Sue SmithSeeing stuff like this and the discussion around it I keep checking if "choose human readable implementations" is a thing yet lol, there are various agent skills etc for simplifying code, but the basic idea that this is an automation for generating code that human beings should be able to continue working with directly still seems rare
We're now at least a couple of years into the knowledge that code review has become prohibitively challenging and yet the overwhelming majority of attention paid to the problem has been on automating code review rather than making it easier for humans, not to mention the multitude of problems that follow shipping
It's like once you've committed to the automation path you must remain on it.. But resilient systems provide ample opportunities for course correction, we seem to be missing that entirely here
Some of my favourite examples of this are in modern road systems (aye fuck cars, it's an intuitive demo) where instead of putting all of the signage into avoiding people getting in the wrong lane, they assume people will get in the wrong lane and provide easy paths to get back on course
Ken Milmore@sue How good are these models at modifying a pre-existing AI generated codebase? Or do users tend to scrap and rebuild from prompts every time a major change is needed?
@kbm0 I think they can be not bad at refactoring etc but it depends on the codebase of course
@sue I'm curious about what would be thrown up by getting a model to describe a pre-existing system in natural language, modifying the description, and then rebuilding from that description. I assume this might behave very differently depending on the scale and complexity of the code base.
When I say I'm curious, I don't mean I actually want to try it...
When I say I'm curious, I don't mean I actually want to try it...