The report claims that Novix was founded by Shi Yu, Tang Jiabin and others, with core members from Google, Microsoft, Meta, Zhipu and Kimi, who had led or participated in projects such as Microsoft Copilot, Kimi K2, AgentOS and AutoAgent. The report calls this mechanism Co-Evolutionary RSI, that is, co-evolutionary RSI.
According to the report's description, in this loop, AI proactively proposes directions, expands the solution space, executes tasks and verifies results; experts continuously feed back what important problems are, what evidence is sufficient for something to hold, what evaluation standards to adopt, and which directions match their own taste and long-term judgment. Every adoption, rejection, correction and feedback becomes a new goal and new standard for the next round of exploration, and further improves the system's task decomposition, tool use, harness and model capabilities.
The report quotes Tang Jiabin as saying: general models tend to propose approximately the same problem distribution, while human expert feedback is the seed variable that determines whether an RSI system is exploring the human unknown. The report also claims that the task distributions, judgment standards, failure records and exploration threads accumulated through collaboration are kept in dedicated spaces, serving only the corresponding users or organizations.
It should be pointed out that the above mechanism description all comes from the team's own account and media relay; the report gives no verifiable criteria for judgment, nor does it explain the substantive difference between "co-evolutionary RSI" and existing AI-assisted scientific research workflows. Therefore, claims at the mechanism level can currently only be regarded as the team's assertions.