A Microsoft researcher built a functional neural network inside the video game Age of Empires II using goats to critique the state of AI science.
The project, detailed on The Decoder, used in-game goats as artificial neurons to create a working XOR gate, a fundamental logic operation. The researcher, Tarun Chitra, aimed to demonstrate the absurd lengths required to run modern AI models, not to solve a practical problem.
The Core Critique: AI Is Overly Expensive
The central argument of this project is that AI research has become too reliant on massive financial resources. Chitra built the goat network to show that the core science of neural networks is simple, but the industry has made it prohibitively expensive.
“The billion-dollar question is whether these huge models are actually necessary, or if we can achieve the same results with far less compute,” Chitra stated.
The goat-based network performed a simple calculation that any modern neural network could do in microseconds. The process, however, required complex coordination of in-game units and resources, highlighting the waste inherent in current AI scaling methods.
How The Goat Network Works
The system operates by assigning each goat a specific role as a neuron. Each goat receives a “weight” value that controls its behavior. When the game is run, the goats collectively process information to produce an output.
- Each goat represents a single artificial neuron. The game’s pathfinding and collision mechanics simulate the connections between neurons.
- The network’s training occurs through in-game interactions. The researcher adjusted goat behavior by modifying the game’s internal logic.
- The final output is a binary result. The system successfully solved the XOR problem, which requires a non-linear combination of inputs.
The project required the creation of a custom map and specific game rules to make the goats function as a network. This complexity underscores the difficulty of even simple AI tasks when resources are limited.
The Deeper Problem: Compute Waste
The project’s primary critique is that modern AI training consumes enormous amounts of energy and hardware. Chitra argues that this focus on scale distracts from fundamental questions about efficiency.
- Training a single large AI model can cost millions of dollars. This creates a barrier for independent researchers and small teams.
- The industry prioritizes “bigger is better” over “smarter is better.” The goat network shows that simple algorithms can work, but the infrastructure to run them is needlessly complex.
- The result is a field that rewards funding over insight. Chitra’s work suggests that the science of AI has been overshadowed by the economics of AI.
The goat network is a form of “conceptual art” that forces the viewer to reconsider the value of massive compute clusters. The underlying math of the network is identical to modern deep learning, but the implementation is deliberately impractical.
The Takeaway for AI Researchers
The message is clear: the AI industry should not accept the current cost structure as necessary. Chitra’s project is a call to action for researchers to explore smaller, more efficient models.
“We are building these huge models because we can, not because we must.”
The project serves as a reminder that the core principles of neural networks can be implemented with any system that can perform basic calculations. The field needs to ask if the current path of scaling is the only path forward.
Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.
What are your thoughts on this? I’d love to hear about your own experiences in the comments below.