The Unknowing Database

Welcome. 2027 is about to enter our domain, and what do we have to show for it? We took the most powerful information machine ever built and used it to destroy the one thing that made information worth anything — the agreement about what it means. The knowledge database became an unknowing database.

How? Simple. We broke two words. Computer and protocol. Let me explain, and bear with me — the first half of this is boring on purpose.

What a computer is

A computer is a device that receives, stores, processes and distributes data. It runs programs. Programs contain algorithms. An algorithm is an exact sequence of instructions that does something to the data in the machine — processes it, alters it, moves it. Exact is the operative word.

computer (n.) — “one that computes; specifically: a programmable usually electronic device that can store, retrieve, and process data.” — Merriam-Webster

algorithm (n.) — “a specific step-by-step procedure for solving a problem or accomplishing some end.” — Merriam-Webster

Until about 2023 that is what we had. And as the owner of a financial startup I was absolutely sure of one thing: that machine would take my data from wherever it was sitting and process it without changes, without opinions, without perversions. Sum a column ten thousand times — get the same number ten thousand times. Which is more than I can say for any human in the company, myself included. That was the entire point of the computer. Not that it was smart. That it was exact.

What a protocol is

Computers that process data with algorithms sooner or later have to talk to each other, so they inevitably grew protocols. A protocol is a formal agreement about how two parties communicate — that it happens, in what order, and with what exact words.

protocol (n.) —
1a. “a system of rules that explain the correct conduct and procedures to be followed in formal situations”
1b. “a set of conventions governing the treatment and especially the formatting of data in an electronic communications system”
3. “an original draft, minute, or record of a document or transaction”
4a. “a preliminary memorandum often formulated and signed by diplomatic negotiators as a basis for a final convention or treaty”
— Merriam-Webster

Look at the order. The dictionary puts people before packets. 1a is how humans behave in formal situations; 1b, the TCP/IP sense, comes second. That ordering is not an accident and it’s the whole argument of this section.

Because the word is old. Greek prōtokollon — prōtos, “first,” plus kolla, “glue” — was literally the first sheet glued to the front of a scroll, the one that told you what the scroll contained. From there it became a rough draft of a treaty, then the official record of a transaction, then a diplomatic document signed by friendly powers, then (by 1896, in English) the formula of diplomatic etiquette, and by 1952 just “the conventions of proper conduct.” (Etymonline has the whole chain. It does not, as of this writing, mention computers at all. Good.)

So for roughly two thousand years a protocol was not a computer thing. It was a people thing. It told humans how to behave in a given situation — who speaks first, who signs where, what the record of the agreement looks like. The computing sense is the baby of the family, borrowed by analogy, and it kept the family trait: a protocol tells the computer how to behave in a given situation. Exactly. Or not at all.

Commands, and the books that used to hold them

Strip any protocol down and what’s left is commands. Strip any computer protocol down and what’s left is the commands one machine sends another to get data back. And the one thing everybody knows about those commands: they get followed to the letter or the protocol doesn’t work. There is no “roughly GET.” There is no “HELO, more or less.”

Humans had the same thing, and we called it a dictionary. Them books explained to human beings how to talk to each other. If you didn’t understand somebody — or didn’t understand how a thing worked — you opened one and found that for the last several hundred years the word dog had meant a furry animal with four paws, a biter installed at the front and a swinger installed at the back. We had an agreement about what a dog was and how a dog was supposed to work.

Now, somebody is already typing “but dictionaries never agreed either.” Correct. Webster and Oxford and Ozhegov have been disagreeing for a century. But they disagreed in print. In a numbered edition, with a year on the spine, saying the same thing to every single person who opened it. The disagreement was public and it was versioned. You could put two dictionaries on the table and argue. Hold that thought — it’s the hinge of this whole piece.

There are basic words you’ll have to understand by observing the world, but once you have a base of 1,000–3,000 words, you can describe practically anything. Take a dog, for example… Try describing one to a seven-year-old. Who cares that you have the full Latin — your poor kid is stuck enduring a lecture about the word dog. But a dog is the easy case. Worst case, you take the kid outside and point.

