How Generative and Agentic AI Shift Concern from Technical Debt to Cognitive Debt
4 points by Corbin
4 points by Corbin
Not to sound too much like Hella Jeff, but we've been warning about the failure of vibecoders to build Naur theories since the advent of vibecoding. One of our prominent vibecoders is personally experiencing this issue.
I told you dog!
build Naur theories
I wasn't familiar with this but it's extremely topical to my journey at the moment because I've been recognizing the same concerns around offloading so much of our work to AI development. Thank you for pointing out the salience of this.
It is one of the tradeoffs. Often agents can help defer some theory building. A little nudge in the prompting can do a lot to convey high level theory, and often you don’t need to be concerned with the line by line level theory.
Yes, I’m saying we have programs that no human understands, but that was always true once authors forgot. Now we have software that we have to learn about because no human wrote every line.
I think this is largely fine because we’ll have more in the way of detailed templates and documents that guided the initial generation of the code.
document not just what changed but why, and create regular checkpoints where the team rebuilds shared understanding through code reviews, retrospectives, or knowledge-sharing sessions.
This is something I've felt strongly about on my teams lately. I started doing a practice called a "system review" where - after finishing a feature delivery that involves any kind of significant architectural updates to a given system - we update our architectural diagrams and have a brief 15 minute overview with the rest of our domain outlining, "Here's what the system looks like. These are the systems that it depends on and the systems that depend upon it. Here's the people who have worked on the system that you can talk to." The goal is to give a really low impact and high level overview to give people a chance to make soft connections between their own work and the system under review. I can't say whether or not it's the most effective thing in the world (after all, couldn't anybody just get the same thing from an AI summary?), but doing nothing feels like it certainly won't help either.