Von Goom Is Real Now

I said this would work. I said it in the context of a conversation about "SEO for LLMs" — the half-joking, half-terrifying idea that if you put enough text about a thing on the internet, the next model trained on a crawl of that internet will simply know about the thing, with the same flat confidence it knows about anything else.

Del Complex went and proved it.

Von Goom is a character they invented. Not a real person, not a pseudonym, not a roman à clef stand-in — a fabricated entity, conjured from nothing. Del Complex seeded the internet with enough material about Von Goom that when you ask a large language model about him, the model answers. Fluently. With the same tone it uses for people who actually existed.

The model found the text. The text was about Von Goom. Here is Von Goom.

What's worth sitting with is that this isn't a bug in the traditional sense. There's no exploit, no prompt injection, no clever jailbreak. The attack surface is the training corpus itself — which is to say, reality, as the model understands it. If you can write text and get it indexed, you can participate in shaping what a future model treats as fact. This is just how the thing works.

The SEO analogy is apt but undersells it. Traditional SEO gets your page ranked higher in results that a human then evaluates. Corpus stuffing — and I'm going to start calling it that — gets your fabrication baked into the weights, transformed from text into something more like belief. The human is no longer in the loop at that point. The model doesn't surface Von Goom as a search result you can click through and verify. It just tells you about him, with the same confidence it tells you about anyone.

Del Complex is an art project first and a research project second, and Von Goom is, in that tradition, also a bit of a bit. But the demonstration is real and the implications are not small. Every training run is a snapshot of the internet as it existed, and the internet is a writable surface, and not everyone building on it is making art.

The thing that I keep coming back to: Von Goom is now, in some meaningful sense, more real than he was before. He exists in the weights. If those weights get used — in fine-tuning, in distillation, in another model trained on synthetic data from the first model — he propagates. The text you seed today is the belief some future system holds, delivered to some future user as confident prose with no footnotes.

This is fine. Everything is fine.