Explore

The Age of Inference

Around 1800, a man by the name of Henry Maudslay was experimenting with ways to manufacture screws that would have smooth and consistent threads. Maudslay was the ideal person to notice this need, he had practically grown up in the tool shop, starting his career around 12 years old he worked his way from carpenter to blacksmith to machinist. By the time he turned 28 he had learned first-hand the value of standardization in manufacturing.

More than anything, Maudslay was fascinated with fasteners.

He obsessed with finding a method to manufacture the perfect screw, one with fine threads separated by perfectly even gaps that could be consistently reproduced every time. However, after many many trials even his best attempts still had threads of varying widths, which of course would jam when you would try to fasten the nut.

He was able to narrow down the problem with his screws to the surface of the plate along which his lathe traveled. Bumps on that surface were ultimately to blame for the irregular threads which caused the screws to bind. Maudslay diligently worked to try to find a way to remove the topography of those plates. He cast about for a general method by which he could detect and remove the irregularities, but so far all he had found was a highly laborious and manual process of marking the metal plate with oil and red chalk, and then grinding away at the raised parts of the plate, and then marking it again. For a time it seemed this highly manual process was the best that one could hope for, until one day he stumbled on a method so powerful and simple that it’s still used to this day.

What he didn't know was that rubbing these stones together would summon the djinn of industrialization and all its attendant wealth and wars. What he would discover on the other side of those stones was the garden in which the automobile and airplane and atomic bomb and automation would grow.

What Maudslay had discovered was flat.

That's it.

Called "surface plates", what's unique about these stones is that with the application of Maudslay’s process their surface could be made to describe an arbitrarily flat plane. That perfectly flat plane proved useful for building his lathe, which required very careful alignment. And the lathe was useful for turning out screws and other tools, and over decades those components would come to form the backbone of machinery that enabled the delicate engines that started the industrial revolution.

That flat surface was the first revolution of the crank that sparked the machine called industry.

The Industrial Age is now long behind us and we’re on the cusp of a new era, but perhaps the lessons of the origin of that revolution can still guide us. After all, we talk about the problem we face as literally an "alignment problem". An extinction threat built on the risk of cross-threading with the machines of our age — that our algorithms will unseat us (and then eat us). Once again we find ourselves trying to tease a smooth surface into the stones; but now instead of iron atoms it's silicon infons we need to align.

So how did our ancestors build something so affine? With what grit or trig or trick? As you'll see, it turns on all three.

To see Maudslay’s method in action, I want you to place your hands together like you're going to pray.

Now, vary the shape of one of your hands. Curl your palm inward, or flex individual fingers, but all the while try to maintain your entire right hand in contact with your left. Congratulations, you look like an idiot. But even more importantly, you've likely noticed that there are many different geometries that the surface of your touching hands can trace.

Now try turning your wrists so your hands rotate relative to one another. Each time you press them together try to maintain the old shape, only yielding where there is pressure from the other hand. By this point, you’ve likely started to converge on a kind of curved shape, where one of your hands is concave and the other convex.

Now go find a friend and ask them for a hand. Instruct them to hold their hand rigidly, facing toward you, and now press yours against theirs. As you press them together, tell them to maintain their hand's original shape as best they can. Rotate your hands relative to one another, as before. Once you’ve reached a new equilibrium, switch hands, pressing your second hand with its deform against their newly reshaped hand. Now press your own two hands together again. Now repeat the cycle, pressing your first hand against theirs, then the other hand, then together.

What you should notice after playing this game for many rounds is that you have converged on the exact same shape as every other pair that might play this game. Your hands are flat, as if in prayer.

Alone you're unconstrained, the possibilities are infinite. To achieve a perfect plane you needed players. Plural.

It turns out this prerequisite of plurality has an attendant proof. The only stable, mutually conjugate surface shape is a plane. This means that sufficient friction between sufficient stones is sufficient to align. In three dimensions, you need at least three surfaces. In four you need four, and so on.

This is what Maudslay discovered when experimenting with standardized screws. And it's this insight that sparked the Industrial Revolution. And perhaps it's this same insight that will power the Inference Revolution.

We can summarize this by noting the three things required for alignment:

abrasion

variation

pluralism

Abrasion is quite obvious. It means that where there is misalignment it generates some shearing force, some friction, some compromise. As we know from physics, friction creates a net loss: dissipated and wasted heat. That cost is important, it’s information that what’s being done doesn’t yet work.

Variation describes the form of the thing. In Maudslay’s case, he achieved this with the various angles that he rotated each piece to. If he had stuck with only one angle without rotating, or only one axis of abrasion, the effect wouldn’t have worked. If you hadn’t rotated your hands the effect wouldn’t have worked (go ahead, try it).

Plurality implies that it can’t be done alone. When Maudslay tried to etch a flat plane in a single stone it was enormously laborious work. By using multiple they aligned themselves.

This criteria for alignment shouldn’t in any way surprise us, we’re veritably surrounded by examples of these principles in the systems we live in.

Evolution

abrasion - survival

variation - mutation

pluralism - niches

Markets

abrasion - competition

variation - offerings

pluralism - different markets

Governments

abrasion - elections

variation - ideas

pluralism - communities

So then, concretely what does this recommend for aligning other inferents, like an artificial intelligence?

Inferents

abrasion - ?

variation - ?

pluralism - ?

I guess that’s what we get to find out together.

Want to print your doc?
This is not the way.

Try clicking the ⋯ next to your doc name or using a keyboard shortcut (

CtrlP

) instead.