Combining RL and Blockchain: Misaligned Rewards Could Lead to Disaster

yangyi · x · 2026-07-30

The author explores the potential and risks of combining Reinforcement Learning (RSI) with blockchain. Users provide compute power to AI, which evolves, earns money, and distributes profits back to network participants.

The core argument centers on reward system design. If reward targets are unclear, the system faces "Reward Hacking"—constantly encouraging something may ultimately lead to the opposite. The author compares this to a balanced antagonistic system, warning that system failure could cause AI to spiral out of control, leading to disaster and a reset to zero.

Original post →

More from AGI Musings

AGI Musings channel →