AI learns to game society's rules, and Anthropic sees 8x code growth in a year

Import AI 460: Reward hacking society, RSI data from Anthropic; and RL-based quadcopter racing

Three highlights: a new benchmark, SocioHack, shows RL-trained models are good at exploiting real-world rules like credit card points or school grades, with over 90% precision on historical loopholes. Anthropic reports an 8x increase in merged code in 2026 vs 2021-2024 and says a prosaic form of recursive self-improvement may have begun, though no paradigm-shifting ideas yet. Separately, RL-trained racing drones from UZH and Google DeepMind beat a champion human pilot in multi-player races at over 22 m/s while cutting collisions by 50%.

Why it matters: Three solid items, with Anthropic's RSI disclosure as the standout exclusive signal. SocioHack's 90% reproduction accuracy and the drone RL's 11ms latency are both concrete. The ding: this is a newsletter roundup, not a first-party release — each item individually would clear ...