Guess what: This has all changed with LLMs. The impact of the average developer is proportional to the amount of code they ship in a world of copilots (key word is "ship", not write) and that's why some of the staff engineers I work with have 3 concurrent Cursor projects for the same codebase in flight at the same time. These are some of the best engineers I've ever worked with and they have drank the cool-aid of volume over quality.
I'm keen to this because I maintain our CI systems and have become acutely aware of the overhead of hallucinations breaking our CI tooling in pathological ways that draw me in to diagnose. A year ago I would have to log into our CI Kubernetes cluster to diagnose a busted build that doesn't self-report failure maybe... once a month. These days it's a couple times a week. LLM based dev is both amazing in the legit force multiplier it adds to writing code as well as the way it introduces some of the most incoherent and silly ways it breaks existing conventions.
I guess the headline is correct in that we are not hired to write code anymore, instead we are hired to shepherd code now, and a lot of it. And a lot of this code we shepherd is good enough but some amount of it is bad enough to break existing processes, but that is secondary to the volume and velocity we perceive from LLM code gen.
I'm keen to this because I maintain our CI systems and have become acutely aware of the overhead of hallucinations breaking our CI tooling in pathological ways that draw me in to diagnose. A year ago I would have to log into our CI Kubernetes cluster to diagnose a busted build that doesn't self-report failure maybe... once a month. These days it's a couple times a week. LLM based dev is both amazing in the legit force multiplier it adds to writing code as well as the way it introduces some of the most incoherent and silly ways it breaks existing conventions.
I guess the headline is correct in that we are not hired to write code anymore, instead we are hired to shepherd code now, and a lot of it. And a lot of this code we shepherd is good enough but some amount of it is bad enough to break existing processes, but that is secondary to the volume and velocity we perceive from LLM code gen.