Update Legend of Elya description (6.36M ternary transformer)

[Scottcjn]

What this changes

Updates the existing Legend of Elya entry, which was out of date. The project has grown from the original ~819K "nano-GPT" into a 6.36M-parameter ternary transformer, and the current description also overstated the runtime environment.

Old:

A 4-layer nano-GPT neural network (819K parameters) running natively on the N64's MIPS R4300i FPU, believed to be the first LLM on Nintendo 64 hardware

New (matches the repo's own measured figures):

A Zelda-style homebrew game whose AI NPCs run a 6.36M-parameter ternary transformer with byte-level inference on the VR4300 MIPS III CPU (1.23 tok/s scalar, 2.19 tok/s via an RSP overlay, measured under the ares emulator). Built with libdragon

The numbers are the repo's measured results under the ares emulator (not on physical silicon), so I removed the unverified "first LLM on N64 hardware" claim.

Disclosure

I'm the author (Elyan Labs / @Scottcjn) — correcting my own entry for accuracy.

[command-tab]

Hi! I just updated the description to be a little more broad and include fewer specifics, since the audience may not be familiar with the significance of number of parameters, the particular type of transformer, etc. It now reads:

A Zelda-style ROM running a transformer for live LLM inference powering an NPC, built with libdragon