The oldest distance

I'm spending my evenings building an AI agent that's supposed to understand one person — me — well enough to actually help. Said out loud, it's a strange project. To explain why I think it's worth the evenings, I have to go backwards. All the way backwards. So this page runs in reverse: it starts now and ends fifty-one thousand years ago.

Start with a fact I keep retelling. For years, AI models mostly got bigger and stayed useless — impressive the way a dictionary is impressive. Then in 2022, researchers trained one to follow what people actually meant, using ordinary human feedback, and human judges preferred it to a model a hundred times its size. Everything you'd call an assistant today descends from that result. Read it again: the field spent years making AI smarter, and the breakthrough was making it listen. The machines became useful the day they started to understand us.

The warning is older than the fix. In 1960, Norbert Wiener — the mathematician who invented cybernetics, the study of how machines steer themselves — had already put the whole problem in one sentence:

The field calls this alignment now, and it keeps some very serious people up at night. Strip off the vocabulary and it's an old, familiar problem wearing a lab coat: getting another mind to understand what you actually mean.

In 1946, one year out of the camps, Viktor Frankl set down what he had watched people survive on:

I believe the everyday version of that: we will accept nearly any feeling — despair included — as long as the people we love are thriving. Proud creatures, noble and warm at the core, and mostly too embarrassed to say so.

Writers had the diagnosis long before the engineers. In 1910, E.M. Forster set two words above an entire novel about people failing to reach each other:

Still the shortest design spec ever written. We are all still failing it.

Rewind further and the machines fall away, but the project doesn't. The phonograph, 1877: a song could exist without a single musician in the room. The photograph, 1826: a likeness that owed nothing to ten years of a painter's training. Machines for copying the world — each one made it cheaper to speak.

In 1677, Spinoza published his method for looking at human beings:

Striven. He doesn't claim he managed it. Three and a half centuries later, with all our machinery, neither do I.

The printing press, 1450, is where the pattern bottoms out: one person's words in front of thousands, for the price of a pull on a lever. Five centuries of tools after it, the accounting is lopsided — every invention made it cheaper to speak, and cheaper to make noise, more copies of the same thing, faster. Not one of them made it cheaper to be understood. That, apparently, we still have to do the old way.

Past the last tool, the trail keeps going. Flutes carved from the bone of a bird, in Germany, forty thousand years old — someone wanted a sound that wasn't a voice. Lions drawn in charcoal on a French cave wall. This is older than farming, older than writing, older than the wheel. Before we could reliably feed ourselves, we were already trying to put the inside of one head into another.

The rewind ends in a cave in Indonesia, fifty-one thousand years back, at the image at the top of this page. Someone pressed a hand to the rock and blew ochre over it. Here is the detail I can't leave alone: the paint never touches the hand. What survives is everything except the person — a hand-shaped gap in the pigment, sized so that another hand fits into it fifty-one millennia later, whoever it belongs to. Southern Africa keeps a proverb old enough to feel at home down there:

Every stop on this page — the press, the camera, the model that finally listened, the agent I'm trying to build — is one more attempt at living up to it. Rewind done. Back to work.