Research

Microsoft Researcher Builds Neural Network Using Goats in Age of Empires II

Adrian de Wynter, a Microsoft researcher, created a working neural network in Age of Empires II using goats to critique AI science, revealing the game's mechanics can simulate computational processes.

A scientist adjusts a microscope in a laboratory setting, focusing on the sample.

Photo: Tima Miroshnichenko / Pexels

A Microsoft researcher has built a functioning neural network within the Age of Empires II game using goats as bits to critique AI science. The project, developed by Adrian de Wynter, highlights the absurdity of current AI research methods by demonstrating how in-game elements can be used to simulate computational logic. Goats are used to represent binary data, with a goat on grass equaling 0 and a goat on a bridge equaling 1. The network uses logic gates and ice ramps to maintain computational integrity, ultimately learning the logical AND function. The project is a deliberate critique of the assumptions made in AI research, particularly regarding the attribution of human-like traits to language models. Source: thedecoder

De Wynter argues that the game's mechanics, such as the in-game market with a gold cap of 9,999, allow for the simulation of a perpetual economic cycle. This enables the game to function as a full-fledged computer, with buildings acting as memory cells and active farms representing the current computational state. The trained perceptron in the game appears as a maze of walls through which goats wander, symbolizing data movement. This concept is used to challenge the idea that language models inherently possess human-like traits, suggesting that the appearance of such traits is more about packaging than internal properties. Source: thedecoder

The study also reveals that more than half of the 315 AI papers examined from mid-2024 to mid-2026 assume language models have human-like traits. De Wynter highlights the circular reasoning in many experiments, where assumptions about model consciousness are used to design tests that confirm the same assumptions. This leads to ambiguous results and reinforces the need for a more observational approach to AI research. Source: thedecoder

Key points

Adrian de Wynter built a working neural network using goats in Age of Empires II to critique AI science.
The in-game market's gold cap of 9,999 allows for a perpetual economic cycle simulating computational processes.
More than half of the 315 AI papers examined assume language models have human-like traits.
De Wynter argues that the appearance of human-like traits in language models is more about packaging than internal properties.
The study highlights the circular reasoning in many AI experiments, where assumptions about model consciousness are used to design tests that confirm the same assumptions.
De Wynter proposes a sober approach to AI research, emphasizing observation over attribution of human-like traits.

Source: The Decoder Read the original →

WRITTEN BY

Maya Chen

AI Research & Breakthroughs

Maya breaks down the latest AI research papers, benchmarks, and technical breakthroughs into plain language.