The “correct” way to use AI for coding (and anything really) is to ask for explanations / tutorials when you can’t find one online, then learn from that.
except the “explanation” frequently will be 100% “hallucinated” bullshit
That’s why I always ask it to cite sources. Basically googld ATP since google is turning to shit and all other search engines still aren’t quite as good
It could very easily use a completely different or hallucinated source.
But a lot of LLM products are now providing source links right in the response. I’ve found them useful, and hopefully they aren’t produced just by feeding the text back in and asking for a link.
For what it’s worth, I’ve been working on (yet another) ActivityPub based micro blogging application and LLMs have been enormously helpful and so far as I can tell, correct. Often it cites the AP specs and its extensions, as well as specific implementations from existing major AP apps. It can show me expected outputs, what responses from my app should look like in response to different requests from other servers, and quickly give context for features like Mastodon’s shared inbox. I’m not having it simply generate code, but I think I’m still moving way faster than I otherwise could. I don’t recall it ever giving me incorrect information.
It’s the first time I’ve used an LLM as a tool this way, and I’m pretty impressed with it. I’m using the assistant made available through Kagi.
Tip: check those citations yourself before publishing with your name on the product. Yeah, they’re usually correct - do you only usually not want to be perceived as a lazy idiot?
I get that. I wouldn’t publish the code anywhere until an alpha is more or less ready and pretty well tested, and yes, I understand the importance of making sure it behaves in an expected, performant and pro-social manner with the existing compatible fediverse apps.
I’m not too worried about it, but thanks for your genuine concern about my reputation. ;) Since I’m the one writing the code, I’m more worried about the quality of that, if anything.
except the “explanation” frequently will be 100% “hallucinated” bullshit
That’s why I always ask it to cite sources. Basically googld ATP since google is turning to shit and all other search engines still aren’t quite as good
It could very easily use a completely different or hallucinated source.
But a lot of LLM products are now providing source links right in the response. I’ve found them useful, and hopefully they aren’t produced just by feeding the text back in and asking for a link.
That’s exactly how those links are produced.
For what it’s worth, I’ve been working on (yet another) ActivityPub based micro blogging application and LLMs have been enormously helpful and so far as I can tell, correct. Often it cites the AP specs and its extensions, as well as specific implementations from existing major AP apps. It can show me expected outputs, what responses from my app should look like in response to different requests from other servers, and quickly give context for features like Mastodon’s shared inbox. I’m not having it simply generate code, but I think I’m still moving way faster than I otherwise could. I don’t recall it ever giving me incorrect information.
It’s the first time I’ve used an LLM as a tool this way, and I’m pretty impressed with it. I’m using the assistant made available through Kagi.
Tip: check those citations yourself before publishing with your name on the product. Yeah, they’re usually correct - do you only usually not want to be perceived as a lazy idiot?
I get that. I wouldn’t publish the code anywhere until an alpha is more or less ready and pretty well tested, and yes, I understand the importance of making sure it behaves in an expected, performant and pro-social manner with the existing compatible fediverse apps.
I’m not too worried about it, but thanks for your genuine concern about my reputation. ;) Since I’m the one writing the code, I’m more worried about the quality of that, if anything.