On machine intelligence, bodies and reality
This is in response to a question posed by Dave over at Cognitive Daily. Dave cited a quote from a book he was reading:
Meaning is rooted in agency (the ability to act and choose), and agency depends on embodiment. In fact, this is a hard-won lesson that the artificial intelligence community has finally begun to grasp after decades of frustration: Nothing truly intelligent is going to develop in a bodiless mainframe. In real life there is no such thing as disembodied consciousness.
Discussion was then opened up regarding the plausibility of the above statement. The resulting comments quickly turned into the classic debate of “what is intelligence?” To be fair, I have no good answer to this, but I don’t believe this matters. In fact, I believe trying to define intelligence is detrimental to understanding how intelligence came to be, and therefore, how to create it.
Intelligence is a nebulous concept we all can innately feel and understand yet so complex we can’t define it. There is not a single definition that fits all realms of intelligence. If your definition is involves self-awareness, then my thermostat is considered intelligent life because it is aware of its own state. If your definition revolves around prediction, Deep Thought was intelligent because it could statically predict your next move, as well as the next 100 moves. Every definition has something that breaks the definition.
So if we can’t define intelligence in general, what can we do?
We look at why humans have this strange thing we call intelligence. Forget trying to define it, look at why we have it. The answer is simple. Intelligence arose in human beings because it conferred an evolutionary advantage to our species. This was an evolved process that gave humans an edge over other mammals.
Higher level intelligence is clearly not needed for basic life. The world is filled with trees and algae and bacteria that exhibit no formal intelligence humans would relate to. Intelligence was not a happy accident that perpetuated through generations, like a figurative appendix for the brain. The fact that humans have advanced intelligence speaks volumes on its importance. We needed intelligence at some point in our evolutionary history to survive, which is why we retain it today.
And that’s the kicker. We needed intelligence to survive. Something about intelligence in our species increased our likelihood of transmitting more genetic material. In this sense, our intelligence is no different than our opposable thumbs. It is another tool at our disposal for replicating ourselves one more generation.
Get on with it, what’s this have to do with AI and physical bodies?
If you have stuck with me thus far, take the reason we have intelligence, survivability, and apply it to machine intelligence. We quickly arrive at our answer to the original question.
A machine mind with no body has no incentive to form intelligent life. There is no selective pressure, no reason to form elaborate strategies to survive and replicate. A body is required for intelligence.
But don’t get hung up on the semantics. Yes, machine intelligence will require a body. It will also require selective pressure so that intelligence will emerge. Intelligence is a survivability tactic and the result of evolutionary pressures. But people get too hung up on intelligence requiring a physical body like a cat or a human. A body is merely a set of inputs and outputs that allow a mind to create predictions and view patterns. Intelligence may arise because of evolution, but it functions because a body feeds it stimuli.
Because of this, a body could take the shape of digital inputs. Perhaps the body of an intelligent digital organism consists of speedometers, accelerometers, altimeters, barometers and radar. Or perhaps the inputs cars on a road and your muscles control the series of traffic lights in a metropolitan city. An intelligent organism could easily live inside a computer, its body being defined as the sum of inputs and outputs at its disposal. You or I could never touch it, but the reality faced by that organism is no less real than my reality is.
It will never know what a “ball” is, how can it be intelligent?
I believe a fundamental flaw in the machine intelligence movement lies in the Turing Test. The Turing Test makes us believe that the only intelligent creature is one that thinks like a human. I will concede a digital, intelligent organism will most likely never experience tossing a ball to another human, the tactile feel of it and the force required to gently loft it.
But this will not make it any less intelligent. Consider an individual blind from birth. This person will never experience vision. I may describe vision in great detail to this individual and they may be able to logically understand what I am saying, but they will never truly understand what vision is. They can’t, it is not part of their stimuli. Their body simply cannot process this information, therefore, their mind cannot. They live in a land of touch and sound and smell but no vision. Their reality is fundamentally different. Are they any less intelligent for not having this stimuli? No.
Similarly, a digitally intelligent organism may have a completely different set of inputs and therefore a completely different reality. Can we say they are not intelligent because of this? No. Just as they won’t be able to fully comprehend tossing a ball, I will never be able to understand what it truly means to have digital ultrasound input and a built in accelerometer.
So God created man in his own image
Human arrogance, I believe, is the cause of the failed machine intelligence movement. We have for too long attempted to craft intelligent life in our own image. Perhaps one day, when we have sufficient computing power and a robotic machine of such quality that it reproduces the same sensations as a human, perhaps then we can create intelligent life that can relate to humans. But in the mean time, I believe it is fully possible to create intelligent life that lives in a reality drastically different to our own. We have been barking up the wrong tree for far too long.
August 30th, 2007 at 9:00 am
Interesting post. Just one minor point.
I agree that the turing test is a stumbling block for many, both inside and outside the artificial intelligence field. But I believe that this is caused by a misunderstanding of the aim of the turing test. The point of the test is NOT that “only things that pass this test are intelligent,” but the weaker claim that “anything that CAN pass this test is intelligent.”
In other words, passing the test indicates intelligence, but failing the test indicates nothing. I.e., you can be intelligent and fail the test, but you cannot be unintelligent and pass it.
August 30th, 2007 at 12:30 pm
The weaker claim is easily defeated though. I would bring up the example of Alice and other chatterbots. At the moment, they cannot endure a turing test longer than a few minutes. But it is conceivable that more advanced bots, programmed with shortcuts to appear intelligent during conversation, will soon be able to pass the test. The reason for this I believe falls back on the idea that an intelligent mind must have a body of some type.
A chatterbot will never understand what a ball is, what colors are or the sensation of pain without experiencing a body. But they could quite easily talk about these concepts of which they have no first hand knowledge simply by exploiting the patterns and rules of our language. Are these bots intelligent because they can manipulate our language in a way that appears intelligent? The turing test says “yes, they are then intelligent because they can pass” while I would contend that they are anything but.
I think the real fault in the turing test lies in its reliance on language. Language is a means to describe the reality we belong in. An organism with machine intelligence resides in its own reality separate from ours. While it may be capable of understanding the grammar of the English language, it will never be fully capable of understanding the underlying concepts. In this way, an intelligent organism could easily fail the turing test, while a simple program could pass with flying colors.
August 31st, 2007 at 2:06 pm
The fact that intelligence developed through natural selection, and that it gave a survival advantage to humans, doesn’t by itself imply that such selective pressure is necessary to develop intelligence.
Zach, I don’t agree that “chatterbots” are anywhere close to passing the Turing test. The issue to me is not whether it is possible to fool an observer, it’s a matter of competence. Can a program communicate in natural language with the full competence of a normal, intelligent human speaker? No existing program, certainly.
Yes, there are certainly philosophical issues about whether a disembodied program could ever really understand, but if there were a program that behaved as if it understood, that would be good enough for many purposes. The question to me is why anyone would care whether something “really” understands, if its lack of understanding is no impediment to its competence.
The reason that we worry that a student doesn’t “really” understand the material, but is only using “tricks” to get the answers is because we believe that someday the tricks will be insufficient. We believe that lack of understanding will eventually show up as a lack of competence. If there is no difference in competence between “real” understanding and “fake” understanding, then why care about the difference? Natural selection certainly doesn’t care about the distinction; if an animal has some strategy that is successful in the vast majority of cases, that’s good enough for evolution.
October 6th, 2007 at 12:14 am
Thank you for sharing!