Similar sentiment here. I taught myself python a decade ago after college, and used it in side projects, during my masters degree, in a few work projects. So it's been handy, but also required quite a bit of time and effort to learn.

But I've been using Claude to help with all kinds of side projects. One recently was to help create and refine some python code to take the latest Wikipedia zipped XML file and transform/load it locally into a PostgreSQL DB. The initial iteration of the code took ~16 hours to unzip, process, and load into the database. I wanted it to be faster.

I don't know how to use multiple processes/multi-threading, but after some prompting, iterating, and persistent negotiations with Claude to refine the code (and an SSD upgrade) I can go from the 24gb zip file to all cleaned/transformed data in the DB in about 2.5 hours. Feels good man.

Do I need to know exactly what's happening in the code (or at lowers levels, abstracted from me) to make it faster? not really. Could someone who was more skilled, that knew more about multi-threading, or other faster programming languages, etc..., make it even faster? probably. Is the code dog shit? it may not be production ready, but it works for me, and is clean enough. Someone who better knew what they were doing could work with it to make it even better.

I feel like LLMs are great for brainstorming, idea generation, initial iterations. And in general can get you 80%+ the way to your goal, almost no matter what it is, much faster than any other method.

That's awesome! That's a lot of data and a great speed increase. I think that as long as you test and don't just accept exactly what it outputs without a little thought, it can be really useful.

I take it as an opportunity to learn too. I'm working on a video library app that runs locally and wanted to extract images when the scene changed enough. I had no idea how to do this, and previously would have searched StackOverflow to find a way and then struggled for hours or days to implement it. This time I just asked Aider right in the IDE terminal what options I had, and it came back with 7 different methods. I researched those a little and then asked it to implement 3 of them. It created an interface, 3 implementations and a factory to easily get the different analyzers. I could then play around with each one and see what worked the best. It took like an hour. I wrote a test script to loop over multiple videos and run each analyzer on them. I then visually checked the results to see which worked the best. I ended up jumping into the code it had written to understand what was going on, and after a few tweaks the results are pretty good. This was all done in one afternoon - and a good chunk of that time was just me comparing images visually to see what worked best and tweaking thresholds and re-running to get it just right.