08-06-2025, 01:48 PM
Yeah, I totally get your point — and I actually agree.
One of the big advantages of slot grammars and curve-line systems is that they're interpretable. You can look at the rules, tweak them, compare them. With GPT-style models, everything’s buried in a tangle of weights and activations, so it’s much harder to pin down what the model is actually learning.
That said, my goal here wasn’t to claim that GPT models are better or smarter — just to see if, given only raw Voynichese sequences, a tiny transformer could pick up on any structure at all. And it seems like it does. The low perplexity suggests it’s finding something regular, even if we don’t know exactly what yet.
But you're absolutely right: without comparing it to existing frameworks, it’s hard to know if it’s finding anything new. That’s definitely the next step — like checking whether the model’s predictions line up with known slot structures, or whether it generalizes in weird ways they don’t.
So I see this more as a starting point — not a replacement for existing theories, but maybe a tool to help test or even challenge them.
One of the big advantages of slot grammars and curve-line systems is that they're interpretable. You can look at the rules, tweak them, compare them. With GPT-style models, everything’s buried in a tangle of weights and activations, so it’s much harder to pin down what the model is actually learning.
That said, my goal here wasn’t to claim that GPT models are better or smarter — just to see if, given only raw Voynichese sequences, a tiny transformer could pick up on any structure at all. And it seems like it does. The low perplexity suggests it’s finding something regular, even if we don’t know exactly what yet.
But you're absolutely right: without comparing it to existing frameworks, it’s hard to know if it’s finding anything new. That’s definitely the next step — like checking whether the model’s predictions line up with known slot structures, or whether it generalizes in weird ways they don’t.
So I see this more as a starting point — not a replacement for existing theories, but maybe a tool to help test or even challenge them.
![[+]](https://www.voynich.ninja/images/collapse_collapsed.png)