Let’s Demystify Generative AI By Looking Under The Cover

As we have seen in recent modern times, AI has grown from humble party tricks like playing cards to diagnosing serious disease from images of strangers. It can create all kinds of things that we define as art, although many people question the art it makes. When we’ve seen films that involve computer soldiers and malevolent robots and read stories about computer programs going rogue and killing off the crew, we are naturally afraid. It’s what we’ve been exposed to and given licence to think via the seeding of fears and anxieties. We’ve also seen clever devices like talking cars and helpful androids, navigation computers that know everything, and digital playmates who teach foreign language and mathematics on the side. We don’t fear these things; we generally want them. This is why a lot of our modern AI is attempting to become these things.  
 
There is a clear and obvious ethical problem at hand in that we can’t make sure that individual people or organisations use AI technology to do things we generally do not want it to do. It’s accepted as fact that military uses are numerous, with data driven warfare becoming a real phenomenon. Does this prevent unnecessary risk or relegate rich human lives to numbers in a mathematical problem? When we use AI to decide what to do with large numbers of individuals, do we risk losing the definition of society in the old sense by creating artificial frameworks based on previous experience and learning? What if someone asks the AI to design a weapon that is so far undefendable? With all forms of law frameworks and crime, it’s not possible to prevent each instance all over the world. Every nation has its own laws and methods that they are entrusted to enforce. Other nations can’t write laws for each other without international organisations that have some level of sovereignty in terms of political decision making, such as the European Union.  
 
The most powerful use of AI for good is in the science and engineering field. With the level of detail an AI can apply to any idea or plan, the system can run through any number of possible combinations and outcomes which it evaluates and then produces a result that fits the best. Generative equations create answers that are not in the database, but they use the database to provide an answer that matches those that do. This means that the output is always new, but it follows any number of patterns that means it fits. It is purely mathematical at its core, with outputs on the computer represented by numbers. In art programs, the numbers represent pixels and their colour. An image with a million pixels on a machine capable of producing a million colours will have a million factorial possible combinations. It simply choses the ones that match the pattern the best. The pixels can become human cells in an organ or weather systems in the sky, given the right amount of information and ability. 
 
 
At the sharp end of the cutting edge of artificial intelligence, we find the science of artificial creative invention. There are three main types of inventiveness that machines are programmed to achieve. The first type, and the one that was the first to be created by computer scientists, is interpolation. Interpolation is the ability to be given a few inputs and create a new output that considers of these inputs and creates some form of average. This average is a novel output, and not one created from the inputs, but it is created because of them. It is a matter of putting the inputs into a numerical form that is represented in as complete way as possible and then giving the system the correct functions to use this data in a way that is meaningful to its original context. The context therefore must be added to the variable information in the database, meaning that one piece of human information might have a dozen or more associated information pieces in the database that allows the AI to understand it completely.  
 
Knowing how it works, it may come as no surprise that computer scientists work with neural scientists to learn how biological minds function as information processing systems. They try to replicate the way nature uses individual neurons as transmitters of information as various chemical signals resulting in various amounts of electronic interaction. By studying the way a tiny brain in a fruit fly works, the process of nature’s mechanics can be placed in a computer system via a series of code and instructions. Translating the neural network of an insect into the functions of a working computer system is entirely dependent on what you are achieving and how the fly's basic functions can be adapted numerically to fit an entirely new model of action. It’s interesting because a fly is instinctive and it can often be predictable, it uses a very basic information system when compared to a higher ape like us, or even a cat or a dog. 
 
The level up from interpolation is extrapolation. When machines can extrapolate, they are able to create something new from the inputs that does not necessarily fit into the original database. It doesn’t use a system of averages and line of best fit, rather it predicts what would be the case in a scenario where this new information was already in the database. It can hypothesise and out the most plausible response without already being aware of what that might be. This is a lot harder. It’s like instead of making a fruit salad by chopping up the fruit you have, it’s planting a seed, waiting for the tree, and building a boat. That’s another level of intelligence. The database of information the system draws on is another level of magnitude bigger and it can plan several steps ahead. These modern art programs such as Nightcafe and Imagine Art do this well, however the real steel is in the medical and science fields where the AI can detect things in results and images that trained professionals miss. It has the mind power to ask if every single pixel in the photo is healthy or if every molecule in the blood test is at the right level even if things don’t appear instantly to the doctor. It can do it before the doctor opens the envelope, which doesn’t get rid of the doctor, just the envelope.  
 
