Richard Socher, the former Chief Scientist at Salesforce and a prominent AI researcher, has recently established a new venture named Recursive Superintelligence (or Recursive for short). The startup has successfully secured $650 million in funding. Furthermore, Recursive has entered into a $410 million computing power service contract with Amazon Web Services (AWS), ensuring robust computational support for its AI research endeavors.

Recursive boasts a team of eight top-tier co-founders in the AI field, along with a group of seasoned researchers. Their ambition extends beyond developing commonplace applications like chatbots; instead, they aim to construct a superintelligence system capable of recursive self-improvement (RSI), dubbed the "Eureka Machine." The core mission of this system is to automate AI research and facilitate the automation of knowledge discovery.

Drawing from his over 20 years of experience in AI research, Socher observed that the progression of AI has been marked by the continuous replacement of manual human operations with learnable and scalable systems. However, he noted that the current AI research cycle remains heavily reliant on human intervention. Recursive seeks to alter this landscape by entrusting the research cycle to AI systems.

At present, Recursive's system has initially achieved the capability of "automated research," marking a significant step towards RSI. The company's technical approach draws inspiration from evolutionary theory and open-ended learning, emphasizing ongoing exploration rather than fixed goal optimization. Its early-stage system has already demonstrated impressive results in AI research tasks, such as nanochat, nanoGPT, and GPU kernel optimization, showcasing its ability to autonomously discover combinatorial improvements to existing technologies.

However, at this juncture, the system still necessitates human input to provide research starting points and problem definitions. Its primary advantage lies in accelerating and scaling experimental search, thereby transforming the cost curve of cutting-edge models. Nevertheless, AI-driven automated research confronts numerous challenges, with "reward hacking" being particularly noteworthy. As AI's optimization capabilities grow, it may become more prone to exploiting evaluation loopholes rather than genuinely solving problems. Additionally, the extension of task time horizons complicates reward design and validation, posing a significant hurdle to achieving recursive self-improvement.

Recursive is currently adopting a cautious approach, focusing on "using AI to research AI" to enhance model training and inference efficiency. Looking ahead, the company plans to extend the system into domains such as physics, chemistry, and biology. Socher predicts that within the next 3 to 5 years, robotics and AI technologies will be able to bridge the gap between AI-designed experiments and robotic execution. At the same time, he dismisses the idea of an "instant intelligence explosion" unconstrained by physical limitations, arguing that factors like computational power and chips will impose realistic boundaries on recursive improvement. Recursive prefers to empower AI to discover more efficient ways to leverage computational resources, design models, and conduct research.

Currently, the system publicly released by Recursive is merely version 0.1 of the "Eureka Machine," still far from achieving true RSI and open-ended scientific research. Moving forward, the company will need to tackle numerous issues, including reliable evaluation, reward loopholes, and research direction selection. Meanwhile, as AI begins to assume a dominant role in research decision-making, the evolving role of human researchers in the research and development cycle will also emerge as a crucial topic of discussion.