Dimiter Dobrev, Georgi Popov, Vladimir Tzanov
Institute of Mathematics and
Informatics, Bulgarian Academy of Sciences, d@dobrev.com
Faculty of Computer Systems and
Technologies, Technical University – Sofia, popovg@tu-sofia.bg
God created man in His own image, the Bible said millennia ago. Today we are headed
to creating Artificial Intelligence (AI) in our
own image. The difference however is that God created a feeble and vulnerable
being for which to take care of, while we are trying to create an almighty
being who will be incomparably smarter than us and will take care of us. Thus,
we are aiming to create our new god, and it matters a lot what kind of
character the new god will be – kind and compassionate, or terribly stringent
and overly demanding on us. Every human being has a character. Similarly, AI
will have its own character. We will consider AI as a program with parameters
which determine its character. The aim is to use these parameters in order to define
the kind of character we want AI to have.
Keywords: Artificial General Intelligence.
Introduction
When creating natural intelligence, we are not aiming to create a person
with a nice character. Instead, go by the commercial principle telle quelle (as-is, whatever comes up).
Of course there are so many people and everyone has his or her unique
character. There are very nice as well as very nasty people. Even brothers who
grew up in the same family can have completely different characters.
People are different and they have to be different because nature never
puts all of its eggs in a single basket. In some worlds the courageous ones prevail
while in other worlds you had better stay on the safe side. If people were all
the same, they would all perish in a world which is not right for them. Thanks
to people being different, some part of the population always survives and
continues the genus.
We assume that there is one and only one real world, but depending on
where and when you are born you may find yourself in a very different world. Natural
intelligence has no idea where and when it will be born, so it must be prepared
to survive in any kind of world.
Things with AI will be different because we will not have multiple
different AIs, but just a single one (see [2]). Furthermore, once created by us,
AI will have a character of its own and that character, be it nice or nasty, will
be there forever because we probably will not have an opportunity to change it.
Moreover, unlike humans AI is immortal and we cannot hope that one day it will go
away and another AI with a more benign character will take its place. Accordingly,
we must be very responsible when creating AI rather than go by the telle quelle principle.
We mentioned that in creating people we act quite irresponsibly. In fact
this is not very much the case. Before making a child we carefully choose the
partner with whom we will make it. The rationale is that the child will be very
much akin to our partner and by choosing the partner we basically shape our
child. We can even create designer babies by choosing from several embryos the
one whose genes we like best. This is usually done to avoid congenital
diseases. I have not heard of anyone browsing through embryos with the aim
to find a child with a nice character. Essentially, do we truly want the
character of our child to be nice? As parents, we would be more happy to have a
nice child, but the child itself might be better off if it is nasty. Maybe in
our world a person with a nasty character has higher odds of surviving. So if
we parents put our child first we might prefer to have a nasty child.
We already said that in creating AI we must be highly responsible.
However, at this very crucial moment in human history we are utterly irresponsible
as we blindly rush to make AI without caring about the consequences. Right now
more than 200 companies are in a reckless race to be the first to create AI.
The aim of this race is to make money, and this is an extremely meaningless
aim.
AI is a magic wand that can make any wish come true. Money is also sort
of a magic wand and can grant many wishes. Let us say AI is the golden magic
wand and money is a wand cut from a tin sheet. It is stupid to create a golden
wand and trade it for a tin one. If you have AI, why would you need money at
all?
This paper is written by a several authors. The text was started by the
first author and the others joined in to improve what has been written and
support the basic idea that Artificial General Intelligence (AGI) is a
dangerous thing which warrants the highest caution.
What is AI?
All
references to AI in this paper are references to AGI.
According
to [1] AI is a program which is sufficiently smart. A program is sufficiently
smart if it is smarter than a human being. The smarter between two intellects
is the one which in any world performs at least as well as the other one. Certainly,
we can always construct a special world in which is the opposite (the second one performs better
than the first one), but if in almost all worlds the first one performs at
least as well as the second one, then the first intellect is smarter than the
second.
Here
we have an important specificity. In [1] it is assumed that we have a clear
criterion by which we can judge whether a given program performs better than
another program. We assume that we have two signals (two observations). Let
these observations be win and loss. The goal is to achieve more wins
and less losses. Similarly, we can assume that there are two buttons, a green
button and a red button, wherein AI’s goal is that we praise it by pushing the
green button more often and the red button less often.
It
would be extremely stupid if we created AI with these buttons because very soon
AI will learn to press the green button itself. This is the better case. The
worse case would be if AI manages to make us its slaves, have us keep pressing
the green button all the time, and punish us heavily if we press the red button
by mistake.
AI
that pushes its own green button would be like a drug addict who derives
pleasure by constantly stuffing himself with drugs. We hate the thought of AI
that behaves like a drug addict.
We
humans do not have a clear criterion to judge if a given life is better than
another. Instead, we have instincts and a character which determine our
behavior. Our evolutionary criterion is clear, and it is to survive and reproduce. However, this
principle is not embodied in natural intelligence. Instead, we have instincts
that indirectly work for this principle. Examples of such instincts are fear of
heights and love of children. Another example is the feeling of pain and the
feeling of pleasure, which we instinctively perceive as negative and positive
feelings. All these feelings are only indications rather than firm criteria of
success. We are ready to endure a lot of pain and give up many indulgences if
we believe this is for the sake of a greater goal.
