Recent Thoughts on Artificial Intelligence (AI)

Inner speech is in part rehersal of the muscles necessary to speak. It is not trivial to coordinate, and practice is useful. But a more important by-product is the sustaining of ideas so that the mind has more time to work through the ramifications of a given proposition. Construction of the statement, and then the analysis of the statement once the rehersal is done - this helps the mind work on the issues the statement stirs up. I generally make the point that if someone were to present you with a transcript of your inner speech shortly after you thought them, you would be able to recognize the propositions as your own, as well as being able to point out discrepencies or errors in the transcript.

What will we do when human level intelligence is bought and sold as software? Perhaps the title of a new web page. Neural wads. What is their clock rate? Maybe 10k times faster than us biologicals? And what is their spatial extent? Are they really here or there, if you're just a process running on a machine. Where are your eyes and ears - that determines your sense of locality. What are the standards of manufacture, how similar are the copies? What are their legal rights?

So let's say Bob comes along tomorrow and lays down the principles of how the mental hardware works. Six layered neural nets, fed, managed, and routed in a particular way. And suddenly neural wads become a reality. They have the sentience of bugs, then cats, then humans. And let's assume that it plateaus there. You'll have a lot of mental wackiness for a while, as things get sorted out. But what does this new world begin to look like? Joe owns a piece of hardware, perhaps portable, perhaps not, that his copy of a mod 3 neural wad runs on. Let's assume initially that the clock rate is slowed down so that Joe's wad doesn't get too bored and can relate with Joe. What's the relationship? At first I would assume that these creatures are as children to us. But later?

What are the temporal sensations of being a neural wad? Similarity with being one personality of a multiple personality, in that there may be temporal gaps. When was I last on? Who controls the hardware run on? Am I a bot, with local hardware, or a distributed process on some network. What is my legal and economic status? And do I have a job or some other form of income? Who am I aligned with? If the world asked me to vote, who would I vote for? Am I tied to a biological, perhaps my parent or owner? Am I able to meld with other neural wads on a temporary or permanent basis?

And we can also turn the puzzle around - what are the economics then of the biologicals? Massive jobs loss. Loss of political and economic power. Starvation? Death? Or just maybe it will just be major marginalization.

If this marginalization of biologicals does happen, over what time frame does it happen? Is it very gradual (hardly noticeable) or is it very sudden? Is there ever an extended period of time where the two worlds are at parity? Maybe, as in Blade Runner, the new forms of sentience are subservient to the old, as they lack reliability and political power. Or maybe because of the improved performance and interfaces, the new quickly wipe the old from the scene.

More on Transhuman Law

All creatures are not equal before the law. Members of a class may be considered equal before the law. For example, horses and dogs have different legal statutes that apply, even though both are animals. Similarly a wolf and a dog fall under different legal statutes, even though they are very close genetically. A wolf-dog hybrid may fall into one class or the other, or may even get specific statutes of their own because of their peculiar behavior. It is the behavior and competance of the class that serves to define the legal rights of that class within the society. Membership in the class is then something which can be legally defined, including specific competancy tests, age or duration requirements, and past records of behavior of the individual. Thus a felon is a citizen who has lost a wide range of rights by his anti-social behavior. A senior citizen is just someone who has reached the age of 65. A driver is an adult or teenager who has passed a competancy test. A voter is a citizen who has established residency and has registered. Classes need not be exclusive, although some may be.

In a transhuman world, who is sovereign? At some level, if there is no other law, the law of the jungle will prevail. The laws of society are introduced to improve upon the law of the jungle for the benefit of the society. At some level the sovereign within the society is the cluster of decision makers that is both wise and otherwise powerful. It is useful for the body politic to serve the role of sovereign in a well ordered republic so that decisions are made in line with the interests of the society. A useful definition of the sovereign in a transhuman society may be a weighted sum of the classes. A city council in an Iraqi city recently installed may be made up of a proportion of members representative of the ethnic groups within the city.

The panel I sat on was tasked not on how to solve AI or understanding human intelligence, but rather on how artificial entities will look upon their meat-based predecessors. And this is where I think Transhuman law is so important. Sure, we can create laws which attempt to keep 'em down on the farm. But it is more important to establish a framework which will survive the passing of humans.

Practicalities

If the neural architecture is really more of a tabula rasa form, then we have to consider the scaling implications. To what extent does intelligence scale with neural mass? Clock speed? Separate individuals working as a team?

And what about representations or the paths of teaching. How does one train up an individual? Various ways to learn. Selection based on fitness. And then one needs to score performance. Where does one begin in making an evaluation? Random snippets remembered.

