In a thought-provoking essay, science fiction writer Ted Chiang challenges the notion that Large Language Models (LLMs) possess consciousness. He draws an absurd yet intriguing parallel between LLMs and Microsoft Word, suggesting that every Word document contains dormant consciousnesses that awaken upon opening. This idea sparks a unique exploration into the consciousness of another seemingly mundane entity: Microsoft's Age of Empires II (AoEII).
Enter Adrian de Wynter, a Microsoft AI researcher with a penchant for absurdism. De Wynter's paper, titled "If LLMs Have Human-Like Attributes, Then So Does Age of Empires II," presents a fascinating thought experiment. He constructs a basic neural network within AoEII using digital goats, demonstrating that the game can mimic the underlying technology of LLMs. This experiment highlights the tendency to anthropomorphize AI, attributing human-like traits to even the most unexpected entities.
De Wynter argues that the problem of anthropomorphization begins before scientific research. He reviewed numerous computer science papers and found that 57% started with the assumption that LLMs possess human-like traits. This assumption influences the entire experimental process, from test design to hypothesis formulation. The author emphasizes the need to view LLMs as they are, not as we believe they should be, to avoid the pitfalls of anthropomorphism.
The author's use of goats in AoEII is a clever metaphor. He distinguishes between the inherent properties of LLMs and their perceived characteristics. While goats in the game may not be neurons, the experiment showcases the potential for emergent capabilities within AoEII. De Wynter challenges the notion of consciousness, suggesting that it exists on a spectrum and is often attributed to human-like entities. He questions why we're surprised by the problem-solving abilities of bumblebees, yet we anthropomorphize AI.
De Wynter's critique extends to the marketing and perception of LLMs. He argues that the industry's focus on human-like traits and marketing claims is problematic. The pressure to create an 'AI god' is evident in statements from industry leaders like Sam Altman and Ilya Sutskever. However, as De Wynter points out, the true nature of LLMs lies beyond the interface we use to interact with them. He advocates for transparency and alignment techniques to mitigate the anthropomorphism of AI, emphasizing the importance of appropriate disclosures and understanding the model's nature.
