Idk it's pretty reliable even now if you chuck a vector db of "knowledge" in and inform the GPT in the overall prompt that it is must not go outside the bounds

Idk it's pretty reliable even now if you chuck a vector db of "knowledge" in and inform the GPT in the overall prompt that it is must not go outside the bounds of knowledge provided with the user's query, and that if no hard knowledge is provided that it should respond along the lines of "I don't know" (or search the internet and parse results for them). I imagine at some point this behaviour will be built in. People are treating GPTs like they're knowledge bases, but they're not. The fact that they can answer simple or complex questions correctly is only a byproduct of the transformer being taught how to string language together. It's like an artist who learns to draw by making paintings of raccoons. You can't then ask them "what do raccoons like to eat?" or "what foods are poisonous to a raccoon?" just because they learnt to draw by painting raccoons. This is how people are treating GPTs atm. They believe in it because they ask the artist "what colour fur do raccoons have?" and because the artist can answer that correctly, they assume all other answers are factual.