Sakana AI, a Japanese startup co-founded by Llion Jones (co-author of the famous Transformer paper), has launched the Recursive Self-Improvement Lab (RSI Lab) to research AI that can repeatedly develop itself—essentially AI that helps create and improve the next generation of AI.
An Idea That Challenges the Mainstream
Instead of competing in the compute war such as buying chips and large data centers like big American labs, Sakana AI believes RSI is a shortcut to reaching the frontier with fewer resources. This idea challenges the foundation of current AI development that typically focuses on scaling models and increasing compute power.
Important Context in the Age of Compute War
The compute war is intensifying. Recently, Google paid SpaceX $920 million per month just to increase its compute capacity. This shows that the cost of training frontier models has become so high that even giant companies need solutions. Sakana AI sees RSI as a potential answer to help AI train faster and reduce dependence on hardware and chips. This could completely change the rules of competition in the AI industry.
Risks and Controversies
However, the concept of self-improving AI remains highly controversial. Many parties view it as mere hype that cannot yet be proven. Additionally, there are safety concerns about AI that can improve itself, particularly regarding control and preventing undesirable outcomes.
Future Directions for AI from Japan
That Sakana AI, a Japanese startup, dares to challenge the American AI development paradigm shows that the AI industry is expanding with new players entering the field. The success of RSI Lab could change the competitive landscape in the AI industry and make Japan another important center of AI innovation. Investing in this research direction, despite its risks, might be a worthwhile bet if successful, as it would help reduce dependence on expensive and limited compute resources—a major problem in today's AI industry.
Reference:
https://the-decoder.com/sakana-ai-bets-ai-that-improves-itself-can-break-the-compute-arms-race-of-frontier-labs/