Hacker News Reader

Comment by collaborative | Hacker News Reader

AI tools are absolutely useless once you get past any initial stage in a software development cycle

They just can't comprehend what needs to be done even if all that's required are 2 lines of code

However, we keep seeing posts from people who don't know html or how to create a chrome extension paternalizing about how we can be much more productive with AI tools. Last I heard a CEO was demanding dev increased productivity or firing some devs because he saw a Youtube video about how easy it is to create a website now

The only usefulness I see is in AI replacing SO. Once SO dries we'll have neither

View on HN

This sounds to me like you have no idea how to properly integrate LLMs into your workflow. Just because they’re not useful for changing large code bases doesn’t mean they can’t be valuable. A few examples of things I’ve used LLMs for:

Writing bash scripts to automate various parts my workflow

Quickly interpreting complex regex

Tutoring me on how to use poorly documented APIs

Proof reading important emails

Everything copilot, writing java docs, auto completing all the cases in a switch statement, etc.

Could you? Do you have to? Would someone pay you to implement an FPGA design, given that you know nothing about it but have access to GPT-4? In practice, I've found that the advantage it might give you quickly ebbs out. It surely helps filling details and blanks though.

>Tutoring me on how to use poorly documented APIs

Can you elaborate on this? I would find it very useful.

My current workflow is to copy a bunch of relevant code into GPT4 and then ask it questions about how to use the API. GPT4 is very good at inferring how to use an API, even from code snippets. The key is that you can paste large amounts of unstructured code and GPT4 will make sense of it.

> Writing bash scripts to automate various parts my workflow

Would you share some of those? Bash is fraught with gotchas and intricacies and seldom have I read Bash code where I couldn’t immediately spot flaws. There’s often something which left unchecked may come back to bite you later.

I’m skeptical the LLM code would be any better, since it’s (presumably) trained on a corpus of subpar code. But I could be wrong.

Give the LLM access to `shellcheck` and have it iterate on the warnings. I think promising results have been shown when you allow LLMs to reflect on their answers and give them access to tools.

Shellcheck can't check everything though. I think using LLMs on difficult to verify code like shell scripts and complex regexes is a pretty bad idea.

I've found them useful as a "smart search engine", e.g. if I don't even know the magic terms to search for.

Most of the time it gives me answers that are completely wrong, but are close enough that I know where to look to find the right answer.

E.g. "Using the rapidcheck C++ test framework, how do you test a function that takes two vectors, one of which must be exactly 8 times as long as the other."

It gets it wrong but it still helped. I definitely wouldn't want to use its output unquestioningly though.

Depends entirely on if the shell script is mission critical. I couldn't care less about the quality of my LLM shell scripts most of the time. My only success criteria is that is that it solves a problem. These are scripts I'm never going to touch again, things like batch converting files, or like one time I had it generate a script to titrate stable diffusion parameters. As long as I can skim the output and see that its correct I'm happy.

You call it a bad idea, but it's saved me hours of work.

> Tutoring me on how to use poorly documented APIs

Good luck with that.

GPT-4 is _incredibly_ bad at Metal (most under-documented GPU api I've ever worked with), but more importantly it's _even worse_ at Vulkan, which is complete opposite - it's meticulously documented.

The stuff LLMs are good at is the stuff that a lot of people use and talk about all the time, in public.

There's just not a lot of public discussion (with examples) about e.g. how to properly use Vulkan in a production application, and the LLM itself can't "reason" within the framework of the detailed specification, so most things it outputs is wrong.

In the long run though, a very likely outcome is that tools that the LLMs have a poor "understanding" of just wither and die, or are relegated to being something only 3.5 people in the world use. I can see a future where literally all software is written in python just because that's the language that OpenAI's LLMs happens to know best.

It has an 8kb (or 32kb if you are lucky) context. Try giving it the relevant documentation examples before asking a question.

the point the grandparent was making was that GPT-4 can teach you how to use an under-documented API. that means the relevant docs are either sparse or non-existent, otherwise there wouldn't be a need to ask GPT-4 in the first place.

No, that wasn’t my point. You don’t need docs, just feed in the raw source code. GPT4 can make sense of code without documentation.

feed it raw source code of what exactly?

I’ve been programming for two decades and you’re just plain wrong here. It’s useful throughout the dev process for many things, the worst part right now is just how slow GPT-4 is to run inference. You’re probably not using it right, aren’t using GPT-4, or something. It’s a multiplier, for specific parts of development - the amount of annoying nonsense I no longer have to spend my limited time on earth thinking about makes me want to cry.

Also got a couple decades, a ChatGPT Plus subscription on top of running my own local instance of Vicuna and access to Bard.

Enjoy it while it lasts. The training set (Stack Overflow) will dry as people move away from it, resulting in model degradation, resulting in not having a useful AI AND not having a useful SO

I admit to using AI, but only for what I was using SO before. Which is 0% of the work I actually get paid for (maintaining existing software that requires expert knowledge). For personal projects, sure, it's useful. Want to code something in a new language? Great. You'll still be a beginner, and will have contributed nothing to the body of knowledge that has trained the model you are using

What? You're jumping to so many conclusions here, it's absurd. You're going to have to go into research for a few years if you want to get anywhere on proving that the capability of the current systems is somehow determined by the presence of StackOverflow. Sounds like you have a very tall tower of speculative assumptions about LLMs today, LLMs in the future, the evolution of the web corpus, and other big, dynamic, absolutely unknowable things.

This is starting to sound like a debate, but I feel that the burden of proof lies more in that LLMs are anything more than a blurry representation of the data they have been trained on than otherwise. I personally don't feel like I've brought Frankenstein to life each time I load a LLM into memory..

I'll be really impressed the day models can generate code by just grokking the language spec and documentation.

It can do that now. Has an 8kb (or 32kb) context. You might need to do a bit of work preparing a concise description and language examples.

Does anyone have any blog posts that elaborate on this?

This is why its important to reverse mentor executive leadership so they better understand best practice usage of LLMs.

I obsoletely see benefits of tools like ChatGPT through the entire lifecycle of a product, however not as a replacement for things like good requirement elicitation, architecture, and product market fit.

In the near future I expect companies to train models on their internal code base/documentation allowing developers to learn about available libraries, API usage etc.

You are expecting it to replace a senior+ dev; that is not yet reasonable.

> Last I heard a CEO was demanding dev increased productivity or firing some devs because he saw a Youtube video about how easy it is to create a website now

reminds me of Donald Trump complaining about the health exchange website costing millions of dollars when he can build a website for $3

Reply on news.ycombinator.com