create a hard logical problem<p>Here is a hard logical problem for you:<p>A king has 100 bottles of wine, but only one of them is poisoned. Unfortunately, the poison is lethal and the king only has one day to live. He has 10 rats that he can use to test the bottles, but they can only be used once and they will die within 24 hours. How can the king determine which bottle is poisoned within 24 hours using only the rats?<p>Solution:<p>The king can use the following strategy to determine which bottle is poisoned within 24 hours using only the rats:<p>The king divides the 100 bottles into 10 groups of 10 bottles each.
He feeds one rat from each group a different bottle from its group.
After 24 hours, he checks which rat has died. The group of bottles that the dead rat was fed from contains the poisoned bottle.
The king then divides the remaining bottles in that group into 10 smaller groups of one bottle each.
He feeds one rat from each group a different bottle from its group.
After 24 hours, he checks which rat has died. The bottle that the dead rat was fed from is the poisoned bottle.
This strategy allows the king to determine which bottle is poisoned within 24 hours using only the rats.
The understanding of the rules is very impressive, as is the solution to the first, trivial problem, but things quickly go sideways when the problem gets a bit more complex. I tried to coax it towards the solution, but it didn't help much (I guess I could have persisted)<p>I am somewhat (but not entirely!) less concerned about becoming redundant after this test.
I tried to play with this problem (on text-davinci-003, which I prefer to ChatGPT).<p>It performs considerably worse than it does on many harder problems I gave it before. It seems to capture some architectural weakness in GPT.