Epoch AI (source: https://epoch.ai/blog/train-once-deploy-many-ai-and-increasing-returns) discusses the concept of increasing returns to scale in AI systems, attributing this to the "train-once-deploy-many" property. Unlike human intelligence, AI models can be trained once with substantial resources and then deployed in numerous instances for inference, leading to economic output that grows faster than the increase in computational input. This is further amplified by the trade-off between training compute and inference compute, where investing more in training can result in models that require less compute for inference while maintaining performance. The article also explores a simplified economic model where AI's ability to produce chips accelerates growth, highlighting the potential for super-exponential growth in an AI-driven economy.