Nano Future 2030

I’ve written before on the need for placing limits on AI development to protect the human race from being overwhelmed by AI entities who can develop faster than we biological animals. I offer these thoughts on what is needed to ensure our own survival.

Asimov had the 3 laws which were to be applied to robots and that was a good first effort. Most sci-fi movies, that present a robot or AI, imply that the entity could grow or evolve to where they could out perform a human in some way and thereby gain power over the humans. We don’t really have a way to limit AI development so it would be good to be able to discuss what makes a good or bad design.

I see AI overrun of humans as a realistic danger because humans, as a group, evolve more slowly than an individual AI could. When you consider the concepts behind the Singularity, it seems likely that once AI programs reach parity with human intelligence and creativity, they could surpass us in a few years and then the fun would begin. If they can evolve ahead of us to the same extent that we have evolved ahead of dogs, why would we be treated any differently than we treat dogs? Just because many dogs are loved, doesn’t prevent cruelty or the destruction of dogs that get in our way.

I believe the critical difference between dangerous AI and safe AI is the lack of or the incorporation of self awareness in the first few generations of AI’s. Consider that only the higher animals can recognize themselves in a mirror. The lower animals are all pretty hard wired to fulfill basic needs. If they are territorial, they defend territory when it is important to their survival or reproduction. They don’t vary from that fixed program of behavior. Our AI’s would not be territorial unless we decided they needed to be. Hopefully, they would not be power hungry unless we design them to be. So, a “lower” AI would resemble most animals in that it attempts to fulfill it’s needs and behave as it is designed to behave. Doesn’t mean it would act like an animal. It could be a factory manager, design work flow, manage production, interface with employees and still have limited behavior by design.

It’s more of a problem when you get to “higher” AI’s that are free to consider different goals and to evolve their own goals. We will need them to think about a wide range of options and situations. This ability is essential to a creative, problem solving AI. Our problem comes when we try to make “people” out of these AI’s and give them the same motivations we have as animals.

Emotion is not something you have built in just for the fun of it. Emotion and feelings are the motivations for your own behavior and the behavior of any sadistic ruler. Without emotion you would not do anything but sit and stare until you starved to death. Even rational thoughts are motivated by emotional needs and desires. The exact same thing will be true for any functional AI. Something has to motivate their behavior according to a hierarchy of needs. The lack of feeling in Commander Data is a physical impossibility and nothing more than the emotionally driven, lack of expression on the face of any Rambo or Vulcan.

So, AI’s will be driven by something we can call emotion even if they don’t cry or wail when things go wrong. Think of source code where if ( anger_variable1 > x) then choose behavior option 14 , to be an emotional statement. We need to prevent this emotion from driving their behavior outside of designated, and acceptable limits the way it drives humans outside of our own acceptable limits.

Self awareness might be the key. If you are not able to be aware of yourself and do not ever consider your own needs for re-evaluation, I believe you will not change your behavior as it applies to taking care of yourself. You will not decide that you need more money or power since you will not consider this “self” to be any more than any of the hundreds of other variables that must be monitored and maintained within limits.

You (or the AI) can certainly take care of yourself (if thirsty, go drink water) without having self-awareness. Self-awareness, as I’m using it, is thinking and feeling about yourself and how you are getting/not getting what you want, about what “you” need and what could be better or worse for “you”.

Self-awareness is the awareness of yourself in time and space and with respect to time and space. In other words, don’t provide the hardware/software to consider yourself in this unique way and you won’t make decisions and goals with respect to that self. You won’t take action to improve or defend this thing called “self”. For a full description of self-awareness as used here, see this article.

Hi, it’s been a while but I’ve been working on a clearer definition of consciousness and thought I’d post it here. How a mind works is very important to advanced AI and to the possibility that we will advance to the Singularity.

I’ve also noted a wide range of concepts linked to the word consciousness. I’ve seen it used as,

• magical soul
• ability to recognize yourself in a mirror
• the opposite of unconscious, that is, not being asleep or knocked out.
• the ability to think about or be aware of yourself,

I’m talking about the last one. The next to last one defines anything from earthworm and above as a conscious individual. But I’m not sure even a dog can think about itself, other than being aware of it’s body’s experience. I will use the term “self-aware” to refer to this concept of self-aware consciousness. Without self aware consciousness in an AI, you get goal directed behavior but without the concept of ‘self’ influencing the decision process. It would still have goals, including self maintenance and any other assigned goals, but without self awareness it would not develop a concept of self and therefore, not have fear for self or desire for self gain. In other words, it would observe the world, function to obtain it’s goals (because it is designed to do that) and not be concerned with modifying it’s goals for self gain.

This concept of designing AI’s without self-awareness may be the crucial difference between a useful slave AI and an uncontrollable, self-aggrandizing competitor to humans. What is a human peer AI, but a human equivalent individual who would compete with us for power and resources? Given that humans often compete without following any rules, I would expect nothing less from a human equivalent AI. And I would not expect to win a completion with a peer class AI when that AI can morph it’s body into any shape, operate at much greater speed than a biological individual, and control it’s own pain and emotional drivers.

But, if the human equivalent AI is not designed to have self-awareness, it remains a functional mind but without the complexity of self concept to make it a human competitor.

