LLMs are terrible at designing AI systems

(seangoedecke.com)

6 points | by saikatsg 7 hours ago ago

5 comments

  • pryanshu89 7 hours ago ago

    "the vast majority of the code that models learn from was written before LLMs were an option."

    Are we heading towards a future where LLMs learn from their own messy output and make code quality progressively worse?

  • uain 5 hours ago ago

    How did you distinguish contradictions from facts changing over time? An old address and a new address can both be valid records. I'd want examples of those cases in the evaluation before deciding whether the checker should be symbolic or model-based.

  • nandanadileep29 6 hours ago ago

    very clearly explained why llms would just complicates very first system design task before creating a system

  • calf 3 hours ago ago

    Just thinking aloud but isn't that equal to detecting a self-contradictory sentence and thus equal to solving the halting problem thus computationally undecidable in general (& in principle, etc.)?