4 ms·

YMMV, but outside of very specialized use cases, it's just not all that useful to the extent that I'd default to using it. I'm building a travel planning app a

YMMV, but outside of very specialized use cases, it's just not all that useful to the extent that I'd default to using it. I'm building a travel planning app and partner and I toyed around with ChatGPT for generation of itineraries and both agreed that the output was generally OK if only using it for identifying destinations, but otherwise mediocre and not at all how a human would plan travel. It tends to be repetitive in how it picks places for a given destination. Often times, it would hallucinate especially when given constraints like distance (because you'd probably want to get a meal near a destination). There are dozens of startups built around this and even one backed by YC. For coding, it's useful to generate small utility functions, but in many cases, I could have written the utility function or found a snippet on SO instead in roughly equivalent time. Would love to know how HN crowd is finding consistent utility out of the service.

spaceman_2020 3y ago
It’s genuinely become a pretty mediocre tool for most things. The text it generates is so formulaic that anyone can spot its AI. It defaults to shallow coding answers. It refuses to answer anything medical or legal without some hardcore prompt engineering. Either its being dumbed down, or OpenAI is genuinely struggling to keep it consistently good. If its the former, it just scares me that a bunch of people have access to a super powerful tool that they’ve intentionally decided to keep from the public. Way too much power in the hands of a few.
GenerocUsername 3y ago
That's a really bad usecase for a LLM. A better use would be to suggest activities or destinations based on qualities of those places and a users interests. It can easily generate rough matches, but you would not want to delve into specific mileage based computations.
- CharlieDigital 3y ago
  > A better use would be to suggest activities or destinations based on qualities of those places and a users interests. For sure, but as spaceman mentions, after a few passes, you can see how formulaic it is. > ...but you would not want to delve into specific mileage based computations. Agree; but it was our own curiosity to see if it could understand geo-spatial locality for this use case. It does have some sense of it likely based on how destinations tend to cluster in the underlying sources, but will obviously include hallucinations in the response as well.

spaceman_2020 3y ago

It’s genuinely become a pretty mediocre tool for most things. The text it generates is so formulaic that anyone can spot its AI. It defaults to shallow coding answers. It refuses to answer anything medical or legal without some hardcore prompt engineering. Either its being dumbed down, or OpenAI is genuinely struggling to keep it consistently good. If its the former, it just scares me that a bunch of people have access to a super powerful tool that they’ve intentionally decided to keep from the public. Way too much power in the hands of a few.

GenerocUsername 3y ago

That's a really bad usecase for a LLM. A better use would be to suggest activities or destinations based on qualities of those places and a users interests. It can easily generate rough matches, but you would not want to delve into specific mileage based computations.

CharlieDigital 3y ago
> A better use would be to suggest activities or destinations based on qualities of those places and a users interests. For sure, but as spaceman mentions, after a few passes, you can see how formulaic it is. > ...but you would not want to delve into specific mileage based computations. Agree; but it was our own curiosity to see if it could understand geo-spatial locality for this use case. It does have some sense of it likely based on how destinations tend to cluster in the underlying sources, but will obviously include hallucinations in the response as well.