Bad Vibes Coding

Which is such wrong-headed bullshit, that I can only conclude they’re either literally incapable of introspection or reasoning, or they’re so deliberately lying that they’re due to be struck by lightning from on high at any moment.

Funny enough, some of the people we're talking about here explicitly say they do not introspect. They might be capable, but they're not interested in doing so.

I wish this author well. They need a hug.

"AI" isn't democratizing, that much is clear. I'll just point out that even if you don't pay a dime for it, the concept itself is anti-"democratizing".

When people say "democratize", the focus is on the short term. Using an LLM, you'll write a 400kloc project. In one week, maybe.

As someone who tried to make LLMs work in the professional environment I'm coerced to use them in, there is no good LLM coder. None. I've never seen a 5000kloc codebase written by an LLM that I've liked. You may not believe this, but LLMs cut so many corners that you would think they'd produce nothing but spheres. I honestly can't get over how impressively terrible they are. It's incredible how people treat the LLM ability to code. In 2026, hearing "LLMs are producing great code" just sounds like "I either don't know how good code looks or I don't care about it at all" to my ears at this point.

If you do somehow, by miracle, create a usable, relatively bug-free project, you'll inevitably rely more on the LLM the next time you need to create something. And thus begins a feedback loop. Use the LLM, rely on it more. Rinse and repeat.

If you do it without an LLM, however, the next time you write a project by hand, you'll be faster, better.

The "democratization" is nothing more than malicious fiction.

All I know for certain is that people are still out there insisting that it’s “democratizing.” That it’s a tool that allows people to do things they weren’t capable of before. That there are all of these creative minds with ideas that they’ve never been able to execute, because they don’t know how to program, or because they’re not accomplished artists.

I think this is a subtle and important point. Maybe I'm expanding this beyond what the author intended, but I think it's true that motivated humans will pretty much always find a way to execute on their ideas, barring some kind of physical or mathematical impossibility. Creative people learn what they need to learn, and combine various sources together in new and interesting ways to get their ideas out of their head. (In fact, constraints have long been known to be fuel for creativity itself.)

So the conclusion seems to be that having these everything-machines just makes it easier for people who weren't already motivated to be creative in a particular domain (writing, coding, music, art, etc) produce bland, mediocre results. Regardless of how you feel about this in terms of productivity, something is definitely lost there.

I found this post to be extremely hyperbolic and un-nuanced. The author is incapable of acknowledging any of the very real reasons why people use LLMs, as if there's only downsides and no upsides.

Have you decided that “ethically-trained” models that only run locally are the answer? We already had those, and they didn’t make much of a big impression on anyone, because they’re simply not useful. Janelle Shane ran a whole blog about them for several years, pointing out how comically inept they are.

I've personally been using local models for some basic voice controlled home automation stuff and they're more than up to the task. They're useful for simple programming tasks too. I don't think you can reasonably point to a blog from "several years" ago to make the point that local models are useless. We've had massive breakthroughs in AI over the past 4 years. An open source model that runs on a small personal computer today (with less electricity draw than playing a PC game) can handily outperform state-of-the-art paid models from two years ago.

A system that isn’t 100% reliable is 0% reliable.

This is the kind of absolutist rhetoric that is laced through the whole article. We have lots of systems that are less than 100% reliable, but still useful: spam filters, spell checkers, automated fraud detection, fuzzers, static analyzers, the list goes on. Hell, most software is less than 100% reliable. My linux kernel has crashed at least once, does that mean it's 0% reliable?

It's appropriate to take issue with all the very real social harms that AI is causing, but pretending like AI has zero useful applications is just out of touch with reality.

The author has an impressive track record. Among many other things, they worked on Grim Fandango!