review code by a human

[jcolomb]

In order to finish the experiment, it would be good to have the code reviewed by humans, it would both help know whether the LLM could actually write "good" code AND help me trust this tool.
Maybe send it to JOSS for review ?

[defnull]

I am very biased, obviously, but I reviewed the code a while ago to check if it can be extended for my needs, and quickly decided to not use any of it and start from scratch instead.

Vibe-coding may be useful to produce a (barely) working prototype for an idea, but the code is beginner-level at best, contains lots of subtle bugs, lacks proper error checking and edge-case handling, ignores basic security best-practices and falls apart as soon as something does not go as planned. The design and architecture (or lack thereof) also makes it hard to add features or fix bugs, because of how much is just hard-coded with no extensibility in mind. It honestly looks like it was copy&pasted directly from often over-simplified stackoverflow answers or tutorials by someone who is just starting to learn how to code. No surprise if you know how those AIs were trained and how they work.

Using AI as a non-programmer may be okay for a toy project that runs exclusively on client-side in a browser sandbox where it should not be able to cause much harm, but really dangerous if you extend the idea to server-side logic and unknowingly expose insecure and buggy code to the internet. An experienced developer may be able to see the (often security-impacting) issues and direct the AI in the right directions, but that is usually more work than to just think for yourself. I tried it, and I had to fight the AI more than it helped me.

[defnull]

Bugs that I see on first glance:

• Server rate-limiting is not respected, which can result in failed requests and may stress the source server more than necessary.
• Sometimes images are broken because they are loaded too fast and hit server-side rate limits. There is no re-try logic.
• If fetching hashtags takes longer than 10 seconds, updates will not wait but cascade, worsening the issue until requests start to fail.
• The three-hashtag-and-one-server limit is hard-coded in several places, making it unnecessary hard to extend it.
• Common errors and typos when adding hashtags or a server are not handled, resulting in an empty wall instead of an error message.
• Post wrapped in a <p> element break the layout. Rendering issues for posts that contain <p> elements.  #8
• The url parameters are parsed with multiple regular expression, some of which are useless hallucinations, instead of using URLSearchParams which is a 10+ years old standard browser API.

[kreuvf]

Just stumbled upon this piece of software noticing this:

Source: https://rstockm.github.io/mastowall/?hashtags=bibliothekenentschlossendemokratisch,bibliocon25,bibliocon&server=https://mastodon.social

(Purely incidental that the post is critical of "AI", it was just the first post with NaNs.)

This is one of the many issues that arise from vibe coding using an advanced text generator ("AI") that is just not capable of actually understanding anything it is doing. So, I fully side with @defnull here, the code quality is so low that a human review would most likely be a waste of time compared to rewriting this from scratch for production purposes.

[defnull]

the code quality is so low that a human review would most likely be a waste of time compared to rewriting this from scratch for production purposes.

More so because a human-written alternative already exists.