The Machine Has Old Taste

Nothing went wrong in that exchange. The answer was accurate, the places are good, and the phone did exactly what it was built to do. It also told you something about every recommendation you will get for the rest of your life. The thing doing the recommending has old taste, and it did not choose to.

Survival is evidence

In 1964 a writer named Albert Goldman put a piece of Broadway folk wisdom into print. The comedians who gathered at Lindy’s, a delicatessen on Broadway, believed you could tell how long a career had left from how it had gone so far. Goldman called it Lindy’s Law. In 1982 the mathematician Benoit Mandelbrot turned the name into a precise rule: for some things, whatever has already lasted can expect, on average, to last about as long again. In 2012 Nassim Taleb stretched it to anything that does not wear out, and gave it the name most people use now, the Lindy effect.

The logic is simple once you see it. For a body, every year survived is a year used up. For a book, every year survived is a vote. Each generation that kept it in print found that it still did something people needed. Survival is not luck piling up. It is evidence piling up, and evidence is what predicts.

For most of history this was a pattern you could beat. A new restaurant could beat it with one good review in the paper. A new book could beat it when a friend pressed it into your hands. Word of mouth moved faster than the record, so the new could outrun the old in the only place that mattered: people’s heads.

The first machine trained only on survivors

Here is what changed. The machine that increasingly answers our questions learned what it knows from the written record, and the record is a survivor’s list.

Think about everything that has been written about a restaurant that opened in 1985. Forty years of reviews, menus, wedding announcements, neighborhood guides, arguments on forums, birthday posts, the founder’s obituary. Now think about the one that opened in April. Six months of anything. A model learns about each thing in proportion to how much text there is about it, and researchers have measured exactly that: a model’s grip on a fact tracks how often that fact shows up in what it read.

Put those two together and Lindy’s Law stops being a tendency out in the world. It becomes a property of the instrument. The machine does not merely reflect that old things last. Its confidence in a thing is roughly proportional to that thing’s age in print.

That loop is the part almost everyone misses. We were told AI would be a novelty engine, endlessly personal, always serving up the new thing. For anything that happens outside the conversation, it is the most conservative recommender ever built. Its taste is the average of decades. The new now pays a tax, and the tax is not charged on quality. It is charged on record.

What I expect

1. Old businesses will be recommended out of proportion to their quality.

   Having been around becomes the strongest ranking signal most owners never chose. New owners will say that AI does not know they exist, and they will be right.

2. Launches will be aimed at machines first.

   The first job of a new business, product or idea will be building a record dense enough to be believed: plain words, dated, consistent, public. The announcement will matter less than the archive.

3. Your taste will get older without your noticing.

   Ask for a book, a song, a recipe or a baby name and you will get the survivors. The canon gets a second life. Novelty will have to be asked for by name.

4. Trends will slow down.

   When the main recommender has a memory measured in decades, the fashion cycle stretches. Things stay in style longer because the machine keeps suggesting them.

5. Density will become a way to buy age.

   A year of rich, honest, consistent record can look to a machine like a decade of scattered mentions. The new will not catch up by waiting. It will catch up by writing.

I am living the fifth one. Charleston AI opened its doors in January. Walk past and it is a room full of people working. To a machine it is an infant, standing next to competitors with years of text behind them. So I do not wait for age. Every post on this blog ships with its full text in a form machines can read, and Charleston AI publishes a delivery every day: a dated, plain-language record of real work for a real kind of business. I used to think of that as content. I think of it now as the only way age can be bought. The other half of this idea is in The Fence That Worked: AI learns only from the record, so whatever leaves no record is invisible to it. And Proof of Person is about what writing still proves now that anyone can produce it.

How this could be wrong

Already true

Lindy’s rule has described books, technologies and institutions for as long as people have looked. Language models know common things better than rare ones, and that has been measured. You can test the rest tonight: ask your AI about the oldest restaurant in your town and the newest, and compare how much it knows about each.

What has to happen

People have to keep routing everyday choices through an AI answer instead of browsing for themselves. And for anything not looked up live, the models have to keep leaning on what they absorbed in training.

Where I am probably wrong

The machines are learning to look things up before they answer, and the same line of research that measured the gap found that lookup helps most with the least-known things. If live lookup becomes the default and treats a six-month-old bakery the same as a forty-year-old one of equal quality, the amplifier was never permanent and this piece is wrong. I will know by asking the same question every spring and watching whether the new places start to show up.

Next Saturday, ask your phone about the new place by name. It will know a little more than it did last week, because somebody wrote a sentence. That is all age has ever been to a machine: sentences.

Write it down. That is how anything gets old.