Curators in the Library of Babel
In 1941, Jorge Luis Borges imagined the universe as an immense honeycomb of hexagonal galleries stocked with every possible 410-page book. Using a fixed alphabet of twenty-five basic characters, the Library of Babel holds every mathematical permutation of letters: every scientific discovery, every historical record, every future prophecy, every coherent translation, and vast, suffocating oceans of meaningless gibberish.
In such a universe, the concept of authorship collapses entirely. No volume on the shelves was composed by a human mind; it simply exists as one point in an exhaustive combinatorial space. For every faithful catalog of the library’s contents, there exist millions of nearly identical catalogs that differ by a single false line, an altered date, or an invented authority. The inhabitants of the library quickly discover that possessing a book is worthless. When a system contains every possible statement, the mere existence of a text offers zero evidence that its claims are true. The scarce, exhausting human task is no longer generation, but discrimination: determining which page accurately reflects reality outside the gallery walls.
The certitude that everything has been written negates us or turns us into phantoms.
— Jorge Luis Borges, The Library of Babel
A modern large language model turns that thought experiment into operational software. The model does not store an encyclopedia of facts or a filing cabinet of pre-written documents. Instead, training across trillions of tokens shapes billions of numerical weights into a vast probabilistic map of human language. When you press send, the model samples from those weights to construct a path through token space, assembling any one of millions of candidate completions in seconds at negligible computational cost.
Like the volumes in Borges’ hexagons, these generated passages arrive without innate verification. Because the model predicts patterns rather than verifying facts, a fluent confabulation and an accurate technical derivation are produced by the exact same arithmetic, rendered in the exact same authoritative voice. The machine makes plausible text infinitely cheap to reach, but provides no internal guarantee of truth.
This fundamentally restructures the economics of technical and creative work. When producing a competent first draft of a memo, a function, or a research summary costs fractions of a cent and takes two seconds, the draft itself ceases to be the scarce or valuable asset. Value concentrates entirely at the two boundaries surrounding the generation loop:
- The specification (upstream): Defining the exact system instructions, operational constraints, and verified reference documents that direct the model toward the right region of probability space before a single token is generated.
- The judgment (downstream): Auditing candidate completions against external ground truth, validating claims with deterministic tests, and ruthlessly rejecting defective outputs.
When generation is cheap, the engineer’s primary craft is no longer drafting text—it is designing the specification and enforcing the verdict.
This site’s own tooling runs on that split, and earlier posts show it at work: the image agent renders nothing until a JSON contract is approved (the spec before the image), and the critic loop pays a second model to reject drafts before a human reads them. The trace below is smaller and stranger. Three stylists share identical knowledge, so the complete answer already exists in three registers before you act; the one human contribution left in the turn is the choice.
ownership and the discipline of selection
When generating text costs nothing, the text itself proves nothing. A fluent paragraph no longer signals effort, authority, or truth. The real signal shifts entirely to ownership: who prompted the machine, what verified sources grounded it, and who takes responsibility for the result. In reliable systems, that chain of custody is the audit trail that makes an answer trustworthy.
Because raw generation is cheap, the engineer’s primary job moves from drafting prose to exercising selection. You train that judgment through a strict, three-step loop:
- Write the criteria first. Define what the output must accomplish and how it will be tested before generating a single token.
- Reject candidates in writing. Generate multiple runs, and explicitly record why the rejected drafts failed. Naming the defect is what turns vague intuition into enforceable system rules.
- Verify against external ground truth. Never ask a model to validate its own factual claims in isolation. A system trained on statistical patterns can invent a plausible justification for anything; check the survivor against an independent source.
Working with language models does not replace human reasoning—it moves your focus upstream into clear intent and downstream into rigorous verification.
To apply this discipline today, write your acceptance criteria before prompting, generate three candidate drafts, and document why you reject two. The introductory module covers the prediction engine underneath, and the glossary names the specific failure modes you will learn to catch.