Hacker News Reader

Pretty much nailed it. Once you’re at about 40k LOC you can just turn off the autocomplete features and use Claude or GPT to evaluate specific high-level issues. My sense is 40k LOC is the point at which the suggestions are offset by the rabbit holes they sometimes send you down, but, more importantly by obscuring from you the complexity of the thing you’re building—temporarily.

View on HN

I expect much of this can be solved with better architecture: smaller, more independent components. Break large code bases up into independent libraries, and LLMs can work again because they need much less code in their context.

Suggesting I restructure my code base to work better with a non-deterministic black box feels like an instance of Conway’s law[1] that should be avoided.

https://en.m.wikipedia.org/wiki/Conway%27s_law

You’re free to avoid it but maybe your competition will do it and may outpace you because of it. Conways law just states that architecture follows organizational structure so it’s only to be expected that ai generated software has an ai influenced architecture.

Yes, I think that’s the other position. But my bet is that significant software projects built today on today’s LLMs will be outpaced by teams in the future who avoided delegating that complexity to them. It’s a bet, but perhaps it’s wrong.

Reply on news.ycombinator.com