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

Can LLMs write better code if you keep asking them to “write better code”?

minimaxir.com

Post reply on HN
Print view View original

31–40 of 461 posts

Re: Can LLMs write better code if you keep asking them to “write better code”?

#31

Post by auc »

xrd wrote: post #19

Wow, what a great post. I came in very skeptical but this changed a lot of misconceptions I'm holding. One question: Claude seems very powerful for coding tasks, and now my attempts to use local LLMs seem misguided, at least when coding. Any disagreements from the hive mind on this? I really dislike sending my code into a for profit company if I can avoid it. Second question: I really try to avoid VSCode (M$ concerns…

I use VSCode + Copilot. For anything more than boilerplate code, I find that Copilot kind of sucks and I use O1 in ChatGPT

Re: Can LLMs write better code if you keep asking them to “write better code”?

#32

Post by botanical76 »

vendiddy wrote: post #13
dash2 wrote: post #5

Something major missing from the LLM toolkit at the moment is that it can't actually run (and e.g. test or benchmark) its own code. Without that, the LLM is flying blind. I guess there are big security risks involved in making this happen. I wonder if anyone has figured out what kind of sandbox could safely be handed to a LLM.

Somewhat related - I wonder if LLMs are trained with a compiler in the loop to ensure they understand the constraints of each language.

This is a good idea. You could take a set of problems, have the LLM solve it, then continuously rewrite the LLM's context window to introduce subtle bugs or coding errors in previous code submissions (use another LLM to be fully hands off), and have it try to amend the issues through debugging the compiler or test errors. I don't know to what extent this is already done.

Re: Can LLMs write better code if you keep asking them to “write better code”?

#33

Post by Kiro »

> What would happen if we tried a similar technique with code?

It was tried as part of the same trend. I remember people asking it to make a TODO app and then tell it to make it better in an infinite loop. It became really crazy after like 20 iterations.

Re: Can LLMs write better code if you keep asking them to “write better code”?

#34

Post by HPsquared »

exitb wrote: post #23
HPsquared wrote: post #7

Using the tool in this way is a bit like mining: repeatedly hacking away with a blunt instrument (simple prompt) looking for diamonds (100x speedup out of nowhere). Probably a lot of work will be done in this semi-skilled brute-force sort of way.

Well, in this case it's kind of similar to how people write code. A loop consisting of writing something, reviewing/testing, improving until we're happy enough. Sure, you'll get better results with an LLM when you're more specific, but what's the point then? I don't need AI when I already know what changes to make.

This method requires a lot less skill, you just need to read it and run it. Reading is much easier than writing!

Re: Can LLMs write better code if you keep asking them to “write better code”?

#35

Post by dale_glass »

dash2 wrote: post #5

Something major missing from the LLM toolkit at the moment is that it can't actually run (and e.g. test or benchmark) its own code. Without that, the LLM is flying blind. I guess there are big security risks involved in making this happen. I wonder if anyone has figured out what kind of sandbox could safely be handed to a LLM.

godbolt exists and can run code, so surely similar principles could be used here.

Re: Can LLMs write better code if you keep asking them to “write better code”?

#36

Post by viraptor »

mrweasel wrote: post #21
cranberryturkey wrote: post #3

its best to tell them how you want the code written.

At that point isn't it starting to become easier to just write the code yourself? If I somehow have to formulate how I want a problem solved, then I've already done all the hard work myself. Having the LLM just do the typing of the code means that now not only did I have to solve the problem, I also get to do a code review.

There's no clear threshold with an universal answer. Sometimes prompting will be easier, sometimes writing things yourself. You'll have to add some debugging time to both sides in practice. Also, you can be opportunistic - you're going to write a commit anyway, right? A good commit message will be close to the prompt anyway, so why not start with that and see if you want to write your own or not?

> I also get to do a code review.

Don't you review your own code after some checkpoint too?

Re: Can LLMs write better code if you keep asking them to “write better code”?

#37

Post by padolsey »

dash2 wrote: post #5

Something major missing from the LLM toolkit at the moment is that it can't actually run (and e.g. test or benchmark) its own code. Without that, the LLM is flying blind. I guess there are big security risks involved in making this happen. I wonder if anyone has figured out what kind of sandbox could safely be handed to a LLM.

Running code would be a downstream (client) concern. There's the ability to get structured data from LLMs (usually called 'tool use' or 'function calling') which is the first port of call. Then running it is usually an iterative agentagent task where fixes need to be made. FWIW Langchain seems to be what people use to link things together but I find it overkill.* In terms of actually running the code, there are a bunch of tools popping up at different areas in the pipeline (replit, agentrun, riza.io, etc)

What we really need (from end-user POV) is that kinda 'resting assumption' that LLMs we talk to via chat clients are verifying any math they do. For actually programming, I like Replit, Cursor, ClaudeEngineer, Aider, Devin. There are bunch of others. All of them seem to now include ongoing 'agentic' steps where they keep trying until they get the response they want, with you as human in the chain, approving each step (usually).

* I (messing locally with my own tooling and chat client) just ask the LLM for what I want, delimited in some way by a boundary I can easily check for, and then I'll grab whatever is in it and run it in a worker or semi-sandboxed area. I'll halt the stream then do another call to the LLM with the latest output so it can continue with a more-informed response.

Re: Can LLMs write better code if you keep asking them to “write better code”?

#38

Post by amelius »

Kiro wrote: post #33

> What would happen if we tried a similar technique with code? It was tried as part of the same trend. I remember people asking it to make a TODO app and then tell it to make it better in an infinite loop. It became really crazy after like 20 iterations.

Reminder: don't ask your t-shirt folding robot to do a better and better job.

Re: Can LLMs write better code if you keep asking them to “write better code”?

#39

Post by ericskiff »

xrd wrote: post #19

Wow, what a great post. I came in very skeptical but this changed a lot of misconceptions I'm holding. One question: Claude seems very powerful for coding tasks, and now my attempts to use local LLMs seem misguided, at least when coding. Any disagreements from the hive mind on this? I really dislike sending my code into a for profit company if I can avoid it. Second question: I really try to avoid VSCode (M$ concerns…

I highly recommend the command line AI coding tool, AIder. You fill its context window with a few relevant files, ask questions, and then set it to code mode and it starts making commits. It’s all git, so you can back anything out, see the history, etc.

It’s remarkable, and I agree Claude 3.5 makes playing with local LLMs seem silly in comparison. Claude is useful for generating real work.

Re: Can LLMs write better code if you keep asking them to “write better code”?

#40

Post by moffkalast »

dash2 wrote: post #5

Something major missing from the LLM toolkit at the moment is that it can't actually run (and e.g. test or benchmark) its own code. Without that, the LLM is flying blind. I guess there are big security risks involved in making this happen. I wonder if anyone has figured out what kind of sandbox could safely be handed to a LLM.

The only proper way to code with an LLM is to run its code, give it feedback on what's working and what isn't, and reiterate how it should. Then repeat.

The problem with automating it is that the number of environments you'd need to support to actually run arbitrary code with is practically infinite, and with local dependencies genuinely impossible unless there's direct integration, which means running it on your machine. And that means giving an opaque service full access to your environment. Or at best, a local model that's still a binary blob capable of outputting virtually anything, but at least it won't spy on you.

Post reply on HN