Now there are loads of fixes for that. But human peer review takes the super-fast out of using AI. You can review faster than write and test, that's for sure, but the best peer reviewing comes from people actually working in the code base, writing real code with their teammates.
Peer reviews by people not working on the code base, are far, far less viable. You'll catch big mistakes, but not mistakes predicated upon a code base you never ever touch or read in depth.
So you're left with lots of fixes we can suggest. I often use another model, eg, fable reviews once in a while on opus work, or I might use chatgpt for an alternative look. But here's the skinny...
Lots of devs do nothing of the sort.
And all current AI does seem to have similar blind spots.
I'm confident opus5.5 could easily re-write the entire codebase. Really, consider its age it cannot be that big of an existing code base. However, are there any devs alive working on that code base?
And do those devs work actively, or have any fears about changing each and every byte? Do those devs understand the difference between 'that was a dumb way to write that' and 'oh, that's actually an oddity so that planes won't smash into each other in the sky'.
So:
* true human review isn't possible unless humans are working actively, and know the codebase
* few do any code review, even different model code review
* no one likely knows why design decisions were made in the software, so the biggest challenge is why was everything done precisely the way it was, which means the current biggest challenge is a series of ramp ups where the coders need to very much know ever aspect of traffic control in depth, then look at the codebase...
Because opus5.5 is good, but context windows aren't going to hold that.
I stand by this firmly, AI=awesome, no devs working alongside to have skill to review=insane.
AI is absolutely a multiplier but people too eagerly try to dismiss existing good practices with "just let AI vibe code it for a week, surely it will be great"
It's delusional
I have no doubt that before the end of this decade, Claude could re-write the air traffic controller software, formally verify it, test it exhaustively, deploy it, and monitor it in operation. It might be able to do it now.
This is not a bubble. Very soon, it will be criminal negligence to run critical infrastructure on insecure, decades old software. Anthropics, or others like them, will be central to the functioning of the computer systems on which modern life relies.