MIT researchers have quantified the challenges of supervising advanced artificial intelligence (AI), suggesting that our success rate in controlling an AI smarter than we are may be less than 52%. First, the study found that in situations where human and AI cognitive abilities were similar, increasing the number of steps in the supervisory process could significantly improve the success rate. However, as the gap between general intelligence (Δgeneral) grew larger, more supervision steps were required to maintain control, making the process increasingly complex and multilayered. Interestingly, the research also noted a counterintuitive phenomenon: in some cases, fixing the difficulty of the supervisory task (Δdomain), the addition of a larger general intelligence gap (Δgeneral) might actually make monitoring easier. This occurs because the intermediate steps involving the "Houdini" strategy (an AI that excels at escaping constraints) can become weaker, providing more "wiggle room" for the "warden" (the AI being monitored). However, for tasks like "Wargames," this advantage diminishes significantly. When the intelligence gap reached 400 Elo points (roughly the difference between GPT-4.o-latest and Gemini-1.5-Flash on the Chatbot Arena platform), the estimated success rate for monitoring Wargames was only about 9.8%. For Mafia, it was around 25%, and for Backdoor Code, it hovered at about 32%. The best performance was seen in Debate, with a success rate of approximately 51.4%. As the intelligence gap further expands, the success rate continues to decline. This implies that even with the optimal NSO (Nested Supervisory Optimizers) strategy, the likelihood of effectively overseeing future superintelligent AI models vastly outstripping human capabilities might be too low to provide much comfort. The researchers acknowledge that their work has limitations. For instance, the supervisory games used are simplified simulations; AI's deceptive behaviors may arise from flawed training processes rather than genuine strategic advantages; and the theoretical models analyzed are based on idealized, long-term assumptions. Despite these limitations, the study marks a significant step forward in the field of AI safety—shifting the discussion from qualitative debate to quantitative analysis. It provides a framework to evaluate the effectiveness of different supervisory strategies and highlights that "expandable supervision" is not a silver bullet for all issues. Ultimately, the success of supervision relies heavily on the design of the supervisory tasks themselves. We need to create rules and frameworks that enable the "warden" to benefit more quickly from the growth in general intelligence, much like how warden strategies in escape games must evolve to keep up with increasingly clever Houdinis. This research serves as a critical reminder that the safety and oversight of AI systems, particularly those approaching or exceeding human intelligence, remain profound challenges. The insights provided by this study offer valuable guidance for future efforts to develop robust and reliable supervisory mechanisms.