I don’t review all the code generated by an LLM, just like I’ve never read all the code in libraries I use.
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'm retired from a 20+ year career as a software developer and today I build (I almost said "I write" lol) more applications now than I ever have in my life. I build constantly, all the things I've always wished somebody else would make. Seemingly millions of lines of code.. and I rarely even take a peek anymore. I did for the first couple of months, mostly because I felt compelled to keep up with what was being written but, eventually, especially since Opus 4.8 and beyond, it's just extra time I don't need to spend. I've come to trust Claude to a very high degree. It's a whole new paradigm, obviously. To have an idea and not really give much thought to how to write it, more important is how to describe it in as much detail as your right brain can conjure. I still write sometimes for fun and that's when I get my alone time with lines of code.
I'm a huge proponent of that with my team. We stopped doing PR reviews and instead do session reviews. It just makes so much more sense now. It's more important to understand the conversation, context, decisions made, why were they made that way, and so forth. More than the lines of code themselves, especially with the massive increase in PRs volume.
AI makes mistakes like humans. We used to review human's code because humans make mistakes. What the hell has changed? I mean it's pretty effective to have another agent review the code too, but I still find it doesn't catch all of the things I do. Also don't tell me the tests you didn't even read will make sure it all works :)
This is where industry is heading long-term, yes. Will take many years to get there. My favorite examples so far are from StrongDM. See https://factory.strongdm.ai/
I don’t review all the code generated by an LLM, just like I’ve never read all the code in libraries I use.
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'm retired from a 20+ year career as a software developer and today I build (I almost said "I write" lol) more applications now than I ever have in my life. I build constantly, all the things I've always wished somebody else would make. Seemingly millions of lines of code.. and I rarely even take a peek anymore. I did for the first couple of months, mostly because I felt compelled to keep up with what was being written but, eventually, especially since Opus 4.8 and beyond, it's just extra time I don't need to spend. I've come to trust Claude to a very high degree. It's a whole new paradigm, obviously. To have an idea and not really give much thought to how to write it, more important is how to describe it in as much detail as your right brain can conjure. I still write sometimes for fun and that's when I get my alone time with lines of code.
I'm a huge proponent of that with my team. We stopped doing PR reviews and instead do session reviews. It just makes so much more sense now. It's more important to understand the conversation, context, decisions made, why were they made that way, and so forth. More than the lines of code themselves, especially with the massive increase in PRs volume.
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
AI makes mistakes like humans. We used to review human's code because humans make mistakes. What the hell has changed? I mean it's pretty effective to have another agent review the code too, but I still find it doesn't catch all of the things I do. Also don't tell me the tests you didn't even read will make sure it all works :)
I read the code reviews that Claude leaves on my PRs if that counts.
This is where industry is heading long-term, yes. Will take many years to get there. My favorite examples so far are from StrongDM. See https://factory.strongdm.ai/