Arkian Forums

Theme

Customize your theme More languages later.
Log in / Register
Arkian Forums›Programming›AI still can't fix modern scum code

AI still can't fix modern scum code

Discussion in Programming

OmegaKV
I went to a "codeathon" type conference where I was tasked with adding a simple feature to an open source project. The feature was just to add a tooltip to a widget which displays some info. I asked where things are located in the project to the ChatGPT web interface so I could get started, and it said it wouldn't find it and that it told me to search with grep. It was useful in other ways though, and eventually I got it almost working, but there was only one problem. Whenever the widget would update, it would make the tooltip disappear. Once I fix this one detail, and make the tooltip always last for a second or two after hovering, I would be done. But this proved much harder than expected, and I didn't get it finished during the conference. I have heard the web ChatGPT is far inferior to Codex, but eventually, after the conference, I went to my office computer where I had a paid business subscription to ChatGPT with Codex installed and I asked Codex to prevent the tooltip from disappearing upon the widget updating, and it was very informative as far as describing the root cause of why it was disappearing (there is something in the JavaFX library that makes the tooltip of a widget disappear whenever the widget is refreshed/updated), but it wasn't able to prevent the tooltip from disappearing. It tried a number of complicated workarounds, involving like a hundred times more code than the original change, but they didn't work. This is just to say AI doesn't seem omnipotent in its ability to fix code yet.
mmHg
I mean what models and effort levels did you use with Codex, and did you use any ai tools/plugins or context files or prompt engineering skills?
OmegaKV edited
> [mmHg](/thread.html?thread=980&post=993#p993) > > I mean what models and effort levels did you use with Codex, and did you use any ai tools/plugins or context files or prompt engineering skills? It uses a good model, don't know the specifics because they do some routing of the models to protect privacy, so I'm not sure if the model name listed is the actual model. My workplace has a lot of money and throws tons of money at this. The average cost of tokens that people in my group use is something like $8k a month. I think it's a good model. As far as explaining the problem, it knows the problem. The issue is the javafx library makes it hard to solve simply. An ideal solution to any difficult software problem would be someone thinking outside of the box to find a simple fix, but most of codex's attempted solutions have been very complicated.
mmHg edited
So Astra on Ultra effort mode? Well, the problem seems to be that Astra is following bad programming practices. The solution for this is (basically) to create context files and add it to your AI project folder in Codex. You can ask AI to help you make context files. It is important that your context files have general instructions for how to do things (e.g., you can tell it to completely avoid HTML5 elements and to use div and span for most things HTML) and lightweight examples (e.g., you give the example of using <div content> instead of using <main> in HTML). I have successfully been able to use context files to make AI write essays in my writing style without being detected by an AI text detector. I never needed it though, because I already finished all of my coursework in philosophy before the fact.
OmegaKV
> [mmHg](/thread.html?thread=980&post=1001#p1001) > > So Astra on Ultra effort mode? Well, the problem seems to be that Astra is following bad programming practices. The solution for this is (basically) to create context files and add it to your AI project folder in Codex. You can ask AI to help you make a context file. It is important that your context file has general instructions for how to do things (e.g., you can tell it to completely avoid HTML5 elements and to use div and span for most things HTML) and lightweight examples (e.g., you give the example of using <div content> instead of using <main> in HTML). > > I have successfully been able to use context files to make AI write essays in my writing style without being detected by an AI text detector. I never needed it though, because I already finished all of my coursework in philosophy before the fact. It knows the problem. I'm not trying to get AI to do a straightforward task that should be clear if it only had the right information, but solve a hard problem that requires real intelligence to solve. The programming aspect is not really that important to me, I just am interested in conceptually how to solve the problem.
mmHg
It knows the problem but context files tell the AI what to do about that problem. Context files are not just merely informational, they are **doctrinal**. Anyway, AI can definitely fix modern scum code.
mmHg edited
About solving hard problems, I mean, I used Astra to help me discover a glitch in a video game (*Army of Ages*) that I was speedrunning. Astra successfully identified a glitch that helped me improve my run time by approximately 25.6%. Let me show you the two runs. First, here is my [glitchless run](https://youtu.be/8cOtUvwnN3A), with a completion time of 4:57. Compare that with my [glitched run](https://youtu.be/fkB7OfX6Aic), which I completed in 3:41. Finally, here is the [Army of Ages speedrunning leaderboard](https://www.speedrun.com/army_of_ages?h=Insane&x=jdrnevlk). Regarding the part of me that is practical, I have also used Astra to successfully solve some really difficult problems with textilesocial, namely drag and drop functionality on mobile devices. I implemented drag and drop to textilesocial so that admins can rearrange forums and categories. After a lot of work and debugging, Astra finally made it work properly and comfortably. Anyway, Astra is a great problem solver. If it were unregulated, I think it would make a masterful computer hacker.
OmegaKV
> [mmHg](/thread.html?thread=980&post=1008#p1008) > > About solving hard problems, I mean, I used Astra to help me discover a glitch in a video game (*Army of Ages*) that I was speedrunning. Astra successfully identified a glitch that helped me improve my run time by approximately 25.6%. > > Let me show you the two runs. First, here is my [glitchless run](https://youtu.be/8cOtUvwnN3A), with a completion time of 4:57. Compare that with my [glitched run](https://youtu.be/fkB7OfX6Aic), which I completed in 3:41. Finally, here is the [Army of Ages speedrunning leaderboard](https://www.speedrun.com/army_of_ages?h=Insane&x=jdrnevlk). > > Regarding the part of me that is practical, I have also used Astra to successfully solve some really difficult problems with textilesocial, particularly with drag and drop functionality on mobile devices. I implemented drag and drop to textilesocial so that admins can rearrange forums and categories. After a lot of work and debugging, Astra finally made it work properly and comfortably. > > Anyway, Astra is a great problem solver. If it were unregulated, I think it would make a masterful computer hacker. I have had AI fix bugs like this in the past so I know what you mean. This one is different, the AI doesn't know what to do. It's not a doctrine thing, it just doesn't know. It knows the problem but doesn't know the solution which should be straightforward but modern scum made it not straightforward. Doctrine is something that helps if you build the software from scratch. AI can't fix hopelessly shitty modern scum's code.
mmHg edited
Give me the open source project and the instructions, and I will take a crack at it.

Log in to reply.