The case for auditable AI reasoning beyond longer chains of thought
A former Google DeepMind employee argues that LLMs typically lack inspectable records of what they know and what remains uncertain.
TLDR
An author who says he recently left Google DeepMind argues that AlphaGo’s ability to search possible futures offers lessons for AI reasoning. He says LLMs typically lack explicit, inspectable records of knowledge, uncertainty and supporting evidence, and that longer chains of thought are not the same as genuine reasoning. He calls for auditable evidence, inference and belief revision to support trustworthy new insights.
Combined views
1.8K
2 Sources, first seen ago
