It’s weird to me how controversial this take is here. It seems obvious that lots of people are learning to leverage LLMs for their dev work and that this isn’t going away. I’m personally skeptical we will ever get rid of human in the loop or even that we will improve output quality much from here, but I don’t think either is necessary for LLM use to become standard practice in software dev.

The output quality seems like it is already good enough for the industry so I don’t think the “ouroboros” problem will stop the trend. Even if LLM-generated code quality doesn’t improve at all from here they will continue to be adopted. I think the jury is still out on what impact LLMs have on learning but I do agree it is not looking good. I don’t think this will stop the trend though, just potentially produce an outcome where even fewer programmers understand what they are actually doing. I can see the risk of that resulting in a scenario where the capacity to keep the LLMs going becomes lost, it seems not very probable though and that instead a kind of stagnation would take over in which the capacity for progress via software development becomes much more limited. Regardless, I don’t think that the trend potentially resulting in everyone becoming too dumb to continue the trend would actually stop the trend before that failure state was reached. I think even knowing that LLMs taking over the software industry could result in the collapse of the industry is not enough to stop the people making these decisions or change the economic forces driving LLM adoption. It is a risk they are happy to take.

Setting all of that aside, my original point was that it is becoming impossible to avoid LLM-generated code and I don’t think we need LLM-generated code to become the majority of code produced for that to happen. Depending on how you want to count things we’re probably already at a point where one way or another you are interacting with code that came from an LLM. I think it’s probably kind of like trying to avoid AWS or Cloudflare and still use the web like a normal person, those days are gone.

I think for too many having code that simply works is enough, and LLM-generated code quality is likely to continue improving over the coming years at least to some degree. Claude Code is already hugely popular and used at a lot of companies. I don’t expect things like that to go away, they certainly won’t be getting worse and currently a growing number of devs apparently find them useful enough. I think it’s probably just a matter of time until the majority of devs are using tools like these at least to some extent. Do you think the trend of devs taking up LLM tools will stall out or reverse for some reason?

From everything I’ve seen, I don’t think you can realistically avoid vibe coded software going forward. We’re fast approaching the day when the majority of all new code is LLM output.