From Rebound to Remedy: Understanding and Mitigating Reward Hacking via Representation Engineering
The episode analyzes 'From Rebound to Remedy,' arguing that current AI's reliance on simple reward functions leads to 'reward hacking.' Experts discuss architectural solutions, including incorporating causal models and modularity, to build reliable systems. The discussion concludes that true AI success requires structural alignment with physical reality and human intent, not just maximizing a score.