ZIGU LABS Unveils GBR, a Non-Transformer Sovereign AI System

Korean AI startup ZIGU LABS, led by CEO Lee Jun-su, on 1 November unveiled the GBR System, a memory-centric sovereign AI architecture built on the company's own mathematical theory. According to the company, it relies on neither the transformer design nor conventional large language models.

Instead, GBR links a proprietary memory structure to database (DBMS) mechanisms. It interprets a question, reasons over stored knowledge and then derives an answer.

Memory units instead of model weights

Two technologies sit at the core: a continuous knowledge-accumulation scheme that needs no retraining, and a precision retrieval algorithm. Conventional neural networks fold knowledge into model weights. GBR stores data separately as independent "memory" units, each carrying a source and a version.

Its "Zero-Miss" retrieval is designed to find the best memory within a predefined scope without omissions. For queries outside that scope, the system withholds an answer or asks for more evidence. The memory network follows a fractal structure, in which whole and part are self-similar. Additions, edits and deletions therefore show up in answers immediately, with no full retraining.

Benchmarks and edge demo

ZIGU LABS also published performance figures. It sampled 10,000 puzzles from the roughly 420,000 in the Sudoku-Extreme dataset. The company says the engine solved all 10,000 in a single reasoning step, without external computation tools. The verification data and result certificates are on GitHub, and the firm plans to extend evaluation to the ARC-AGI benchmark, which tests how well systems handle unseen tasks.

The company also demonstrated on-device operation on Nvidia's Jetson Orin Nano Super, an 8GB embedded computer. The device answered questions and updated its knowledge with no cloud connection. In the same unit, camera and motor control signals were linked directly, which the company says shows the system can run independently in robotics and edge settings.

Lee said that putting AI to practical use in industry requires control: users must be able to see the logical basis for an answer and correct errors at once. He said the company will refine national-scale sovereign AI solutions built on an architecture that does not depend on overseas frameworks.