Last one on AI for a while

I've written a few posts on AI lately. They have proved surprisingly popular, more so than my usual guff about some old song or other that I like. Although when I say popular, what I mean is they've had a lot of traffic and comments. I'm not trying to suggest the coming omnipotence of AI is popular, because it is far from that amongst my modest readership. But indulge me one more time. I read about a very basic limitation of one AI engine, and so had to try it out for myself. How's your counting, because ChatGPT's isn't always very good...

Oh dear. The correct answer is four, as any fule kno. So what's gone wrong? If I ask how many "z" are in "zzzzzz" it correctly answers six, so it's probably not to do with repeated letters. Maybe for words that it doesn't immediately recognise (and there are many variant spellings of pizzazz) it looks to the dictionary definition and counts from that? Or from the pronunciation pɪzæz?

Maybe I could teach ChatGPT to count...

Which, to my programmer's mind, suggests it is probably missing the last "z". Think for loops versus while loops, maybe. Either way, seems I'm still asking the question wrong. Let's try again. If I was programming this task, let's say in JavaScript, it's a doddle to iterate through a string and count how many times a certain letter appears. Look, I did it here if you want to learn some basic code - as you can see, it correctly returns 4. But what does ChatGPT make of it?

Which is really, really disappointing. For whilst it correctly understands and describes the purpose of the function, rather than run through the code it shortcuts straight to what it thinks it already knows...

This is a very human response though - to have a preconception of what is right and assume it to be true, rather than retest one's understanding. Confirmation bias, in other words, and we all do it at times. I hope, for our sake, that ChatGPT is not exhibiting egocentric bias... Whatever, it seems that the real skill in these early days of AI is asking the question correctly. And then, just maybe, being sceptical about the answer.

Post script

ChatGPT, backed by Microsoft, struggles, as per the above. Bard, developed by Google, gets it right first time, as below. Not only that, it also offers a snippet of code to prove its answer, inadvertantly demonstrating that my programming skills will soon be obsolete too...