ELYZA founds ELYZA RSI Research for AI-driven AI development

ELYZA is an AI company spun out of the University of Tokyo's Matsuo Lab, working on large language models and their deployment. It joined the KDDI group in April 2024, a move it frames as giving it GPU resources to pursue research and product releases at scale.

In its own account of the field, LLM accuracy has risen fast, simple tasks plateaued about two years ago, and the current front line is science use and complex multi-day tasks. It also points to a wave of open models, largely from China, arriving two to three months after the frontier. Against that backdrop, it describes recursive self-improvement as already beginning at companies such as Anthropic and OpenAI, and lays out four stages, from humans doing everything to an improvement loop that runs entirely on AI.

The stakes hinge on how quickly the evaluation side can move. ELYZA has full automation in the harness but only semi-automation in evaluation, and it says full autonomy is stage 4. Its planned extension from LLMs into vision-language models and vision-language-action models for robots is likely to be judged on the same safety conditions, which its executives described as a matter of isolation and the ability to stop the system.