Building the fundamentals of neural networks using goats from 'Age of Empires II'—what's the purpose?
Adrian de Winter, a principal researcher at Microsoft and a researcher at York University, has published a paper reporting an experiment in which he used goats as signals within the game '
[2605.31514] If LLMs Have Human-Like Attributes, Then So Does Age of Empires II
https://arxiv.org/abs/2605.31514
A Perceptron in Age of Empires II
https://adewynter.github.io/notes/aoe2-circuits
If AI Is Sentient Then So Is 'Age of Empires II'
https://www.404media.co/if-ai-is-sentient-then-so-is-age-of-empires-ii/
De Winter likened the position and movement of goats to signals of 0 and 1, and created a NAND gate as a mechanism for them to move over grass and bridges. NAND gates are fundamental components for constructing complex calculations, and this mechanism can, in principle, represent the calculations of neural networks.
You can see how the goat-based circuit actually works in Age of Empires II in the following video. The grass rails are set to 0 and the bridge rails to 1, and the goats carry signals to produce the result of a NAND operation from the two inputs. The ice rails are a waiting mechanism to prevent the order of calculations from being disrupted, and after calculating AND(0,1), a negation is added to create the NAND output of 1.
Furthermore, De Winter created a 1-bit perceptron within the game and trained it to perform AND operations. A perceptron is the simplest neural network that classifies inputs as either 0 or 1, and in this video, it performs the classification by combining two XNOR gates and an AND gate. The goat not only represents the input values and weights, but also serves to signal the next circuit when each calculation is complete.
The following video shows the circuit used to train the perceptron. Given the correct label, input value, and current weights, the circuit on the left first performs a judgment, then calculates the error in the subsequent circuit and returns new weights. Bamboo forests are used as XNOR, forests as AND, baobab trees as OR, and deep water as NOT, with the game's terrain itself becoming the symbols in the circuit diagram.
De Winter states that these circuits do not demonstrate that goats possess intelligence, but rather that they show that neural network flows can be created in a game-like representation that differs from that of a typical computer.
People often judge whether an LLM possesses human qualities based solely on the appearance of natural conversation. Even if an LLM responds empathetically to the input 'lonely,' that alone doesn't mean they have empathy. It's necessary to clarify what to measure to determine empathy and understanding.
De Winter argues that even if LLM were moved into ' Age of Empires II ' and could respond to the same input with the same text, the impression people receive might change. While empathetic responses in a normal chat screen are easily perceived as human, knowing that the output is produced by a goat moving around might make it harder to perceive empathy or understanding in the same response.
De Winter stated, 'The purpose of this paper is to formally demonstrate how easily we define 'human-likeness' and how we sometimes make overly strong claims about the capabilities of LLMs. We need to stop assuming that LLMs will behave like humans simply because they are trained in natural language.'
Ultimately, de Winter does not conclude whether LLMs possess consciousness or empathy. Instead of making assumptions about the presence or absence of human-like qualities, he suggests that we should measure what kind of output is produced in response to an input, and under what conditions that behavior is maintained, and limit the scope of conclusions to the experimental conditions.
Related Posts:
