Statistician Argues LLM Alignment Is Impossible Since Models Are Distributions

Statistician Gerard Sans argues that alignment presupposes a fixed object to align, but LLMs are mathematically distributions whose behavior varies with context, making alignment impossible with current techniques.

2026-09-18 ~ 2026-09-18 · 2 related posts

• Alignment debate: models are context-dependent distributions, so alignment is impossible — gerardsans · 2026-09-18
• Statistician argues alignment is impossible: LLM outputs are distributional, not fixed traits — gerardsans · 2026-09-18