Arkian Forums

Theme

Customize your theme More languages later.
Log in / Register
Arkian Forums›Programming›Some thoughts after trying AI software development

Some thoughts after trying AI software development

Discussion in Programming

OmegaKV
I am using AI to try to replace a software application I use on a daily basis. The challenge is what AI generates is essentially a huge black box, and you don't know what parts of it work and what don't when it is first generated. To mitigate this, rather than try to fix every detail, it may pay to first break up the problem into layers which can more easily or more generally tested empirically, and have AI generate these layers. And then ideally these layers would be useful for other applications, so that testing done for one application will result in benefit for other applications. It makes more sense now than before to have thinner layers, because the number of projects will increase, which means the number of opportunities that a layer can be reused will increase. And more effort will be devoted into strategizing how to organize a problem into layers, in a way that would be most useful to you in the long term.
OmegaKV
After looking at the code it generated, AI is like having a personal modern scum programmer slave. The code works, but it does some dumbass things, like gives member functions parameters in cases where it should have no parameters and just use the data from the class. Also, I see it pointlessly deconstructing an object and replicating the functionality that already exists in the object in a different way.
fschmidt Admin
Which model are you using? Anyway, the reason I am making a skill is to make AI program like me, not like modern scum.
Reactions
OmegaKV
> [fschmidt](/thread.html?thread=1799&post=1813#p1813) > > Which model are you using? > > Anyway, the reason I am making a skill is to make AI program like me, not like modern scum. GPT 5.5 with medium reasoning
mmHg edited
> [OmegaKV](/thread.html?thread=1799&post=1823#p1823) > > GPT 5.5 with medium reasoning GPT 5.5 is a horrible model. GPT-6 Sol is better for half the cost. GPT-6 Luna is slightly weaker than GPT 5.5 but it is at least 50x cheaper. With an optimal skill, surely even Luna can do a lot of good work.
OmegaKV edited
Today I used codex to deploy some software. I have written scripts in the past to attempt to automate deployment of that software, but I could never remember how to use them, modifying them is a pain since testing changes requires redeployment of a software, and there is always some uncertainty about it working on given the possibility of edge cases. So I instead just told codex what needed to be done, told it to ask me questions if it is unsure about something or if there are anomalies, and told it to take notes into SKILLS.md for the next time I ask it to do this. It worked quite well. It's nice to be able to give advice, make corrections, mention things you neglected before, as you go along, without really needing to holistically test something like a deployment script.
mmHg
Codex is nice; I like it better than Claude Code.

Log in to reply.