The work in creative AI systems at this time is in the inventive field that can be termed Inferpolation. This is a stage beyond extrapolation because the result is completely new and independent from the constraints applied by the original data. Using a door handle to open a beer is perhaps a simple example of this in action, where we use a thing that is completely out of context to achieve a goal that it was never designed to achieve. Putting the device on another device, like the iPhone, is another example. You can’t get a computer that knows everything about mp3 players to suddenly know about phones, what would it take to program an mp3 player to philosophise on other purposes it could have? You’d have to give it a lot of internal knowledge and an ability to draw upon it to create hypothetical situations. The system needs to link multidisciplinary sources in hypothetical scenarios to test unknown theories without instruction. Humans do this all the time, but can a computer? Could a piece of software intuitively open a metaphorical beer with a symbolic door handle? Knowing how we do it, and why, is the possibly best way to learning how to program the machine to imitate it.  
 
Humans are not like machines in which we can actively process lots of information at the same time. We rely on our intuition to learn what our subconscious is telling us. Th subconscious is a lot more like the machine. It doesn’t rely on big picture models and imagined gap filling; it simply has experience built in over millions of generations of evolution. The conscious mind must infer all kinds of things, there is no way the brain can absorb every detail our eyes, ears, skin, and so on are telling us. It takes short cuts based on what we’ve learned. The imagination fills in the gaps so when we see a smoking chimney, we already see the roof, and the building, the details are built up as we look more, yet the basic construction can be inferred from the size of the plume, the location of the building, and the material the chimney is made from. Every new information nugget can be used like a seed to grow the bigger picture before we’ve had a chance to study it. This is why we fail to notice many things if they’re out of context and we’re giving our attention to one particular focus. A large database of interconnected ideas and knowledge is required for this to work. It takes a vast reservoir of known examples that we use our subconscious to process. A machine will have a database of information that is interlinked to itself with associative data, and it uses it at the speed of electricity.  
 
How does this database work and how is it acquired? The resource of information that a machine must use for its decision making is defined in two distinct ways. Expert systems are given a set amount of data to use that is designed by a human being. These machines do exactly what they are told or what their set rules allow them to do according to the rigid parameters in the program. The best-known example if this in action is the AI opponent in a video game. The difficulty can be set which enables the character to do more or fewer moves in its efforts to defeat you. It then is given a list of available moves and instructions on how to execute them. You, the player, outsmart the computer or you don’t. In the old days, they never learned from your tactics although they knew how to adjust over time to throw you off. Throw in a random variable or two to make it more like real life, and you get a good game each time. It’s not that intelligent, it’s just doing what it’s told extremely well. The techniques of programming machines to beat human opponents were perfected with chess computers. A set list of available moves is given, and the machine tests each one in turn to decide the most effective move.  It can be taught how to forward plan by several moves and anticipate the human interaction for various outcomes so it can assess the best likelihood of a win. This was achieved in the 20th Century and was the result of years of programming work across several projects.  
 
When the outcomes are so various that it’s impossible to account for them all, a creative and intuitive element is required. The real world is not like chess, we have infinite squares and moves each, and there are billions of us. How do we ask an AI to think about and generate responses that account for this? The answer is in the other model of artificial intelligence. The learning model is not the expert system. In fact, if you ask a learning model to play a game of something as simple as connect 4, it probably won’t win for at least 50 games. This is because it is deliberately given zero initial information. Instead, the system is given the task of learning for itself to fulfil a goal. It is then given a positive reinforcement for a correct response and a negative reinforcement for an incorrect one. It’s treated much more like a child or a pet, who given the right guidance, can grow up to be a fully functioning and responsible member of the household. The learning system is given the information sources that it then uses in hypothetical and scientific examination ways that are programmed into its learning model. These elements are under human control, however the end results and the ways it chooses to apply the information are not. Provided we do not give it goals that are counter intuitive to humanity, it won’t look for them. What is stopping it from determining harmful actions on its own, to fulfil a different goal, like an end to poverty? Killing everyone surely would end poverty in a computer’s eyes unless it already knew that was never to be done. This means that this powerful tool must have ingrained contingencies that it can’t work around regardless of its goal and the line of best fit it has determined. How do we know what these are? Perhaps teaching it the laws that human beings must live by would be a good start, and a procedure that results in shut down if one is about to be broken. Now, where have I seen that before?  

It's probably a good idea to grab yourself some professional tuition on AI engineering and use from skilled computer scientists. Check them out.
​