An AI researcher at Microsoft recently published a paper describing a neural network implemented and trained on an unusual substrate: goats and terrain in the video game Age of Empires II. The author’s point is that anthropomorphic attributes should not be so lightly awarded to AI. It’s too easy to look at the functional power of modern LLM-based systems and the mystery of the linear algebra under the hood and treat it with undeserved reverence. And de Wynter’s trick of showing that goats in video games can produce the same results serves nicely to break the illusion.
But, perhaps the most fundamental property of complex systems—to be honest, the only one I’ve internalized, as I am far from an expert in this area—is emergence, the idea that novel phenomena can exist in a complex combination of elements which do not present the phenomena in any individual subparts or many smaller combinations. It’s a fairly intuitive property if you drill down into small enough subparts. We humans have many properties that a different collection of biological cells would not possess.
In the same way, LLMs are immensely capable task solvers. But if you drill them down into granular enough units, they’re all 0s and 1s, and basic computer hardware gates processing sequences of these. Those bits and gates cannot solve problems themselves, and it’s not clear exactly why combinations of them can sound a lot like a human being, enough to pass the famous “Turing Test” and in some circumstances be virtually indistinguishable from a human in natural language communications.
The Turing Test is named after Alan Turing, who is also the namesake for the Turing Machine, which distills the theory of computing down into its most basic principles: an infinitely long tape that can store 0 or 1 symbols, and a set of rules about the current state of the machine that guide it in whether any 0 or 1 it encounters should be left alone or swapped. Alan Turing shows that such a machine can solve any computational problem of any complexity; the concept often serves as a computer scientist’s introduction to mind-bending concepts of infinity. Part of de Wynter’s argument is that the AoE II setup he uses is “Turing-complete”, or able to solve any problem that a Turing machine can solve, and thus any problem.
Dealing with mathematical infinities creates great theoretical arguments and ugly practical realities. Adam Becker’s book, More Everything Forever, is a criticism of certain worldviews that seem to require infinite progress to be plausible. Rapid, progressive growth is a hallmark of emerging technology, and it’s easy to get caught up in it, particularly when doomsayers call for the end only to have the perceived obstacles leapfrogged over. Where the real limits lie, whether it’s for AI computing power or space travel, are hard to assess with any certainty. But, nothing can deliver more of everything forever.
Like infinite progress, infinite regress also cannot continue forever. Every parent has played the “why” loop game with their children—for any assertion, the child can ask “why?”, and it is, in the moment, a logically valid question. The more intellectually serious the parent, the more rounds they will play the game. (If you haven’t gotten at least as far as explaining the sky is blue because of the wavelengths of different colors of light, you’re not really trying.)
The story of infinite regress, dating back centuries in various forms, is often told with infinite turtles. A letter written in 1599 tells of a Hindu parable that the world rests on seven elephants, whose feet in turn rest on the back of a tortoise, all supporting the world. In the 19th century, philosopher William James, who often referenced this parable, also told of an old wives’ tale—recorded in print in 1838—of the world as resting on top of a rock, itself on top of a rock, with “rocks all the way down”. Somehow, these fables have been conflated, and now the notion of “turtles all the way down” has become well ensconced into modern culture.
If we want to get anywhere, we have to start somewhere. And wherever we start, it will be an imperfect, and incomplete, picture. Any set of assumptions can be picked at and prodded until there’s nothing left. To spend time quibbling over the methodology and the assumptions inherent in an approach might lead to journal articles and academic tenure, but will not produce outcomes that influence and improve the world around us.
In my book I center abstractions as the foundational unit of our abilities as intelligent, sentient beings. It is our comfort in working with these abstract ideas as we form them in our minds, and not drilling down too far to break everything, that makes us able to change the world to suit our wills. We look at the turtle we’re standing on, we test to see if it’s sturdy, and from there we climb.
What has made modern AI so powerful is that we have given it the same ability. As I write in Chapter 5: A Step Forward for Abstractions, the transformer approach encodes a functional capacity to identify and manipulate abstractions of inputs in useful ways, one that resembles what we humans do. If the AI doubted its own model weights, then it wouldn’t be able to perform inference or, well, anything.
Of course, de Wynter is right that we are too quick to anthropomorphize, and that we shouldn’t make undue assumptions. But in the same way some philosophers seem to argue that consciousness cannot exist outside biological systems—a point of view I’ve written in response to separately, here—he seems to me to be looking too far down the infinite stack of turtles, and as a result his theory seems unlikely to change the minds of those set on seeing spirit in silicon. Instead we need to select a useful level of abstraction, a middle turtle in the stack of cognition, and look there for the functional capabilities of consciousness. I continue to believe that the capability of representing and manipulating abstractions, itself, is a central piece.