Leading AI developers are exploring the concept of recursive self-improvement (RSI), a process in which AI systems could autonomously enhance their own capabilities and create improved versions of themselves. Once a futuristic idea, RSI is now viewed as a more tangible goal by many researchers and companies. For example, Anthropic, a prominent AI firm, has reported that its AI model, Claude, is already playing a major role in its own development, contributing to 26% of the company's research and development efforts. However, these systems are not yet fully autonomous in this process. Definitions of RSI vary across the industry. Some companies consider any AI feedback that contributes to model improvements as a form of RSI, while others define it as complete self-directed enhancement without human intervention. Anthony Aguirre, president and CEO of the Future of Life Institute, has warned that as AI systems take on more of the improvement process, they can operate much faster than humans, raising concerns about the potential for uncontrolled or runaway superintelligence. Meanwhile, John Thickstun, a computer science professor at Cornell University, notes that a form of RSI has been happening for years, as previous AI models have helped build newer, more advanced ones. Researchers have made incremental progress in using AI to train new systems, though major breakthroughs remain elusive. Andrej Karpathy, co-founder of OpenAI, has worked on projects where AI assists in the development of new models, but these have not yet led to dramatic improvements. Aguirre believes that AI companies are now closer than ever to achieving significant leaps in this area, citing Anthropic’s progress as an example. OpenAI is working on an automated "research intern" that can perform specific research tasks under human supervision, with plans to develop a fully automated AI researcher by March 2028. The company highlights that while RSI could help align AI models with human values, rapid development must be approached cautiously to ensure human control and informed decision-making. Elon Musk of xAI has mentioned that the Grok series of AI models is being improved with less human input, and he suggests the process could be fully automated by 2027. In contrast, Microsoft’s Mustafa Suleyman is promoting the idea of “humanist superintelligence,” which focuses on AI that benefits humanity within clear ethical boundaries, rather than pursuing unbounded autonomy. Discussions around slowing AI development for safety reasons have also influenced the pace of RSI progress. Anthropic has called for a coordinated slowdown, which would depend on the participation and transparency of global competitors. OpenAI acknowledges uncertainty about safely achieving full RSI, emphasizing the need for advancements in alignment and safety to match the growing capabilities of AI models. Despite these challenges, the company still sees RSI as an important goal, noting that automated AI researchers could also play a role in ensuring the safety of future systems.