Finetuning Borges
A researcher uses Kimi K2.5-Thinking to recreate 'Pierre Menard' as if written by a 2026-era AI.
Researcher Linch has undertaken a novel AI experiment: fine-tuning the Chinese open-source large language model Kimi K2.5-Thinking to generate Jorge Luis Borges' seminal short story, 'Pierre Menard, Author of the Quixote.' The ambitious goal was not mere transcription or creating a 'Borgesian' pastiche, but to have the 2026-era model produce the exact text as an original act of AI authorship. This required navigating significant technical constraints, including the model's 256K-token context window, which proved insufficient for containing Borges' entire life history and influences as a system prompt.
After considering and rejecting complex methods like machine unlearning and sparse autoencoders to isolate a 'Borges feature,' Linch pursued a path where the model's output is inherently 'contaminated' by its modern training data. The successful excerpts generated by Kimi have, in Linch's view, surpassed the quality of Borges' original prose. The AI's version of a key philosophical passage is praised for its 'elegant writing and machine innocence,' with Linch boldly claiming the output is 'far better than the vast majority of Borges’ own (honestly quite mid) fiction.' This project challenges notions of originality, authorship, and the aesthetic potential of fine-tuned LLMs.
- The project fine-tunes Kimi K2.5-Thinking, a Chinese open-source LLM, to generate Borges' 'Pierre Menard' as an original AI work.
- Technical challenges included a 256K-token context window and the philosophical goal of avoiding simple mimicry of Borges.
- The researcher claims the AI-generated excerpts set a new benchmark for LLM fiction, surpassing the perceived quality of the original story.
Why It Matters
Pushes boundaries of AI authorship and fine-tuning, questioning originality and the aesthetic value of machine-generated text.