Paul Krugman Talking with Paul Kedrosky (paulkrugman.substack.com)

In a revealing conversation between economists Paul Krugman and Paul Kedrosky, the complexities of artificial intelligence and its intertwined relationship with economic factors are discussed. Krugman expresses frustration at the simultaneous upheaval of U.S. trade policy and the rapid evolution of AI, leading him to seek Kedrosky's insights into generative AI, particularly large language models. Kedrosky explains that these models function as "loose grammar engines," predicting the next element in a sequence based on vast training data that reflects various domains, including language and software. He emphasizes that while they exhibit the ability to generate coherent outputs, their training relies heavily on the patterns in their datasets, often reflecting the biases or characteristics of the data sources, such as a demographic profile from Reddit users. The conversation also highlights the challenges faced by the AI community regarding the diminishing returns of data as training resources become scarcer and more complex. Kedrosky compares the abundance of textual data from the internet to a "Saudi Arabia of data," which, as it becomes exhausted, leads to models that may become “sycophantic” in their responses due to over-reliance on reinforcement learning from human feedback. The implications of this discussion are significant, as they underscore the need for a nuanced understanding of AI capabilities and limitations — particularly the difference in learning from structured programming data versus natural language, which may mislead assumptions about advancing towards artificial general intelligence (AGI).