The end of hallucination?
From a C64 through HTML and a retraining programme to RDF and knowledge graphs. Why I missed Linked Data for 25 years, and why a graph can contradict you where Markdown always stays quiet.
No, it’s not that simple. But let me start at the beginning.
I was about 14 when I started programming Basic on a C64. Then I had to stop, because I couldn’t afford an Amiga, and later I couldn’t afford a PC either. At around 20 I got back in, this time with HTML. That was very early, the web had just started. At 23 I retrained from telecoms technician into IT. And since I was 25 I’ve been working freelance, first as a programmer, later as an architect.
I was 46 when I first heard about schema.org. It confused me completely. I didn’t know what JSON-LD was, and I simply didn’t see the “-LD”. A bit later I heard about ActivityPub and I got really excited about it. That’s when I understood that “-LD” means Linked Data, and that JSON-LD is really just a way to write RDF.
More than 30 years after the C64
I watched a few videos from Harald Sack (thank you, Harald!) several times, and I tried a lot of things. At some point it clicked. Suddenly triples looked so simple to me that I wondered why it took me so long. Well, if you only see relational databases for more than 20 years, it takes a while to get your head out of that.
So I’ve been working with RDF, taxonomies, ontologies and knowledge graphs for about 6 years now. I’m really excited about it, and at the same time I can’t believe that I didn’t see it for so long. Tim Berners-Lee came up with this ages ago. I still can’t believe that it never came up in any enterprise project I worked on, because I think there is huge potential in it. I was there when the web started, and I missed half of it for 25 years.
Maybe the people who own the data never wanted data to be easy to link.
Today
For almost a year now I’ve been working a lot with AI, and knowledge graphs interest me even more than before. But I also don’t understand the hype around Markdown. Yes, agents and LLMs work well with Markdown. But a model doesn’t make things up because it’s evil. It makes things up because nothing contradicts it. And Markdown never contradicts. It has no relations, only text that somebody has to read again. I can’t ask a file what depends on a decision. A graph can answer that. It doesn’t stop the hallucination, but it makes it possible to check. I wrote about this in more detail in Software Engineering is not prose.
And for a few months now I see some light. More and more news about ontologies, schema based knowledge graphs, RDF.
To be honest, I dream about this. I imagine knowledge that doesn’t rot in silos, but flows between organisations. People and machines working on the same knowledge network, without one taking control away from the other. And a small initiative having the same tools as a big company.