> High risk would be reading legal documents and recommending what the lawyer should do. Explored this myself, and it's a definite no. GPTs cannot do this reli

> High risk would be reading legal documents and recommending what the lawyer should do. Explored this myself, and it's a definite no. GPTs cannot do this reliably (we've tested multiple LLMs). Simply put, it's because it's a rule-extraction problem, and not a generative problem, and rule-extraction situations are typically too sensitive for a "99%" solution like LLMs. Basically, that 1% failure rate is enough to screw any gains you get from the 99% successes. For years, all I've ever heard from the DS space was to always be mindful that the model you apply to a problem fits the nature of the problem. The past few months, this has somehow become a lost wisdom. "LLMs for everything" For this reason, HN comments lately have had my pupils permanently fixated to the top of my eye sockets.