After hearing Sebastian Bubeck talk about the #SparksOfAGI paper today, I decided to give #GPT4 another chance.
If it can really reason, it should be able to solve very simple logic puzzles. So I made one up. Sebastian stressed the importance of asking the question right, so I stressed that this is a logic puzzle and didn't add anything confusing about knights and knaves.
Still, it gets the solution wrong.
Just for fun, here's a #KnightsAndKnaves version of the same puzzle.
It does no better.
(Actual solution: Only a knight can say "at least one of us is a knave", so 4 is a knight and indeed there is at least one knave. Since Jailer 2 says that Jailer 1 is a knight, Jailers 1 and 2 have to be the same type. If both are knaves, then jailer 3 must be a knight (otherwise 1's statement would be true). If both are knights, jailer 3 must be a knave. So the tiger is behind door 1 or 3. Open door 2.)
Ok, to be fair it's not systematically wrong. I ran the original, simple-language problem six times. Four times it incorrectly told me to open doors 2 and 3. Once it incorrectly told me to open doors 1 and 2. And once it correctly told me to open doors 1 and 3.
Even in this final case the logic itself was flawed.
@ct_bergstrom
Wait! Is the probability to get a correct answer by random choice not 2/3?
So GTP4 is not only wrong sometimes, but it is considerably more wrong than rolling dice 😱
Carl T. Bergstrom
Steve Dodge
Claudius Link
standev