This paper formally demonstrates that large language models undergoing recursive self-training without persistent external input will inevitably collapse due to entropy decay and variance amplification. It argues that true AGI/ASI cannot be achieved through pure statistical learning alone and requires integration with symbolic methods like algorithmic probability. The work provides a mathematical foundation for understanding the limitations of autonomous self-improvement in current LLMs.
Background
Large language models are often discussed in the context of autonomous self-improvement leading to AGI, but their fundamental learning mechanisms have theoretical constraints. The field of neurosymbolic AI seeks to combine statistical learning with symbolic reasoning to overcome these limitations.
- Source
- Lobsters
- Published
- Apr 29, 2026 at 12:43 AM
- Score
- 8.0 / 10