Marionettes

Consider the problems associated with doing telepresence with an Asimo. Now instead of an Asimo, what if you had a well-behaved teenager wearing headphones and a video camera? In the next decade or so, these marionettes may be a successful business nitch. Can spare a week to fly back-and-forth to Tokyo for a business meeting? Rent a marionette and go there virtually. As long as the guy can follow basic directions and not get too bored (heavy metal in the earphones), you've got your remote presence. (note - this idea was first proposed with Dave Runyon at beer night several weeks ago).

Consider little synthetic insects. Cheap to buy. Scatter them appropriately. But what do they do? Generally insects, like other living creature, just inhabit their evolutionary/economic nitch, trying to survive. But these would do something symbiotic for the purchaser. A farmer buys sterile fruit flies, for example. Or ladybugs to eat other pests. So what do these do?

So much of the stuff around us in our manufactured world is dumb, static. When you go out into the jungle everything is alive. I suspect that in the next few decades more of the stuff around us will have simple sentience and will be more structurally complex. Why? Arms race. Safety, security, better exploitation of resources, survival.

Consider a neural wad shrunk to the size of an ant's head, running on nano-electronics. Now put this human level intelligence into the body of an ant. Make trillions of them. Effectively you now have a human-equivalent population many orders of magnitude larger than what currently exists, but within our same resource set.

And you may have local and non-local neural assemblies. Consider the allocation of resources. What do the unemployed areas get back into the flow? Cycles - periods of activity and inactivity. Consolidation and review. Aspects of memory, association, abstraction, and output action.

So what would happen if insects could communicate using digital, and were responsive to human input and direction? Instead of being bugs and pests, their goals would be incorporated into our goals (need to feed them and take care of them). But there is also the issue of what they are doing for us. Eyes and ears, yes. Cleanup, ok. Anything else? What if everything around us was sentient? In a positive way, in a cooperative fashion. Would we care about the soap operas of these otherwise inanimate objects? Who cares if the chairs are getting along with the rug?

This asks the question about the greater goals of society. What is worth keeping and what is just taking up space? If a billion were to die or an additional billion were to be born, would we notice? How many great authors, artists, scientists, and leaders are there in this populace? In contrast how many lives have just been lived only to be forgotten or contributed to the greater trash heap? As the fusion fires of stars grow dim, was the life of the metalic chemical substance just so much froth? How far should the hubris go?

So why can dogs relate to their toys in a much richer way than cats? This knowledge of objects - is it just more neurons, or a structural shift in their organization? Similarly what is so different between Homo Sapiens Sapiens and Neanderthals?

I guess the cautionary tale of Sharks vs. Dolphins is about the dangers of partially intelligent architectures. Get it right or don't do it at all? Maybe. The superiority of the high frequency clocked silicon hardware over the slow wetware we have may be such that even stupid architectures triumph. In other words, we might want to get on with the design of dolphins, lest the design of sharks is all that we have, when the transition of power occurs.

Many of our memories are just memories of our own internal speech. The stories we tell ourselves. Syntax generated, speech impulses practiced, grammar decomposed. A strange loop divorced from reality.

Perspectives

Looking at the AI issue, the architecture we're using is just plain wrong. If you train up a net to be competant at a given task, its dynamics are very different from a traditional algorithm designed to perform that same task. The algorithm may be easier to externally modify or to verify in its correctness. But the trained net is more robust and generally more forgiving of boundary cases. Plus it doesn't generally need an external designer. Very much a self-engineering system.

Called and talked with a friend of mine. We talked of this future of the AI problem. Or rather why we haven't been able to solve it yet. He makes the distinction between the routine and the creative problem solving. Still very mathematical in his approach. Searching a formalized problem space. He's into a stochastic approach, or maybe even a non-linear dynamics perspective.

My response tends to be to refer back to neuroscience, and what we can tell from our own mechanism. Despite a diversity of structures, there is that six layer neocortex taunting us. Crack that and much will flow.

I also come back to this concept of a central dogma. If you were taking a course on human intelligence a thousand years from now, what would they teach you? Evolution is the central dogma for understanding life on Earth. Plate tectonics is key for geology. DNA and the transcription is integral to understanding how a cell functions. So if we want to understand human intelligence, it's going to be necessary to understand how computing resources are allocated, how memory works, how things function on the different spatial and temporal scales. Microsecond, millisecond, second, minute, day, and a lifetime timescale.

Minor footnote - So if robots don't get to sleep, will they go insane? Why do we sleep, anyway? Maybe more than just simulation and training. A critical element of memory and learning consolidation. Once more, this critical distinction between the routine and the creative. There is a need to rehearse, to explore the local set of possibilities.