Post

Writing code has become a lot faster, but understanding code hasn't. What can we do about that?

cloudhead@cloudhead.io

The biggest problem with vibe coding is understanding the state of the code base at any given time. The reason to write code yourself, or to review all the code written by an agent is to have a clear mental model of the program. This is still a major bottleneck.

Don't say "there's no need to understand it", the code is not just the code, the code defines the program and user experience fully. Not having a clear grasp of this is a problem, because the agent doesn't make better decisions than you with regards to your product.

8/1/26, 11:36 am0 reposts1 like

My working hypothesis is that we need higher level language for domain modeling, one that is more like spec defining expectations and helps you just like type system help you think about all possible states and relations across component.

I am also inclined to think algebraic effects are part of this equation as a way to narrow down possibility spac and definition constraints more clearly

I think acornprover.org is a good example in this direction where how LLM came to prove is not important because expectations are well defined and so are capabilities. However this is very narrow use case and I think we need something like that but more general purpose

acornprover.orgThe Acorn Theorem ProverPerfectly rigorous mathematics, as simple as possible.

Another analogy I’ll make is how say when one write in JS they don’t think about what machine instructions VM generates or how it manages memory necessarily. I think same applies we need different language for modeling problems that LLM can take and generate instructions almost like JIT