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

71–80 of 229 posts

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

#71

Post by wpietri »

corethree wrote: post #47
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…

The over-hype is causing it to look under-hyped. 48% of the code being freaking correct from 0% about 5 years ago is huge. With everyone and all the news talking about it constantly inevitably people are going to start rolling their eyes. This is bias. 48% is half way to 100%. Once it reaches 100% you don't have a job. You realize that right? It's halfway their to taking your job and you're underwhelmed. The bias is…

> Once it reaches 100% you don't have a job. You realize that right? It's halfway their to taking your job and you're underwhelmed. The bias is on your side.

What this study looked at was feeding StackOverflow questions to LLMs and then looking at the quality of the code. If you think a programmer's job is just turning an english-language description of a function into isolated code that never gets modified, I don't know what to tell you. In my opinion, anybody like that should not have a job today, never mind the future.

A proper professional programming job involved AGI-level understanding of human users and their needs as embedded in a social context. Plus the ability to create novel solutions. Plus the ability to use code as a collaborative medium to make a code base that is sustainable over the long term by their colleagues.

LLMs are not even 1% of the way to replacing professional developers.

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

#72

Post by wesleywt »

hedora wrote: post #46
thomasfromcdnjs wrote: post #24

Earlier quoted context omitted.

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 pa…

Meh. My coworkers do what you’re doing, and then I end up rewriting their stuff. (AI generated stuff generally passes code reviews because it gets the benefit of the doubt.)

Meh. I started coding professionally in the age of AI. My productivity is way higher than it should be. I don't let the AI "write" the code for me. I use it as a companion that I asked question of and suggestions.

It suggests solutions to problems, I add the complexity. So far my code has been both clear and performant. This should not be the case at my level.

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

#73

Post by rvz »

lockhouse wrote: post #29
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.

There you go. But even before that 8 years ago, ConvNets, AlexNet, etc were getting all the hype and then came the adversarial images and issues which weren't addressed and then that hype fizzled out very quickly.

Now the AI bros here are attempting to sell us their new snake-oil in the form of a stochastic parrot which promises to be the solution to everything.

Well, unsurprisingly it is another unexplainable AI black box which still requires the human to check every single output so that it doesn't hallucinate something incorrect, which it does almost all the time with a lack of transparent explainability as to why it hallucinated.

So the fact is, it is already overpromising and under-delivering for serious use-cases. Just like FSD (Fools Self Driving) was when that was over-hyped and with little to no Tesla Robo-taxis on the road.

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

#74

Post by pmcf »

This link is being copied to slack channels everywhere with “I told you so” statements and something about “You kids and your GPTs” and “Back in my day”

Meanwhile copilot is becoming a superpower to those who figure it out.

And it hasn’t even been a year! I’m going to need a lot more popcorn.

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

#75

Post by wesleywt »

master_crab wrote: post #39
thomasfromcdnjs wrote: post #24

Earlier quoted context omitted.

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 pa…

I feel like using suggestions for line by line composition might lend itself to better review by programmers. As opposed to ChatGPT which spits out dozens of lines of code based on a request and therefore requires more involved editing after the fact to understand what’s happening.

Yes, GPT is an aid not a programmer. You should still be doing the programming while ChatGPT gives you quick solutions to problems without having to look up Stack Overflow.

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

#77

Post by toyg »

redhale wrote: post #68
thomasfromcdnjs wrote: post #24

Earlier quoted context omitted.

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 pa…

This! I wonder if it's a language difference thing (I use C# mostly), but my experience with Copilot has been nothing short of mind-blowing. It's not writing every line, and when writing truly new feature code it's less useful. But here are two patterns that I've noticed it is especially and consistently good at as a potential place to look for initial value if you are skeptical: - If you have any kind of repeated pa…

Languages like C# and Java have always required a lot of boilerplate and repetitive (sorry, "patterned") code. They are perfectly suited to GPTs.

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

#78

Post by toyg »

drewfis wrote: post #67
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…

Chat GPT is a game changer for a lot of my work. I make it write stored procedures in MS SQL along with the calling code in .NET, C#, and Dapper. The generated code will likely have a few small issues, but it still makes me way more productive. It allows me to work ~10 hours a week less and enjoy my life.

Until your boss/customer gets wind of it, and claims back that time.

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

#79

Post by damowangcy »

API misuses is one thing, the more concerning outcome is the misuse of AI itself.

Just like how there is "common misuse pattern of x language", there is also "common misuse of copilot/chatgpt4".

From my observation, those who are successful with code generation tools are using it as an assistant, they already know the language quite well, AI is there just to help, most of the code are boilerplate code or common patterns.

Those who complaint are usually using it to do more than just that, relying it to generate working piece of code for a particular function without knowledge of the language. Currently, we use StackOverflow for this purpose, when we asked question, we don't just say, "hey, code please" but we are trying to understand why and how things work instead. This knowledge will then help us to code contextually, which current AIs are incapable of. The plus side is that code generation is faster.

Also the point of the paper is not to prove that GPT-4 is unusable but that it can be further improve. They didn't just manually determine misuses but did it via an API checker that they designed. So, in the future, AIs are just going to be more reliable not less.

We are not at a point that AI is common enough to talk about misuse yet though, most don't even have access to CoPilot or ChatGPT 4.

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

#80

Post by raincole »

corethree wrote: post #47
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…

The over-hype is causing it to look under-hyped. 48% of the code being freaking correct from 0% about 5 years ago is huge. With everyone and all the news talking about it constantly inevitably people are going to start rolling their eyes. This is bias. 48% is half way to 100%. Once it reaches 100% you don't have a job. You realize that right? It's halfway their to taking your job and you're underwhelmed. The bias is…

38%.

And 38% is not "a third way to 100%". I'm sure you know what diminishing return is.

Post reply on HN