> In-game constructions of NAND gates and a perceptron (forward prop and training) as described in in 'If LLMs Have Human-Like Attributes, Then So Does Age of Empires II'.

Interesting concept

> We begin by proving that Age of Empires II is functionally- and Turing- complete. Then we build a perceptron and a circuit to train it in-game. With that, we argue that changing the substrate (representation) of an LLM also alters the perception of their attributes.

This is fun, but I don't think it's particularly surprising. A substrate being turing-complete alone is enough evidence that you can train and run a perception on it, assuming the available memory is sufficient.

> We then show that research in LLM anthropomorphic attributes cannot be done starting by assuming that these attributes exist (or not) in the system; even if you aim to conclude that they do not exist. This assumption can happen even when you do not make it explicitly! It also shows that there are ways to do good, sound research without needing to make that assumption.

I... don't see how this follows? I wanted to see how this argument unfolded, but it seems the arxiv link on this page is broken? It just links to arxiv.org and the rest of what is on this linked page doesn't seem to cover this second assertion at all.