- cross-posted to:
- technology@lemmit.online
- technology@lemmy.world
- cross-posted to:
- technology@lemmit.online
- technology@lemmy.world
The Linux creator is interested in AI, but the hype means he “basically ignores” it.
The Linux creator is interested in AI, but the hype means he “basically ignores” it.
10% is being generous. The parts of “AI” that aren’t just an expensive way of getting exciting-looking but unreliable results are mundane things like autocorrect, image upscaling models, handwriting recognition and such: unglamorous statistical learning in narrowly constrained domains nobody would claim is on the verge of becoming sentient and spawning the Singularity.
LLMs can be very useful if you understand how they work. The danger is when you assume that its correct.
This. Sorry but I’m a web developer and one of my colleagues obviously uses it without checking if it is correct, then bugs me or others when he doesn’t understand why it doesn’t work as expected. It is frustrating as hell and I’ve explained it to him multiple times:
Over prompt the AI if you are going to use it. Long lengthy prompts that are very succinct but give as much context as possible.
It is highly preferable to check other sources first like Stack Overflow. Even Medium articles can be better than using AI sometimes.
Type out what the AI output rather than just copy and paste. As you type line by line, explain to yourself what is happening.
Question everything. Do you think this code will work. Why will it work?
Test the code. If it doesn’t work as expected, trouble shoot it.
Don’t be afraid to scrap the whole thing and start over. Even open another prompt and try again if you really think the AI can answer the question (there are many cases where your problem is just too specific and the AI can’t).
He does none of these things. I swear he is the laziest developer I’ve ever met, and I’ve met my fair share.
That sounds like more work than just doing the thing.
I have copilot in my IDE and it’s basically worse than the ‘dumb’ autocomplete that just builds off the existing codebase. Copilot knows all the code on GitHub, but doesn’t seem to know anything about the local codebase.
As I suspected, I’m better off just doing the fucking thing myself.
Don’t blame all of AI because your coworker is lazy.
Also its best not to create really long prompts as that can confuse it. Instead, do your job. Its good for smaller things but it can’t replace a human.
I’ve found that just asking “did you make that up again?” after every response improves the quality of code Chat GPT produces. It seems to pick up fairly quickly on methods it just invented.
I dunno, it helped me with excel formulas quite a bit. I only had to use non ai sources multiple times to correct the errors the ai kept giving me, but it still set the ground work for standard forum users to fix.
Honestly, AI is super useful if what you need doesn’t require much creativity. For example, if I want to know how to setup an eDiscovery case in Exchange Online then I can probably trust Copilot to tell me the correct answer.
If I need to write a block of code to check for XSS or CSRF, then no, AI is not going to be trustworthy.
If it’s something that’s been done and written about dozens of times before, an LLM will probably barf up something resembling it.
Nah, unfortunately it’s been fed stack overflow. So instead of having a wealth of alternative solutions to a problem it just tells you your question is a duplicate and refuses to help.