Microsoft Principal Researcher Adrian de Wynter, in an academic paper titled "If Large Language Models Were Human, Then Age of Empires 2 Would Be Too", used Age of Empires 2's scenario editor to construct a 1-bit perceptron using in-game goats, grass, and bridges. He argues that the underlying computational logic of large language models (LLMs) like ChatGPT is fundamentally the same as the game's goat trigger mechanism.
The Technical Implementation of the Age of Empires 2 Neural Network: The Computational Roles of Grass, Bridges, and Goats
(Source: CryptoCity Screenshot)
Wynter used the custom scenario trigger system in Age of Empires 2 to build a NAND gate out of game elements: grass represents the digit 0, bridges represent the digit 1, and goats act as bits. When the logic gate is triggered, goats serving as input signals are removed, and a new goat is spawned on the output track.
Through this mechanism, Wynter implemented a 1-bit perceptron in the game, which is the most basic building block of modern neural networks. Wynter states that this goat-driven network is essentially the same as the underlying technology driving ChatGPT, differing only in complexity and scale.
Wynter's Statistical Data on Anthropomorphic Bias in 300 Computer Science Papers
In his paper, Wynter analyzed over 300 computer science papers from the last two years and revealed the following figures:
· 57% of the papers, at the outset of experimental design, pre-assumed that LLMs possess human traits such as anxiety or morality.
· Among the papers that assumed these traits, 77% ultimately concluded that the traits do indeed exist.
· Wynter characterizes this result as confirmation bias, arguing that the researchers' initial anthropomorphic assumptions directly determined the final scientific conclusions.
· Wynter calls on the scientific community to instead use null hypotheses that do not presuppose human traits in their experiments.
Big AI's Commercial Narratives and the Market Mechanism of LLM Anthropomorphism
Wynter points out that the business strategies of AI companies reinforce the public's perception of LLM consciousness. He notes that consumers tend to pay more when they can resonate with a product, a principle that applies from toasters to LLM subscription services.
In his paper, Wynter cites three cases: OpenAI CEO Sam Altman has repeatedly hinted that building LLMs is a path to "god-like AI"; former OpenAI scientist Ilya Sutskever has discussed with employees regarding the company's models as "god-like consciousness"; and Anthropic CEO Dario Amodei has told the media that he "cannot be sure whether AI is conscious." These statements have all been cited by Wynter in his paper and are not his personal interpretation of the three individuals' positions.
Frequently Asked Questions
#### Did Wynter build a complete AI model in Age of Empires 2?
Wynter implemented a 1-bit perceptron in the game, which is the most basic building block of modern neural networks, not a complete AI model. The purpose of the paper is to demonstrate that the same underlying computational logic exists in both the game's goat trigger mechanism and the underlying technology of ChatGPT, thereby illustrating that people's perception of ChatGPT's anthropomorphic traits primarily comes from the text interface and psychological expectations.
#### What do the statistical figures of 57% and 77% represent?
According to Wynter's statistical analysis in his paper, among over 300 computer science papers from the last two years, 57% pre-assumed that LLMs possess human traits such as anxiety or morality; among these papers, 77% ultimately concluded that the traits do indeed exist. Wynter interprets this as confirmation bias—the researchers' initial assumptions directly influenced the experimental design and final scientific conclusions.
#### How does science fiction author Ted Chiang describe the problem of over-anthropomorphizing LLMs?
According to the article, Ted Chiang once wrote that blindly believing a language model possesses consciousness is akin to believing that every time you open a Microsoft Word file, you awaken multiple individuals lurking within the conversation history. He uses this analogy to illustrate that the anthropomorphic perception of LLMs is a cognitive illusion.
