[sixtyj]

I have talked with some friends who are long-time programmers (20+ experience). Even they (all) admit that they use Claude Code, OpenAI Codex, Google Antigravity or AI Studio - you name it.

So in future everything’s gonna be “agentic”, (un)fortunately.

Everytime I write about it, I feel like a doomsayer.

Anthropic admits that LLM use makes brain lazy.

So as we forgot remembering phone numbers after Google and mobile phones came, it will be probably with coding/programming.

[reconnecting]

OK, let's say there are two categories of software now.

One is where the human has a complete mental map of the product, and even if they use some code generating tools, they fully take responsibility for the related matters.

And there is another, emerging category, where developers don't have a full mental map as it was created by an LLM, and no one actually understands how it works and what does not.

I believe these are two categories that are currently merged in one Show HN, and if in the first category I can be curious about the decisions people made and the solutions they chose, I don't give a flying fork about what an LLM generated.

If you have a 'fog of war' in your codebase, well, you don't own your software, and there's no need to show it as yours. Same way, if you had used autocomplete, or a typewriter in the time of handwriting, and the thinking is yours, an LLM shouldn't be a problem.

[slibhb]

> And there is another, emerging category, where developers don't have a full mental map as it was created by an LLM, and no one actually understands how it works and what does not.

I work with a large number of programmers who don't use AI and don't have an accurate mental map for the codebases they work in...

I don't think AI will make these folks more destructive. If anything, it will improve their contributions because AI will be better at understanding the codebase than them.

Good programmers will use AI like a tool. Bad programmers will use AI in lieu of understanding what's going on. It's a win in both cases.

[seanmcdirmid]

> And there is another, emerging category, where developers don't have a full mental map as it was created by an LLM, and no one actually understands how it works and what does not.

Are the tokens to write out design documentation and lots of comments too expensive or something? I’m trying to figure out how an LLM will even understand what they wrote when they come back to it, let alone a human.

You have to reify mental maps if you have LLM do significant amounts of coding, there really isn’t any other option here.

[hluska]

That sounds exactly like using a new library. The absolutists miss so much and contribute so little.

[holistio]

We never really have a complete mental map.

"Oh, this library just released a new major version? What a pity, I used to know v n deeply, but v n+1 has this nifty feature that I like"

It happened all the time even as a solo dev. In teams, it's the rule, not the exception.

Vibing is just a different obfuscation here.

[Tade0]

I don't think so. At the end of the day it's just a tool.

Case in point: aside from Tabbing furiously, I use the Ask feature to ask vague questions that would take my coworkers time they don't have.

Interestingly at least in Cursor, Intellisense seems to be dumbed down in favour of AI, so when I look at a commit, it typically has double digit percentage of "AI co-authorship", even though most of the time it's the result of using Tab and Intellisense would have given the same suggestion anyway.

[hypercube33]

Weirdly I actually hate tab complete in Cursor but love it with GitHub Copilot. I feel inversely about their agentic chat however.