1:18:37 for me, looking up specific facts/raw information ("What/where did [specific Church Father] say/write about [specific thing]?") has been the biggest boon of this new tool, the only way to really trust it, and the best way to bypass its non-sentience-ness (asking a robot for moral/theological "answers" or, seen in your experience, for a totalizing definitive when there exists a multiplicity of views and explanations). an unbiased/unbought, vastly more precise, and agent-ized version of Google that is really good at finding a needle in a haystack if you know how to ask it the right question in the right way.
The current version of ChatGPT still hallucinates a fair bit--so it will often misattribute quotes/ideas or just make something up completely. There are ways to hook LLMs to a specific corpus of data to draw answers directly from in context instead of trying to indirectly imply a response from its trained model weights. But to your point--it is extremely "autistic"/left-brained, and so asking pinpoint questions that aren't open ended is the best use case. I realize that even though my objections were specific, the model treats them in a general way (perhaps because there are a lot of priors that need to be sussed out first before the actual answer) which is I think why it was becoming repetitive.
1:18:37 for me, looking up specific facts/raw information ("What/where did [specific Church Father] say/write about [specific thing]?") has been the biggest boon of this new tool, the only way to really trust it, and the best way to bypass its non-sentience-ness (asking a robot for moral/theological "answers" or, seen in your experience, for a totalizing definitive when there exists a multiplicity of views and explanations).
an unbiased/unbought, vastly more precise, and agent-ized version of Google that is really good at finding a needle in a haystack if you know how to ask it the right question in the right way.
The current version of ChatGPT still hallucinates a fair bit--so it will often misattribute quotes/ideas or just make something up completely. There are ways to hook LLMs to a specific corpus of data to draw answers directly from in context instead of trying to indirectly imply a response from its trained model weights.
But to your point--it is extremely "autistic"/left-brained, and so asking pinpoint questions that aren't open ended is the best use case. I realize that even though my objections were specific, the model treats them in a general way (perhaps because there are a lot of priors that need to be sussed out first before the actual answer) which is I think why it was becoming repetitive.