> We're essentially subject to the whims of someone who is letting us do something for as long as we play nice.
Isn't that so the spirit of the times when working with LLMs? If one asks for a Prisma schema or something else out of GipPTy meant to be fed into a traditional parser, we're up the the whims of the attention blocks and layers to write out something that doesn't become a parse error. One can turn the temperature down, fine-tune, or self-host a model but does that guarantee the syntax will be correct (much less the semantics)?
Isn't that so the spirit of the times when working with LLMs? If one asks for a Prisma schema or something else out of GipPTy meant to be fed into a traditional parser, we're up the the whims of the attention blocks and layers to write out something that doesn't become a parse error. One can turn the temperature down, fine-tune, or self-host a model but does that guarantee the syntax will be correct (much less the semantics)?