News for hackers. Stuff that matters.
Re:Experiencing Decreased Performance with ChatGPT-4 by carabiner on Friday July 07, 2023 @01:37PM (#36635341)
You can't evaluate them with a single prompt, single execution! Any given output is just a sample from a range of possible outputs, but all of them (ALL) are considered plausible returns. You have to think of it as returning a distribution as a random variable (in high dimensional space) but unfortunately there's only 1 sample from it at a time. Maybe some day they'll let us control a random seed.

There are no comments on this comment yet.