I lay out a general design for the kind of process that I suspect may be necessary for the development of strong artificial intelligence.
Some problems in complexity are NP Hard. There are no shortcuts: there exists no polynomial time algorithm that would allow you to see the future of the system before that future actually comes in its own good ‘time’ so to speak. The only way to see is for the full system to walk through the entirety of its history with full fidelity to the details of its interactions and structure. It has occurred to me that perhaps the development of any strong AI would require a process of development similar to this, even if it were done in silico. And therefore the development of a strong AI requires that the AI have its own full history and evolutionary path, in which it would pass through a series of stages in the same way as the human brain has passed through its own.
Maybe this does not entail that the process take 500 million years the way the brain has, but it may require that some computational runtime or informational/entropic kind of equivalence to substitute for the years must be followed through in the process. There should be some kind of energy and dissipation cost associated with generating it.
Some modern development processes have made use of fitness functions and an iterative/evolutionary process for producing new technologies. Assuming the reader is familiar with some of these methods, the immediate analogue to natural selection is obvious. What may be less clear from what I mentioned above is that to develop something very complicated, with an exceptionally long history like the brain, it might be necessary to first select not for typically human brain properties like language, abstraction, theory of mind, etc. but to, in the course of this evolution-like simulation, first select for more basic things.
For example, we may begin with some kind of smaller neural network on the order of 100 components (like a worm), and subject to a fitness function for mobility like light seeking. Later on more sophisticated behaviors would be required, using a more complex set of constraints. In each subsequent stage the process selects for more complex tasks and augments the network, building from the prior one. I imagine that something like this; a living system evolved through a virtual ecosystem could at some point, generate a very advanced piece of artificial intelligence and perhaps even necessarily would do so (1) .
- The U.S. military had the stealth technology used in the F-117A for more than a decade before it was revealed to the public. Sometimes I wonder what sorts of things they have in their possession today that we won’t know about for another 10 years, a technology that seems so improbable and fantastic and on the boundary of science fiction you would have trouble convincing anyone they could have possibly done such a thing. This is one of those kinds of things.