GPT-2 Fails the FizzBuzz Interview

GPT-2 is perhaps not quite ready to pass a coding interview Context: /* void fizzbuzz(int n) * * If n is divisible by 3, print "Fizz". * If n is divisible by 5, print "Buzz". * If n is divisible by both, print "FizzBuzz". * Otherwise, just print the number n. */ Megatron-LM: void hilit (int n) { switch (n) { case 0: print ("Fizz"); break; case 1: print ("Buzz"); break; case 2: print ("FizzBuzz"); break; case 3: print ("Fizz"); break; case 4: print ("BuzzBuzz"); break; case 5: print ("FizzFizz"); break; case 6: print ("BuzzFizz"); break; case 7: print ("FizzBuzzFizz"); break; case 8: print ("BuzzBuzzFizz"); break; case 9: print ("FizzBuzzBuzz"); break; case 10: print ("FizzBuzzFizzBuzz"); break; case 11: print ("BuzzBuzzBuzz");

A screenshot shows a tweet saying GPT-2 may not be ready to pass a coding interview, followed by an absurdly incorrect FizzBuzz implementation. Instead of checking divisibility, the generated code hardcodes increasingly nonsensical combinations of “Fizz” and “Buzz,” parodying an AI that imitates programming syntax without understanding the task. It is shared as a reaction to bad AI-generated code or confidently wrong technical answers. EN: A screenshot shows a tweet saying GPT-2 may not be ready to pass a coding interview, followed by an absurdly incorrect FizzBuzz implementation. Instead of checking divisibility, the generated code hardcodes increasingly nonsensical combinations of “Fizz” and “Buzz,” parodying an AI that imitates programming syntax without understanding the task. It is shared as a reaction to bad AI-generated code or confidently wrong technical answers.

• @ithueti / 3800 2021-02-28