Barriers to Complexity-Theoretic Proofs That "AGI" Using ML Is Impossible (arxiv.org)

A recent paper by van Rooij et al. claims to demonstrate that achieving Artificial General Intelligence (AGI) through machine learning is intractable from a complexity-theoretic perspective. This assertion is significant as it challenges the feasibility of developing human-like intelligence via data-driven methods, a core goal within the AI/ML community. However, the authors of the critique argue that the proof hinges on an unverifiable assumption regarding the distribution of (input, output) tuples in the dataset, which could undermine its validity. The discussion highlights two critical barriers to strengthening the proof: the necessity of a clear definition of "human-like" intelligence and the recognition that specific machine learning systems possess distinct inductive biases that significantly impact any analysis. Furthermore, attempts to refine the proof by examining subsets of the data encounter complications related to defining these subsets effectively. This ongoing debate underscores the complexities of proving theoretical limits in AI and emphasizes the importance of foundational definitions in advancing our understanding of machine learning capabilities.