Meaning Becomes an Address

01Every address so far has been a name

Think about how you have found anything, ever. A library book sits where a cataloger decided it belongs; Melvil Dewey published his decimal system in 1876 so that every book in a library would have exactly one place. A file lives in the folder where someone dragged it. An email is found by the words its sender happened to use. Even a search engine, for most of its life, matched the words you typed against the words on the page.

Every one of these is a name. Someone other than you chose it, and to find the thing you have to reconstruct their choice. When you cannot, the thing is, for practical purposes, lost. It still exists. It is simply unaddressable by you.

Vannevar Bush saw the problem in 1945. In “As We May Think,” he complained that artificial systems of indexing are why we cannot get at the record, because the mind does not work that way. It works by association. Bush named the problem clearly. He did not foresee this solution. His imagined machine linked things together by hand, one trail at a time.

02Three kinds of address

There have really only been two ways to give a thing an address, and now there is a third.

Figure 01

Three ways to address a record

• The name

  Chosen by whoever filed it: a title, a folder, a label.

  To find it, you must guess what they chose.

  Good for people who remember.

• The fingerprint

  Computed from the exact bytes, as a hash. Identical content, identical address.

  Change one character and the address changes completely.

  Good for proving identity.

• The meaning

  Computed from what the content is like, as a direction in a space of thousands of dimensions.

  Similar content lands nearby.

  Good for finding.

Two sentencesThe seller shall leave the boat lift.
The seller shall not leave the boat lift.

Fingerprint (first 16 characters of SHA-256)3680c358ec25ace7
05c19db91ea8c2cd

Meaning addressClose neighbors.

The fingerprint is the opposite of the meaning address, and the contrast is the clearest way to understand both. A fingerprint is perfect at saying “this is exactly that.” It is useless for “find me something like this,” because it was designed so that nothing is like anything. The meaning address is built the other way around. It gives up exactness to get closeness.

03Where the meaning address comes from

The idea is older than the machines. In 1954 the linguist Zellig Harris argued that differences in meaning go along with differences in where words appear. In 1957 J. R. Firth, working independently, put it in one line that is still quoted in every course on the subject: “You shall know a word by the company it keeps!”

Gerard Salton turned that into engineering. His SMART retrieval system, begun in the 1960s, represented documents as points in a space of word counts, and his 1975 paper with Wong and Yang made the vector space the standard way to think about search. In 2013 word2vec made those spaces learned rather than counted. In 2021 a model called CLIP placed images and text in one shared space, so a photograph and a sentence describing it could sit side by side.

So “meaning” here has a precise and modest sense. It is the company a thing keeps. Two passages get nearby addresses when they are used alike, talked about alike, surrounded by the same kinds of words and situations. That turns out to be enough to find almost anything.

04Why nearby is cheap, and why it isn’t

This is where the geometry from this morning’s piece comes back. In thousands of dimensions, random directions are almost always nearly perpendicular. Unrelated things are, by default, far apart. So when two things point in nearly the same direction, that is not an accident of crowding. It is a signal.

But here the old curse returns too. Finding the exact nearest neighbor in high dimensions is expensive. A careful 1998 study found that the clever indexing trees that work in two or three dimensions do worse on average than simply checking everything once you pass roughly ten dimensions. What made meaning search cheap was not beating the curse. It was declining to fight it. Modern methods find a neighbor that is almost certainly among the nearest, in time that grows roughly with the logarithm of the collection in practice, by settling for approximately nearest instead of exactly nearest.

High dimensions are generous to anyone willing to accept “nearly.” They punish anyone who insists on “exactly.”

That is the same lesson as the landscape. It keeps turning up because it is the same geometry.

05The turn: the address moves from the filer to the thing

Here is what actually changes. For all of history, the address was a decision, and the decision belonged to whoever filed the thing. The meaning address is not a decision. It is computed from the thing, the same way for everyone. Nobody has to file it. Nobody has to name it well. Nobody has to remember what it was called.

That quietly retires a job humanity has done since the first archive: deciding where things go. It also retires a kind of power. Whoever named a thing controlled how it could be found. Under meaning addresses, the thing speaks for itself.

Now the danger. Look again at Figure 01. The two sentences about the boat lift say opposite things, and they are neighbors. Meaning addresses capture what a passage is like, and a sentence and its negation are very much alike. Published tests of retrieval systems find most of them handle negation about as well as chance. For finding the contract, that closeness is exactly what you want. For deciding what the contract says, it is exactly what you must not trust.

