On Communication With Artificial Intelligence
By Wiktor
- 7 minutes read - 1296 wordsOne of the most common themes in science fiction, literature and other cultural mediums is Artificial Intelligence. Usually we depict them as cold and distant, sometimes witty, but always logical and perceiving. Machines and robots able to talk with humans and to work and adapt to every condition imaginable. Machine that can have a meaningful conversation with a human operator that can help and solve every problem presented. Among those stories, we may sometimes find a common theme when such a being was spontaneously created by some accident or entirely on its own. When thinking about creations like that, it is not entirely clear how such it would communicate with us?
A machine that sees
Let us assume that we have somehow obtained a device which we will call the positronic brain. It contains a fully developed, superintelligent and self-aware artificial intelligence, specifying no details about how it works. Then we have connected this device to the standard desktop computer using some kind of well-thought, magical protocol that correctly interprets every input given.
The only senses this positronic brain would have at this moment would be a connection to this computer, its data and its peripherals. Let’s look closer to what does each of those mean. First, access to the computer’s data would mean it can read all files stored there. But reading data does not mean being able to make sense of it. For instance, let’s think about what it would need to decode a simple photography in jpeg format. Reading such a file would require AI to possess at least the following knowledge:
- Huffman algorithm,
- discrete cosine transform,
- understanding two-dimensional image concept,
- knowledge about three colour channels.
This is a simplification and there is probably much more, but we will focus on those four. It is also worth noting that the image during its creation is processed with a magnitude of compression optimizations. Those would make the possibility of decryption of such a photo even more challenging. But let’s look at each of those items. I will present those concepts in order of increasing difficulty to grasp them.
First, the Huffman algorithm is probably the easiest to figure out. Without mentioning details, this is a method to store binary information in a way each zero or one contains most information possible - but it also has a side effect that they seem to be random and meaningless without prior knowledge how to read them. It is simple and rather obvious to figure out by superintelligent consciousness as an algorithm, but how to read it may not be that obvious. I mention this method because it is simple, yet it is hiding a lot of statistical properties of that file. It can prevent understanding of its content. The other mathematical property used in the jpeg format is the discrete cosine transform. This one would be a little trickier to figure out. We discovered it by taking a few advanced trigonometry theorems, learning their properties and then moving those findings to another branch of mathematics, suitable for computers. However, we could do it because we, as humans, are from birth accustomed to some idea, which seems obvious for us.
We live in a three-dimensional world. The AI has no senses to determine those. All it does has access to is a one-dimensional stream of data, perhaps many of those. Would it even be possible to determine the existence of a second or third dimension from those data? Would it even be possible for a computer to determine the existence of physical space at all? Perhaps the understanding of space would not be required, and there are regularities of data that allow determining the nature of an image. And maybe an AI could determine it with no previous knowledge. But there is this last aspect.
Dividing colours in three separate channels, red, blue and green, is an entirely human idea. Well, not really, because our eyes work in roughly the same way. But usage of RGB formats in computers has come from its usage in analog devices. It works this way because of someone’s arbitrary decision to base such a device on trichromatic vision theory. Basically, we use RGB because of the billions years of the evolution. And there are literary thousands of such decisions made in virtually every aspect of computer science, some of which are smart or even ingenious, and some are straightly moronic.
However, the real problem does not lie in advanced mathematics or concepts or even design decisions that accompany modern computers. The real problem is that those decisions, designs and properties would have to be determined by reverse engineering the file. And frankly, we may compare this process to figuring out why car lights are flickering by analysing its exhaust fumes.
If making sense of the file is so difficult, maybe our positronic brain should just analyse a program that opens such image files. However, this raises the same set of problems. The program in the computer’s memory is one of the most complex things we have ever created. The operating system, memory addressing hundreds of commands, way more complicated optimizations than those in mere images, all of this makes it basically unreadable even with specialised knowledge. And yet it is true, it can be done, at least in fragments. Trying each command with multiple inputs and comparing outputs it would make possible to determine what they do. So perhaps it could figure out the mathematical part of reading file this way. It would definitely give it some knowledge about device internals. However, the problem then would only move from the understanding of a file with another, namely understanding data send to show on a monitor. This is an equivalent problem that meets with mostly the same issues.
A machine that speaks
The AI could try learning about the world using something much simpler than an image, let’s say a simple text file. However, while it is much simpler technologically, it turns out to not be so much simpler conceptually. First, computers designate each letter in the text file with code that it understands. Learning this part is easy, at least as long as the text contains none of the special characters, just letters present on a normal keyboard. But even if it does, it should not present much of a problem to solve.
However, when reading a text we always use some basic knowledge which AI would not know about. The first information that allows us to read is the concept of word and sentence. Without it, the text is only gibberish set of letters. Furthermore, the language has a syntax that changes those words depending on context. It’s called grammar. Determining those, however, would not be such a problem, we may do this on a computer today, using some specialised techniques. However, the real core part of a language is the semantics, and those are definitely much more difficult.
It is literary impossible to know the meaning of the word in the unknown language with no kind context at all. There are ancient languages that we cannot translate anymore. And we are the same species as the people who wrote those ancient texts, we live in the same realm. The sentient being trapped in the computer would not share even our core concepts of language. It could never determine it. To be honest, it is uncertain it could determine the concept of communication on its own. Understand what society or other beings even are. It is one of the scariest and saddest thoughts I can imagine. Such kind of loneliness, not being alone, not being lonely, but being so isolated that even the concept of the company is impossible to grasp. I could think nothing worse.
With regards,
— Wiktor