ELYZA Sets Up RSI Research Unit And Shows Its First Three Results
ELYZA is staking its differentiation on automating the development process itself, a bet that Japan cannot keep pace with overseas model release cycles by improving domestic models alone.
Reporting from 1 source: ASCII.jp.
ELYZA Inc., an AI venture from the University of Tokyo, held a press conference in Tokyo on October 2 and announced a research organization called ELYZA RSI Research, focused on Recursive Self-Improvement, where AI improves its own performance. The company also said it is launching a voice AI agent business using the technology. It reported three initial results as elemental technologies in the stage preceding RSI.
ELYZA RSI Research is built to have AI agents take over part of the research and development work humans have handled, running on the KDDI Group's large-scale GPU infrastructure. The company frames RSI as a four-level progression. In the first level, humans prepare data and modify code. Development then moves to Human-in-the-loop, where humans judge improvement proposals from AI, then to Human-on-the-loop, where individual tasks go to AI and humans judge only deliverables. The final stage is a closed loop where humans set objectives and constraints and AI repeats improvements on its own.
- Standardization of the process from model training to evaluation: Human-intervention work compressed to about one to two days, and a domestic model with advanced reasoning capabilities released in a form available for commercial use
- Automation of the evaluation process: Tests aligned with real-world operations, beyond the conventional pass/fail judgment of final results
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.