MirroS | Let every unknown become the starting point of the next round of evolution

Andrej Karpathy's AutoResearch demonstrates a minimal RSI (Recursive Self-Improvement) loop: AI modifies code, runs experiments, evaluates results, retains valid improvements, and then moves on to the next round. This mechanism works because the software sandbox is a "pre-written world" with fixed rules, limited states and clear goals. What the AI needs to do is to find better answers in a known world.

But the real world has never been fully documented.

Robots do not know what may appear in a room; advanced manufacturing systems cannot assume that digital twins are always accurate; scientific experiments may also produce results that cannot be explained by existing theories. In these scenarios, failure does not necessarily mean that the AI has made a wrong choice, but may indicate that the world it uses to find answers is itself missing a part of reality.

This is exactly the problem that Physical RSI aims to solve. It does not only improve the Actor that performs tasks, but also must synchronously expand the World Model that the Actor relies on to act. The two co-evolve, and the progress of each will expose the next boundary of the other.

What we need is not a World Model that only generates realistic videos, but an Agentic World Model that can understand reality, construct simulations, put forward hypotheses, and continuously correct itself based on real feedback. Such a World Model grasps the objects, relationships and causal laws in reality through abstraction, and then uses language and code to turn an accidental event into a reproducible, verifiable and transferable environment.

For example, a quadruped robot gets tangled in a cable. The system must first judge: is it that the world model cannot explain the relationship between cables, tension and motion? Or has it understood the problem but does not know how to get free? The former is an environment gap that requires expanding the world model; the latter is a skill gap that requires improving the Actor. New understandings and strategies ultimately need to be verified back in the real world.

Thus, "unexpected events" are no longer just errors, but become the entry point for intelligence evolution:

Encounter → Unexpected Event → Diagnosis → Abstraction → Simulation → Solution → Verification → Internalization

The same loop can be applied in embodied intelligence, advanced manufacturing, scientific discovery, and any field that requires continuous interaction with the incompletely known reality.

The first half of RSI is to make AI a better problem solver.

Its second half is to enable AI to continuously expand the world it can understand and solve.

True intelligence does not mean continuous optimization in a fixed world, but making every unknown the starting point of the next round of evolution.

Mirror the ever-changing world,

Build an OS that enables the co-evolution of the world model and intelligence.

MirroS, born for the unknown.