←
Martynas Jusevičius10,622 posts
Martynas Jusevičius avatar

Martynas Jusevičius

@namedgraph

Systems thinking behind agent-ready Knowledge Graphs Founder of @atomgraphhq

◷ Joined March 2009▧ 599 Photos and videos
Martynas Jusevičius avatar
Martynas Jusevičius@namedgraph·Nov 3

I've argued that the Semantic Web doesn't need new standards. It just needed an application architecture, and it turns out RDF, SPARQL and XSLT are enough. But that's for the read-write Web. To make it actionable for AI agents, I think we need another language. Hear me out.

Post attachment
Martynas Jusevičius avatar
Martynas Jusevičius@namedgraph·Nov 13

Post attachment

The web had a declarative rendering engine in 1999. React and web components later spent years rebuilding parts of it in JavaScript. XSLT matches templates to data and composes them. It runs on the server and in the browser. LinkedDataHub 6 renders every page using it.

Martynas Jusevičius avatar
Martynas Jusevičius@namedgraph·Nov 13

The web had a declarative rendering engine in 1999. React and web components later spent years rebuilding parts of it in JavaScript. XSLT matches templates to data and composes them. It runs on the server and in the browser. LinkedDataHub 6 renders every page using it.

In many web apps, the interface is built twice: server templates for the first page, front-end components for the browser. LinkedDataHub 6 renders both with the same XSLT 3.0 stylesheets, Saxon on the server and SaxonJS in the browser. There is no separate front end.

Post attachment
Martynas Jusevičius avatar
Martynas Jusevičius@namedgraph·Nov 11

Most web applications run on the web, but their data doesn't follow its rules. By 1999 there was a standard for that data, RDF, and one for rendering it, XSLT. Web 2.0 went with JavaScript instead. That is why AI agents now need an adapter for every application.

1:51
Thorbjørn Rønje avatar
⟳ Martynas Jusevičius retweeted
Thorbjørn Rønje@thorbjornron·Nov 8

A proprietary knowledge graph becomes a very interesting asset when AI agents can actually use it. Every customer interaction, transaction and workflow can potentially improve the information available to the system. The difficult part is getting the underlying data right. I suspect a lot of AI companies will discover that their real engineering challenge lives there.

Martynas Jusevičius avatar
Martynas Jusevičius@namedgraph·Nov 5

Lithuania publishes procurement, EU funds, tax arrears, donations, lobbying and Seimas votes separately. A government that takes open data seriously would link them. Somebody had to, so I built linkeddata.lt: 35M RDF triples that AI agents can investigate autonomously.

Post attachment
Martynas Jusevičius avatar
Martynas Jusevičius@namedgraph·Nov 3

It's already in LinkedDataHub 6. The assistant generates Web-Algebra plans that the runtime executes. It's a missing piece of the agentic Web: LLMs describe what needs doing, without managing every step. github.com/AtomGraph/Web-Alg…

In LinkedDataHub 6, the assistant is available on any page. Ask in plain language. Read-only requests run immediately. Here, it finds the five customers who ordered the most and presents them as text and a chart. Changes to data wait until Execute is pressed.

0:43
Martynas Jusevičius avatar
Martynas Jusevičius@namedgraph·Nov 3

That's the idea behind Web-Algebra, a JSON DSL I've built for composing HTTP requests, SPARQL queries, transformations, loops etc. An LLM writes the whole workflow as one JSON document. The runtime executes it and returns the result. No intermediate results pass through the LLM.

Martynas Jusevičius avatar
Martynas Jusevičius@namedgraph·Nov 3

Right now, an LLM typically calls one tool, gets the result, figures out what to do next, calls another tool, and so on. Why involve the LLM in every step when it could describe the whole workflow upfront and let something else execute it?

Martynas Jusevičius avatar
Martynas Jusevičius@namedgraph·Nov 2

AI makes vulnerabilities ever cheaper to exploit. So I've built Lean models of LinkedDataHub's HTTP and access control, and test the server against them. Low-code and AI app builders ship new code with every app. LinkedDataHub apps are data, so they all share one proven core.

Post attachment
Charles Vardeman avatar
⟳ Martynas Jusevičius retweeted
Charles Vardeman@CharlesVardeman·12 Jan 2025

Aaron Swartz was in the process of writing "A Programmable Web" in 2013. The unfinished manuscript was published after his death and is still relevant today in the age of Generative AI. The original was published by Morgan & Claypool under a CC-BY-SA-NC license and is available in wikicommons. upload.wikimedia.org/wikiped…