Many people understand that technology is nearly perfect in its ability to come to an answer. With just a basic knowledge of how computers work it becomes clear that any answer that a computer returns is just a result of manipulating numbers and returning the end result of that manipulation. Assuming there weren’t any technological glitches along the way, this would mean that the same input will always result in the same output, no matter how many times you tried it. This is true of all levels of computers, from a simple calculator to the smart device that you are reading this on. Despite the dramatic difference in possible utilities of your devices, they still work in the same way on a fundamental level.
The part that becomes surprising to people is trying to explain that AI works the same way. It’s easy to believe that AI is somehow smarter and is making logical decisions based on information it has learned from. However, although it is making those decisions they shouldn’t be mistaken for smart decisions. At the core of what an AI is doing is taking in inputs, manipulating the information, and then sending outputs. Notably, if the inputs are exactly the same, so will the outputs be.
The only thing an AI is doing is taking in information and finding a way to manipulate that information in a predictable way to achieve our desired outcome. Frankly, this isn’t a novel idea if we take it outside of computers. To compare it to the real world, it would be equivalent to designing an assembly line (maybe to build cars as an example). As raw materials and resources come in, they need to be processed very specifically and in a particular order to result in a complete product on the other side. Through the design of the assembly line, it’s possible to devise ways to ensure perfectly replicable manipulations to the input to achieve a desired output. AI is doing the same thing but in the digital world instead of the physical. However, instead of through smart behaviours, the AI creates these replicable manipulations through trial and error. As elaborated in my previous article – Why Machines May Never Be Smarter than Us – the advantage that a machine has is that it can perform countless more manipulations than us, allowing it to test nearly all methods of manipulation, eventually leading it to find possible solutions as well as optimized solutions quicker than a human could.
Where this is all building to is to say that AI is already performing as expected. It is doing what is has been designed to do and it does it nearly perfectly. The problem isn’t with AI itself, it’s with us and limitations we impose on AIs, specifically in terms of the information that AI has to learn from.
A commonality across all versions and types of AI is that they require data to learn. Sometimes this can simply be raw values given to the automation and it can try to solve for complex patterns in the data. Other times automation requires very specific types of information in which the input and output is already known, and all the automation has to do is solve how to manipulate the input to get the desired output. Regardless of these minor differences, the commonality is that AI requires data. Because the learning of the AI is dependent on the data, that data must also be accurate and cover a wide range of possible inputs and outputs. The consequence of not doing so is that the AI will likely make many errors as it will be presented with scenarios it has never seen before.
The obvious solution to the data problem is to just have more data. Given that the number of technological devices we have has been growing exponentially, so has the data that we have access to with which we can teach our AIs. However, it is not always clear that just because we have the data it is valid and ready for us to use that data to teach AI. For example, if we wanted our self-driving cars to be able to identify what a house is (presumably to teach it to not crash into houses), a thought that someone may have is to go to google maps to extract images of houses so that we can teach the AI to identify a house. Although there are many pictures that would be ideal for this purpose, there are also many pictures that would not be useful. For example, included in the catalog of house pictures from google maps would also be an arial view of houses. This would not only not be useful to the automation, it would likely be detrimental to the learning of the Automation as it would be learning information that it would likely never have to use while driving.
Although in an ideal world we would be able to provide our AIs with every piece of information we have and have all that data accurately labeled, this just isn’t the reality of the world. To give AIs every piece of information would lead to automation that either never learns to do a single task or is so complex that it would exhaust all computational power and still not complete before the expiry of human life. For as long as we continue to use brute for methods of trying to solve AI they will continue to be imperfect as the data we give them to learn will never capture the full picture of the complexities of the world.
The take-away from this is that AI is already doing what we have asked of it and it does it exceptionally well. For AI to progress any further it isn’t the AIs themselves that we have to improve, it’s the data that we have collected that we are able to give that AI to learn. As we create more data (and more accurate data), AIs will naturally become smarter because they will have access to new information that didn’t previously exist.