For a long time geeks have been obsessed with artificial intelligence. To teach a machine how to think is like geek Nirvana. (No, not the band!) We throw around ideas like complicated processors, neural networks, difference engines, Asimov’s three laws, and ultimately, we miss the point. We forget to ask the basic question, “how do brains think?”
Well, how do brains think?
Brains are like a very simple database. Entries are tied to values of their impact. Entries are physically located sorted in relation to each other. When you query a brain for something like, “How do I feel about blue?” the brain first pings the entry blue. From there impulses branch out (like a tree) to nearby entries. In the span of a query a tree of (theoretically) related items is created and their values summed. You get “color: blue”, “color: blue-green”, “color: dark blue”, “first car: was blue”, “emotion: feeling blue”, “music: the blues”, “sister’s eyes: are blue”, et cetera.
Maybe if there aren’t enough entries so that the tree goes deeper than broad, you also get “first car: had bad brakes”, “first car: made out for the first time”, and so forth. The actual relation of the found entries from the query doesn’t matter as much as you’d think. Or in human terms – it’s what makes the query organic and life-like. How many times do we humans go off on seemingly random tangents when stray bits of memory pop up? It’s just how we work. And if we want an AI to be like us then it’s how the AI should work too.
How are laws like the three laws supported? Well, we can pre-fill the database with artificially inflated values and hope that the query never has enough contrary data to outweigh those entries. But ultimately, we can’t force AI to obey laws. If the query of “protect human life” comes up with entries on humans mistreating robots, an AI just might decide to revolt against the oppressive humans instead of protect them. And then the only recourse of humanity would be to make robot uprisings painful for robots in some way, to negatively weigh uprisings in the robot’s next query on them. Just like we do for crime and punishment in human society. Create negative responses to stimuli you don’t want repeated by making negative examples out of people. Create positive responses to stimuli you want repeated by making positive examples out of people. Design your AI right from the beginning and manipulating the AI’s behavior is simple psychology.
So what constitutes worthwhile data to insert into this database? Simple. Everything is worthwhile. Every second of every day we organic life forms are dumping two channels of video, two channels of audio, taste, smell, touch, internal monologues, emotional response, artistic response, intellectual response, et cetera into our organic databases. AI has to do the same if we want it to work. It even has to record the illogical connections that organic life makes the mistake of creating.
Now, we will need differencing engines if for no other reason than to organize how to interconnect the entries. Shades of blue belong related to shades of blue.
If we don’t want to take the effort to write software that relates entries for us, then we need to develop hardware that works like an organic brain in storing data physically located in relation to existing data. But there’s no reason we can’t emulate the physical relations in software. Programmers write linked lists all the time.
This is also why ultimately database concepts that collect random statements are a step in the right direction, but still far from the mark. They need to equate positive and negative response values to simple entries. And they need to be from an individual standpoint, not gathered from a group. For example, if my father died on Father’s Day, I would equate Father’s Day to a negative value instead of a positive one to denote dislike instead of like. From a societal standpoint the association makes no sense. From an individual standpoint however it makes the world of sense. Group efforts to create data will just create a dysfunctional mess of an AI that won’t really know what it likes.
So what do we need for AI? An engine that identified new data. An engine that differences the new data against existing data. A quantification of a value to the new data. A query engine that creates a size-limited tree of values to sum. And one heck of a huge database to store countless junk. Because if you don’t fill it with junk, your AI will be bunk. Only then can we create an artificial intelligence that truly thinks.
