Hacker News Reader

Comment by b800h | Hacker News Reader

I wonder whose code this was trained on? I guess it's interesting that this is quite a popular search result on Github:

https://github.com/search?q=plane+spotter+language%3AJavaScr...

Exactly. The only interesting thing here is that it managed to convince this person that "they built" anything. Also it seems there are tons of blogs about this exact task. I guess it's slightly interesting that there now exists a way to find example code snippets besides Google, but it's funny that the supposed insight here is that the important thing is implementation of novel ideas when that's exactly the opposite of what has been demonstrated

Outside of any attribution issues, I think it's a cool capability to be able to get a concise how-to vs having to search for and read a blog. I've asked chatgpt for recipes, I know I could find them online and that neither I nor gpt are inventing anything, getting a concise reply is just the most convenient format.

I completely agree with you. LLMs are great at combining the content of multiple sources, which saves a lot of time. OP was able to say "OK, you've created the app, but I'd like it to use a more modern design." It's not that a dev can't do that with a few Google searches, but it certainly takes longer. Same thing with recipes, you can say, "Give me a recipe for gazpacho that replaces the tomatoes with something else for my friend who's allergic to nightshades."

Exactly… likely copy pasted from someone’s github repo with 0 attribution. A brave new world…

OP here. I think it's a bit reductive to think that GPT simply copy pastes. It sometimes does. But my experience working with it nearly daily since December, is that it's also quite capable at reasoning. My sense is it most likely saw the API endpoint in another context, learned how it was parsed, and then adapted it to my situation. Perhaps it even parsed the documentation page itself.

Since attribution isn't possible in ChatGPT, we'll never know. It's a great way to avoid all IP claims via plausible deniability.

This topic has been beaten to death though, and it isn't changing. Companies have little incentive to have explainable AI since it would open them to infringement claims.

That is absolutely not how GPT works.

If the code has been done hundreds or even thousands of times, what meaningful attribution could possibly be added? Like, we don’t add attribution to each for loop even though it’s not our original idea.

https://github.com/konstantintzt/planetracker/blob/9a1f407cb...

At least this person does not fetch all planes and then iterate over all of them but sets boundaries in which area to get all the planes.

Both the GitHub project you linked to and the webpage in the posted article use a bounding box; has the webpage changed in the last two hours?

No it hasn’t. ChatGPT from the get go used a bounding box. There is no way you could query all planes on Earth in one query on a free API

Reply on news.ycombinator.com