Self Simulation

Artificial Intelligence (AI), the idea that we might someday create something as smart or smarter than we are, is a breathtaking possibility. The emergence of such an AI might well be as important to the story of mankind as making contact with intelligent aliens. Some of our most popular books and movies have vividly depicted our hopes and fears surrounding the possibility that machines could achieve conscious independence from us.

Yet, for all of our ambition, effort, and angst, we’re scarcely closer to real AI then we were when the field was born a half century ago. A diversity of approaches have been tried with little success, and the one thing they all have in common is the assumption that intelligence can be designed. This is the principal conceit, and the biggest obstacle, in the field of Artificial Intelligence. AI researchers have dared to believe that they can understand intelligence well enough to literally describe how it works. By turning this description into a set of instructions a computer can follow, the computer becomes intelligent. That’s the idea, anyway, that intelligence can be reduced to a logical framework.

Oxymoron: Designed Autonomy

But designed intelligence is bound to underwhelm, for the simple reason that to design is to cheat. When we design an AI, it exhibits our intelligence and pursues our goals. In other words, the intelligence and motivation is external to the entity that acts on it. In contrast, true intelligence emerges intrinsically, of its own accord. It is self-motivating, because it determines its own goals.

As an example, today’s chess-playing programs, which are entirely designed, have no choice about whether to play chess or not. They have no autonomy, because they are not free to determine their own goals. Chess playing software is motivated from the outside. It is compelled by its user.

A chess playing AI that we might consider truly intelligent would be self-motivated. Its chess-playing ability would emerge from within, and not be specified (designed) by us. This means, therefore, that it would have to learn the rules on its own, just like we do. It would have to practice and hone its strategies. But most importantly, it would have to want to do all these things! The only way we can consider the AI to be autonomous is if it can choose whether to play or not.

Current AI efforts are far more sophisticated than today’s chess programs. Many of them can be said to “learn” in the sense that their creators have bestowed on them a way to correct their own behaviors. Other AIs can generate novel behaviors because their creators have plugged in an “evolutionary algorithm” that allows the AI to try things not specified directly by their creators. These are indeed important advances, and very clever.

But these advances were all designed. That’s still a problem because when you design, you are what causes the AI to operate, even if what you’re telling it to do is very abstract. Even if a designed approach succeeded wildly at being intelligent, this issue of autonomy would always be a matter of debate.

The Prison of Design

There’s also the issue of designing the AI into a corner. The question here is, if my AI design allows it to try new behaviors that I didn’t specify, and it can improve these behaviors over time, isn’t that intelligent? It is, as long as the AI remains inside the domain it was designed for. The completely-specified chess program is only competent in the extremely narrow domain of chess. It is useless at checkers.

A more sophisticated (hypothetical) AI might have the ability to “learn” to play more than one game, by trying strategies we don’t specify until it happens upon something workable, improving all the while. This sophistication widens the domain in which the AI can be competent, but the domain is still fairly narrow: it’s still only competent at board games, for example. Our design will always be limiting, because even if we don’t specify such an AI’s strategies, we specify the context in which it operates - a context it cannot escape from.

A New Approach Emerges

If a designed approach is doomed to never achieve general intelligence, what is the alternative? How do we build something without designing it? The answer is that we must set up an environment in which an AI emerges.

Emergence is the name for a state of affairs in which some phenomenon cannot be strictly described by breaking down the behavior of its individual parts, but only by examining the large-scale organization of those parts. The classic example of this is water. Water feels wet, it has surface tension, and it floats as a solid, and yet none of these properties can be predicted by knowing everything there is to know about a single water molecule. It’s the organization of large numbers of water molecules that gives rise to the properties of water as we know them.

The universe abounds with emergence. A few examples:

• Water emerges from H2O molecules. Clouds emerge from water.
• Termite and ant colonies emerge from the behavior of large numbers of those insects.
• Flocks of birds emerge from lots of individual birds.
• Brains emerge from a large number of neurons.
• A single cell emerges from the interactions of a large number of molecules.
• Traffic emerges from the collective behavior of individual vehicles

In any instance of emergence, you have a micro-level description and a macro-level description. In the case of water, the micro-level description is at the molecular level, and the water itself (a raindrop, for example) is the macro-level. What makes emergence so interesting is that the rules of interaction at the micro level cannot be used to predict the rules of interaction at the macro level. Even if we understood perfectly well how a single driver would behave in any conceivable scenario, it would be impossible to predict the behavior of traffic as a whole based solely on that. That’s because traffic is about the large-scale dynamics of collectives of drivers, and the rules that govern the behavior of traffic as a whole are of a different order than the rules that govern the behavior of individual drivers.

