Import AI 469: Science AI; RSI simulator; and Zuck's technological pessimism

DiG-bench is a new benchmark featuring 70 handcrafted, text-based interactive games with hidden rules and objectives, designed to measure an artificial intelligence system’s capacity for creative intuition and discovery in novel environments. The games span seven difficulty tiers. While individual humans achieved complete success, current frontier models struggle heavily, with only models like Opus 5 and Fable 5 managing a twenty percent success rate on the hardest tier. Researchers suggest that human parity on this benchmark will likely arrive by mid-2027, potentially unlocking recursive self-improvement.

To help people understand recursive self-improvement, Paradigm Research released the RSI Simulator, a browser game where players manage an AI company by balancing compute, researcher investment, and data licensing. Additionally, an AI startup called Inherent developed Faraday, a supervisory model built on top of Qwen-3.6-27B that controls frontier models to improve their scientific research capabilities. Trained on a dataset called Replica consisting of historical machine learning papers with missing results, Faraday outperformed standard frontier models like Opus and GPT on a majority of replication tasks, pointing toward systems that can autonomously design experiments and advance state-of-the-art science.

In a recent essay titled The Future is for Everyone, Mark Zuckerberg outlines Meta's philosophy of aggressively proliferating superintelligence tools to every individual to empower personal growth, invention, and education while avoiding concentrated power. However, the essay does not address whether systems capable of superhuman invention will exclusively serve individuals who are less capable than the AI itself, leaving open how such technology will alter global power balances.