Artificial Superintelligence: Exploring Self-Improving AI

There is no superintelligence project yet. The nearest work is a self-improvement loop: it swaps in a new model only when that model beats the current one on a fixed 100-question exam.

• Last promotion: 12 September 2026, a score of 97 against the previous model's 90
• Since 6 September the loop's evaluations have not run through; we are tracing the cause