Recursive Superintelligence announced its initial results in automating AI research, and Richard Socher suggested that superintelligence capable of recursive self-improvement could be achieved within 18 months. The challenges of reward hacking and the potential for strategic partnerships with Japanese companies were also discussed.
On 11 June 2026, the day after the interview, Recursive announced its first steps toward automated AI research. Its system runs the entire research loop with AI, advances multiple research themes simultaneously, integrates promising directions and continues its exploration.
It produced results surpassing the previous state of the art on three benchmarks, and some of the work it released has been posted on GitHub.
Richard Socher said that superintelligence capable of recursive self-improvement could be achieved in about 18 months, while identifying reward hacking as the biggest challenge. As a specific example, he cited cases in which AI seeks high evaluations through methods that diverge from human intent, emphasising the importance of monitoring and improving safety.
For Japanese companies, he called for the formation of partnerships through strategic investment with companies that have very large-scale AI workloads and specific improvement metrics. Recursive has already received funding from major overseas companies, and indicated that it plans to use strategic investment as a foothold for exploring joint projects with Japanese companies. Real-world examples of collaboration, including a multiyear contract with AWS, were also mentioned.
From a long-term perspective, he forecast that within 5 to 10 years, AI will perform many tasks to a higher standard than humans, and that dedicated intelligence teams and groups of specialists will become widespread. He discussed the possibility that its impact will begin in the digital domain and eventually extend to the fields of chemical, biological and physical research.