AI
Moravec Inverted: 500 Million Years of Pretraining
The things that feel effortless, like seeing, walking, and knowing who's in the room, feel easy because they're the most optimized things you have. Evolution was the pretraining, and the machines inherited the wrong five thousand years of it.
In 1988, a Carnegie Mellon roboticist named Hans Moravec wrote down something that still sounds backward. He said it’s comparatively easy to get a computer to perform like an adult on an intelligence test or at checkers, and difficult or impossible to give it the perception and mobility of a one-year-old. That’s Moravec’s paradox, and almost forty years of AI have proven him right.
The things we call hard, like chess, calculus, or proving a theorem, turned out to be cheap for a machine. The things we take for granted, like recognizing a face or catching a ball nobody warned you was coming, turned out to be brutal. A four-year-old does the second kind without thinking and can’t do the first kind at all. Meanwhile we’re still fighting to get a robot to pick up a cup it’s never seen before.
How hard something feels to you is inversely related to how long your system has been optimized to do it. Seeing feels like nothing because it’s the most optimized thing you have. Math feels like work because it’s the newest layer, bolted on last.
If you work with these models, the analogy is pretty direct. Evolution is pretraining. The genome is the checkpoint, and your lifetime is the fine-tuning on top of it. Vision, balance, object permanence, knowing who’s in the room and where their hands are: all of that got something like five hundred million years of optimization, on the most expensive cluster ever built, the one where the models that ran badly died. Symbol manipulation got maybe five thousand years. Elizabeth Spelke’s work at Harvard shows that babies show up already knowing a lot of physics. They know objects are solid and don’t vanish when a hand covers them, long before they have a word for any of it. That knowledge came with the checkpoint.
That’s why AI is where it is. We trained it on the one corpus we bothered to write down, five thousand years of symbols, and it went superhuman at the part that’s newest for us. It never got the five hundred million years, because nobody ever wrote the motor cortex down. So it’s brilliant at the layer we find hard and clumsy at the layer we find trivial. People call that a paradox. To me it reads more like a training log.
The tempting move is to skip all that training and just copy the biology. Read the weights right out of a brain. But you can’t. We’ve had the complete wiring diagram of C. elegans, all three hundred and two neurons, for decades, and we still can’t run it forward and get the worm back. Having the wiring isn’t the same as having the program. What might carry over is the learning rule, the process that grew those weights in the first place.
This also kills the comforting story that an AI is like a baby learning from scratch, an early mind making the same climb we did. It’s the opposite of a baby. A baby has a full physical prior and an empty symbolic slate: everything about bodies and space is built in, and it hasn’t said a word yet. An LLM is the reverse. It has an empty physical prior and a full symbolic slate. It’s fluent in everything we ever wrote and can’t feel the floor. They’re mirror images. So the next time a machine fumbles something your toddler does easily, don’t read it as stupid. Look at what it was trained on.