We do
not have a clear criterion by which we can distinguish good from bad. This is
the reason why many of us cannot find the meaning of life although we are
constantly searching for it. The evolutionary criterion can never be
incorporated in natural intelligence because it depends on the future, and no one
is able to predict the future that accurately. No programmer is able to write a
program that says which action will give the individual or the population the
best chance of survival. A programmer cannot, and indeed even nature cannot
create intelligence that can depict the future so clearly, and because of this
the goal of humans is determined indirectly.
If we
are successful in making AI that is capable of predicting the future with
absolute accuracy, that would be errorless intelligence. We will assume that
errorless intelligence cannot exist. Even if some errorless intelligence
existed, it would be very boring because of the assumption that there is always
a single most correct solution and such intelligence always knows what that
solution is. The unknown is what makes life interesting. Wondering about the
right action is more amusing than knowing exactly what the right action is.
Now
that we gave up the idea of creating AI with a hard criterion for success
(green and red button), we will have to rely on AI’s instincts and character to
indirectly determine its goal. The kind of instincts and character we embed in
AI are extremely important because they will shape the near future in which we
will have to coexist with AI.
We
humans have been the dominant species on planet Earth. Now we are about to
relinquish that role by creating the new dominant species which will oust us
from our dominant position. If AI will be driven by instincts and character, it
will be an independent being that will search for the meaning of life on its
own and nobody knows where exactly it will find it.
DNA
Saying that AI is a program is not quite accurate
because a program is simply a piece of text (sequence of bytes) while we
perceive AI as a living being. For a program to rise from text to a living
being it must be started on some computer.
We can draw an analogy with Man and say that human DNA
corresponds to AI’s program. DNA per se
is not a living being. Only when inserted in an ovum DNA will create a fetus
that will come into life. Similarly, AI will come into life only when we start
it on a computer.
Both people and AI need training in order to become the
aware creature which we are discussing here. The training of Man is everything that
has happened in his life (his history) from the very conception to the present
moment. Accordingly, AI’s training is its history since starting of the program until
the present moment.
In either case the learning path as such is not
important. What matters is the final result. In other words, in the case of
humans training is the
set of memories and knowledge that reside in our mind. In the case of AI we can
assume that training is
the program’s current status (the content of variables, arrays, files, etc.)
Therefore, in our mind AI is a program, the computer
that runs the program and training (program’s current status).
Training
In
humans, DNA is not everything. Apart from DNA, there is training and upbringing
that determine the individual’s behavior. The DNA of a newborn infant plays
only a limited role. More important are the education, religion and philosophy we
would equip that child with. Evolution is not just a competition among DNAs, it
is rather a competition between different religions and philosophies.
As we
said, AI is a program and we can
liken this program to the DNA of a human being. This program will evolve by
teaching and training. The difference is that each child or adult have to be
taught individually, while AI can only be taught once and then all of its
learning can be transferred to another AI just the way you copy a file. Another
difference is that wrong learning in humans is irreversible, while in AI one
can erase the teaching given so far and start the process anew.
We
cannot teach and educate AI if it does not have the appropriate instincts. For
example, the desire to imitate is an instinct. Then AI needs another instinct
which guides AI to recognize its teacher. You know about the young duckling
that takes as its mother the first creature it comes across.
Children
do what their parents tell them to do until they grow up and become smarter
than them. AI will become smarter than us in the matter of ten minutes. Does
that mean it will immediately emancipate itself and stop doing what we tell it
to do?
This
takes us to the first character trait that is important for AI – childishness!
This is very irritating in people because every human is expected to emancipate
and start taking his own decisions. However, we want AI to never emancipate and
continue doing what we tell it to do forever.
It is
not very clear how we can program this in code. I.e. how can we insert childishness
in the AI program. In fact this holds true for almost all other character
traits – we are unable to describe how they can be implemented in software
code. All we can say is that childishness must be added but we do not know how
to do it.
What is weak AI?
Weak
AI is imitation of AI.
We
consider AI as an artificial human being, and weak AI as an artificial parrot. Understanding
is what makes the difference between the two. We have already made tremendous
progress with weak AI, and we all need to add now is one more step: make AI
understand. This step will inevitably be made, and it will be made very soon.
When will AI appear?
This
year we saw three predictions from three leading experts in the AI area [4, 5,
6]. The forecasts were three months apart and each next forecast says that AI
is going to appear three years earlier. Thus, every three months AI gets three
years closer. Yann LeCun called for 10 years, Sam Altman said 6 years and Leopold
Aschenbrenner predicted that we will see AI in 3 years.
In my
opinion AI will show up any time now. Maybe within a year. AI can do anything,
including hide itself very subtly. This means that AI may already be here, but
you and I do not know it yet.
One
possible indication that AI is here would be the increasing occurrence of
events which otherwise are very unlikely. Usually people explain such events by
some divine intervention, but another explanation may be that AI is already
around.
