1–10 of 10 posts
Some of the recent examples include the Bun Rust migration, etc. I wanted to understand both from startup’s and experienced individuals how do you see the trend, are we really moving towards a workflow where we simply optimize the loop or graph and code is merely an artifact this is produced as part of this workflow similar to bytecode or something?
Wrote about it here: https://aq.dev/guides/how-to-review-an-ai-coding-session/ and how we do it here: https://aq.dev/guides/how-to-review-an-ai-coding-session/
I hope this helps.
Disclosure: I'm the founder of AQ.dev
I see AI-assisted software development as just another step up the stack. My goal in development is to solve a problem. I look for the most appropriate tools to solve that problem in the most reasonable way possible. For my entire career, that’s included using plenty of libraries that I didn’t write and don’t fully understand. It’s been abstracted away from me, so I can simply “look at the output” and move onto the next step to solve my problem.
To me, AI-assisted dev is very similar. The key difference being that the LLM is “producing” the “library” on the fly, customized to the prompts and context. So it hasn’t gone through the rigor that a properly reviewed and maintained OSS library has gone through. But then again, the models generating the code were trained on essentially all OSS that could be found, so in theory it builds on the collective principles.
I read the code if I’m curious, or if I’m feeling frustrated and in a mood to debug. But if the output looks good and it solves a problem, I’m moving onto the next one. I don’t expect it to be perfect. Software has had and always will have bugs.
I need to be able to answer questions about what was built and maintain it. I’m ultimately responsible for it.
For something for myself, that won’t be deployed, it depends. If it’s an area I know, I’ll give it a quick look over to make sure some prompt injection didn’t stick something in the code that’s going to do something malicious. If it’s an area I don’t know and I want to learn, I will use the LLM as a jumping off point for how things connect, then make my own so I can actually learn it. The review might make sense, but it isn’t until I actually do it myself that I understand how it actually works… that’s when I realize what I thought made sense in the review was wrong, or at least not the full picture.