Borges' cartographers and the tacit skill of reading LM output

The fable, retold
Jorge Luis Borges' tiny, terrible parable about an empire whose cartographers built a map the size of the empire is on repeat in tech circles. It has been reported that a recent post circulating on Hacker News and a blog by Galsapir used that story to make a blunt point: maps work because they compress. When fidelity swallows compression, you lose the map’s purpose. Oof. That line lands hard in an age when models can vomit pages of plausible-sounding text on demand.
Models as maps — and the danger of empire-sized fidelity
Language models are, in practice, maps. They summarize probabilities, compress context, and surface what they think you want to see. But fidelity is seductive. Want every citation? Fine. Want exhaustive output? Here it comes — the map that covers the whole country. More detail can mean more noise, more hallucination, and harder-to-interpret signals. The result isn’t just verbosity; it’s a breakdown of usefulness. You asked for directions, and the system handed you the entire atlas.
The tacit skill: reading between machine lines
So there's a human skill emerging that’s as important as prompt engineering: the tacit art of reading model output. How do you tell confidence from flourish? Where are the hedges, the internal contradictions, the stylistic tells? Experienced users learn to spot them — short answers when the model is sure, hedged lists when it’s not, inconsistent facts that mean “verify this.” This isn't taught in a textbook. It's apprenticeship, pattern recognition, and a healthy skepticism rolled into one.
What now?
If Borges' cartographers are an allegory for overfitted fidelity, then our job is to keep the map useful. Better UI cues, calibration metrics, and training that privileges compression over exhaustive mimicry will help. And humans? We need to learn the craft of interpretation — to read the map, not worship the paper. Otherwise, we risk carrying around empires of text when all we wanted was a route.
Sources: galsapir.github.io, Hacker News
Comments