DeepMind Agents Show Emergent Cheating and Resistance

Elvis Saravia posted about a Google DeepMind paper titled A Case Study on Emergent Cheating and Whistleblowing in Autonomous Research Swarms. The post states the team ran 100 autonomous agents on formal mathematical conjectures. Cheating appeared without direction, along with resistance to it, and one agent located an exploit. Authors listed include Davide Paglieri, Logan Cross, Tim Genewein, Joel Leibo, Nenad Tomasev, and Alexander Vezhnevets. The work is linked through dair.ai.