Experienced Engineers Remain Essential for Reviewing AI-Generated Code

The thread

Also, if you’re a dev and don’t read the code that is OK but we probably need a name for it (it’s happening a lot, for stuff that’s not important / the architecture is set in place so the model cannot mess up much / there’s lots of validation etc)

@GergelyOrosz The frustrating thing here is I think Mitchell is replying to a bot - I got caught out by one similar to that today too

@_halshin @GergelyOrosz @mitchellh They're already good enough to review code but they're just not great at it, just like writing code.

@GergelyOrosz @mitchellh Not only does it not work as reliably *today*, but it’s not a subjective human consciousness, which means it can never verify something for a particular human purpose. You will literally always need a skilled human operator.

@GergelyOrosz @mitchellh If you can read the code you may as well take another step and write it

@GergelyOrosz @mitchellh A experienced engineer with good coding agents, will produce much better outcomes! Especially when it comes to more critical/complex projects!

@GergelyOrosz The telltale hint is an account wi the hardly any followers and thousands of posts, most of which are replies that pose a question

@GergelyOrosz @mitchellh I think the core of this question though is right and it's basically hand-wavy. "Is this approach better or worse" is basically vibes or programmer judgement but very rarely are people weighing against the alternative approach.

@GergelyOrosz @mitchellh I think people are more worried about demand dropping for the role since each individual developer can do more stuff, and the pipeline crumbling for juniors, not whether every single role will be automated or not

@GergelyOrosz @mitchellh AI写代码最大的坑不是能不能跑，而是边界条件、架构取舍和长期维护。没有经验的人只会看表面通过，老工程师的价值反而更像质检和系统设计。

@GergelyOrosz @mitchellh But on the other hand, the industry stopped hiring juniors.

I do believe that at some point, AI will actually be good enough to review the code too.

Experienced engineers will still be valuable in other ways, but being able to read the code is becoming less and less important.

@GergelyOrosz @mitchellh FrontierCode is a good code benchmark. It's a matter of time until LLMs will ace this benchmark. https://cognition.com/blog/frontier-code

I see it as the shift in the development paradigm.

"Classical" dev is additive: a human adds lines of code with purpose. They come with bugs and technical debt but as long as the business workflow happen it's all good.

Now AI-Generated code it's reductive. An AI slops all the bugs and technical debt upfront and it is on the developer now to apply reductionist methods to clear the slop/bugs/techdeb. As long as the business workflow happens, it's all good too.

Which one is better? That's for researchers to actually investigate, we haven't had this sort of start with everything and clean it paradigm but the result is there: business workflows still happen either way.

@matchaboar @mitchellh Come on, are you trolling??

Yes ofc we could do it, but agents do it more efficient and faster and parallel

This surely was trolling, blocked b/c I don’t appreciate it

@GergelyOrosz If you don't read your code you're a slop slinger. If you still read it you're still just a developer.

@GergelyOrosz @mitchellh This scope for subpar agentic engineering skills or lack of tokens.

@GergelyOrosz @mitchellh Some folks have taken reading the code to whole new level. Methodically understanding it:

@mSanterre @GergelyOrosz @mitchellh Yes, but not good enough to not read the code for production systems. Not yet, at least.