In my last article on coding with AI, I was very impressed by the output produced by AI code gen tools and the speed at which these tools allowed us to prototype. Here is that article for reference: https://www.explosion.fun/blog/post/rebuilding-my-website-with-ai/
This time I was much more bullish on the capabilities of AI tools, and I employed them during the making of my React-Native Fitness application. Though these tools were markedly less effective in this category. React-Native I suppose forms a significantly smaller pie of ‘all the code in the world’ than web development or NextJs does.
I can see this being a real problem for any subset of programming that doesn’t get as much attention as NextJs. Perhaps programmers would still have a job in the AI driven utopia future.
I should delineate how I used AI for this experiment and the tools I utilized. I started off with Augment Code, an AI coding helper said to be as good as Cursor AI. When I used it though, I didn’t think it was, so I switched to Windsurf – a strong Cursor competitor.
Windsurf though has apparently lost it’s place after OpenAI acquired Windsurf, since Anthropic had stopped supplying Claude to Windsurf. Windsurf was so much worse at programming without Claude. Later Anthropic agreed to be present on Windsurf again, though as users we would have to BYOK (bring your own key) to use Claude on the Windsurf editor. So I found myself deleting Windsurf a few hours after the experiment started.
I decided I would stick to Cursor’s free tier or just do it myself. The free tier was doing much better than Windsurf was performing.
I used Stytch, Google’s AI UI design platform to design a few panels and they came out really good. They provide an option to copy paste the design onto Figma, though that feature didn’t work as smoothly as I would have wanted.
Right around this time, Google released their Gemini CLI to the public and made it Open-Source. So I figured I would try it out since it managed to design UI panels really well. Maybe it could do the app quite well too. Gemini CLI did not manage to contribute meaningfully to the application.
The poor performance of AI code gen tools in making mobile applications was giving me confidence about the future software job market. I pray my confidence lasts.
Is AI Net Negative?
I’ve programmed some more.. the code base grew.. I let the AI run rampant on the code base and now the AI tools have no idea what’s going on.
Here is a panel and some other artifacts the AI created. At the top you can see a white bar with (tabs) written in within it. That shouldn’t be there.. So I asked Cursor/Claude why it was there. After some analyzing it agreed that it shouldn’t be there.. AI added that and now didn’t know how that got there.
It was something similar with every component AI had built, everything needed re-writes, large scale re-writes. I believe it did contribute significantly at the very start, but it can only do the absolute basics.
I just have trouble seeing this as the reason for the current decline in hiring.