Hacker News Reader

When you ask again the prompt includes the context of your previous question and correction. When I ask the prompt doesn't have that context so the model fails to give me the correct answer.

I'm using the default free model in the app, based on GPT4.

View on HN

Yup - well this is where my suggestion is to change the GPT architecture. You can think of having a logic program function as the 'frontal lobe' of the general pre-trained auto-associative model. This 'frontal lobe' region would try to come up with logical sequences to go along with it's internal auto-associative representations. Of course - the logic programming piece is just one approach - maybe chain of though or chain of reason prompting could work here too as many humans I think use this chain-of-reasoning approach themselves. Logic programming to me would function as a suggested shortcut.

Reply on news.ycombinator.com