Follow
@JamesGleick We're doomed.
Most of my experience with AI is using it to accelerate coding by asking it for an implementation of a straightforward, easily described algorithm, simply to save the time of coding it myself.
I then waste a bunch of time pointing out to it that not only does its code fail to meet my specifications in obvious, fundamental ways, -- which it clearly "understood" given that it accurately paraphrased those requirements back to me -- it fails to match its own description of what it was giving me. It apologizes, agrees with me, gives me another implementation that's just as wrong, sometimes in exactly the same way.
Rinse, repeat a few times. Then I give up and write it myself anyway.
@pieist @JamesGleick I find it's pretty good if I keep the problem simple and let it write one module at a time. But yep, I've boxed myself into "it's going to get it right this time" through too many iterations a few times.
I don't know if it's improving, or I'm learning how to write prompts, but it happens less and less for me.