Our New AI Website Builder

Maybe I'm wrong, but I can't help but feel this is a solution in search of a problem.

I feel like "building a simple website" has been a solved issue with templates for decades now. The only thing you need to add is the text, and for a useful website, you're already going to have to be typing 90% of that into the prompt anyway - most of what an LLM is going to add will be more of a value-subtract than a value-add.

Sure, that still leaves the tweaking and customization, but I feel like that's the part most people enjoy the most? Humans love decorating.

Maybe I'm seeing this the wrong way, and I'm forgetting truly non-technical folks exist, and this is for the people who would otherwise be forcing their nephew to help them make a basic website, and that's the role the LLM is playing here, as a conversational interface. I think the marketing copy for this announcement is total bullshit, then (plastering "AI" all over the announcement is more for marketers than customers), but I can at least see that use case.

I have no idea about the quality of the sites this produces, but I think your comment about non-technical folks existing is on the money. There are many people who really don't understand how to properly structure even a simple website, and being walked through a conversation with a series of pertinent questions will be a much more satisfying process.

Going from a basic template to all the little changes you need to make to build the basic shell you need for a new site is still a lot of work. Think of all the best practices cruft required for the pages typically linked to from the footer. A sporting good site is different than a directory blog is different than a single serving social media aggregation site.

I think it's plausible for AI to help with the tedious setup stuff and get you to the part where you start making it your own.

I am one of them. Without help from a capable and kind developer friend back in 2004 I never ever would have been able to create mine.

I signed up out of curiosity and it looks like it's supposed to complement the manual editing UI so you can get things done without digging through menus. After the initial prompt and wizard, it's a chat box that sits in the corner of a typical WP admin page.

But in its current state, it seemed pretty broken to me. I just wanted it to add text to the top of the front page, and it kept saying "I couldn’t find the block you mentioned. Describe where it is on the page or select it and try again." no matter how many different ways I attempted to describe it.

> Sure, that still leaves the tweaking and customization, but I feel like that's the part most people enjoy the most? Humans love decorating.

Even if that were true (which I heavily contest), people might like the idea of "I want the sidebar nav moved to the right side", not opening up template.css and template.html and figuring out which html/css they need to change.

The LLM is the thing that lets us do the fun part.

But let me disabuse you of the claim that technical people enjoy fiddling with html/css/design especially on their Wordpress website when they just want to make some changes, and somehow nontechnical people are the only ones who might have to circumvent all that fun-having by letting an LLM do it for them.

It's like saying that you don't see how LLMs could be useful to software developers because don't they enjoy writing code? Aside from the answer being no, most code isn't fun to write, you're forgetting the goal day to day is to get something done, not dick around with your Wordpress theme or software Jira tickets because it's fun.

This is a solved problem even in the sense that nearly all the Wordpress competitors already offer this.

Jira tickets: depends on what metrics your management is using?

It may well be your life goal.

another term for it is "investor driven development".

An AI website will look about as unique as a Facebook profile in 2004. Even the most non-technical people can tell a designer "make it look like this other website that I like". An AI won't be able to understand that, because it can't surf the web, locate the site and analyze what makes it good in the eyes of the customer.

> I feel like "building a simple website" has been a solved issue with templates for decades now

I did this recently for a friend and 1000% no. It wasn't easy to find a good template or edit it, and things we tried (including various builders) were a massive pain in the arse.

I asked sonnet for a site and had it right in a few minutes. I asked for changes and they just worked. It wasn't a complex site but it was drastically easier, quicker and more fun than dealing with the nuts and bolts of it all.

> Sure, that still leaves the tweaking and customization, but I feel like that's the part most people enjoy the most? Humans love decorating.

Absolutely not.

Choosing customisation? Sure!

Making the customisations? Nope. I'm sure some do, but I and many others I think just want a thing.

Just asking for some changes and seeing them was great.

> Even the most non-technical people can tell a designer

The vast majority of people who want to make a site don't have a designer, and if they could get one the comparison is something that's near instant and costs pennies.

> An AI won't be able to understand that, because it can't surf the web, locate the site and analyze what makes it good in the eyes of the customer.

Perhaps ignoring the "what makes it good in the eyes of the customer", although I'd argue that point for many things, these systems often can surf the web, can locate sites, can take images as input and already know many major themes and major sites.

> An AI won't be able to understand that, because it can't surf the web

We've given the AI the ability to scrape websites, so I'm not sure that holds.

> although I'd argue that point for many things, these systems often can surf the web, can locate sites, can take images as input and already know many major themes and major sites.

But that's not what's being offered here, and my bet is it won't be. It's more lucrative for businesses to put out LLM freemium shovelware and claim it replaces people than do the enormous work of developing the custom models that actually can.

Scrapes are cached, it is too expensive and too long to do them on-demand based on a live query. Lots of sites change designs frequently and don't output semantic code in the html source either. And good luck getting the customer to prompt correctly for things like background images, gradients, carousels, animations and page transitions.

We are actually solving this problem by building the Huggingface Hub[0] for browser automations -automatically tested and updated around the clock, and a homegrown distributed automation framework [1] that can leverage your own browser, your community’s or your team’s up to millions of orchestrated browsers.

It’s not just for AI, we created a custom package format[2] for automations that can run in CLI as a one-off, a REST API server, an MCP server with a single command..

0: https://herd.garden/trails

1: https://herd.garden

2: https://herd.garden/docs/trails-automations

Those are not custom models, image input is standard in flagship models and searching is offered by multiple providers.

What did Sonnet build it on, that you ended up using it?

One of the drawbacks of all frontier models (with Sonnet being the most extreme) for web dev is their strong preference for React in any kind of situation, even for Wordpress-blog equivalents where it makes no sense.