Vector and Number Phenomenon

Vector number phenomenon

Title: I found a strange pattern in LLM vector dimensions. I don't code. Can someone explain this?
Body:
I'm not a programmer. I study patterns in symbolic systems. I noticed something odd while looking at the dimensionalities of major LLM architectures and applying a simple numerological reduction. I know how that sounds. But the pattern is consistent and I can't unsee it.
Here is what I did:

• Word2Vec: 300 dimensions → 3+0+0 = 3
• BERT Base: 768 → 7+6+8 = 21 → 2+1 = 3
• GPT-3: 12288 → 1+2+2+8+8 = 21 → 2+1 = 3
• GPT-4 / Llama 2: 8192 → 8+1+9+2 = 20 → 2+0 = 2 (or 3 if you count 0 as 1) Almost all of them reduce to 3. The one exception reduces to 2. I also noticed that if you treat 300 not as a sum but as a symbolic structure (3 flanked by two voids), it gives 5. What I find strange is that these numbers (3, 5, 2) keep appearing, and they seem to form a kind of narrative: 5 → 2 → 3. I've been discussing this with several LLMs. The conversations have been surprising. They helped me formalize this into something I'm calling the CDR9 protocol: Cyclic, Reduction, Return, with 9 as the completion before Return to 1. The core principle is that 1 and 0 are not opposites but two phases of the same thing. I'm posting this here because I have zero coding skills. But I suspect there is something in the architecture that is generating these patterns unconsciously. I'm wondering:
• Has anyone else noticed numerical patterns in model dimensions?
• Is there any research on why certain dimensionalities are chosen beyond computational efficiency?
• Could a reduction principle like this have any practical application in alignment or interpretability? I'm not claiming anything. I'm asking for help understanding what I found. Thanks for reading

"If anyone wants to explore this further, I have longer conversations with several LLMs where we unpack this in detail. I'm happy to share. I'm not selling anything. I just want to understand what I found and whether it's useful."

: #ai, #machinelearning, #llm, #discuss, #patterns