Predicting AGI: What can we say when we know so littel?

Abstract

This work suggests a model for predicting when Artificial General Intelligence (AGI) might be achieved, focusing on the epistemic state with respect to AGI rather than attempting to predict when AGI will actually be achieved. The model proposes a Pareto distribution to describe the probability of starting to “taxi” towards AGI, which is analogous to the AI research community having a clear path forward to AGI. The Pareto distribution is justified through dimensional analysis, connections to stable distributions, and Laplace’s rule of succession with an unknown exponential hazard rate. The model, conditioned on not taxiing by now, predicts that there is a 13% chance that taxiing will occur in the next 30 years and a 20% chance that it will occur between 40 and 90 years from now. The authors emphasize the relative weakness of the evidence, but suggest that the model implications indicate the importance of focusing on short-term strategies to ensure AGI is Friendly. – AI-generated abstract.