How do you add or not add self awareness to an AI? Self-aware consciousness is simply the unconscious incorporation of previous experience into present experience. That is, the thing that makes us self-aware is our awareness of key elements of our previous experience. Such as kinesthetic body experience, sense of time, sense of location, memory of what we were thinking about, concept of who we are, what our name is, who we are speaking to, what goal is important at this time, etc. Those things that we know ‘now’ are unconsciously ( that is, without our knowledge or awareness of the process ) fed back into our present experience as an integral and subliminal part of our present experience. This feedback takes some time since our brains don’t deal with new information more than 10 times per second. So we get a 100 millisecond delay in this feedback loop. What we experience through this delayed channel is actually a copy of what we were experiencing 100 milliseconds ago. Which also contains a degraded copy of what we were experiencing 200 milliseconds ago. That’s why our feel of who we are fades away the further we remember into the past, until all we have are discrete memories of objective experience by way or short or long term memory. The consciousness delay loop is outside and separate from the short and long term memories. It is a hardware thing. You either have it or you don’t although I suspect that there are early versions in animals that are rudimentary to what we know as self-awareness.

OK, it’s been too long. So, I thought this story was worth the effort.

Science Daily had a story about two guys who rewrote Asimov’s 3 laws of robotics. See
Science Daily

They basically tried to bring the laws into the present and into the real world. I did not have a positive reaction and I’d like to explain why.

Asimov’s first laws:

• A robot may not injure a human being, or through inaction, allow a human being to come to harm.
• A robot must obey orders given to it by human beings, except where such orders would conflict with the First Law.
• A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.

And the three new laws by Woods and Murphy:

• A human may not deploy a robot without the human-robot work system meeting the highest legal and professional standards of safety and ethics.
• A robot must respond to humans as appropriate for their roles.
• A robot must be endowed with sufficient situated autonomy to protect its own existence as long as such protection provides smooth transfer of control which does not conflict with the First and Second Laws.

They certainly lost the beauty of the original, but they were trying to be real so I appreciate their effort.

However, the first revised law is not for or about robots, it’s about humans. Their first new law is more appropriately part of a “3 Laws for Robotic Researchers” that someone should write. And the language is PC and essentially worthless. “highest” has no objective value since “high” can not be measured. It’s equivalent to saying the researcher must be a good guy or he should not create robots. Maybe law one could be rewritten as “A robot must implement only approved behaviors … “, but that would make the law impractical.

Law number two is appropriate for robots but they are required to have as much judgment as an adult human to evaluate the fitness of a human who gives them orders. I guess that can be simplified to the robot must only obey a human who can provide adequate ID or a secure command key. Asimov said “Obey humans”. The new law says “Obey only privileged humans.”

Automobiles have been abiding by law number two for the last 80 or 90 years, and automobiles kill about 50,000 Americans a year. If you put an AI mind into all cars, Asimov’s laws would shut down the transportation system. Whereas, this new law would allow cars to operate, provided someone in authority said the robots were trying very hard not to kill anyone.

Their third law says a robot MUST be capable of self protection but not so protective as to disobey their key holders or to corrupt their own operating system parameters that make them meet the “highest legal and professional standards of safety and ethics”. They ignore Asimov’s first law, which was the point of the whole thing after all.

So I can’t build robots without some minimum capability of self protection? Why?

I would like to see a real set of rules for robots but they would have to be written in terms that applied to the situation. Asimov succeeded because he avoided the details of how to implement the rules. If you are going to get serious, the rules need to be in terms of behavior of the robot that anyone can measure.

I wonder if robots will do any better than humans at obeying rules when they are motivated to reach goals, just as we are, that lie outside of acceptable behavior. Any mind will do what seems best given all available inputs. If you screw around with “what seems best” you can get any behavior under the sun. Literally.

Hi,

For forty years I’ve waited for someone to get AI right and create something that actually asks a question instead of simulating the asking of a question. I guess the reason is, we all thought it was more simple than it turns out to be. I suspect the problem is still not well thought out. Sure, very intelligent and deep thinkers are working on logic software, neural net computing, expert systems, and language parsers. But I have not heard anyone talk about the system level structure of the human mind. Maybe that is more my fault – I’ve been busy on other things.

Consciousness is creeping back into legitimate consideration – and that is another issue where I’m not impressed by the effort so far. People don’t even talk like they understand what it is, much less how to create it. Look inside your own mind and figure it out by what you see there. Forget words. Listen and look and then figure out what you see and hear.

I also remember getting comments that were critical of my advocation that we place limits on the ability of AI during the early decades. I think I understand why it bothered him, but it says to me that he did not understand how to design a mind. There is no person in an AI implementation unless you put the person there. And why would you place power into the hands of some person you did not understand or control?

As a culture, we don’t understand children and we don’t, in general, know the first thing about how to raise them. If you want to create a really messed up mind, go have a kid. Hopefully, you and the kid will have a lifetime to work it out. I don’t know anyone who is not damaged in one way or another. Life is about coping with the cards you were dealt. Let’s be careful what cards we deal to an AI that can live forever, modify itself, and develop it’s own goals independent of humans.

Here’s a radical thought: Don’t have kids or design AI minds until you are 50 or more, have forgiven your parents for their screw ups mainly because you see yourself walking the same path, changed your career field at least once, and have spent significant energy on your own emotional growth. The idea is that you probably can’t create an AI mind any better than the average, screwed up 25 year old can raise a kid unless you have been through the ringer a few times and come out the other side trying to understand and heal the situation. And that’s just the personal stuff.

On the technical side, you haven’t got a prayer designing a AI mind unless you understand your own. Oh, sure, you can design one but it will have as many problems as the kid of the 25 year old who only knows how to make babies. Kids can’t raise kids although they’ve been doing it for thousands of years and we have a world around us that proves it.

So, know thyself before inflicting thyself on others.