[tymonPartyLate]

I'm currently tinkering on a customer support bot with langchain and gpt3. The bot can answer questions about services and their terms, it can use tools to make bookings and perform some taks like scheduling appointments, in a conversational manner. It's becoming clear to me that subtle changes in the prompt can lead to bullshit answers and gpt making up facts, despite being specifically told not to do so. If the prompt reaches some complexity threshold, the output quality goes down visibly. I learned that I have to split the bot into subtasks, each having different, smaller, prompts. So, yeah, I believe prompt engineering can be a thing. At least for a while, until the models become smarter at understanding what we want from them :)