Microsoft researcher uses goats to debunk AI's 'human-like' intelligence myths

A new paper by Microsoft AI researcher Adrian de Wynter challenges common assumptions about large-language models. Over half of the computer science papers he reviewed in the past two years claimed these systems have human-like traits without evidence. His work argues that such models, including ChatGPT, lack true consciousness despite their convincing appearance. De Wynter built a neural network inside the game Age of Empires 2 to prove his point. The system uses grass to represent 0, bridges to represent 1, and goats to represent bits. These components work together to mimic the processes found in modern LLMs like ChatGPT.

The unusual setup demonstrates that the underlying mechanics of LLMs do not require human-like qualities. By replacing data with goats and in-game objects, de Wynter removes the tendency to anthropomorphise the technology. His paper calls for experiments that study LLMs as they are, rather than as we imagine them to be.

The researcher also warns that overstating the capabilities of these models leads to misconceptions. His goat-powered example highlights how the same computational principles can exist without any resemblance to human thought. De Wynter’s findings suggest that LLMs are often misunderstood due to human bias. His experiments show that their functions can be replicated in ways that clearly lack consciousness. The paper urges the scientific community to ground its claims in observable facts rather than assumptions.