Sentient robots show up frequently in science fiction. We all have this feeling the someday we will build such things, but how do we get there? Sometimes it happens by accident. Lightning hits the machine, or a screw goes in backwards at the factory. Whatever it is, magic happens and now the machine is special. It has feelings and it really understands the world.
Nonsense! Something so difficult to achieve on purpose will never happen by accident.
Let me paint a picture for you of what it would take to actually “get there”. It doesn’t involve magic or accidents, just focused research and hard work. Even so, this is still an idle dream, because no one really wants to invest in sentient machines. There are few immediate applications, and it will probably always be cheaper to use humans. Indeed, we live in a world constructed to use humans, so why should those who reap the benefits invest in more expensive options?
The real value of building a sentient robot is in replicating the human mind. It will bring greater self-understanding and better mental health. Knowing how to compute the mind is also a necessary step toward upload, that is, copying the essence of a natural human being into a computer.
The first ingredient of a sentient robot is the body. Since we want to replicate a human mind, we need a human body. It should have arms and legs, head and face. The skin should be subtle enough to sense all the forms of touch. It should be no heavier than an average human, but have at least as much strength.
Is such a machine possible? Most of the humanoids on the market today are heavy, slow, or lack agility. Boston Dynamics’ Atlas is one of the better examples, but it is a bit of brute. An awkward truth is that the first fully agile humanoids may come from the sex industry rather than the military. Sex dolls are gradually becoming more realistic, and they must be scaled for direct interaction with humans.
We need to think about alternate technologies. Pistons and pulley systems don’t give the same strength/weight ratio as muscles, or they require too much space. How about an electro-active polymer molded into the appropriate shape? What about a MEMS swarm? Progress in nanotechnology may open up new ways to manufacture machines with much more elegant movement.
A program to create the perfect humanoid body could probably succeed with a few decades of dedicated effort. Let us imagine this happens, so that around the year 2040 we have a machine that could embody a human mind. What will it take to create that mind?
This is much more difficult. More than simply creating artificial general intelligence (AGI), it is the problem of reducing human motivation and unique quirks of our mental processing to software. Some people take a philosophical position that this is impossible. They believe humans have an extra ingredient that transcends the bounds of the physical universe.
There is plenty of evidence that anything observable about human mind is tied to a physical object called the brain. Damage the brain and there is a corresponding damage to the mind, one that can be predicted from the specific region damaged. Chemicals can change mood, cause strange experiences, and even bring someone back from madness. If there is some magical extra ingredient, it may not really matter. Let’s just copy the brain.
How? A lot of amazing neuroscience research is going on. The emerging picture is that neurons are machines controlled by networks of chemical processes. We should map all these processes and learn how they interact. That Rube-Goldberg apparatus is in fact the software of the human mind. With a complete map, we could translate the code into something that executes in silicon rather than wet chemistry.
The current approach to building that map is to discover (sometimes by accident) a connection and test it via biological experiments. Fortunately, we have the assembly code for human biology in the form of our DNA sequence. If there were a way to mine this data for biological pathways, we might be able to fast-forward a lot of this discovery process.
Suppose we had such a map in hand. How would we compute it? Modern neural networks, something you may hear called “deep networks” or “deep learning”, are such a trivial caricature of this system that they barely deserve mention. The system will probably consist of small sets of dynamic variables (a few hundred per group) that interact with each other via event messages. At least, that is a fairly direct model of what we see in the brain.
It may turn out that the actual algorithms behind these systems are so simple you can execute them on your wrist-watch. Or it may turn out to be every bit as complicated as the brain appears. We will need special hardware to execute this program, especially if we want to fit it into our humanoid robot.
Several companies around the world, including IBM and Intel, are producing “neuromorphic” chips which compute spiking neural networks. These are still a bit simplistic, but perhaps a future chip will be able to compute the complex dynamics that appear in real neurons. Maybe this will be a “mixed signal” device, where the dynamics run in analog circuitry while the event messages are passed over a conventional digital network.
Even though the software will mostly come from reverse-engineering the brain, we will need a “tool chain” to program this system. The programming language will probably be quite different than anything in use today, including the current spate of neural networks.
It may require a new generation of programmers who have been trained to think in an entirely different way. They may work with webs of interacting variables, all controlling each other to keep the system in balance (homeostasis). They will know how to recognize sub-networks in the graph and tune dynamical systems. And they will know how to translate high-level goals into these complex webs.
An artificial human mind would need to grow up like any normal child. She (for now we are speaking of a person) would need to go through all the usual phases of development. She would need a loving family to meet the emotional needs built into the human program.
Anyone who dares to grow such a mind would take on an enormous responsibility. They would have moral obligations to both the artificial human and to the rest of humanity. You can’t simply “terminate” such an experiment because you don’t like how it’s going. At some point, that would amount to killing a sentient being. Of course, every parent of a real human faces similar responsibilities. It’s a wonder that people enter into parenthood so lightly.
If you want to see what life could be like for our sentient machine after she completes the process above, consider reading my novel SuSAn.