Why SIMMBALL:from simms to symbols

Imagine a table. It is very large, perfectly flat, and has no friction at all. On it, countless tiny magnetic marbles roll, collide and bounce off the walls. Nothing ever slows down. This is the careenium, a thought experiment Douglas Hofstadter describes in I Am a Strange Loop.11Douglas R. Hofstadter, I Am a Strange Loop (New York: Basic Books, 2007). Our account of the careenium is a paraphrase and simplifies a great deal. The book's version is richer and worth reading in full. The marbles are called simms, short for "small interacting marbles."

Watch for long enough and what you see begins to change. The simms are magnetic, so now and then a few of them stick together. Clusters form, grow, drift, run into other clusters, and sometimes hold together for a long time. Hofstadter calls these large clumps simmballs. The pun is on "symbol," and it is not an accident.22In this essay, simm and simmball in lower case are Hofstadter's marbles. SIMMBALL in capitals is us.

A symbol happens one level up

Nothing new has been added to the table. Every simmball is only simms, and every simm obeys the same dull physics it obeyed before. With unlimited patience you could describe the whole table marble by marble, and the description would be complete. It would also be almost useless.

At some point the better question stops being "where is simm number 40,312 going?" and becomes "what is that big simmball doing?" The clusters push one another around. One drifts into another and changes its course. An account in terms of simmballs is shorter, predicts more, and is closer to what an observer actually wants to know. That is the sense in which a new level has appeared. Not a new substance, but a new level of description that earns its keep.

Hofstadter takes one more step, and it is the step that matters most to us. Once the simmball level exists, it becomes reasonable to say that simmballs cause things: that a large cluster sweeping across the table is why a particular simm ends up where it does. This is what he means by downward causality. The small parts still do all of the pushing. But the best explanation of the pushing lives upstairs.33"Downward causality" is the phrase Hofstadter uses; related ideas are discussed elsewhere as "downward causation." Whether higher levels really cause anything, or whether this is only a convenient way of talking about lower-level events, is a long-running argument in the philosophy of mind. We use the phrase in the weaker sense here and expect to return to the stronger one.

The careenium is meant, loosely, as a picture of a brain. Neurons are the simms. Thoughts, concepts and the sense of being an "I" are the simmballs. We do not have to decide whether the picture is complete to find it useful. Its central claim is modest and, we think, hard to argue with. Meaning is not stored in any single part. A symbol happens one level up, out of many small interactions, and once it happens it has a say in what the parts do next.

Why name a company after a clump of marbles

We build small things. A tool that does one job. A web app with a narrow purpose. An experiment that may not survive the month. Each one is a simm: simple enough to understand completely, small enough to replace.

What we care about is what they become together. Small tools used side by side start to share habits and a way of seeing. Some of them will clump. Most will not, and that is fine. The name is a reminder to judge the work at both levels. Is each piece honest and well made? And is anything larger beginning to form?

It is also a reminder about the kind of AI we want to build. The word we keep using is "human-ish," and we chose it with care. We are not claiming that our software is conscious, or will become conscious, or that a large enough pile of parameters eventually wakes up. We do not know what that would mean, and we are wary of anyone who answers too confidently.

What we mean is narrower and, we hope, easier to check. A human-ish machine can surprise the person using it, in a way that is useful rather than merely random. It can notice that two ideas from distant corners belong together. It can wonder, in the working sense of asking a question nobody handed it. And it can notice its own patterns: see that it keeps reaching for the same move, and try another.

Arthur Koestler had a name for the second of these. In The Act of Creation he called it bisociation: the moment when two frames of reference that normally never meet are suddenly held together, and a joke, a discovery or a work of art falls out.44Arthur Koestler, The Act of Creation (London: Hutchinson, 1964). Koestler applies bisociation to humor, scientific discovery and art alike. A one-line summary does not do the book justice. Koestler's idea is a cousin of Hofstadter's, not a twin. Both point the same way, though. Creativity has less to do with one brilliant part than with what happens between parts, and with a level that can look back at itself.

That last phrase is where the strange loop comes in. Roughly, a strange loop in Hofstadter's sense is a path through levels that unexpectedly lands back where it began, and his book argues that an "I" is a loop of this kind. A system that can represent its own patterns, and be changed by that representation, is doing more than clumping. We do not expect to build that in one step. If we build it at all, we expect it to form the way simmballs form: from small parts, interacting, over time.

Two levels of the same work

This is the first essay on the site, so a word about what else is here. There are two kinds of things. There are products: services and web apps we build and run ourselves. And there are essays: notes from the places where AI meets IT, philosophy, aesthetics, the humanities and the work of starting companies.

It would be easy to treat these as separate departments, one that ships and one that talks. We think of them as two levels of the same work. The products are the simms: concrete, testable, occasionally broken. The essays try to describe, in words, what is forming one level up while it forms. Sometimes an essay will explain a product. More often, we hope, an essay will change what we build next. That is downward causality on a very small budget.

So expect some unevenness. Some products will be tiny. Some essays will turn out to be wrong, and we will say so when we find out. The site is set like a working paper because that is what it is: a document still under revision, with its figures and footnotes in plain view.

Back to the table

Go back to the table for a moment. Billions of collisions, none of them meaningful. Then, without anyone deciding it, a cluster holds together long enough to be worth naming. Nobody put it there. It happened one level up.

That is the whole bet, stated as plainly as we can. Make the small parts carefully. Let them interact. Watch for the moment a pattern starts to push back on the parts that made it.

From simms to symbols.