If we assume that the brain is a kind of computer, artificial intelligence is the process of reproducing its functioning. Based on this hypothesis, it’s easy to dismiss the possibibility of above-human intelligence by arguing that we can only program what we understand, which would means the intelligence in the machine is bounded by our own. But it’s also very easy to refute this limitation by arguing that we encode learning processes in the machine. These learning processes would be working at a scale and speed that we can’t match. The machine will beat us.
This later argument definitively seems to hold if we look at recent achievements in deep learning. Computer achieve some tasks that very much ressemble some form of intelligence. Looking more carefully, it’s however questionable whether we should speak of intelligence or simply knowledge. Techniques like deep learning enable computers to learn facts based on large amounts of data. These facts might be very sophisticated, ranging from recoloring images correctly to impersonating the artistic style of a painters. But the computer isn’t intelligent because no reasoning really happen.
This leads actually to an interesting question about intelligence. How much of intelligence is simply about predicting things based on experiences? If an object fall, you predict its position in the future to catch it, based on other experiences with falling objects. If someone asks you a “what’s up?”, you can predict that they expect to learn about what’s going on. With GPT-3, which works according to this principles, you can almost have a conversation. I say almost, because we also see the limit the approach. There are some classes of question that don’t work, like basic arithmetic.
Current artificial intelligence is able to learn, either by analysing large quantities of data (deep learning) or simulating an environment and learning what works and what doesn’t (reinforcement learning). But we’re still far from sentient, thinking machines.
If we assume that our brain is some kind of computer performing a computation, there’s however nothing that prevent us from replicating it. With this line of thought, it’s only a matter of time until we “crack” the nature the intelligence and will find the right way to express this computation. When this breakthrough will happen is unknown – maybe in a decade, maye much later – but there’s nothing that make it impossible. With sufficient perserverence, this breakthrough is inevitable.
Speaking in terms of computation and data, a system becoming smarter can happen in two ways. The first one, is what we have now. Systems that learn over time through the accumulation of data. The computation remains the same though. A deep learning network is programmed once (by human!) and than trained on large quantities of data to adjust its paramter. But maybe a second class of systems exists: system that self-improve by changing their computation. Systems able to inspect and change themselves do exist and are called reflective systems. In such a system, data can be turned in computation and computation into data. The system can thus modify itself.
Some people believe that with artificial intelligence, we risk beeing outsmarted by a “explosition” of intelligence. Systems of the first class learn within the bounds of the computation that defines them – however complex this computation is. The possibility of an explosion is limited. With systems of the second classes, we’re free to speculate, including the possibility of an explosion of intelligence. Such a system could outsmart us and lead to superintelligence.
If we assume that our brain is a computation: is it self-improving or not? Children acquire novel cognitive capabilities over time, which at least give the illusion of self-improvement. But maybe these learnings are only very complex form of data accumulation. Also, the boundary between reflective and non-reflective systems is not black and white. A fully reflective system can change any aspects of its computation, whereas a non-reflective system processes input data according to fixed rules that never changes. A system that is able to infer and defines some rules for itself would fall in between both categories: the rules can change, but to an extent that is limited to some aspect of the computation. The adaptive nature of neural networks could, in some way, be seen as a limited form of rule changing: the rules are fixed, but the “weight” given to them change over time due to feedback loops.
Learning requires data provided by an environment. We’re able to learn only because we interact with the world and other people. If we were to replicate the computation in our brain and the learning process that takes place, we would also need to simulate the environement. The computational complexity of all this is probably enourmous. Maybe we can replicate the computation in our brain, but not the environment, or only limited forms of it. In which case, it’s hard to tell what kind of intelligence could be achieved.
Depending on the computation and environment that we simulate, the intelligence won’t resemble human intelligence much. The algorithm of AlphaGo learns in an enviornment that only consists of the Go rules. We can not even image what this world would be like. Assuming that the artificial intelligence is human-like misjudges the nature of human intelligence. Intelligence is not one quantity that we can weight based on clear criterium. Intelligence has many facets and is contextual.
For some facets, like arithmetics, machines are for sure already superintelligent.