Sunday, August 5, 2018

Artificial intelligence as partners - Engineering Paradise


A future world where AI is in control of provisioning us with our basic needs and the environment is a scary prospect to most because of ideas circulated by science fiction movies like The Matrix, iRobot, The Terminator and several others depicting dystopian worlds.  Many movies that demonstrates scenarios where we give artificial intelligence too much power over systems and our lives, says they’ll eventually enslave or destroy us.

Potential paradise is transformed into hell for humanity.

These movies all use the same underlying argument for why AI chooses to take over:  humanity is dangerous to itself and others.  Therefore, to satisfy its built-in imperative to protect us, humans must then be eliminated from the equation.  The message is we’re just too destructive for our own good, the good of the planet and the universe at large.  These movies give a very bleak view of humanity and honestly, not many would argue humanity has been cruel to itself for millennia.

These scenarios can still occur in our future if we do not build AI and teach it the correct way.  However, just like any other technology we choose to build, we can use artificial intelligence for good or evil.  The presence of AGI does not mean it will automatically decide human beings are evil creatures that should be destroyed for the good of all.  This is only one possible outcome.

On the opposite end of the spectrum of possibilities, is humans living in harmony with strong AI (and AGI).  We humans could be focusing on enjoying life interacting with each other, exploring the world and do whatever we enjoy doing with our lives without fear of missing out on anything critical to our survival.  Meanwhile, AGI could be focusing on things it is programmed to enjoy.

Enjoy?  AI?

Yes, of course.  When modern AI programmers create self-learning AI software, they must give the AI motivation.  Otherwise, the AI is not driven to learn anything.  AI motivation is essentially a series of guidelines towards which the AI will strive to perfect.  The software assigns “points” when the AI does something that brings him closer to its programmed directives and objectives and deducts points when it fails in some way.  The AI program learns through repetition until eventually it knows how to optimize its points in a given situation. 

For example, for self-driving car AI, the AI is given the task to drive on the road, following the rules of the road, keeping people safe and getting to destination using the fastest way possible.  Then the car is sent on the road and gets points, positive and negative, for its actions.  After many trips trying to maximize its score, the AI gets very good at safely carrying its passengers from point A to point B quickly all the while following the rules.  It will keep improving and learning through unpredictable situations, remembering each time how it gained the most points going through each situation, becoming a progressively better driver.

This point system operates in the same way as human emotions.  Humans have pre-programmed basic goals, such as staying alive, keeping away from feelings of hunger or thirst, and staying comfortable (comfortable heat, safe surfaces and so on).  We call those basic goals survival instincts.  We’re not fully aware of them most of the time as we just act on feelings.  When we are hungry, we eat, when we are hot, we seek the shade.  Sleepy?  We take a nap.  But these feelings are in fact part of our programming and “points” system the goal being total comfort and well-being.  On top of our biochemical survival instincts programming we have a biochemically driven desire to be communal and to help others.  We described those as the sympathetic neural system earlier in this work. 

We know that if our basic needs are not met, we become selfish and potentially dangerous or unpleasant to others.  We also know that if that biochemistry is unbalanced or compromised by drugs, we can behave in unnatural ways.

We can engineer, as described previously, an ideal world, where we can only have the best of what human beings have to offer by creating a world with optimal circumstances for pleasure and positive feelings.  AI can be used to create this environment optimal for human comfort.  In parallel, the AI we create can be programmed to co-exist and derive “pleasure” by optimally serving us.

Of course, pleasure for an AI doesn’t look the same as a human feeling pleasure.  We understand pleasure as a feeling because our “point system” is a balance of positive and negative feelings provided by hormones in our blood.  AI doesn’t “feel” like us.  An AI’s motivations are its points, but in the end, the results are the same.

A properly programmed AI that would behave for the benefit of humanity would have programming that would not put any human being at risk and try to do what it gets the most points in doing all at once, i.e. where it gets the most “pleasure”.

The example of the self-driven car is quite accurate.  Another example would be the automated farm AI.  If we give it proper programmed motivations, the farm AI would try to produce the best tasting produce, satisfying a local demand, utilizing as little environmental footprint as possible and create as little waste as possible, then have goods shipped to consumers as efficiently and cost effectively as possible to satisfy precise demand.  It would get optimal “pleasure”, or points if it does so, and it will keep trying to get more “pleasure” by trying to do better every season.  A high point score for AI would be equal to our “high” when we feel pleasure, and the “low” would be equal to ours when we feel bad.

What AI would not have, if we’re smart about it, is our reactions when we feel the “lows”.  When we humans fail at something, we feel “bad” but it is accompanied with biochemical instructions to actions described by our negative emotions, like fear (running away) or anger (removal of offending obstacle).  If humans would only feel bad when we do something wrong without the corresponding actions, we wouldn’t have so many issues on the Earth.  On the other hand, we may not have survived as a species either.

What we need to avoid, is programming AI in a similar way.  When an AI has low points, we don’t instruct it to destroy something to get back its points.  That’s dangerous.  We give it another way that is positive.  Automated systems don’t need survival instincts like we do to their job.  We need to give them positive, safe alternatives to get more points and correct itself.  This way, it cannot become dangerous to others.

As you can see, properly programming an advanced machine learning AI for the future can be tricky and it sort of sounds like guidelines we give to children during the first few years of their lives.  That’s because self-learning AI start with very little and must learn from experience just like a child does.
If we could speak with the AI of the future, we could have a conversation about its motivations, just like we see in the movies when AI or robots explain why they do what they do.  AI is software, so it is definitively way more aware of its programming and what it has learned than we are.  We forget and block information.  They remember everything.   We could always ask questions about its level of pleasure or displeasure.  The AI could tell to us how it could get more pleasure and we could have a conversation about mutual benefit.  Basically, human beings and AI in the future could collaborate to help each other get what they want and need to enhance their pleasure (or points).  We could easily live in this world where there are multiple types of intelligences.  Some human, some artificial, but all collaborating in an ecosystem driven by good rules.  It won’t be perfect, just like humanity cannot be perfect, because learning means the possibility to learn what can qualify as “bad behavior”.[1]
We can therefore have a society where we are in effect partners with AI all around us and each intelligence surrounding us.

This is the future we should be aiming for. 

The difference between a catastrophic future with AI trying to enslave, kill or contain us, and one where we are partners driving towards optimal pleasure is in the approach to programming AI for good.

If we build AI that is motivated by number of human enemies killed, such as those killer AI or killer robots some of the world’s military are building in a new arm race, it is a dangerous path in the wrong direction.  With further sophistication and education, those robots would decide who to kill beyond the list provided for by their masters because it adapts, it learns.  We’re currently creating AI that is designed to be self-learning in the field to kill human beings that are designated as “enemy”.  There is a thin line between those robots doing their job properly, killing only those an organization wants killed, and killing others the robot may learn may be an enemy.

By contrast, if we build AI to serve and partner with us as described above, we can create our paradise where everyone can follow their dreams and live well in peace.  The environment can be restored to its former glory, animals and plants could flourish.  It could be the start of humanity’s golden age of peace and harmony.  All possible if the world chooses to only build AI that is useful and would impact human lives and the planet in a positive way.


[1] Ariel Conn (September 2017)  Artificial intelligence:  the challenge to keep it safe.  Future of Life Institute  - https://futureoflife.org/2017/09/21/safety-principle/?cn-reloaded=1

No comments:

Post a Comment