If LLMs Have Human-Like Attributes, Then So Does Age of Empires II
20 hours ago
- #Turing-complete
- #LLMs
- #anthropomorphism
- Research often attributes human-like traits such as morality or language understanding to LLMs, but this may be incorrect.
- A simple neural network trained on Age of Empires II shows that any entity in a powerful substrate (like LEGO or Boston) could also exhibit such attributes.
- LLM's anthropomorphic attributes are empirically non-unique: some properties stay constant, but interpretations change with the substrate.
- Discussions about these attributes need explicit measurement criteria; otherwise, interpretation depends on representation.
- Assuming these attributes exist or not generally leads to circular or uninformative conclusions.
- Proposes a 'null' assumption of LLM non-uniqueness instead of anthropomorphism for experimental setup.
- Addresses objections, surveys the field, and proves Age of Empires II is functionally- and Turing-complete.