MULTIPITA: Reorganizing Compute Without Compressing Identity

How NINMENI approaches the computational cost of native character-identity modeling
Large language models usually make an implicit compromise: text is compressed into tokens or other larger units before most of the expensive computation happens.
NINMENI starts from a different premise.
In NINMENI, a normalized character is assigned one fixed identity, an NMU. Identity is not merged into a learned subword unit, dynamically repartitioned, or replaced by an unknown-token fallback. Computation may be reorganized, but the identity substrate remains intact.
That creates an obvious engineering problem:
If every character remains an individual identity, how do we make the computation affordable?
MULTIPITA is the answer NINMENI is currently exploring.
The Problem Is Not Identity. It Is the Cost of Computing Over Identity.
NINMENI's current instantiation uses a fixed registry of 10,240 identity slots. Every character retains its own identity, state position, output, and training target. The computational system therefore cannot simply reduce sequence length by introducing a new learned segmentation unit.
So the engineering variable becomes computation.
The question is not:
How can we compress the sequence?
It is:
How can we reorganize computation around the sequence without changing its identity?
That distinction is the foundation of MULTIPITA.

papers.ssrn.com