4 ms·

It uses AI for some things you can't really use free APIs for: - Turn a colloquial name for a neighborhood into the exact latitude and longitude (I'm sure ther

by blixt 2y ago

It uses AI for some things you can't really use free APIs for: - Turn a colloquial name for a neighborhood into the exact latitude and longitude (I'm sure there's some API that can kind of do this but I don't know a free one that is as accurate as GPT) - Adding to the previous point, if you say "Hogwarts" or "Tatooine" it will dutifully give you the weather of those too thanks to AI [1] - Most importantly, write a Dall-E compatible prompt to generate an image with the actual weather conditions (and sun/moon etc) [1]: https://x.com/blixt/status/1724099967361462714 https://x.com/blixt/status/1724099967361462714

gandalfgreybeer 2y ago
Of these, only the first one actually has any real value. And as you've mentioned there are likely less "expensive" ways to do so. By expensive, I mean this is essentially using a sledgehammer to crack a nut (especially the Dall-E call).
- blixt 2y ago
  I don't understand how I would get a unique but accurate prompt that makes Dall-E produce an image that besides the weather/lighting conditions also has a randomly picked iconic scene for the location, people in the picture that have the correct clothing and doing reasonable activities, etc. I'm trying to understand the criticism here, so please give the GPT a shot with a few different locations (you can also do neighborhoods or fictional locations) and describe how this is a nut that can be cracked with non-LLM solutions: https://chatgpt.com/g/g-QIydQSFRm-sol-mate https://chatgpt.com/g/g-QIydQSFRm-sol-mate
  - itishappy 2y ago
    Google images? You could run a classifying model if you wanted to segment by time of day and weather.
  - stetrain 2y ago
    If your goal was wanting to use AI image generation to create such an image, this makes sense. I think people are just questioning the assumption that such a feature is a desirable or necessary feature of a weather display. I definitely was a bit surprised to open the repo to a weather display and see the need for an OpenAI key without an explanation of which features were AI-powered. But of course this is a personal project, so "Generate a representative AI image using the location and weather information" is a perfectly valid, and cool, challenge to set yourself and achieve.
itishappy 2y ago
From your first point, I'd expect you to be using a service that provides the weather for a given latitude and longitude. From your second point, I'd have expected that to fail for Hogwarts or Tatooine. Taken together, it sounds like you're blindly trusting the coordinates generated by AI, and will happily generate correct-looking results for any input. I'd call that hallucinating the weather. If I type in "Springfield" I'm gonna get results but I'm not gonna know if they're mine. This may be an unreasonable level of concern for a fun app.
- blixt 2y ago
  I do use a service for getting the weather (backend: https://github.com/blixt/sol-mate https://github.com/blixt/sol-mate) Yeah the second part I consider a bonus. Especially because I can use street names or slang for locations that no rigid API supports. > Taken together, it sounds like you're blindly trusting the coordinates generated by AI Actually I guess I do "blindly" trust the AI's coordinates more than a traditional location search service. Not sure how much you've tested it yourself, but I haven't gotten a single false latitude/longitude pair from GPT-4 unless the location was truly ambiguous, in which case it would be one of them. But as a European, I've definitely experienced rigid APIs happily giving me the weather for "Stockholm, NY" which apparently does exist, but wouldn't have been my first option... But yeah, it is just a fun app, so I wouldn't want to make it more rigid either way!
  - itishappy 2y ago
    Have you stress tested it? "Springfield" is a very common name in the US. I'd love to know how it handles "Sandwich" as well. Again, probably overboard for a fun app!
    - blixt 2y ago
      They're fair questions, and yeah it will have to disambiguate in such cases. I tried it now and it picked: - Springfield in Illinois - Sandwich in Massachusetts (the exact coords: https://www.google.com/maps/place/41%C2%B045'31.0%22N+70%C2%B029'36.2%22W https://www.google.com/maps/place/41%C2%B045'31.0%22N+70%C2%...) I think that people underestimate how much non-AI technology fails in these cases. And the recovery path is to retry with more specificity in either case.
rideontime 2y ago
Moistere Vapeers, very believable.

by blixt 2y ago

gandalfgreybeer 2y ago

Of these, only the first one actually has any real value. And as you've mentioned there are likely less "expensive" ways to do so. By expensive, I mean this is essentially using a sledgehammer to crack a nut (especially the Dall-E call).

blixt 2y ago
I don't understand how I would get a unique but accurate prompt that makes Dall-E produce an image that besides the weather/lighting conditions also has a randomly picked iconic scene for the location, people in the picture that have the correct clothing and doing reasonable activities, etc. I'm trying to understand the criticism here, so please give the GPT a shot with a few different locations (you can also do neighborhoods or fictional locations) and describe how this is a nut that can be cracked with non-LLM solutions: https://chatgpt.com/g/g-QIydQSFRm-sol-mate https://chatgpt.com/g/g-QIydQSFRm-sol-mate
- itishappy 2y ago
  Google images? You could run a classifying model if you wanted to segment by time of day and weather.
- stetrain 2y ago
  If your goal was wanting to use AI image generation to create such an image, this makes sense. I think people are just questioning the assumption that such a feature is a desirable or necessary feature of a weather display. I definitely was a bit surprised to open the repo to a weather display and see the need for an OpenAI key without an explanation of which features were AI-powered. But of course this is a personal project, so "Generate a representative AI image using the location and weather information" is a perfectly valid, and cool, challenge to set yourself and achieve.

itishappy 2y ago

From your first point, I'd expect you to be using a service that provides the weather for a given latitude and longitude. From your second point, I'd have expected that to fail for Hogwarts or Tatooine. Taken together, it sounds like you're blindly trusting the coordinates generated by AI, and will happily generate correct-looking results for any input. I'd call that hallucinating the weather. If I type in "Springfield" I'm gonna get results but I'm not gonna know if they're mine. This may be an unreasonable level of concern for a fun app.

blixt 2y ago
I do use a service for getting the weather (backend: https://github.com/blixt/sol-mate https://github.com/blixt/sol-mate) Yeah the second part I consider a bonus. Especially because I can use street names or slang for locations that no rigid API supports. > Taken together, it sounds like you're blindly trusting the coordinates generated by AI Actually I guess I do "blindly" trust the AI's coordinates more than a traditional location search service. Not sure how much you've tested it yourself, but I haven't gotten a single false latitude/longitude pair from GPT-4 unless the location was truly ambiguous, in which case it would be one of them. But as a European, I've definitely experienced rigid APIs happily giving me the weather for "Stockholm, NY" which apparently does exist, but wouldn't have been my first option... But yeah, it is just a fun app, so I wouldn't want to make it more rigid either way!
- itishappy 2y ago
  Have you stress tested it? "Springfield" is a very common name in the US. I'd love to know how it handles "Sandwich" as well. Again, probably overboard for a fun app!
  - blixt 2y ago
    They're fair questions, and yeah it will have to disambiguate in such cases. I tried it now and it picked: - Springfield in Illinois - Sandwich in Massachusetts (the exact coords: https://www.google.com/maps/place/41%C2%B045'31.0%22N+70%C2%B029'36.2%22W https://www.google.com/maps/place/41%C2%B045'31.0%22N+70%C2%...) I think that people underestimate how much non-AI technology fails in these cases. And the recovery path is to retry with more specificity in either case.

rideontime 2y ago

Moistere Vapeers, very believable.