From Semantic Web to LLMs: A Second Chance for the Open Web
I’ve been revisiting some old ideas about the web—and trying to rebuild them with modern tools.
The web was supposed to be open.
Built on protocols, not platforms.
Data anyone could build on.
Identity not owned by a single company.
Somewhere along the way, that vision faded.
We moved toward a platform web, where a handful of companies shape how we communicate, discover, and create.
What went wrong?
It wasn’t just business models. It was also how the web was built.
The early web was open, but it wasn’t very understandable to machines.
That’s what the semantic web tried to fix.
The idea was straightforward: add structure so machines could understand meaning, not just text.
In practice, that meant:
- RDF (Resource Description Framework)
- OWL (Web Ontology Language)
- explicitly defined relationships between entities
“The Semantic Web provides a common framework that allows data to be shared and reused across application, enterprise, and community boundaries.”
It was ambitious, and in many ways correct.
But it didn’t scale.
Not because the idea was wrong, but because it depended on humans manually structuring the internet. That was never going to work.
What changed?
Instead of trying to define meaning upfront, we started learning it.
Machine learning, and later large language models, took a different approach.
Don’t define relationships.
Learn them from data.
LLMs don’t use ontologies in the traditional sense, but they build something similar internally—a learned representation of how concepts relate.
In simple terms:
The semantic web tried to make the web machine-readable.
AI is making it machine-understandable.
But there’s a catch
Most AI systems today live inside closed ecosystems.
The data might be public, but the systems are not.
So we risk repeating the same pattern:
Open inputs → closed platforms → controlled outputs
Why the fediverse matters
This is where the fediverse gets interesting. Not as a Twitter replacement, but as infrastructure.
Protocols like ActivityPub define how decentralized social systems can communicate.
“ActivityPub is a decentralized social networking protocol based upon the ActivityStreams 2.0 data format.”
It shifts things back toward:
- open data flows
- user-owned identity
- composable systems
It’s messy and fragmented. But it’s open.
A different path forward
What if we combine the openness of the fediverse with the ability of AI to extract meaning?
Instead of asking people to structure the web, we let machines do it on top of open protocols.
That starts to look like a web that is both open and machine-understandable.
What I’m exploring
I’ve started experimenting with this idea by building small tools on top of Mastodon and ActivityPub:
- pulling raw social data via API
- applying AI to extract meaning
- gradually turning that into structured knowledge
Not as a product, at least not yet, but as a way to explore where this could go.
Closing thought
The semantic web didn’t fail.
It was an attempt to solve the right problem under the wrong constraints.
Now we have better tools and new protocols.
It might be time to try again.