So every record will need two addresses. One for finding, which is near. One for trusting, which is exact. Find by meaning. Confirm by the original.

06What I expect to see

1. The folder dies as a job.

   People will stop naming and filing things. Folders will survive as views generated on request, “everything about the dock,” rather than places someone had to put things by hand.

2. Every record carries two addresses.

   A meaning address for finding and an exact one for trusting. The costly mistakes of the next few years will come from systems that use the first where they needed the second, and retrieve “shall” when the document said “shall not.”

3. The search box becomes a description box.

   Searches will stop being keywords and become sentences about the situation: “the one where the roofer said it was the flashing.” The skill of guessing the right word will quietly stop mattering.

4. Records gather by situation, not by app.

   The photo of the dock, the text about the dock and the contract clause about the dock will surface together, because they are about the same thing. Which program created each one will become a detail.

5. Being describable beats being named.

   A business found by what it clearly does will beat one known only by what it is called. The advantage moves from owning a name to being unmistakable about a meaning.

07Where I ran into this

This month I have been working on a transaction manager realtors can set up for themselves, with a local tool that washes names, prices and addresses out of contracts before any AI sees them. Working on it, something struck me. Take every name out of a contract and it is still findable. Its name address is gone. Its meaning address is intact. “The one where the sellers left the boat lift” still points straight at it.

That is useful, and it is a warning. The same fact that makes a washed contract findable means that washing out the names does not wash out what the deal was about. This morning I wrote that high dimensions give AI its strange abilities. This is the one of those abilities that will touch every drawer, inbox and filing cabinet you own.

08The ledger

Already true

Photo libraries already find pictures from a plain description of what is in them. Search by meaning is already the standard engine underneath AI assistants that read your documents. Images and text already share a single space, so a sentence can find a picture.

What has to happen

Meaning addresses have to be assigned when a thing is saved, not later by a separate tool, and they have to work across programs rather than inside each one. The exact address has to travel with every result, so the original is always one tap away.

Where I am probably wrong

“Meaning” is a generous word for what these systems capture. They know what a passage is like, not whether it is true, and they are weak at “not.” Approximate search sometimes misses the very thing you wanted and cannot tell you it missed. And a meaning address is not anonymous: research has shown text can be substantially reconstructed from its embedding. If people come to treat “found it” as “confirmed it,” the two-address discipline fails, and this becomes a story about expensive mistakes rather than easy finding.

09Back to Tuesday

You type: “the roofer who said it was the flashing, not the shingles.” It is the first result. You never learned what it was called. You never needed to.

You only had to know what it meant.

Background

• Melvil Dewey, A Classification and Subject Index for Cataloguing and Arranging the Books and Pamphlets of a Library, 1876.
• Vannevar Bush, “As We May Think,” The Atlantic, July 1945.
• Zellig S. Harris, “Distributional Structure,” Word 10, 1954.
• J. R. Firth, “A Synopsis of Linguistic Theory, 1930–1955,” in Studies in Linguistic Analysis, 1957, p. 11.
• G. Salton, A. Wong and C. S. Yang, “A vector space model for automatic indexing,” Communications of the ACM 18(11), 1975.
• Tomas Mikolov et al., “Efficient Estimation of Word Representations in Vector Space,” 2013.
• Alec Radford et al., “Learning Transferable Visual Models From Natural Language Supervision” (CLIP), 2021.
• Roger Weber, Hans-J. Schek and Stephen Blott, “A Quantitative Analysis and Performance Study for Similarity-Search Methods in High-Dimensional Spaces,” VLDB 1998.
• Yu. A. Malkov and D. A. Yashunin, “Efficient and Robust Approximate Nearest Neighbor Search Using Hierarchical Navigable Small World Graphs,” IEEE TPAMI 2020.
• Orion Weller, Dawn Lawrie and Benjamin Van Durme, “NevIR: Negation in Neural Information Retrieval,” EACL 2024.
• John X. Morris et al., “Text Embeddings Reveal (Almost) As Much As Text,” EMNLP 2023.
• John Rector, “Every Trap Is Near the Floor,” September 2026.
• John Rector, “Dimensions Before Parameters,” July 2026.

More at johnrector.me.