While building web and mobile applications with AI (specifically Lovable) I have concluded that AI is best for building MVPs. Generating new ideas quickly is what it is really good at. They are really good at starting a project but as the project progresses they slowly become more and more unreliable, I suspect this is due to the amount of memory -how much it can remember. So if you’ve reached the edge of the AI’s ability to remain sane on the solution, you’ll have to roll up your sleeves and get your hands dirty.
Conclusion is AI is over-eager(always happy to completely rewrite huge chunks of code.. bad ai bad), over-confident, and quick to assume but is capable, never complains or get bored and is forgetful.
There are 2 interesting but contradicting fact about AI and coding.
First, the AI can write useful code. The second is that the AI can also get completely lost, fall into a rabbit hole, chase its tail, and produce unusable garbage.
Fall into a rabbit hole
I tried to put multiple changes in a single prompt but realised it is just like a human where it makes mistake and even cause overlapping problems if you do multiple changes at one go. It is better to make one change at a time commit check code change and results then rinse and repeat. This gives you important points to restore to.
Chasing tails
Lovable tries to stop tail chasing by just totally preventing it from doing what it already did previously. I try to reduce rabbit hole but just using a different AI, try a different prompt or use a new chat as a reset.
AI can’t maintain code
ChatGPT struggles with editing pre-written code (including code it wrote) —in fact, it often won’t do it at all. Even if you use a nearly identical prompt, the output can be very unexpected. The only way to mitigate this is either you scope it down to a single line or a specific logic of code or you get it to figure out the solution and then you modify the existing code with its idea.
