ELYZA launches ELYZA RSI Research to chase self-improving AI

ELYZA's pitch rests on a shift it says has already happened in AI research. By its account, LLMs hit an accuracy ceiling on simple tasks back in 2024, and the frontier has since moved to "AI for Science" — using AI for drug discovery, materials and mathematics — and to "Long Horizon" tasks that span multiple steps a person would need hours or days to finish. Benchmarks have followed, with advanced reasoning, mathematical proof, coding, task completion and reliability now the main talking points.

The company points to concrete progress behind that shift. On September 3, 2026, OpenAI released GPT-6 Astra, which it says markedly improved "Computer Use" — the ability to operate a PC across multiple steps. Chinese firms, meanwhile, are pushing ahead with developing and open-sourcing foundation models, which ELYZA describes as arriving with roughly a two-month lag. Against that backdrop, ELYZA says AI-driven AI development has been accelerating since early 2026, and RSI is what comes next.

Soneoka also framed agent development as more than a foundation-model problem, saying harnesses — the surrounding systems that handle complex processing — and evaluation mechanisms matter just as much, and that humans still build and tune both today. That gap is where the stakes sit: whether self-improvement arrives on ELYZA's timeline hinges on how fast those human-built pieces can be automated, and on whether regulation or frontier-lab caution keeps the resulting models out of general public use.