Where exactness kills

So take something harder. Something a life hangs on. The controls of an airplane.

It is utterly fatal not to know the exact definition of every control on the machine you’re flying — and the industry has buried people over exactly this, including, not so long ago, over a control nobody in the cockpit had a definition for. Yes, there are many models and they work differently. Fine. But if you don’t know how to trim, you’re giving everybody in the cabin an unpleasant jolt if you’re lucky, and if you’re unlucky, a one-way roller coaster to the end of their lives.

Knowing the exact definition of a word matters precisely as much as knowing the exact definition of a command in a protocol. Sometimes they’re the same thing.

Enter 2026, the LLM, and the so-called knowledge base

All knowledge is the same, but some knowledge is more samer than other.

Here’s the party trick. Open a couple of your LLMs — and don’t tell me you have only one — and ask each of them for the exact and full definition of the word set. Then look. You’ll get wildly different answers that match nothing, not even each other. Off they go in wild abandon. (Dictionaries, by the way, have fought over set for a century — it held the record for the most senses of any English word until run overtook it. But they fought in editions. With dates. Which is the point.)

So what are you left with? Different definitions of the same word. And if you and I have different definitions of set, please explain how we’re going to talk about a set. We can’t.

Take politics. Everyone’s throwing terms around, calling the other side this or that, and nobody knows the definition of the word they’re using to belittle the other side. And when they finally go to find out? Nobody opens a dictionary. Everybody pulls out an assistant — and every major assistant now ships with memory on by default, so it takes what’s in your profile and tints the text, pads the text, adjusts the text before handing it back. Three hundred million people, three hundred million private definitions of the words they’re calling each other. Not public. Not versioned. Not on the table.

So nobody knows what the information means. There is no protocol. There is no algorithm anymore, because you cannot have an algorithm in a world where the definition of every command changes with the viewport. You can still process enormous amounts of data — you just can’t say what the output means. You can’t have a reliable source of information, because you took a system that by design cannot produce an exact result and asked it to produce exact results. Then you act surprised when it doesn’t.

No, I’m not saying stop

Am I saying AI should be stopped forever and never used? Of course not; that’s stupid as hell. Sometimes you don’t need exact results. Sometimes you want to model something on noise. Sometimes the noise is the result you wanted to see.

But in any way, shape or form, the programmer’s job has changed, and not in the direction the bootcamps think. It is no longer to know a specific language and follow its ways. It is to walk the infrastructure and draw a line: this part needs set-in-stone algorithms, so write them formally, in a way that can still be changed later — and that part doesn’t need an algorithm at all and will, in fact, be handled better by a model. Drawing the line is the job. Everything else is typing.

So welcome to the world where you’ll be managing data flows. I suppose that’s what we call programming now. Fine. But do manage them. Don’t let the model creep into places a model should not exist. And verify its output against them books called dictionaries.

Dead weight

One more thing, and then I’ll let you go.

Pretty much everything you read online now is generated. Not the spammy corners — the bulk. Product descriptions, how-to guides, news summaries, reviews, the forum answer that reads a little too well. The dictionary in your browser tab is one product update away from being generated too. And I can already see the next step: the site that generates its content on demand depending on who’s visiting. Reads your profile, decides what you’d like the word to mean, serves it fresh with a confident tone and a citation that was also generated a second earlier. Two people open the same URL and read two different articles, and neither ever finds out. It’s coming. Don’t worry. Or do worry, actually — because there will be nothing to point at when you argue about what you both read.

So my biggest treasure now is my hard-copy books, and I have plenty. A book printed in 1975 can’t be regenerated overnight. It doesn’t know who’s reading it. It doesn’t check my profile before telling me what a word means. It says the same thing to me, to you, and to the seven-year-old — which is exactly the property we just lost everywhere else.

I even snatched up a pile of dictionaries from the seventies at a yard sale. The seller thought he was unloading dead weight. I think I bought the last reliable protocol I’ll ever own.

Those books are going to be quite expensive in five to ten years.