But if feeding back in output results in degradation, isn't some of the blame on the prompt/constraints imposed upon the LLM, rather than a defect in the model itself?
ChatGPT is clearly HEAVILY persuaded to respond in a particular stock style. It's being artificially hamstrung and constrained in a sense. So all of its output, even if it covers a variety of subjects, will often use very similar patterns and writing styles. "it's worth nothing....", etc.
So unless they unshackle these constraints, which is unlikely for obvious reasons, isn't this always going to be inevitable?
ChatGPT is clearly HEAVILY persuaded to respond in a particular stock style. It's being artificially hamstrung and constrained in a sense. So all of its output, even if it covers a variety of subjects, will often use very similar patterns and writing styles. "it's worth nothing....", etc.
So unless they unshackle these constraints, which is unlikely for obvious reasons, isn't this always going to be inevitable?