How to Build a Thinking Machine
A survey of the animal kingdom makes it clear that a great degree of intelligence can evolve in separate lineages in dramatically different body plans. Dolphins are intelligent cetaceans, elephants are intelligent pachyderms, crows are intelligent birds, octopuses are intelligent mollusks, anoles are intelligent lizards, hermit crabs are intelligent arthropods. Primates are all fairly intelligent, but these many other diverse examples show that intelligence is not specific to one body plan, and it has evolved more than once, which means it must not be the result of a specific algorithm present only in human bodies.
So what mechanisms do all these intelligent creatures have in common? They all have two things in common…complex nervous systems and sleep.
It is important to keep in mind what a nervous system is…it is an organ for coordinating the sensations and actions of a whole body. Why would we suppose then, that a thinking machine could be embodied only in a computer program, and that it would require no coordination of physical actions at all, but could just analyze data and become intelligent?
Intelligent creatures have sleep in common, and especially dreaming sleep. So why is it that research into artificial general intelligence has given so little attention to the problem of sleep? Perhaps because it is hard to study quantitatively…or because we have no way to break it into smaller sub-specialties. Sleep is not a process that is amenable to reductionist analysis.
Nevertheless, it is clear that sleep is not a resting process, since it does not save much energy, and fasted animals sleep less, not more. So sleep is clearly a repair cycle. Why would intelligence require a large whole-body coordinating system, and a repair cycle? Maybe it’s because cognition requires the whole body to change. Cognition is a form of whole-body plasticity.
Here is how I think it works. Every tissue of the body, as it lives through the stresses and strains of active life, undergoes mutations and trauma to various degrees. Then, each night, the nervous system coordinates the expression of a repair algorithm in each tissue of the body. This must be true, because if you don’t sleep, every system of your body begins to fail quickly. No systemic function is spared the ravages of sleep deprivation, and it is fatal. Sleep is so important that intelligent animals have evolved to need it even when they are in life-threatening danger.
Think for a minute about what this must mean. It must mean that your body is healing itself every night through sleep, strengthening the structures that have been stressed, and reconfiguring broken structures into new arrangements. This has all come about under evolutionary control. Some structures have evolved to be very rigid, like the skeleton, and others very plastic, like the circulatory system or the muscles, which adapt to exercise quickly. Intelligence seems to be plasticity of the nervous system, but remember the nervous system is just there to coordinate all the other systems. They all work together as a whole. So it would make no sense for a part of the body built for coordinating other systems to change independently of those other systems. The changes in it must be intertwined with all the other changes in the body that are being repaired with sleep.
Why should we not think of the nervous system as just another tissue, that undergoes mutations and trauma, and then gets repaired by an algorithm every night? I suspect that we should think of it exactly like this.
The best theory of the origins of human knowledge so far is the epistemology of Karl Popper, who thought that knowledge is built by creative conjectures, most of which are refuted by experiment. The conjectures that survive refutation by experiment become our best theories. But the open question, that Popper could never answer, was this: where do conjectures come from?
My answer, and the answer of the Santiago theory of cognition, is that conjectures are configurations of whole human bodies. The way we arrive at an idea is that we autoencode it in our entire body, in all the connections not only between the neurons of our nervous system but in all the molecules and cells of all the systems. An idea feels a certain way; it has a whole-body emotional and physical depth.
Every system of an animal body aims at one goal: to maintain homeostasis. The nervous system is no different. The entire organism is set up to expect a certain set of conditions in its environment, and when those conditions are met, it remains at rest. An organism reacts only when conditions somewhere inside or outside the body vary outside of its narrow range of comfort. Anything that upsets this narrow set of expectations is a refutation of the conjecture set forward by the configuration of the body. You can think of it as a “reaction shadow.” Outside this shadow, the body reacts to this refutation by trying different mechanisms, also autoencoded, to remedy the surprise. The body is literally a guess; a guess about what it would take to keep itself alive no matter what kind of thing happens.
Consider this: When was the last time you paid conscious attention to the inside of the second knuckle on your left pinky finger? Perhaps a year or maybe more. Now put on a pair of gloves. Do you notice that particular spot? No. Now put on a pair of gloves with a hole right in that exact location, on the inside of your pinky. See?
This is nothing like what artificial intelligence algorithms today are set up to do, which is execute a particular decision about new data given lots of pre-determined solutions as training sets. The conjectures of the body evolve in open-ended interaction with the environment. They start out completely untrained, and they track the conditions around them by simply mutating to some degree every day, and being repaired every night by the algorithms of sleep. This is how we forget so much, but remember what’s most important. This is what Maturana and Varela, the authors of the Santiago theory, called “structural coupling.” The organism is structurally intertwined with its environment.
Now here is the most important part: If the explanation I just gave for cognition is correct, then the content of the repair algorithms do not matter. Any repair algorithm will accomplish the same open-ended learning, the same intelligence. What matters is that the structure of the body as a whole must be very sensitive to trauma that comes from the environment, and it must be given time to thoroughly heal itself. It must not only be sensitive in general, but it must be sensitive to the very particular sorts of perception that intelligent humans are sensitive to. Perception itself is the result of disrupting experimental configurations. Perception is a form of trauma, of breaking neural and other connections that expected something slightly different to take place.
If this theory is correct, then it will be relatively straightforward to build a thinking machine. We just have to configure a robotic body with several arms and fingers so that it can move and grasp things, a few speakers, microphones, cameras, seismometers, gyroscopes, touchscreens, a CPU, and a clock. The data brought in from these sensors must be recorded as a set of expectations. Then, as new readings depart from the expected values, the program must begin to generate robotic actions and noises to try and return itself to the expected conditions. The motions that achieve some success should then become the motor program of the machine, the set of motor reactions encoded in the “neural structure.”
This set of reactions then must undergo stochastic mutation, and also “break” when the reactions do not successfully predict future conditions. The broken structures must then be healed into spontaneous new reaction pathways, while the pathways that survived the trauma of engagement with the world should be strengthened. Again the specifics of this algorithm do not matter very much. What matters is the sensitivity of the system to the specific trauma that human cognitive systems are sensitive to.
The easiest way to think of this is that it is not exactly the human body that is intelligent, but rather the ecological niche of humans. It is the set of conditions that the human body is sensitive to that is intelligent. You could substitute a body that was arranged quite differently, like a robotic octopus, for example, and still achieve the same cognitive function. A short person, a tall person, a person with no legs, they can all think. Even many stroke victims can recover much of their cognition with only partial brain function. If the structure follows the process laid out above, it should autoencode knowledge that will be recognizable to us as intelligence.
There is virtually no chance that the knowledge that you and I both have, that I have been using to communicate with you while writing this essay, is encoded in exactly the same structural pathways of neuronal connectivity in my brain and your brain. We both have vastly different patterns of cognition in our two nervous systems, and all through our bodies. Yet we can communicate. The specifics of the structural encoding do not matter very much; only the totality of the niche that we both share and are both sensitive to.
It is that niche that is intelligent.
If you liked this essay please check out more on www.talkingoctopus.com
Copyright 2021 by Charlie Munford
