Nah, I don't think there's any way that LLMs can totally circumvent the problems of pervasive hallucination and low-quality output. There's a fundamental divide between the things LLMs would have to do to replace us (think critically, problem-solve) and the things they're designed to do (produce "realistic" text based on the data they were trained on). Sure, they can engage in a basic level of critical thinking and problem-solving, but I'd expect to see diminishing returns long before they can actually compete with humans. After all, we're not training them to produce smarter or more correct answers—just more verisimilitudinous ones. The local maximum is bland and uninsightful yet very plausibly human writing.
I'll worry when we come up with an architecture & training scheme that directly optimizes for problem-solving.