The Color of LLMs

In 1937, Jorge Luis Borges wrote a characteristically engimatic historical note and literary reflection on «Ramon Llull' s Thinking Machine.» It was the year when the idea of «thinking machines» picked up momentum with Harold Aiken’s mechanical Mark I early computer that became the electronic Mark IV. 85 years ahead of ChatGPT (first released in 2022), Borges is looking back some 650 years at Ramon Llull"s great arts which includes an ars inveniendi that inspired rational speculation ever since. This is one five essays that look though the lens of Borges at contemporary learning machines. It explores how Jorge Luis Borges' account illuminates the nature and limitations of contemporary large language models. The author recalls that Jonathan Swift’s Gulliver’s Travels already describes an engine that mechanically assembles books from broken sentences, and Borges associates it with the combinatorial tradition of thinking machines. This tradition, including Llull’s disks, Raymond Queneau’s cut-up book, and the Infinite Monkey Theorem, relies on generation by chance and recombination. It is contrasted with the algorithmic tradition, represented by the Antikythera mechanism, astronomical clocks, and the calculating machines by Schickard, Pascal, Leibniz, and Babbage, where problems are solved by rules. The author notes that astrology failed because it depended largely on correlation rather than computation, and that modern LLMs are also based on correlations learned from human artifacts. At the same time, LLMs contain a combinatorial element guided by prompting. A parallel is drawn with Oulipo: literary constraints, such as writing a novel without the letter «e,» foster creativity, whereas LLM prompting directs and constrains searches. The analysis of Borges' color motifs — «red,» «zenithal,» the tiger, Blake — shows that Borges' text is a subtle critique of simplistic thinking machines and an affirmation of human creativity and intertextuality, which cannot be reduced to mechanical recombination and correlation.