Prompts are boring. Also, avoidable.

GenAI and LLMs altered a way of working of many developers. Typically, we all started by manually fine-tuning prompts and learning how to better express needs to model. That approach (even though working and useful) is not perfect - models can be updredictable in understanding their task.

Recently, best way of prompting evolves through “Show, not explain” approach into Automated Prompt Engineering.

Due to recent advantages as “Meteor” framework, human is responsible less for a prompt itself, but more for providing training data.

How far we can get in avoiding prompts? Can we use pure loss functions and hope LLM generates best prompt for himself?