Truth is not a direction: a Tarski attack on LLM probes

This article highlights the limitations of using embedding space directions to determine truth in large language models (LLMs). It demonstrates that the idea of a single 'truth' direction is fundamentally flawed, echoing Gödel's findings that truth cannot be fully captured by formal systems, which has significant implications for AI safety and the development of trustworthy AI systems.

• The concept of a 'truth' direction in LLM embeddings is mathematically flawed.
• Attempts to classify truth in LLMs resemble Gödel's limitations on formal systems.
• Understanding these limitations is crucial for advancing AI safety and trustworthiness.