This is why it seems so mind-boggling (and frustrating) when we wind up in a traffic jam that forms for no discernible reason. We are not able to see the big picture. Seen from above, however, sporadic traffic jams can look like the transitions between liquid and gas (see this article, for example). The rules of those “phase transitions” emerge at the macro level. There is a logic of a different order at play, not comprehensible at the lower level.

I call this causal orthogonality. It’s not that the cause/effect relationships at the micro level have no effect on the macro level (that would be a violation of basic physics). Rather, those micro level causes and effects (e.g. the way people drive) have such a tiny influence at the macro level, that the emergent phenomena (the traffic jam) is better described in terms of high-level causes and effects that represent the statistical behaviors of large numbers of the lower level phenomena (description of traffic in terms of “liquid/gas transitions”, which is based largely on traffic density). The two levels are practically independent from one another.

Breathing Artificial Life into AI

The key to autonomy lies in this ubiquitous principle of emergence. The reason designed AIs are not autonomous, as discussed above, is because our designs specify the rules of cause and effect that govern the AI’s state and behavior. In an AI that emerges from its environment, however, the rules of cause and effect emerge with it. We would not specify those rules, and we would not be able to predict in advance what an emergent AI will do. As a result, we would truly have the potential to create autonomous agents.

The basic idea is to design an environment in which large numbers of small ‘bits’ interact according to rules of cause and effect that we specify. The hard part is tailoring our design to support the emergence of some kind of persistent entity (like a cell from a soup of molecules). Once we have this in place, we would need to demonstrate that the complexity of the system increases on its own, because only then would the new “creatures” have hope for getting more complex (and potentially, smarter).

This is an Artificial Life (A-Life) experiment. We’re basically figuring out how simple, persistent entities can emerge on their own and reproduce, getting more complex all the while. This is analogous to modeling bacteria. We probably wouldn’t consider bacteria to be intelligent, not in the AI sense, but this is where we need to start, because the most important consideration is that our creations have autonomy, or the freedom to determine their own goals and behaviors.

Now imagine for the moment that we have succeeded at this - we’ve created “artificial bacteria”. We haven’t specified them, or how they behave, but there they are, and getting more complex over time. The direction we need to go in to make this relevant to AI is that we’d like these things to eventually get smart. However they do it, we don’t care too much, because we’re not designing them.

To increase their intelligence, over time, we make the environment increasingly more difficult (through competition, scarce resources, etc.), so that the “bacteria” are forced to get smarter and smarter, or they die. We could certainly make some specific changes to the environment in the hopes of evolving specific advances. Perhaps at some point we could encourage the evolution of a strategy of our bacteria clumping together, and specializing. We basically want to evolve our simulation through higher and higher levels of complexity and intelligence.

It would certainly be a long road until human level intelligence could be achieved (at least 50-100 years, I would think). But the wonderful thing about this framework is that you’d have all these incremental milestones, as beings of increasing complexity and capability emerged. We would have a real measure of progress, instead of the recurrent cycle of hype and disappointment we have today.

A New Way Forward

This approach neatly dispenses with most of the open problems in AI, which are really problems associated with designing intelligence. For instance, traditional AI researchers grapple with how their AIs should represent knowledge internally, and philosophically with how their AIs come to know what that knowledge means (aka the symbol-grounding problem). An emergent AI, however, structures knowledge however it wants to. The meaning it makes of that knowledge is directly related to how that knowledge helps it survive, reproduce, or achieve other goals it forms for itself. Many AI problems relate to how an AI can deal with logic or natural language, but this is like worrying about how to make space suits for a walk on the moon when you haven’t even invented rockets yet.

AI has seen more hype, hubris, and failed expectation than almost any other scientific or rational endeavor. Of course, most fields of scientific inquiry deal with phenomena that can be dissected and analyzed. Yet, intelligence and consciousness are no such objects, so it’s ironic that the vast majority of research in AI to date has proceeded on the basis that intelligence can be snared in a net of logic.

A promising avenue of exploration thus opens up when we hold forth that intelligence and life emerge. But accepting this requires the tacit acknowledgement that the study of Artificial Intelligence cannot proceed until we have solved the basic riddle of Artificial Life.