So, I started working on an app.
I have written novels.
I have written a children's book.
I have written some icon cut out books.
I have made many images.
I have many many videos.
I have edited videos.
I have tried to grow an audience.
I have tried to sell the things I've made.
I have started gigs on fiverr.
I have decided that I didn't like the experience of having to do all of these things on many different platforms and am attempting a way to allow people to have an account with multiple personas, generate content, sell and distribute that content or monetize the channels they already have.
I've been using AI to write the code. I'm using hermes and a variety of models. My best experience so far is with opencode go. That's the least expensive plan I can find. I think Nvidia has some free stuff and Gemini is supposed ot have some great starting packages, but I've been very happy with opencode go.
So, I have a little money going to Nous Research every month for my subscription there. I have a little money parked on openrouter so I can get access to some of the free inference that gets tested there frequently. I have my opencode go plan.
I'm using hermes agent through nous research, have been extremely happy with it, and use the more or less free models from opencode go.
These things are fairly competent. I would equate them to a good junior developer. what I have noticed is that I have to provide a lot of the thinking around how it operates and if I don't then it gets off track and builds stupidly. I noticed in one of the notes that it was running into problem on line 7219. That told me one of the files was monolythic and had 7219 lines or more. I had to refactor the whole thing and spent a day or two doing that.
I use grok to help me. Grok is one of the more intelligent models I've found and I'll do periodic checkings with Grok when I'm building.
The good news is that this developer cuts development costs down to ~$10 a month.
The bad news is that it doesn't really make any of the random smart decisions that human developers do. So, it's way more ahnds on. I could pass a good vision onto my dev team with minimal instruction. Now I have to provide vision, put together a plan, and double check everyting it's doing for random acts of stupidity. I'm farly certain if I didn't have 10 years of experience doing this that I wouldn't be able to make anything useful with these tools, and the jury is still out if I'm just spinning my wheels or actually making something.
I've been having a good experience with deepseek flash 4. It had a hard time with some of the visual things so I would switch to qwen 3.6 or kimi 2.6, but for most stuff Deepseek Flash is fast, seems to work well, and doesn't give me lots of random hangups and I have a high use limit with it.
Anyway, if you're thinking about AI you can get a RTX 3090 for $1000-1500 on facebook marketplace and do this stuff locally (I've seen some qwen rigs on smaller setups), and otherwise you can snag some cheap inference at opencode go.