Microsoft Researcher Exposes AI's 'Intelligence' Illusion with Goats in Age of Empires II
A Microsoft researcher has created a working neural network using goats in Age of Empires II to critique the methods used in AI research, highlighting the illusion of intelligence in language models. This thought-provoking experiment reveals that the perceived intelligence of AI models is largely a product of packaging and user interface.
In a surprising move, a Microsoft researcher has successfully built a functioning neural network within the map editor of the classic strategy game Age of Empires II, using goats as bits to process information. This unorthodox approach was designed to challenge the conventional wisdom in AI research, particularly in the field of language models. By creating a working neural network with such an absurd setup, the researcher aims to demonstrate that the intelligence attributed to AI models is often an illusion created by clever packaging and user interface design.
The experiment involved using goats standing on grass or bridges to represent 0s and 1s, and then building logic gates using the game's scenario editor. The resulting mini-network, consisting of two XNOR gates and one AND gate, was able to learn the logical AND function. This achievement may seem trivial at first, but it has significant implications for the field of AI research. It shows that, in theory, any computer can be replicated using an idealized version of the game, making it as powerful as a full-fledged computer.
This is not just a theoretical exercise; it has practical implications for the development and perception of AI models. The researcher argues that if a language model can be replicated using goats in Age of Empires II, it can also be replicated using other unconventional means, such as Lego bricks or even a city's population. For instance, the 667,000 people living in Greater Boston could, in theory, be used to run the math behind a language model by texting each other computational steps on their phones. This thought experiment highlights the fact that the intelligence attributed to AI models is not inherent to the models themselves, but rather a product of how they are presented to users.
The researcher's critique of AI science is not limited to the experiment itself, but also extends to the broader community of AI researchers. An analysis of 315 AI papers revealed that more than half of them make the mistake of attributing human-like qualities to AI models. This mistake can have significant consequences, as it can lead to unrealistic expectations and overestimation of AI capabilities. In contrast, rival AI models from other providers, such as Google's LaMDA or Facebook's LLaMA, have also been criticized for their lack of transparency and tendency to anthropomorphize AI.
The implications of this research are far-reaching, and they challenge the conventional wisdom in the AI community. For developers, this means that they need to be more careful when designing and presenting AI models, avoiding the temptation to attribute human-like qualities to them. For businesses, this means that they need to be more realistic about the capabilities and limitations of AI models, and not overpromise what they can deliver. For everyday users, this means that they need to be more critical when interacting with AI models, recognizing that the intelligence they perceive is often an illusion created by clever design.