moves to the next branch as soon as it finds a condition that evaluates to false. This makes sense as an and statement cannot evaluate to true when any of the conditions within it evaluate to false.
  • For multiple conditions in a branch separated with or:
    – it moves on to the next condition within the branch whenever the current condition evaluates to false, as the branch can still evaluate to both true and false.
    – it stops evaluating and moves to the next branch as soon as it finds a condition that evaluates to true. This makes sense as an or statement only requires one of it’s conditions to be true to evaluate to true.

All of this is good, and logical. But that’s not how conditions with AI (studio) evaluations seem to work. They seem to operate with some kind of fuzzy logic where it evaluates all conditions in all branches, and eventually decides which branch is most true.

This is problematic for a couple of reasons:

  • It’s expensive, as it requires needless computer power (and tokens)
  • It’s inaccurate, as rules can’t be designed with gates where you can assume it only reaches branches further along in the rule when it has sucessfully concluded that earlier branches should not be applied. (hence “Otherwise if”)

For consistency, accuracy, and efficiency I argue that AI evaluation of rules should follow regular conventions for if, else if, and, and or statements.

Another way of putting it would be that I think AI conditons should just evaluate to a hard true or false, and these should be respected without requiring re-evaluation at the end.

2 Likes
2

@Forum-team, is there any chance you could reach out for more info about this? Is there a reason why this isn’t possible?

I guess a workaround would be the same one as we have to use in some other cases of rules where we make a section change the “gate.” The first rule ends with changing the section so things can reset, then the section rule takes over upon arrival in the new section. Of course, this is not ideal (hence, workaround), and perhaps is insufficient in others.

Interested to hear what’s at play here…

Thanks,

Larry

3

Many thanks for this @Jan-Rienk and @lpb

We’ll share this with our team and get back to you once we have an update. In the meantime, please note that you can report this directly to the AI Studio team by using the “Show reasoning” option in the AI activity feed. This sends the details straight to our developers along with the relevant logs.

For more information, feel free to check out this article.

2 Likes
4

Thanks!

:white_check_mark: