> According to Karpathy, vibe coding typically involves accepting AI-generated code without closely reviewing its internal structure, instead relying on results and follow-up prompts to guide changes.

What you are doing is by definition not vibe coding.

There are "Open" (anyone can enter) and "Women" (only women can enter) events.

If you did not have the "Women" category, then you would see only men play in these events as there are no active women in the top 100.

> provided you remember to turn on lossless in apple music

I love that you ended your rant with an audiophile myth :)

Lossless isn't even diminishing returns better. People can't tell the difference in an ABX test.

KDE plasma is the best DE that exists right now (once you configure it to mimic gnome 2).

> once you configure it to mimic gnome 2

Why is it better than Gnome 2 then? This is what I prefer (it's called Mate now).

I was a Mate user for ages. It's great. Unfortunately, the lack of development is starting to show. For example, no fractional scaling for 4k monitors.

I've configured KDE Plasma to look almost identical to Mate (the defaults are similar to Windows, nice, but I prefer the Mate layout):

- top panel / bottom panel

- desktop switcher bottom right

- task bar on bottom

- desktop button bottom left

- clock top right

- app indicators top right

- app icon launchers on top bar

- app menu top left

It's not just layout, either. Gnome can be configured to do much of this, but it just feels terrible. Task bars can't be dragged to re-order. Desktop switchers just have numbers instead of contents. Animations are slow and annoying. Etc. Etc.

> For example, no fractional scaling for 4k monitors.

I don't have a 4K monitor, but what does `xrandr --output HDMI1 --scale 0.8x0.8` do? I have a 1024x768 monitor and do to all the useless whitespace in modern programs, I scale into the opposite direction.

But I agree Gnome lost the plot completely, and sadly Gtk too. Which is a pity, because I prefer GTK+ to Qt, but they deprecated so much useful Widgets and the alternative given is 'just don't do that'.

You can use xrandr to scale the desktop, but it's not the same thing.

You can render at 2x in Mate, and then scale it down slightly (ie 1.25x1.25) with xrandr, but taking a large image and scaling it down using a cubic filter won't look as sharp as real fractional scaling.

The command you gave is upscaling, which will be worse than 2x + downscaling.

Real fractional scaling scales the sizes of elements before rendering. This results in the sharpest image and there is no resizing/filtering in the loop.

Then I don't understand how this actually works. Doesn't this require support by the underlying UI kit? Because after the UI kit, there will already be pixels and scaling that will always be blury.

Yes, it needs to be supported by the UI kit. That's why GTK needs to support this, not just Gnome. Gnome uses GTK4, which supports this. Mate uses GTK3, which does not.

KDE Plasma uses QT6, which also supports this.

So applications using e.g. Motif or some custom UI toolkit have completely wrong sizes then?

> Gnome uses GTK4, which supports this. Mate uses GTK3, which does not.

Damn. I tend to build against an older GTK3 version, because GTK deprecated so much good stuff, but that means my programs won't work correctly for that. I need to look whether it's easy to backport this.

I'm not sure about GTK3. But I know that KDE plasma has a setting for legacy X11 applications (I think all this fractional scaling stuff requires wayland). It can render them as-is or scale them up (bitmap scaling).

Gnome 1 had the best design of all the Gnome versions.

Try again with Opus 4.5

Try again with Sonnet 4

Try again with GPT-4.1

Here I thought these things were supposed to be able to handle twelve lines of code, but they just get worse.

The goalposts have been moved so many times that they’re not even on the playing field.

Nahh, just trying to make it concrete. I could instead just ask which model they used instead.

"High" the the reasoning level. The context window never changes.

You're right! Still learning the details of this agentic stuff; I was pretty late to the party.

I have experienced much of the opposite. With an established code base to copy patterns from, AI can generate code that needs a lot less iteration to clean up than on green fields projects.

I solve this problem by pointing Claude at existing code bases when I start a project, and tell it to use that approach.

That's a fair observation, there's probably a sweet spot. The difference I've found is that I can reliably keep the model on track with patterns through prompting and documentation if the code doesn't have existing examples, whereas I can't document every single nuance of a big codebase and why it matters.

This is confusing.

TFA says they added an index to Agents.md that told the agent where to find all documentation and that was a big improvement.

The part I don't understand is that this is exactly how I thought skills work. The short descriptions are given to the model up-front and then it can request the full documentation as it wants. With skills this is called "Progressive disclosure".

Maybe they used more effective short descriptions in the AGENTS.md than they did in their skills?

The reported tables also don't match the screenshots. And their baselines and tests are too close to tell (judging by the screenshots not tables). 29/33 baseline, 31/33 skills, 32/33 skills + use skill prompt, 33/33 agent.md

Good catch on the numbers. 29/33 vs 33/33 is the kind of gap that could easily be noise with that sample size. You'd need hundreds of runs to draw any meaningful conclusion about a 4-point difference, especially given how non-deterministic these models are.

This is a recurring problem with LLM benchmarking — small sample sizes presented with high confidence. The underlying finding (always-in-context > lazy-loaded) is probably directionally correct, but the specific numbers don't really support the strength of the claims in the article.

I also thought this is how skills work, but in practice I experienced similar issues. The agents I'm using (Gemini CLI, Opencode, Claude) all seem to have trouble activating skills on their own unless explicitly prompted. Yeah, probably this will be fixed over the next couple of generations but right now dumping the documentation index right into the agent prompt or AGENTS.md works much better for me. Maybe it's similar to structured output or tool calls which also only started working well after providers specifically trained their models for them.

It was a fake story he made up to help in his job search. Don't expect any of the details to add up.

A 2020 Tesla Model Y and 2026 Tesla Model Y are at least as different as a 2020 Nissan Rogue and a 2026 Nissan Rogue.

So buy nissan stock?