ChatGPT bombs test on diagnosing kids' medical cases with 83% error rate

There's two layers of issues as I see it.

First is that most people are just not yet well-versed in prompt engineering. So it takes a while -- even as someone that is doing this every day -- to find the best prompting approach to get the desired outcome. Sometimes if I step away from a prompt for a few days and come back, I'll have better ideas. But this is doing it every day. Most folks just don't have this patience and are now conditioned with Google Search whereas GPT requires much more intent to get better results.

Most normal folks are not going to be in the right mindset to iterate and build optimal prompts for their use case because they are not really using GPT in that way; they are using it to get some task done.

Second is that some use cases -- like this case of diagnosing medical cases -- will require more advanced techniques to be used in combination like RAG, web search, collaborative agents, function calling, and so on. Basically, it still needs some level of design and engineering for it to be usable.

It is definitely not there yet for normal people and why I think there's still a moderately sized window where startups can close that gap by building that "structure" for specific use cases that make them viable. Essentially, vertically-oriented GPTs.

Well, it's not a tool for normal people to use _to diagnose their kids illness_. They can still ask it to come up with a poem about a science subject to read to their kids, or they can use it to have a conversational first dive into the surface layer aspects of a topic they're interested in (I've done this many times in physics and the results have never hallucinated, even when asking for book recommendations, just don't ask it to do a determinant).

Its definitely some kind of tool for some large aspect of society to use. At this point i think we've done a good job of dispelling the whole "agi is here and is about to make my job obselete" type of existential fear in the general populace. You don't see people talking like they did when it first came out anymore.

I do think that knowing what type of structure to add is essential. For instance, i basically always start with assigning it a role as a trusted high level scientist or clinician or whatever, I'm an interested student at some particular level tailored to the level of detail I'm seeking, and I'm about to ask you questions about this topic, etc.

Normal people are worse communicators than they realize, so that statement makes perfect sense to me.