Modern humans have experienced a slow evolution in living standards, with significant economic growth only occurring in the last couple of centuries. Steven Lundsburg notes that after a long period where most people lived near subsistence levels, per capita income in the West began increasing at about 0.75% per year. This shift has been particularly pronounced in recent decades, as global income levels have also started to rise. The article highlights research from METR by Kwa and West, focusing on the capabilities of large language models (LLMs) in completing software engineering tasks. A key finding is that these models have a “doubling time” of about 6 months, which means they can effectively handle tasks of increasing complexity at an accelerating rate. The graph referenced shows improvements plotted on a logarithmic scale, indicating a consistent upward trajectory in task completion capabilities as new models are released. Multiple factors influence the accuracy of these findings. The reliability of the models significantly affects the intercept on the graph, meaning that a model's performance can vary greatly based on the benchmark used. The authors also discuss how the type of task and the inherent challenges of real-world applications create a "messiness tax" that can impact the models’ effectiveness. On the other hand, factors affecting the growth rate include the exponential increase in resources like compute power and data. The discussion also raises questions about the limits of LLM capabilities, especially when it comes to physical tasks or the necessity for real-world data collection. The article concludes with consideration of future implications, such as whether AI will reach a point of recursive self-improvement, potentially leading to rapid advancements in model creation. While uncertainties remain regarding the overall impact and effectiveness of AI, the data suggests a compelling trend towards increased capabilities in specific domains, particularly software engineering.