Thermodynamic Measure of Intelligence

The paper proposes a framework for measuring intelligence based on thermodynamic principles, defining it as the amplification of rare but valid futures through recursive self-simulation. It establishes a connection between the fidelity of internal simulations and the ability to achieve high levels of rare-valid lift, suggesting that effective policies within these simulations are crucial for optimizing intelligence. This framework allows for a universal measurement of intelligence across various systems, including LLMs, highlighting the importance of self-referential modeling in intelligent behavior.