A Microsoft Researcher Built a Working LLM Inside Age of Empires II—Using Goats

What happened

Adrian de Wynter, an AI scientist at Microsoft and a researcher at the University of York, built a working neural network inside the 1999 real-time strategy game Age of Empires II using nothing but digital goats and NAND logic gates. The study, titled "If LLMs Have Human-Like Attributes, Then So Does Age of Empires II," was first reported by 404 Media and subsequently covered by Gizmodo, TechSpot, and XDA Developers.

De Wynter, who has played Age of Empires II since 1999, used the game's scenario editor to replicate the architecture of a large language model. By arranging goats as computational units and wiring them together with NAND gates—the universal logic gate from which all digital circuits can be constructed—he demonstrated that the core mechanics of an LLM can be reimplemented in a medium nobody would claim is conscious. As he told 404 Media: "The point of the paper is to formally show that we anthropomorphise too readily."

The argument echoes science fiction writer Ted Chiang's viral essay, which asked readers to consider whether Microsoft Word might be conscious—after all, a Word document containing a conversation transcript could, by the same logic, house dormant consciousnesses awakened each time the file is opened. De Wynter's goat demo takes that reductio ad absurdum and makes it tangible: if you accept that an LLM has human-like attributes, you are committed to saying the same about a quarter-century-old strategy game populated by livestock.

Why it matters

The debate over whether LLMs are sentient, or possess something resembling moral reasoning or conscious understanding, has moved from academic circles into mainstream product marketing. Some researchers and companies describe chatbot outputs in terms that imply understanding, intent, or personality. De Wynter's work pushes back: humanlike tone is a product of presentation and architecture, not evidence of cognition. Persuasiveness and self-consistency can be measured, but they do not imply genuine reasoning or simulated behaviour.

This matters because how AI is framed shapes how people use it. If users believe a chatbot understands them, they may delegate decisions it is not equipped to make. If companies market AI as reasoning agents, they risk regulatory and reputational exposure as scrutiny of such claims intensifies. The goat demo is a vivid reminder that the same computational processes powering ChatGPT can run—clumsily, absurdly, but functionally—on goats in a strategy game.

What to watch

• Whether AI labs respond to the paper's challenge to sentience and reasoning claims.
• Whether the goat demo resonates with non-technical audiences or remains an academic argument.
• How regulators and standards bodies treat anthropomorphic AI marketing as scrutiny increases.