What happens when AI starts building itself?

Recursive Superintelligence, a San Francisco-based startup, recently emerged from stealth with $650 million in funding. Founded by Richard Socher, known for You.com and Imagenet, the company brings together a team of prominent AI researchers, including Peter Norvig and Tim Shi. Their ambitious goal is to create a recursively self-improving AI model.

This AI would be capable of autonomously identifying its own weaknesses and redesigning itself for improvement, a long-standing aspiration in AI research. The company distinguishes its approach through "open-endedness," aiming for truly recursive self-improvement where the entire process of ideation, implementation, and validation of research ideas is automatic.

Socher explained that this AI would initially focus on automating AI research, eventually expanding to other research domains, including physical ones. He highlighted the particular power of AI working on itself, developing a new sense of self-awareness regarding its shortcomings. An example of open-endedness is "rainbow teaming," where two AIs co-evolve, with one continuously attacking the other to identify vulnerabilities, leading to a more robust and safer AI.

The company envisions a future where computing power becomes the most critical resource once such a system is established, as the speed of improvement would directly correlate with processing capacity. Socher anticipates releasing products from this research within quarters, not years, emphasizing the company's aim to deliver impactful products alongside groundbreaking research.