ELYZA RSI Research targets AI developing AI

ELYZA, which handles AI development within the KDDI Group, had been working on recursive self-improvement before formally establishing the research unit. The company's stated motivation is speed: CEO Yuya Soneoka said overseas frontier AI from OpenAI, Anthropic and Google has been improving rapidly, and that the underlying loop of AI developing AI has been replacing work previously handled by people.

The results ELYZA cites are operational rather than theoretical. In inquiry handling, task completion rose from about 48% to about 75%, and inference costs fell by about 67% at the same accuracy.

Soneoka also described where frontier AI competition has shifted. Simple tasks such as summarization reached a peak around 2024, he said, so OpenAI, Anthropic and Google are now competing on harder work — AI for Science in drug discovery, chemistry and math, and long-horizon tasks that span multiple steps a human would need hours to finish. He pointed to examples such as roughly 10,000 AI agents working on an unsolved math problem and OpenAI's GPT-6 Astra operating a PC to carry out tasks. The release pace has followed: new models arrived every three months through 2025, he said, and are now appearing every month.