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…
NINMENI tackles the computational expense of modeling each character's unique identity in large language models. Rather than compressing text into tokens, NINMENI maintains a static identity for each normalized character. This identity, called an NMU, remains unchanged even as computation is restructured. The result is a different approach to computational organization.
The core challenge lies not in preserving character identities themselves, but in the cost of computations performed on them. NINMENI currently employs a registry of 10,240 identity slots, with each character retaining its own identity, position, output, and training target. The computational system cannot simply reduce sequence length by introducing new learned segmentation units. Instead, it must rethink how to organize computation around the sequence without altering the characters' identities.
This distinction forms the basis of MULTIPITA, an exploration NINMENI is currently pursuing. The key is not how to compress the sequence, but how to reorganize computation around it while maintaining the characters' distinct identities.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
