So instead of refreshing that less used knowledge I just ask the AI to do it for me. The implications of this vs searching MDN Docs is another conversation to have.
They don't think like graphic designers, or like programmers. It's not easy for beginners. It's not aimed at ease of implementation. It's not amenable to automated validation. It's not meant to be generated.
If there is some person for whom CSS layout comes naturally, I have not met them. As far as I can tell their design goal was to confuse everyone, at which they succeeded magnificently.
Before 2017, the web had no page layout ability.
Think about it. Before the advent of Flexbox and CSS Grid, certain layouts were impossible to do. All we had were floats, absolute positioning, negative margin hacks, and using the table element for layout.
> They don't think like graphic designers or like programmers. It's not easy for beginners.
CSS is dramatically easier if you write it in order of specificity: styles that affect large parts of the DOM go at the top; more specific styles come later. Known as Inverted Triangle CSS (ITCSS), it has been around for a long time [1].
> It's not aimed at ease of implementation. It's not amenable to automated validation.
If you mean linting or adhering to coding guidelines, there are several; Stylelint is popular [2]. Any editor that supports Language Server Protocol (LSP), like VS Code and Neovim (among others), can use CSS and CSS Variables LSPs [3], [4] for code completion, diagnostics, formatting, etc.
> It's not meant to be generated. Says who? There have been CSS generators and preprocessors since 2006, not to mention all the tools which turn mockups into CSS. LLMs have no problem generating CSS.
Lots of developers need to relearn CSS; the book Every Layout is a good start [5].
[1]: https://css-tricks.com/dont-fight-the-cascade-control-it/
[2]: https://stylelint.io
[3]: https://github.com/microsoft/vscode-css-languageservice
It's not the fault of CSS that most developers don't learn to use it correctly. That's like blaming the bicycle when learning to ride one.
Frankly, it's not a priority for most of them to learn CSS; they don't see it as a "real" programming language; therefore it's not worth their time.
It's not like blaming the bicycle, that's the whole point of my analogy to programming languages. Like I said, learning a new programming language in a few weeks of regular use is a common experience. This also happens with bikes, because you can try a few things, lose balance, make a few intuitive adjustments, and iterate easily.
This just doesn't work with CSS. There are so many pitfalls, corner cases and reasoning is non-compositional and highly contextual. That's the complete opposite of learning to ride a bike or learning a new programming language.
You literally do need to read like, a formal specification of CSS to really understand it, and even then you'll regularly get tripped up. People just learn to stick to a small subset of CSS for which they've managed to build a predictable model for, which is why we got toolkits like Bootstrap.
Edit: this also explains why things like Tailwind are popular: it adds a certain amount of predictability and composition to CSS. Using CSS was way worse in the past when browser compatibility was worse, but it's still not a great experience.
If I do this a few more times it might even stick in my head.
These days I use display: flex; so much that I wish the initial value of the display property in CSS should be flex instead of inline;
Wade through ads
Skim a treatise on the history of centering content
Skim over the “this question is off topic / duplicate” noise if Stack Overflow
Find some code on the page
Try to map how that code will work in the context of your other layout
Realize it’s plain CSS and you’re looking for Tailwind
Keep searching
Try some stuff until it works
Or…
Ask LLM. Wait 20-30 seconds. Move on to the next thing.
I did that a lot initially, it’s really only with the advent of Claude Code integrated with VS Code that I’m learning more like I would learn from a code review.
It also depends on the project. Work code gets a lot more scrutiny than side projects, for example.
Aaand done. Very first result was a blog post showing all the different ways to do it, old and new, without any preamble.
Given how often people acknowledge that ai slop needs to be verified, it seems like a shitty way to achieve something like this vs just checking it yourself with well known good reference material.
If you haven’t been convinced by pure argument in 2026 then you probably won’t be. But the great thing is you don’t have to take anyone’s word for it.
This isn’t crypto, where everyone using it has a stake in its success. You can just try it, or not.
But there is an interesting looking-glass effect at play, where the truth seems obvious and opposite on either side.
You: Hey ChatGPT, help me center a div.
ChatGPT: Certainly, I'd be glad to help! But first you must drink a verification can to proceed.
Or:
ChatGPT: I'm sorry, you appear to be asking a development-related question, which your current plan does not support. Would you like me to enable "Dev Mode" for an additional $200/month? Drink a verification can to accept charges.
Microsoft’s dedication to infusing OpenAI tech into everything seems like a play to cut even those tepid brains out of the loop and capture the vehicles of planning and production. Training your workforce to be dependent on third-party thinking, planning, and advice is an interesting strategy.
An article about the best open weight models, including Qwen and Kimi K2 [3].
Maybe in 3-5 years this will work on consumer hardware at speed, but not in the immediate term.
DeepSeek-V3.2-Exp has 37B active parameters, GLM-4.7 and Kimi K2 have 32B active parameters.
Lets say we are dealing with Q4_K_S quantization for roughly half the size, we still need to move 16 GB 30 times per second, which requires a memory bandwidth of 480 GB/s, or maybe half that if speculative decoding works really well.
Anything GPU-based won't work for that speed, because PCIe 5 provides only 64 GB/s and $2000 can not afford enough VRAM (~256GB) for a full model.
That leaves CPU-based systems with high memory bandwidth. DDR5 would work (somewhere around 300 GB/s with 8x 4800MHz modules), but that would cost about twice as much for just the RAM alone, disregarding the rest of the system.
Can you get enough memory bandwidth out of DDR4 somehow?
Come on!
As a bonus, I pay attention to what the AI did and its results, and I have actually learned quite a bit about how to do this myself even without AI assistance