AGI Starts with the Stomach

    I feel bad that I didn't write down the scientist that said "AGI must start with the stomach" but he is so right. 99% of AI (ANN/GOFAI/DL/Neurosymbolic/ et al) starts with sensory input and 99.99% attaches the AI to motor activity if it has something to do with robotics, where that motor activity is a direct result of everything sensory. In current AI, motor activity is always secondary to sensory input and most often, AI doesn't even put motor activity as a requisite. 

Let me flip the script and instead of starting with sensory input, let me start with motor activity.

The reason AGI starts with the stomach is because when we are hungry, we need to move. As an infant, we begin sucking on anything that represents a nipple. As we get older, we grab anything that we know we can eat and put it in our mouths. Even older, we go to the kitchen and make us something to eat. The drive to squash the hunger pain is the reason we learn. To eat in any instance, requires us to refine our motor skills. Our intelligence is built on the necessity to sustain our lives and to sustain our lives, we must move. I have yet to see any animal stay stationary in one place absorbing its environment through its sensors and be able to survive; however, the current state of AI is precisely this process.  

I have stated a few ideas in the past to explain the idea of motor activity produces general intelligence.

1) When you see a baby flailing, most people just believe it is due to lack of coordination. I believe it is due to the baby being preprogrammed to move in order to hone its motor activity to gain control to quench its desires. These movements as an infant through toddler is the basis we gain intelligence. We will put anything in our mouths as a toddler but we learn what is good to put there and what is bad pretty quickly. If we had no means to put that something in our mouth, we would never gain the intelligence needed to decipher what is good and what is bad. 

2) The human today is the same human that existed 20,000 years ago. What is the difference between our needs 20,000 years ago and today? Nothing at the most basic levels. We need to eat, sleep, defend and procreate. How we eat, sleep, defend and procreate may have different methods today than it did 20,000 years ago but its how we move in our environments to gain those basic needs that defines our intelligence. To eat 20,000 years ago, my movements would have been to fashion a spear and go hunt animals. Today, it might be to hone my finger movements to order a pizza on-line. In either scenario, I have gained the intelligence through movement to be able to stop my hunger pains.

3) It has been said many times, the only way we can sense our environment is through movement. If I blindfold you and have you just directly touch an object with one finger, it is extremely unlikely you can tell me what it is. If you can move your finger(s) around the object, you will be able to figure out very quickly what is in your hand. For sensory input to work, it requires movement of some sort. There are many who argue that qualia does not take in account any movement but I disagree. One needs to experience their environment in order to define any qualia. We cannot define something like "love" if we have not had encounters through movement of that which triggers a "love" response. Again, if all your life you were just stationary with only sensory input and no movement, how would you be able to even know or understand any qualia?

4) We go to the kitchen to make us something to quench the desires we have whether its food and/or drink. The difference between AI (et al) and AGI is that we learn through our general intelligence how to satisfy our desires no matter the environment. The classic tea making scenario illustrates this precisely. I can learn very fast and easily how to make tea in my own kitchen. I know where the pots are, where the tea is at, how to use my faucet to fill a pot to boil water, how to turn on the stove, how to get my cup to pour the water and tea bag into it, etc. Furthermore, I could most likely train a robotic AI system to do this same task with no issues. The divergence comes when I go to your house and you ask me to make tea. Your kitchen is completely different, your pots, tea, faucet, stove, etc are all different. Through my abilities learned as general intelligence, I can easily overcome any differences and make tea. However my robotic AI system will fail immediately and would have to be totally trained on how to make tea at your kitchen. The general intelligence required for us to be able to make tea in anyone's kitchen comes from our movement and the activities defined by those movements. Could we make tea with someone just telling us what to do without us ever being in a kitchen or even knowing what tea is? Perhaps, but this wouldn't be general intelligence, it would be following a program of steps. 

To conclude, I define all general intelligence as the steps required for us to move through an environment to be able to perform a task related to the internal desires at hand. I know how to make tea at your house because it is the same steps I use at my house. Yes, the stove maybe different but I still must turn it on to boil the water. I have to turn on a stove at my house, your house or the house in another town if I wish to have some tea. The precision to turn on the stove could be different in different environments but the general step of "turn on stove" is a step in all cases. We could further generalize to "create heat enough to boil water". This is a generalization of the general step. If we were camping, this step might be to start a fire. Starting a fire or heat in all cases is a step required to boil the water so when we get to that point in the tea making process, our ability to move and create that step is what defines generalization. 

Why aren't more researchers taking the motor-sensory-cortex approach and still stuck on the sensory-cortex-(maybe motor) approach? My guess is because 1) they are stuck in the DL rut that all Comp Sci people learn and believe defines intelligence (which is false), and 2) AGI is much harder. By using networks defined by connectomics, we can already put into place an AGI framework because nature has already provided it.