Post

I'd argue it is impressive that some of the generated front-ends are better than half-bad, considering these factors. Also, models have visual capabilities for while now. What models definitely lack is the "living process", how and why things get created, and the evolutionary pain that led there.

0 reposts6 likes

Exactly pointing out the issues. In my experience it’s great at writing functionality until it gets complex (context loss issue). It’s quite bad at design/visual stuff, despite models understanding “visual stuff” but that’s more Photoshop like visuals, not websites.

I do not think complexity is the issue here, at leat not how we understand it. Models clearly understand the mechanics, also what gets perceived as "good" (or palatable) and they excel at tackling what we think as complexity. Models have no problem holding a whole website in their context.

What I want to point out, what models lack, is what we call intention, motivation, experience through countless failures. They have to derive what we want, what our goal (their mission) is from a few tokens we hand them in the prompt. Therefore the results are fairly impressive IMHO.

Another way to look at this is that all the garbage those models get trained on gets filtered through an unrelenting training process, which rewards only the "better" outcomes. That a few silly tokens in the prompt interpolated/inferred with the outcome of this training process yield these results.

If you pay a lot yes, otherwise not. I had it for a couple of big app features/services that they lost context and broke their own previous code in the same session. But we’re not talking about normal websites here but complex services.

Those failures, if they are not due to older/subpar models, may be a failure to communicate properly with these models -> "context engineering". If anything LLMs excel at grasping complexity our brains cannot, at least not consciously/"on demand". See e.g. Mythos/security fixes.

Can you site which LLM you have been using along with any other tools for this article? I find a few MCP servers have improved design to code outputs like figma MCP but the results massively vary per model and still not anywhere near 100% accurate.

based on when i experimented with claude my hunch was that their visual model turns images into a relatively limited number of tokens plus it does OCR to capture text. it was terrible at spatial relationships... which is exactly what you need for front-end stuff lol