Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

A study on robustness and reliability of large language model code generation

arxiv.org

Post reply on HN
Print view View original

21–30 of 229 posts

Re: A study on robustness and reliability of large language model code generation

#21

Post by keyle »

makestuff wrote: post #11
keyle wrote: post #8

Ok but what are we calling API misuses? We collect 1208 coding questions from StackOverflow on 24 representative Java APIs. We summarize thecommon misuse patterns of these APIs and evaluate them oncurrent popular LLMs. The evaluation results show that evenfor GPT-4, 62% of the generated code contains API misuses,which would cause unexpected consequences if the code isintroduced into real-world software. Should have u…

I have been trying to use ChatGPT to make a VueJS website (I’m a backend dev). What I have noticed with VueJS specifically is that it has little to no knowledge of the composition API, and when I ask it to help build a component it is pretty good at the scaffolding. The component will generally work; however, it is pretty useless for CSS styling. I probably should try Copilot as I bet that would work better. Also the…

Copilot probably wins because it holds more of _your_ code against the desired output.

I find ChatGPT is really great at getting something going, think scaffolding wise. But if you enter the realm of gnarly stuff, it can get completely delirious with answers full of red herrings and sends you running in circles very quickly.

Re: A study on robustness and reliability of large language model code generation

#22

Post by tadfisher »

version_five wrote: post #9
EGreg wrote: post #5

Earlier quoted context omitted.

I wonder what would happen in 39 years if an AI which is designed to feel and emulate humans enters human competitions and beats them at everything. Like can Liutenant Commander Data be excluded from human competitions? Yes. But can a replicant from Blade Runner? I guess the answer is that they will not feel the need to compete. But what if they do? Level 1: https://en.m.wikipedia.org/wiki/The_Measure_of_a_Man_(Star_…

If there's a human only competition and the AI "identifies" as a human, you'd expect it still wouldn't be let in. Anything else would be absurd, right?

More damning of competition as a concept than the AI, methinks.

Re: A study on robustness and reliability of large language model code generation

#23

Post by stephc_int13 »

I used Copilot for a few months, and I've toyed with different versions of Chat-GPT, for code.

Not impressed so far. With some effort and improvement this kind of tech can be used to fill boilerplate and glue code, and that is very nice as this is something that can help to stay in the flow by avoiding boring work.

But, the problem is that it is too unreliable even for this style of simple tasks, sometimes it works directly but it is too often plain wrong. Frustrating to use, not enough value, so I will try again later.

Re: A study on robustness and reliability of large language model code generation

#24

Post by thomasfromcdnjs »

Gigachad wrote: post #14

I feel like AI for programming has been so overhyped. I've attempted to use ChatGPT for programming so many times, and almost every time it's just wasted my time. Giving me outright lies or generating stuff that looks right but doesn't work. And just throwing out it's answer and starting from scratch is faster than fixing its output. I've only found it useful for explaining basic terminology or concepts for a topic I…

I've heard this from many people now but I cannot relate at all. Copilot has changed my entire strategy of programming (which I have been doing for 20 years)

It writes lots of code before I even think about what I was going to do in that file. Half the time it just works outright, and if not, with minor changes.

It is fantastic when I have to build functionality in languages I am not great at.

I regularly copy and paste code from my colleagues PR's and paste it into GPT-4, it explains it perfectly and gives tips on how to improve it (which I add as comments to my PR review)

Re: A study on robustness and reliability of large language model code generation

#25

Post by capableweb »

Gigachad wrote: post #14

I feel like AI for programming has been so overhyped. I've attempted to use ChatGPT for programming so many times, and almost every time it's just wasted my time. Giving me outright lies or generating stuff that looks right but doesn't work. And just throwing out it's answer and starting from scratch is faster than fixing its output. I've only found it useful for explaining basic terminology or concepts for a topic I…

Which version of GPT are you using?

I've found GPT4 moderately helpful, like getting help from someone with a wide but shallow experience of lots of things.

> I can't easily fact check it

Why not? How did you find out things about basic terminology or concepts you're not familiar about before GPT? Apply the same methods, although you just have to fact check rather than coming up with what to check, so removes a small initial discovery, at least for me.

Re: A study on robustness and reliability of large language model code generation

#26

Post by malux85 »

Gigachad wrote: post #14

I feel like AI for programming has been so overhyped. I've attempted to use ChatGPT for programming so many times, and almost every time it's just wasted my time. Giving me outright lies or generating stuff that looks right but doesn't work. And just throwing out it's answer and starting from scratch is faster than fixing its output. I've only found it useful for explaining basic terminology or concepts for a topic I…

Yeah, {zero, one}-shot is frequently wrong, to get good results you need to ToT or GoT (Tree of thought, graph of thought) which is currently only useful in more automated codewriting systems (like I'm building with https://atomictessellator.com) and not really useful in a co-pilot scenario.

Re: A study on robustness and reliability of large language model code generation

#27

Post by sen »

willsmith72 wrote: post #13

It's an interesting idea, but why is this written so weirdly? "the users of LLM code generation services are actually the developers that are most vulnerable to these code that seems right -- They are always novice developers that are not familiar with the APIs that LLMs generate code for them" It looks like they could've benefited from an LLM doing a pass-over. Are they saying it's only "novice developers" using LLM…

Just looks to me like the people who wrote this don’t speak English as a first language? It’s not difficult to parse what they meant to say.

Most of the users who are likely to turn to LLMs to help them code, are people who aren’t very experienced programmers. Hence needing help. These same people are more likely to fall for mistakes in the generated code as they don’t have enough experience to notice the errors.

LLMs are very good at sounding confidently correct about things that are subtly wrong. You need experience in the domain to notice these issues usually.

Re: A study on robustness and reliability of large language model code generation

#29

Post by lockhouse »

Gigachad wrote: post #14

I feel like AI for programming has been so overhyped. I've attempted to use ChatGPT for programming so many times, and almost every time it's just wasted my time. Giving me outright lies or generating stuff that looks right but doesn't work. And just throwing out it's answer and starting from scratch is faster than fixing its output. I've only found it useful for explaining basic terminology or concepts for a topic I…

AI for nearly everything has been very overhyped. Paintings of people have extra fingers and other strange artifacts. Using it for writing prose is hit or miss. Full self driving cars are thwarted by rogue traffic cones.

I think we'll have AI some day, but today it's just not at the level that all the hype claims it is.

Re: A study on robustness and reliability of large language model code generation

#30

Post by Retric »

Gigachad wrote: post #14

I feel like AI for programming has been so overhyped. I've attempted to use ChatGPT for programming so many times, and almost every time it's just wasted my time. Giving me outright lies or generating stuff that looks right but doesn't work. And just throwing out it's answer and starting from scratch is faster than fixing its output. I've only found it useful for explaining basic terminology or concepts for a topic I…

Yea, I had a thought early on that programmers might start poisoning the well with these LLM by uploading a lot of broken code. But I think people have been unintentionally doing so for years.
Post reply on HN