Alien Intelligence · Part four
The Dough Pushes Back
What a body brings to the work before anybody puts it into words.
It’s still dark when the baker gets to the bench. There’s a note taped to the flour bin that says Work the dough until it feels right. Yesterday the apprentice asked her what right meant. She tried a few words, then put the apprentice’s hand on the dough and said, “Lift it a little. There. Feel it pulling back?”
That was a better answer than the one she’d been trying to give. It also points to a different kind of intelligence from the one I’ve been describing in this series. At the bread bench, the baker’s body can change what matters, her movement can produce the evidence, and her hands can learn a skill she still struggles to explain.
Before anyone asks
The baker planned to shape all the loaves before breakfast. Halfway through, she notices she’s hungry enough to keep looking at yesterday’s bread. She sets down the scraper and eats a slice. The order sheet still calls for the same loaves. Something inside the person doing the work has changed which job comes first.
Anthony Goldstone and colleagues showed food images to the same volunteers in a fasted state and after breakfast. Fasting made the higher-calorie foods gain more in reported appeal than the lower-calorie foods. The pictures were the same. The state of the viewer changed how they were valued. Read the fasting experiment.
Ask a chat AI to schedule the baker’s morning and it can account for breakfast if it knows when she last ate, or if she says she needs a break. It can even ask. But the pressure to eat originates with the baker. A robot can be built to divert itself to a charger when its battery runs low. The designer has to connect the battery reading to the decision about what to do next. The baker’s need enters the job before she has said a word about it.
Give the AI the order sheet alone and it may come back with a perfectly plausible plan that has her at the bench for four hours without a break. Give it the constraint that matters and it can find a gap between batches. Her body supplies part of the assignment.
A hand makes evidence
Take a photograph of the dough at rest. It shows a pale lump on a dark bench. Now lift the edge the way the baker did. How far does it stretch before it tears? Does it pull the rest of the lump along with it? Those facts become visible during the lift. The baker can also feel resistance before anyone watching sees much change.
Susan Lederman and Roberta Klatzky asked blindfolded participants to match objects by properties such as texture or hardness. People used different movements for different questions. They rubbed a surface to learn about roughness and pressed to learn how it yielded. The motion was part of how the person found out. Read the experiments on touch.
A vision model given a still image can draw on what similar dough usually looks like. To know how this piece responds to a tug, it needs something that records a tug. In one robotics experiment, researchers equipped a robot arm with a rolling pin and combined camera and tactile measurements to shape soft materials such as Play-Doh. That gives the machine something a chat window starts without. It can act on this piece and use the result to guide another movement. Read the robotic shaping experiment.
Over a kitchen counter I’d ask the AI which observation would help distinguish the possibilities. Then I’d take that observation. “When I lift the edge, it stretches about two inches and tears” gives us something the photograph never had.
The hands learn first
On the first morning the apprentice tries to lift the dough, it sticks to her fingers. The baker shows her how she loosens an edge. A week later the apprentice can feel a piece begin to stick and release it before it folds over on itself. If you ask what changed, she might say, “I’m lighter with it now.” That phrase leaves most of the improvement in her hands.
In a study by Suzanne Corkin, Henry Molaison, the patient known as H.M., improved at tracing a figure while watching his hand in a mirror despite profound difficulty remembering recent events. The performance changed even when he couldn’t recount the practice that produced the change. Knowing how can show up in a movement before it shows up in words. Read Corkin’s motor learning study.
With a language model the order can run the other way. It can describe why you’d change the angle of your fingers, and it can explain that change to a beginner, before it has ever controlled a hand in a kitchen. Those words can improve somebody else’s practice. The ability to produce them and the ability to fold a batch cleanly have to be established through different work.
If I want help teaching, I might ask the AI to turn the baker’s demonstration into a sequence the apprentice can follow. If I want to know whether the apprentice has learned, I’ll give her another piece of dough. The new piece might be wetter or colder. Her next movement tells me more than her ability to repeat the directions.
The next morning the baker sticks a fresh note to the bin. This one says, Lift an edge and see what comes with it. That’s a sentence the AI could have written. There’s a reason she put it beside the flour, where somebody can try it.