[How Far Is AI From 'Improving Itself'? Scale Launches RSI Bench] Beating AI News Flash: Scale AI has launched RSI Bench, specifically designed to test whether AI can conduct AI research like human researchers. The AI is directly provided with papers, code, and computational resources, then tasked with designing experiments, running training, modifying methods, and seeing if it can improve upon the original AI. The initial tests used Claude Opus 5 and GPT-5.6 Sol. Both models are already capable of running experiments, tuning parameters, and iterating on solutions, and in some tasks, they indeed surpassed the given baseline. However, current AI still struggles with 'inventing.' When challenged with the latest research, most attempts remain confined to existing methods and parameter tuning from the papers, with very few genuinely novel ideas proposed. [Original Article Link]