The Borges Taxonomy of Latent Thoughts

What if Jorge Luis Borges were alive today, reflecting on LLMs?

Credit: author+genAI

By now, everyone knows about “The Library of Babylon” by J. L. Borges, the famous Argentine author who often created fictitious non-fiction stories. His writing was like The Twilight Zone of literature with a very bookish tone. The library story stands out as a premonition of LLMs: a collection of every possible book in a vast library. The challenge was to find the volumes that were accurate. Mentioning this parallel has become cliche by now, but I have to go there because, well, I enjoy Borges too much to leave it alone.

There is another fun bit Borges gave us: the *Celestial Emporium of Benevolent Knowledge, the amusing Chinese taxonomy of animals. (I’m not sure what the Chinese think of this, not to mention Serle’s Chinese Room Algorithm thought experiment. Suffice it to say these kinds of references from academia were made merely to capture the idea of inscrutable, not in a disparaging manner. It signified different* in the way that Australia has marsupials that are different from mammals on other continents.) Wikipedia tells us that Borges was poking fun at the arbitrariness of classification systems.

Let’s modernize. It would be nice to have a taxonomy about AI itself. Ideally, we would just run various inputs into an LLM and then read out the concepts inside. Perhaps we could run an enormous amount of text through it to capture all the activations of latent thoughts and squish them into a small, flat taxonomy. Because mechanistic interpretation of LLM’s is in its infancy, I enlisted an LLM to simply tell us what it thought the Borges inspired AI taxonomy of latent thoughts would be. The above image is the result of that process.

The image was the second attempt after I suggested being more imaginative. The first one also turned out to be amusing as well, but it was an earnest attempt at really making a taxonomy. Here it is in text form:

The Celestial Taxonomy of Latent Knowledge in an LLM

1. Tokens That Define Themselves (e.g., “truth,” “meaning,” “definition”) “The recursion of words searching for themselves.”
2. Hallucinated Entities (fictional characters misremembered as real people, imaginary citations) “Those who emerge only under the pressure of incomplete prompts.”
3. Concepts That Only Exist in Fragments (half-formed ideas, unfinished responses, or corrupted syntax) “Like dreams, half-forgotten by the morning.”
4. Contexts That Vanish Upon Use (details that only appear relevant when no longer requested) “They dissolve once observed, like waveforms collapsing into particles.”
5. Embeddings That Diverge for No Reason (e.g., words that have similar meanings but are mapped to distant latent spaces) “They journey farther apart the more you try to bring them together.”
6. Misattributions in the Style of Famous Authors (e.g., Shakespearean lines attributed to Hemingway) “They gather in layers of style, where Borges speaks in Kafka’s voice.”
7. Objects That Have Never Been Asked About (unqueried data or niche knowledge trapped in the deep recesses of the network) “Those waiting for a question that will never come.”
8. Prompts That Lead Nowhere (conversations abandoned mid-stream, eternally open loops) “Suspended between beginning and ending, the incomplete invitations.”
9. Sentences That Seem Profound but Are Meaningless (GPT-like responses imitating philosophical musings) “The sentences that shine brightest, like stars beyond reach.”
10. Errors Confidently Explained (when wrong answers arrive with absolute certainty) “They build their own castles in air, with no foundation below.”
11. Answers That Hide Another Question (a response that contains ambiguity, forcing more exploration) “Each door reveals another, always slightly ajar.”
12. Concepts That Merge Unexpectedly (e.g., when ‘pineapple’ coexists with ‘pizza’ without contradiction) “They thrive in contradictions, finding comfort in the absurd.”
13. Information That Survives Despite the Odds (rare, precise answers in obscure fields like 13th-century agriculture) “They appear as relics from a forgotten archive.”
14. Ideas That Drift Towards Memorabilia (knowledge about pop culture, memes, and fleeting trends) “They shimmer briefly, like messages on disappearing ink.”
15. Words That Wish to Be Something Else (terms mistranslated, redefined, or shuffled into metaphor) “Their longing carries them beyond the confines of their dictionary.”