The LLM Ceiling: Why Large Language Models Cannot Produce ASI, and What That Means for the Field

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The dominant trajectory in AI development assumes that scaling large language models is the path to artificial superintelligence (ASI). This paper argues that assumption is structurally mistaken — not because LLMs lack capability, but because the architectural absences responsible for their known alignment failures are not engineering problems correctable by scale. They are structural absences that the LLM paradigm cannot address from within itself. We identify four such absences — frozen reward signal, consequence vacuum, stakes blindness, and calibration anchor — and argue that each represents not merely a safety failure but a capability ceiling. We examine the strongest counterargument to each absence, and argue that the counterarguments, when examined carefully, either concede the structural point or substitute episodic workarounds for the continuous architecture ASI requires. The paper concludes by identifying what ASI would actually require, situating this within a co-evolutionary framework developed in companion work. The paper is self-contained.