Issue 460 of Import AI covers four main topics: (1) SocioHack, a benchmark testing how RL-trained AI systems can 'game' real-world institutional rule systems like credit card rewards or school grading; (2) Anthropic sharing preliminary data suggesting prosaic recursive self-improvement is underway, with an 8x increase in code merged in 2026 vs. 2021-2024; (3) University of Zurich and Google DeepMind research showing RL-trained quadcopter drones outperforming a five-time Swiss national drone racing champion via multi-agent self-play on a single RTX 4090; and (4) a Nature study demonstrating that state-controlled media measurably biases LLM outputs, with models responding more favorably to authoritarian governments when prompted in the country's native language.
