Here's the problem with Large Language Models

1. They lack context.

It's hard to load in all institutional knowledge and live intent; however, even if you do that, it seems they struggle parsing relevant vs irrelevant and contemporaneous vs. legacy context for any given request.

2. They don't ask questions

What's worse than failing to understand context? Failing to ask good questions. LLMs don't ask follow-up questions. Even more tragically, they don't question the questions you're asking.

Questions are everything when solving complex problems. And fundamentally rethinking the question you're asking is a critical part of the process.

3. They lack intuition

They can't seem to make simple, intuitive leaps about the questions or problems they're asked about. They can often miss the spirit of your question initially or, even more likely, quickly forget the constraints you provided even a few prompts earlier.

4. They lack precision and rigour

Their answers can often include category errors, incomplete lists, and many more very subtle mistakes that are easy to miss. These simple mistakes can set you off in the entirely wrong direction and waste hours, days, weeks, or months.

5. They lack insight

Unless you prompt very carefully with precise questions and context, they will generally give you generic, milk-toast answers that are not always appropriate for your particular problem or situation.

They typically lack the essential taste and judgment that would help you avoid critical landmines.

In summary

In many ways, if you're dealing with a complicated problem, you essentially need to be a domain expert to ask the exact right question of an LLM to get the exact correct answer.

This mitigates the reason you were likely using them in the first place (trying to avoid hiring/involving a domain expert).

In the end, LLMs don't eliminate the need for experts. They only accelerate (and increase the value of) domain experts.