Re: reminding the scope for this AI KR CG

Milton, Dave, all,

Following Milton’s point that “the core elements must be well defined” 
and that this is missing from the blue bubbles, we’ve just completed a 
small training run in K3D to construct such atomic elements explicitly, 
without tokenization.

Very briefly, for a restricted domain (ASCII + math glyphs), we define:

F = executable visual RPN programs (form space)
M = execution/semantic RPN programs (meaning space)
E = procedural embedding space ℝ^D
and build atomic units as:

A = (c, f, m, e) with c ∈ Σ, f ∈ F, m ∈ M, e ∈ E
In the current run we implemented:

148 atomic units total
72 dual‑program “stars” where each character has both:
a visual RPN program that actually renders the glyph on GPU, and
an execution RPN/bytecode program (e.g. e as Euler’s number, + as ADD, ^ 
as POW)

Cross‑modality here is compositional: we store visual and mathematical 
programs in the same atomic unit and retrieve that composite object, 
rather than projecting everything into a single token embedding space.

There is no natural‑language tokenization step in the LLM sense; form 
and meaning live in separate, well‑defined program domains (visual RPN 
and execution RPN), with natural language sitting on top rather than 
being the primary representation.

Fusion happens via the 3D contract (the star), not by collapsing 
everything into one natural‑language vector space.


A short write‑up of this proof‑of‑concept, including the set‑theoretic 
definitions, metrics (148 units, 72 dual‑program, ~2 minutes training, 
~2.2KB per unit), and example stars for e, +, and ^, is here:

https://github.com/danielcamposramos/Knowledge3D/blob/main/TEMP/W3C_AIKR_ATOMIC_UNITS_PROOF_NOV19.md

This is still very early and deliberately narrow in scope (one small 
domain of discourse), but I hope it’s a useful concrete example in the 
space you’re both describing:

Milton’s requirement for constructible atomic elements and domains of 
discourse;
Dave’s emphasis on structured but not purely formal KR, where plausible 
reasoning layers (PKN‑style) can sit on top of explicit, 
machine‑readable foundations.

If anyone is interested in the implementation details, I’m happy to take 
that to a separate thread or offline so we don’t overload this one.

Best regards,
Daniel