A Bear Case: My Predictions Regarding AI Progress

Abstract

The current AI paradigm based on large language models (LLMs) is approaching a plateau in capabilities that will not lead to artificial general intelligence (AGI). While LLMs will continue to improve, these improvements will face steep diminishing returns and decouple from the intuitive metric of “getting generally smarter.” Recent models like Grok 3 and GPT-4.5 confirm this trend with minimal improvements over predecessors. LLMs excel only at eisegesis-friendly problems (where humans do the work of making their outputs useful) and in-distribution problems (tasks similar to their training data). They lack true agency and struggle to maintain focus across long inferential distances. Despite impressive benchmarks, real-world application reveals their limitations. By the 2030s, LLMs will be deeply integrated into the economy as useful tools, but will not achieve autonomous general intelligence or replace broad-scope human labor. However, at some unknown point—likely in the 2030s—someone will develop a different approach to AI, potentially leading to transformative consequences. Until then, AI progress will be characterized by incremental improvements masking fundamental limitations, with hype consistently outpacing actual capabilities. – AI-generated abstract.