Google has proposed "recursive self-improvement (RSI)" as a next-generation artificial intelligence (AI) technology that goes beyond artificial general intelligence (AGI). By setting RSI—where AI designs and improves better versions of itself—as a new goal for AI progress, the idea is to close the gap between massive AI investment and revenue. However, the industry says that, for now, RSI has limits to fully improving itself without human intervention, and it will likely take considerable time to realize in practice.
According to The Information and other foreign media on the 5th, Jasjit Sekhon, Google DeepMind chief strategy officer (CSO), said at the "Agentic AI Summit" on the 3rd (local time) that RSI has now become the new focus beyond AGI, which used to be the goal of the AI industry. RSI is a concept in which AI improves its own performance through trial and error and even automates the improvement process. Beyond AGI, which implements general intelligence at or above the human level, the essence of RSI is that AI continuously boosts performance by improving even its own learning methods.
In fact, not only Google but global AI corporations such as OpenAI and Anthropic are focusing on RSI as the next arena of AI competition. Anthropic co-founder Jack Clark said on his X (formerly Twitter) in May, "I think there is more than a 60% chance that RSI will be realized before 2028," adding, "A world is coming where AI creates itself." OpenAI has also drawn up its own roadmap to "build a fully automated AI researcher" by Mar. 2028.
The reason global AI corporations are eyeing RSI is clear. If AI can improve algorithms and develop new models on its own, it can drastically cut research and development (R&D) expense and time. Development work that human researchers carried out over years could be repeated and improved by AI in seconds using supercomputers, accelerating performance gains exponentially.
This ties directly to the industry's biggest recent concern: profitability relative to investment. While global AI corporations pour astronomical funds into securing data centers and semiconductors, AI service sales have not kept pace, fueling an "AI peak" narrative. Google will invest about $200 billion (about 285 trillion won) this year in AI infrastructure and equipment and has signaled additional expenditure next year. RSI is drawing attention because it has the potential to dramatically speed up AI development and efficiency, easing these profitability concerns.
At the summit, Sekhon, the CSO, said, "The revenue currently generated in the AI field still does not cover the capital expenditure we are committing," adding, "Investment is continuing, but there is a risk of an 'AI air pocket' where sales do not follow." He added, "RSI is a key element underpinning the current investment logic of the AI industry."
Buoyed by such expectations, some AI corporations are shifting strategy to focus on RSI research rather than developing large language models (LLMs) for commercialization. A prime example is "Ineffable Intelligence," founded last Nov. by David Silver, a former DeepMind researcher. The company raised $1.1 billion in investment at launch from top venture capital firms (VCs), marking the largest seed round in European startup history. The fact that massive capital poured in despite the company not even disclosing a concrete product roadmap shows the market's expectations for RSI.
However, the industry expects it will take considerable time to achieve full RSI. For now, AI is closer to rapidly automating research and development tasks than improving itself. Anthropic's AI coding tool "Claude Code" already writes most of a team's code, and some engineers assessed that if performance improves further, the top-tier model "Mythos" could replace mid-level developers who handle complex projects without supervision. Still, it shows limits in long-term project management, priority setting, and determining research direction, leading to the assessment that it remains far from full RSI.