Hacker News Reader

Comment by softwaredoug | Hacker News Reader

I find when I do a lot of AI coding, or any "blocking" task with AI, I inevitably end up scrolling social media. And I feel dumber.

I'm left wondering whether I should have just hand-coded what I was doing a bit slower, but kept my attention focused on the task

View on HN

Waiting for the LLM is the best time to do the deeper review of the last output.

I like to fire the model off to do exploratory implementations as I refine the existing work.

This sounds nice, but what I've run into is that the model fails to write changes if the code has changed under it. A better tool, where it takes a snapshot at the start of each non-interactive segment, and then resolves merge conflicts with my manual changes automatically, would make this much easier.

I run the model against its own dev branch and either cherry-pick commits or do merges the old-fashioned way.

I'm using Aider though, which makes this easy: it's just another tab in the terminal.

The worst part is when you find out your vibe coded stuff didn't actually work properly in production and you introduced a bug while being lazy. It's really easy to do.

Do (did) you feel the same when you wait for code to compile? Or wait for CI/CD?

The earlier days of programming had more "blocking" since compilation was quite slow. So the issue is obviously that "blocking", but social media.

That is both a sweeping generalization and plainly wrong. The "much earlier" days of programming had blazing fast compilers, like Turbo Pascal. The "earlier" days had C compiler that were plenty fast. Only languages like C++ had this kind of problem.

Worst offenders like Rust are "today", not "earlier".

I am old enough to remember a time when compilation could last minutes or even one hour, depending on what you compiled, and it was in the late 1980s.

Certainly, but that depended on your own choices in technologies. This condition is not time-dependent. You can inflict yourself the exact same condition by choosing Rust.

Compile times are faster now for C and C++ right? Some of it due to further compiler optimizations but mostly due to higher CPU power.

Still, you seem to be arguing that the choice should be Pascal instead of Rust. There is a reason why we choose these new languages: language features. Compile time is a lesser consideration.

No, I'm arguing that having more "blocking" time is not a function of time (early days of programming), but a self-inflicted choice. Yes, choosing C in the 80s meant self-inflicting yourself with "blocking" time, but there was the choice of Turbo Pascal, or Forth or whatever. Plenty of great choices that ran really fast on those old machines.

Do I mean that one should choose Pascal today? No, compiling C code today is really fast and has practically no "blocking" time. But you can still inflict yourself "blocking" time if you want, with languages like Rust.

Are things clearer now?

In C we had to resort to tricks like precompiled headers to get any sort of sensible compilation and it still lasted a minute for a decent library.

C++ was/is even worse what with generation of all the templated code and then through the roof link times for linker to sort out all the duplicate template implementations (ok, Solaris had a different approach but I guess that's a nitpick).

I have not worked on any large project in Pascal, but friends worked with Delphi and I remember them complaining how slow it was.

So, in my experience, it really was slow.

I used to work on a project that could take 30 mins+ to compile the entire project.

Nearly every time, your problems were detected _early_ in the process. Because build systems exist, they don't take 30 minutes on average. They focus on what's changed and you'll see problems instantly.

It's _WAY_ more efficient for human attentional flow than waiting for AI to reason about some change while I tap my fingers.

If you need an attention sink, try chess! Pick a time control if it's over 2 minutes of waiting, and do puzzles if it's under. I find that there's not much of a context switch when I get back to work.

I'm having the same problem. LLMs really take me out of the task mentally. It feels like studying as a kid. I need to really make a concerted effort; the task is no longer engaging on its own.

yet you kind of feel "too slow" if you don't use one?

It depends. For a task I know well the LLM is often much worse. If I'm being asked to do something brand new, the LLM does speed me up quite a bit and let me build something I might have gotten stuck on otherwise. The problem is that although I did "build the thing," it's not clear I really gained any meaningful skills. It feels analogous to watching a documentary vs. reading a book. You learned _something_, but it's honestly pretty superficial.

Because when you simply “read” you’re not necessarily learning. The illusion of knowledge is real, where you nod in agreement that you understood something but when it’s your turn to do it, you have no idea how to. You need to do something yourself to actually learn it, and it involves struggling, frustration, eventually insights etc.

As someone with focus problems, I find it more productive to have a conversation with ChatGPT (or Claude) about code. And avoid letting it make major changes. And hand code a lot with Copilot.

How slow the AIs respond provides some opportunity to work on two task at once. Things like investigate a bug, think about the implementation for something larger, edit code that experience has told you it will take just as much or more typing to have the LLM do it.

It's less cool than having a future robot do it for you while you relax, but if you enjoy programming it brings some of the joy back.

Human attention doesn't work this way. This type of task switching makes both tasks fairly inefficient.

"What was I doing again!?" is a big problem

They're not that slow! You want me to believe we've gone from programmers being so fragile that disturbing their 'flow state' will lose them hours of productivity, to programmers being the ultimate multitaskers who can think and code while their LLM takes 10 seconds to respond? /s

You could do other work while the AI does one task :)

That’ll likely degenerate into “I want my AI to do dishes and laundry so I can code, not code so I can do my dishes and laundry”

That's even worse, because I'm task switching between two intensive tasks

Sorry no, nothing is worse than doomscrolling social media.

Checking Facebook while you have a few minutes available isn't the worst thing in the world. Calm down.

Wait, people still use Facebook? :)

no, but zuckerberg has a fake HN profiles.

Reply on news.ycombinator.com