From an SRE standpoint, the things I wonder about AI-written code changes include:
Does this code change accomplish the stated goal in a way that minimizes future maintenance work? Is it going to break in a way that makes me need to do more work in the future to fix it? Or is it robust against not only new traffic, but also future changes? Am I gonna get paged at 2am and wake up my partner and be frazzled the next day?
Does the agent have goals in life other than writing code and maintaining the system? Human tech workers have lives and families and hobbies; at a very salient level, they want to minimize the time they spend on work, in order to maximize the time they can spend snuggling their partner, playing hockey or Magic, petting the cat, cooking tasty food, or doing other life things.
Human engineers want to write a good code change, in part because they value good code, but also because they want to do things other than writing code changes; and they have a gears-level understanding of how the quality of the change I’m writing affects whether I can chill with my friends at the party this weekend, or whether I’m going to get paged to fix the service and not be able to enjoy the party.
I do find that a lot of the problems of AI labor seem vaguely related to the fact that the model doesn’t really experience the consequences of its work.
From an SRE standpoint, the things I wonder about AI-written code changes include:
Does this code change accomplish the stated goal in a way that minimizes future maintenance work? Is it going to break in a way that makes me need to do more work in the future to fix it? Or is it robust against not only new traffic, but also future changes? Am I gonna get paged at 2am and wake up my partner and be frazzled the next day?
Does the agent have goals in life other than writing code and maintaining the system? Human tech workers have lives and families and hobbies; at a very salient level, they want to minimize the time they spend on work, in order to maximize the time they can spend snuggling their partner, playing hockey or Magic, petting the cat, cooking tasty food, or doing other life things.
Human engineers want to write a good code change, in part because they value good code, but also because they want to do things other than writing code changes; and they have a gears-level understanding of how the quality of the change I’m writing affects whether I can chill with my friends at the party this weekend, or whether I’m going to get paged to fix the service and not be able to enjoy the party.
I do find that a lot of the problems of AI labor seem vaguely related to the fact that the model doesn’t really experience the consequences of its work.