Search the Community

Hi there, I received my OpenAI beta test account recently, after waiting for that more than a year now. Finally I was very much excited about the AI capabilities, as I watched the amazing Youtube series of GPT3 Leta from Dr. Alan Thompson, who is experimenting with the DaVinci model and avatars to provide a real human chat experience. This machine will easily pass the turing test IMHO. Well, but when I first checked the Playground, I found the GPT3 models were not so much evolved as I expected. I also tried out the REST API, with the same disapointing outcome. GPT3 often gives completely wrong answers, mostly doesn't understand anything at all, not even a name or very basic texts, and all in all its far from capabilities of the Leta series I mentioned above. So maybe this is the case because I have not provided enough or wrong training data, but there is also no real explanation howto provide information, and especially what data would be really needed, to enable GPT3 to deliver most useful results. In the meantime during waiting for GPT3 access, I already tried GPT-J and tried others, but they all were not especially "intelligent" neither in understanding normal texts nor in providing real world answers (like: who won the latest nobel price in physics ?). Now I would say GPT3 is on a same level as GPT-J, at least when it comes to a unspecific chat application. What is Dr. Thompson's secret to make it so clever ? I assume that all GPT3 models use the same training data ( it was mentioned that training data was used only from the time until year 2018 ), so that also Dr. Thompson's machine should be using the same data. I wanted to ask if anybody here in DP had also some AI experiences, and maybe could squeeze better results out of GPT3 ? Would be great to exchange some tips and hints. howto handle AI best.