Originally posted on Proof of Concept

My experience with the World Wide Web was like growing up with a childhood friend. The early days of building personal websites were what some would call Web 1.0—the static web. I’d observe my brother building HTML websites with Notepad, inline styles, and all, then publish them to Geocities. The Social Web (Web 2.0) brought user-generated content, web applications, and social networks. Because of these new technologies, people deviated arom personal websites in favor of simple about.me pages, popular blogging platforms, and exclusively publishing on Social Networks like Facebook.

We’re in the midst of Web 3.0—whatever you want to call it. Blockchain technologies and decentralization are capabilities often associated with this phase. However, Artificial Intelligence and Machine Learning are included i this era as well Let’s focus only on the implications of what AI/ML brings to the new web, particularly Large Language Models (LLMs). Because of its capabilities of data, fine-tuning, and generative capabilities, there are two likely scenarios of how LLMs can be used to create web experiences.

Scenario 1: website builders equip themselves with AI capabilities: Products built for making websites will add AI capabilities to their product offerings. Every SaaS product seems to be doing this now to make their product Copilot or Generative AI-ready.

Scenario 2: LLMs build websites: With experiences like Designer GPT and Grimoire, empowering LLMs to build websites is already possible today.
With products like Designer GPT, LLMs have generate websites for people.

I believer Scenario 2 will be adopted quicker than Scenario 1. In fact, glimpses of it is happening already. When The Browser Company launched the iOS app, they created a unique experience for the form factor. Instead of yet another browser, they created an experience focused on search. One of the key features is, “Browser For Me,” which has the app construct a simple web experience based on the search prompt. Perplexity recently launched Perplexity Pages, which builds a wiki-like website based on your searches.

Though I don’t think it makes sense for Arc and Perplexity to have customization (a different topic), you can see how other products can take the concept of “build it for me” for web experiences generated by prompts.

There are still bottlenecks that need to be solved to have generative websites be more widely adopted. There are a few core areas in building and maintaining a web experience:

• Content: Copywriting, creative assets, and information architecture
• Design: Layout, style, and aesthetics
• Tech: Frameworks, hosting services, and how the site is published/deployed
• Management: Maintenance, SEO, and other growth levers

As I say to Design Systems teams: it’s easy to generate something net new than it is to maintain it. This is a similar challenge with websites generated by AI. It gives you a first version to work from but you still need to maintain and host it.

What LLM-powered site building looks like

I believe there will be a lot of people in the world who will continue hand-crafting sites (myself included) because we have the ability and desire to do so. Let’s acknowledge that while speculating what a site-building experience looks like when it is more derived from a generative prompt.

Good web design starts with content—something the majority of people can create, on the other hand, is a skill that needs to be developed—one people might not bother to learn.

Authoring experiences will move closer to distribution

Professionals are always going to want their own space for their power tools. However, the rest of people often prefer convenience: not having to learn something new, reducing friction, and anything else that accelerates the job to be done. What we’ll see is authoring experiences more infused with publishing tools. For example, if a person is writing a blog post, they might generate creative assets, adjust layouts, and make stylistic changes in the content management system vs. an external authoring tool like Figma—faster editing that leads to faster publishing.

Settings and profiles facilitated by AI Agents

My bet is AI tooling not going to be a monolithic LLM—a god-like AGI. Instead, it’s having a hive of tiny LLMs interacting with one another to collaborate. Take the example of making changes to a web host provider. Instead of logging into Dreamhost and upgrading hosting services, there might be a chat service I build or interface setting I can change to invoke the changes on the hosting services.

A spectrum of customization continues to exist

Ben South said it well: “The average person doesn’t want to customize their software—they want the best defaults.” Even before AI became the talk of the town, the spectrum of customization exists in building web experiences and other types of publishing, with the majority of people simply wanting a thing. LLMs are decent at building these now and the workflows to maintain and publish will get there.

There will always be room for people who care about hand-crafting experiences with maximum customization. My belief for AI in tooling is the capabilities should always be lowering the floor, not the ceiling Elegant AI-powered experiences help people without the skills to do things they only imagined being able to do. While doing that, it enhances and makes work more efficient such as fixing random syntaxes and nuances we find frustrating. Design and building websites will be content-focused with Dynamic Interfaces doing the heavy lifting, so we can focus on curation and publishing.