A Harness Has a Tribe

The one-page convention of LLM Wiki became a movement. Strip the hype and what is left is small, modest, and everything else is built for one tribe.

In April, Andrej Karpathy posted a gist. In about a thousand words it described a way to have a language model keep a personal wiki: put your sources in one folder, let the model write and maintain a second folder of interlinked markdown pages, and now and then ask it to check its own work. It was a good idea, plainly stated, and it came with an admirable disclaimer, which was that this is a pattern and not an implementation and the machinery is your problem. Within months it had thousands of stars, a Google specification, and a comment section that reads like a trade show floor. I used it in earnest, on the work I actually care about, which is long and entangled reading in the humanities, and every version I tried fell over. This is about why.

First let’s say what the thing is, because the noise has made it hard to see. LLM Wiki is a file format and three verbs. The format is markdown with a little structured metadata front matter at the top of each file. The verbs are ingest, query, and lint, and the gist suggests that even these are advisory, that everything in it is optional, and that your agent will fill in the rest. It is not novel and never pretends to be. It is markdown and wikilinks and a folder, the furniture of every note-taking system since the index card, plus one new player: a model will now do the filing you and I were always going to skip. That is the whole of it. It is modest and useful, and it is about as conceptually deep as a well-labelled filing cabinet, which is a compliment, because any good infrastructure is shallow and acknowledges it.

So why the commotion? Part of it is that the pain is real and old. Everyone who reads or intends to read all those open tabs, and year-old bookmarked pages, has tried to organize what they read and at some point has quit. We quit not just because the reading is hard but because the bookkeeping is: the cross-references rot, the summaries drift, nobody wants to update the index. That is exactly the drudgery a model performs without complaint, so the idea reads as a cure for an ailment people already have. Part of it, frankly, is the byline.

But the most instructive convert is Google, which in June announced the Open Knowledge Format, an open specification that, in its own words, “formalizes” the LLM-wiki pattern. Read the post and watch what the pattern quietly becomes: no longer a tool for a person reading books, but a way to describe BigQuery tables, metrics, and join paths so that agents can consume them, shipped with a reference implementation that plugs conveniently into Google’s own Knowledge Catalog. Google did not join a movement so much as notice a warm pool of free attention gathering around a markdown convention and lay a boat ramp from it to the cloud. The word “open” is working very hard in that announcement, as it usually is.

Here is the part the whole parade walks past. Because the format does nothing, everything that decides whether it works is the machinery it does not specify. Call it the harness because it is. And a harness is never neutral, which is an old lesson in the knowledge organization and classification: every scheme for ordering knowledge is a claim about what knowledge is, wearing the costume of a neutral container. Dewey’s decimal system presents itself as a universal map of all human thought, and turns out to have room for a hundred fine distinctions among Protestant sects and a single crowded shelf for everyone else ¹, because it is the mental furniture of one nineteenth-century American gentleman. Jorge Luis Borges made the joke eternal with an invented encyclopedia whose animals include, among other categories, ‘those that belong to the Emperor’, ‘embalmed ones’, and ‘those that at a distance resemble flies’ ². Every schema is somebody’s ³. It fits the world of its author and pinches everywhere else.

Now look at who is actually building the LLM-wiki harnesses. Read that comment thread top to bottom and you are listening to a single culture talk to itself: engineers, storing what engineers know. The knowledge these systems are tuned to hold is the specs and commit hash, the deprecated endpoints, the architecture decision, the compliance control, the state of a codebase between sessions. Their idea of a fact is a value that is either current or stale. Their idea of a contradiction is a defect to be located and closed. Their idea of maintenance is reconciliation toward one consistent truth. The most elaborate implementation I tried came with community-detection algorithms and a four-signal relevance graph, and underneath the finery it was still a beautifully engineered instrument for remembering engineering. Another author in the thread described a system that shreds every document into six-hundred-character fragments. For his facts, that is fine. For an argument, it is a wood chipper.

This is why they broke for me, and it was not just because I fed them too many files. Serious work in the humanities runs on precisely the things these harnesses are built to eliminate. A contradiction between two thinkers is not a defect in my book. It is frequently the object of study, the tension I am there to hold open, not to launder into a tidy consensus that describes nothing. A claim is not current or stale. Augustine is not superseded by a rebuttal filed last spring. The value of such a corpus lives in the connective tissue, the argument and the lineage and the disagreement and the reading, which is exactly the tissue a fact-store dices. And when a system whose deepest instinct is resolve toward one truth meets a body of work whose entire life is sustained, meaningful disagreement, it does not fail loudly. It does something worse. It hands you a clean, confident, immaculately cross-referenced synthesis that has quietly ironed the actual thought flat. A wiki that has achieved internal consistency across the humanities has usually earned it by amputating the parts that mattered.

The reflex here is to say the tooling is young and will mature. I do not think maturity is the axis. A finer carpentry of the same assumptions yields a better cabinet for facts, not a harness for thought. The implementations are not unfinished. They are finished, for someone else. They encode, faithfully and often well, an epistemology in which knowledge is a pile of facts converging on consistency, and for what I need that epistemology is not a limitation but a category error.

So, past the stars and the specification and the boat ramp to the cloud, the bottom line is this:

LLM Wiki is a good, honest, minor idea, a format and three verbs, that has been inflated into a movement it never asked to be, and the movement, for all its noise, is building for one tribe and calling it the world. The format and verbs are fine. The harness is everything, and the harness has a tribe, and it is not mine. Serious intellectual work will need its own: one that treats a disagreement as a structure to preserve rather than a bug to close, that models the interpretive lens as carefully as the fact, that can tell a stale value from a living argument. I don’t see anybody around that pool going build it, for the simple reason that you build a harness for the work you actually do. Which leaves the inconvenient conclusion that whoever needs such a thing will have to build it themselves, from their own epistemology up. That is hard work, and from my point of view a worthy cause. It is also the only part that was ever going to matter.

Footnotes