Updated thoughts on AI/LLMs

At some point these have started to become useful to unleash in my codebase. Since my last post about this, I've tried cursor, claude code, aider, and of course been using the web based interfaces. The best way to get utility from these is to get a feeling about what they're capable of through using them. Sometimes they save you enormous time by generating large swathes of code. Sometimes they just waste your time creating a bowl of spaghetti you need to toss. Knowing where to unleash them, what to say, and which model to use for a given task is going to give the most gains. One great thing I found is that they have become pretty good for generating UX. Bolt/Lovable/Replit are all basically free UX designers now. Even web based claude sonnet 3.7 has given me things I've implemented in my webapp. The gains are pretty enormous if you know how to assess the output and not let technical debt build up. Hoping that within a few years, we will have the equivalent of today's sonnet 3.7 and DeepSeek R1 that can run at a decent speed locally on a laptop. My enthusiasm has peaked today when from an experience I just had telling emacs aider to create a react native screen based on my web app code, and it did most of it very well. I didn't even load up the CSS in the context, but I'm assuming it used it because it managed to nail down most of the design. I also just updated the landing page on MotorSaif yesteday to be a blend I got from both lovable and bolt, and I paid nothing for it. Currently I feel the various models will get you 60% there if you want something specific, and you have to know how to do the other 40%. My experience so far is that if I keep prompting to attempt to get the other 40%, it gets ugly very fast. Update a few weeks after writing the initial post: A lot of the code generated had to be heavily manually refactored. While it was really interesting being able to "see" something right away, the ergonomics of maintenance and adding more features was completely missing. I'm still using it for things here and there, but mainly keeping it on web based chat instead of using them directly in my IDE. As my project becomes more mature, it is becoming less and less useful and in order to be able to actually pinpoint and maintain my software in prod by providing hotfixes and growing features in a timely manner, I need things to be done in a specific way. I'm using it more focused laser-like instead of letting it generate larger pieces, with the complete understanding of every line that gets inserted. This is an enormous time saver but can also be a foot gun if wielded improperly, and only experience will let one know whether he's about to ship an injured foot into prod.