Getting the Right Answer from ChatGPT – O’Reilly

A few days in the past, I used to be fascinated about what you wanted to know to make use of ChatGPT (or Bing/Sydney, or any related service). It’s straightforward to ask it questions, however everyone knows that these massive language fashions continuously generate false solutions. Which raises the query: If I ask ChatGPT one thing, how a lot do I must know to find out whether or not the reply is appropriate?

So I did a fast experiment. As a brief programming undertaking, quite a lot of years in the past I made an inventory of all of the prime numbers lower than 100 million. I used this checklist to create a 16-digit quantity that was the product of two 8-digit primes (99999787 occasions 99999821 is 9999960800038127). I then requested ChatGPT whether or not this quantity was prime, and the way it decided whether or not the quantity was prime.

Study sooner. Dig deeper. See farther.

ChatGPT accurately answered that this quantity was not prime. That is considerably stunning as a result of, when you’ve learn a lot about ChatGPT, you already know that math isn’t one in all its sturdy factors. (There’s most likely an enormous checklist of prime numbers someplace in its coaching set.) Nonetheless, its reasoning was incorrect–and that’s much more attention-grabbing. ChatGPT gave me a bunch of Python code that carried out the Miller-Rabin primality take a look at, and stated that my quantity was divisible by 29. The code as given had a few fundamental syntactic errors–however that wasn’t the one drawback. First, 9999960800038127 isn’t divisible by 29 (I’ll allow you to show this to your self). After fixing the plain errors, the Python code seemed like an accurate implementation of Miller-Rabin–however the quantity that Miller-Rabin outputs isn’t an element, it’s a “witness” that attests to the very fact the quantity you’re testing isn’t prime. The quantity it outputs additionally isn’t 29. So ChatGPT didn’t really run this system; not stunning, many commentators have famous that ChatGPT doesn’t run the code that it writes. It additionally misunderstood what the algorithm does and what its output means, and that’s a extra severe error.

I then requested it to rethink the rationale for its earlier reply, and obtained a really well mannered apology for being incorrect, along with a distinct Python program. This program was appropriate from the beginning. It was a brute-force primality take a look at that attempted every integer (each odd and even!) smaller than the sq. root of the quantity beneath take a look at. Neither elegant nor performant, however appropriate. However once more, as a result of ChatGPT doesn’t really run this system, it gave me a brand new checklist of “prime elements”–none of which had been appropriate. Apparently, it included its anticipated (and incorrect) output within the code:

      n = 9999960800038127
      elements = factorize(n)
      print(elements) # prints [193, 518401, 3215031751]

I’m not claiming that ChatGPT is ineffective–removed from it. It’s good at suggesting methods to unravel an issue, and may lead you to the suitable answer, whether or not or not it offers you an accurate reply. Miller-Rabin is attention-grabbing; I knew it existed, however wouldn’t have bothered to look it up if I wasn’t prompted. (That’s a pleasant irony: I used to be successfully prompted by ChatGPT.)

Getting again to the unique query: ChatGPT is nice at offering “solutions” to questions, but when you could know that a solution is appropriate, it’s essential to both be able to fixing the issue your self, or doing the analysis you’d want to unravel that drawback. That’s most likely a win, however it’s important to be cautious. Don’t put ChatGPT in conditions the place correctness is a matter until you’re keen and capable of do the arduous work your self.