When I plug in the training prompt from the technical report (last page of the paper) into the free version of chatgpt, it gives a response that seems very similar to what five-thirty nine says. This is despite my chatgpt prompt not including any of the sources retrieved.
Have I interpreted this wrong, or is it possible that the retrieval of sources is basically doing nothing here?
EDIT: I did another experiment that seems even more damning: I asked an even simpler prompt to chatgpt: “what is the probability that china lands on the moon before 2050? Please give a detailed analysis and present your final estimate as a single number between 0% and 100%”
The result is a very detailed analysis and a final answer of 85%.
Asking the same question to five-thirty nine “what is the probability that china lands on the moon before 2050?”, I get a response of pretty much the same detail, and the exact same final answer of 85%.
I’ve tried this with a few other prompts and it usually gives similar results. I see no proof that the sources do anything.
When I plug in the training prompt from the technical report (last page of the paper) into the free version of chatgpt, it gives a response that seems very similar to what five-thirty nine says. This is despite my chatgpt prompt not including any of the sources retrieved.
Have I interpreted this wrong, or is it possible that the retrieval of sources is basically doing nothing here?
EDIT: I did another experiment that seems even more damning: I asked an even simpler prompt to chatgpt: “what is the probability that china lands on the moon before 2050? Please give a detailed analysis and present your final estimate as a single number between 0% and 100%”
The result is a very detailed analysis and a final answer of 85%.
Asking the same question to five-thirty nine “what is the probability that china lands on the moon before 2050?”, I get a response of pretty much the same detail, and the exact same final answer of 85%.
I’ve tried this with a few other prompts and it usually gives similar results. I see no proof that the sources do anything.