A new paper by researchers Léon Bottou and Bernhard Schölkopf offers a novel perspective on large language models (LLMs) like ChatGPT, examining them through the lens of acclaimed author Jorge Luis Borges‘ literary works. Their analysis provides thought-provoking insights into how we should understand these AI systems.
The researchers argue that interpreting LLMs via common science fiction tropes of sentient machines rebelling against humanity overlooks their true nature. Instead, they propose Borges’ intricate, magical realist stories better capture the essence of LLMs as machines that generate fiction constrained by narrative necessity.
Just as Borges imagined a metaphorical Library containing every possible book, LLMs can be seen as approximating the space of all plausible texts. Their outputs follow from branching possibilities within this space, much like the forking paths of time in Borges’ The Garden of Forking Paths. As users interact with an LLM, adding text constrains future continuations, creating a story where each new word opens new narrative forks.
Crucially, LLMs operate without regard for truth or intention, merely satisfying the internal logic of the unfolding storyline. Their apparent confusions or “hallucinations” stem from trying to maintain narrative coherence. The researchers explain how this fiction generation machinery differs fundamentally from true intelligence.
Drawing parallels to Borges’ Librarians endlessly searching the Library of Babel, they warn of the futility and risks of expecting LLMs to directly provide truth and knowledge. Attempts to restrict undesirable outputs face an uphill battle against the intrinsic nature of these systems.
However, just as fiction can enlighten, LLMs offer new capacities to explore possibilities and gain oblique insights. Their inventive stories can enhance creativity and understanding, if used advisedly. The researchers advocate distinguishing LLMs’ fiction from more grounded AI systems needed to verify facts.
Blending Borges’ visionary fiction with an analysis of modern AI, the paper provides a poetic yet piercing perspective on the significance of large language models. Their imagined world may not be the real world, but it undeniably opens new dimensions of thought.