A Microsoft researcher has constructed a functional neural network within the venerable real-time strategy game Age of Empires 2, utilizing in-game units and terrain to represent computational elements. The project, spearheaded by AI researcher Adrian de Wynter, aims to challenge the growing tendency to attribute human-like sentience to large language models (LLMs) simply because they process and generate natural language.
De Wynter detailed his findings in a paper provocatively titled 'If LLMs Have Human-Like Attributes, Then So Does Age of Empires II.' He explained his penchant for extreme demonstrations, stating, "I have this tendency to dial up things to 11 when I really think I need to make a point," adding that "absurdism is pretty standard in philosophy and theoretical computer science."
The intricate construction was built using Age of Empires 2's scenario editor. De Wynter ingeniously represented binary digits using game elements: grass signifies '0,' bridges represent '1,' and goats function as the bits themselves. This system allowed him to build a working NOT AND gate and a 1-bit perceptron, a fundamental type of neural network. The result is a system that, while visually baffling to a casual observer, mirrors the underlying processes powering sophisticated LLMs like ChatGPT. He chose Age of Empires 2 specifically because it is a less conventional medium for such a demonstration compared to games like Minecraft, which have seen similar player-driven computational projects.
Videos available on de Wynter's GitHub page showcase the goat-powered network in operation. He posits that the abstract and non-human nature of the components—goats and grass rather than natural language—prevents observers from perceiving the output or behavior as inherently human. This visual and conceptual disconnect is central to his argument.
"The point of the paper is to formally show that we anthropomorphise too readily, and that sometimes the claims we make with regards to LLMs capabilities are too strong," de Wynter stated. He emphasized the distinction between what LLMs are intrinsically (defined by their internal weights and operations) and how they are perceived by humans.
De Wynter is concerned that the assumption of LLM sentience or human-like consciousness, without rigorous proof, can have significant negative implications, particularly in scientific research. He noted that in his experience reviewing over 300 computer science papers in the past two years, more than half began with the premise that LLMs possess human-like qualities.
"I propose that we need to stop assuming that LLMs behave like humans just because they were trained with natural language," de Wynter urged. "Instead, we should perform experiments that allow us to see LLMs as how they are, not how we believe they should be." His goat-powered network serves as a tangible, albeit unconventional, argument for this approach, advocating for empirical observation over subjective interpretation when assessing artificial intelligence.