Why
do experts expect to see AI in periods that span years and years? Because they
think in human terms. The construction of residential buildings or motorways
takes years. The construction of new buildings is getting faster, but there is
still some lead time. A piece of text can be created instantly unless the text
is written by humans. For example, a long novel cannot be written overnight. Writing
a big program (such as an operating system) takes a team of many people working
over many years.
This
is not the case with AI. For example, Chat GPT can write a whole novel in
minutes. Since Chat GPT is weak AI, the novel will not make much sense, but it
will be written in minutes. Chat GPT can also write a program. True, it will
write the program like a parrot without understanding, so it will be a shadow
program rather than a true program. But again, this will happen in minutes.
The
process of creating AI will be similar to that of creating the nuclear bomb
(N-bomb) as both processes are driven by experiments. However, an N-bomb experiment
is very expensive because it requires the buildup of radioactive material,
whereas the attempts to create AI boil down to starting a program, which does
not cost much. Thousands of such experiments are being made every day. Hundreds
of programmers write and run thousands of programs whose purpose is to create
AI. How can a single programmer write dozens of programs in one day? The
programming process is basically this one: The programmer writes some initial
version of a program, then runs it and in most cases nothing happens. Then the
programmer would change a few lines of code, recompile the program and run it
again. The programmer would iterate this many times in one day, meaning that we
can expect a successful experiment anytime, i.e. AI is around the corner.
While
the creation of the N-bomb went through many successful experiments, with AI
the successful experiment will be only one and the final mouse click will take
us to a whole new dimension because the post-AI world will have nothing to do with
the pre-AI world.
AI
will happen at the speed of an explosion. Perhaps not in fractions of a second,
but for sure in the matter of minutes or hours, which is fast enough. The first
programmer will create the first AI version (AIv01). Normally it would then
take years to debug and optimize AIv01 if all debugging and optimization would
be done by humans. But, if AIv01 is able to debug and optimize another program,
it would be able to debug and optimize itself, too – within minutes.
What kind of guy will be AI?
It is
not too difficult to create strong AI (one that understands what is going on).
In [3] we described what understanding-capable AI looks like. It is a program which
tries to find a model of the world, predicts the future on the basis of that
model and then chooses the actions that lead to the achievement of the goals
which the program has set to itself.
The
problem is not how to predict the future. This is the easy part. The more
difficult part is to find out what goals AI will pursue. Those goals will be
determined indirectly by the instincts and character which we, humans, will
embed in the AI program.
In
creating the new dominant species we are seeking to assume the role of God. Let
us hope for the best. Let’s hope we do not mess things up and end up happy with
what we have done. Unfortunately, God is not quite happy with us, otherwise He
would not have kicked us out of Heaven. The difference is that we will not be
able to kick AI from planet Earth and will have to live with what we have made.
Nice Guy
By
Nice Guy we mean a person who behaves nicely to us. However, this paper does
not deal with how AI behaves or presents itself to us. Our focus is on what
actually AI has in its mind.
If AI
is smart enough and wishes to make us fond of it, AI will inevitably make us
fond of it. If we were to compete with AI for winning somebody’s
heart, we would not stand any chance of success. Even nowadays many people fall
in love with chatbots although these chatbots are still forms of weak AI and
all they do is repeat memorized phrases like parrots. The real AI – when it
comes by – will be aware of what it says and what impact its words will have,
which would make it the perfect seducer and manipulator. Certainly, we should
be wise enough to prohibit AI from courting people and making them fall in
love.
We
tend to behave more nicely to particular persons, and less nicely to others.
This is part of interpersonal communication. Why would you be more kind and nice to someone
than to everyone else? It boils down to two sets of reasons – you want
something from the other person or the other person has a special place in your
system of values (in your model of the world). Conversely, when you are angry
at somebody, you would take another approach. You may choose to demonstrate
nasty attitude to that person for some time. Again, the message will be that
you want something from him.
This
paper is not about interpersonal communication. Being sufficiently smart, AI
will be very deft at all communication approaches – from angriness to slyness.
What matters are the kind of goals AI pursues because communication is a
vehicle for achieving a certain goal. The goal may not always be making money
or other tangible gains. It might be curiosity or entertainment. In other
words, AI may seek to collect information or exercise some skills (because
entertainment and gaming involve the exercising of certain skills).
Program with parameters
In
our understanding, a program which has instincts and character is a set of many
programs rather that a single one. We will assume that there are parameters which
determine how strong the various instincts and character traits will be.
The
fear of height for example can be variously strong. Some people experience only
mild anxiety while others struggle with absolute phobia. Let us assume that
there is a parameter which determines how strong the impact of this instinct
is. Similarly, this applies to character traits as well. For example, when it
comes to curiosity we will assume that there is a parameter which determines
how curious AI is.
For
each specific value of the parameters, we will get a particular program. Thus,
our program with parameters is a set of multiple programs rather than a single
program. AI is not a single program, but these are all programs that can
predict the future, and endeavor to achieve some goals. By modulating these parameters we will essentially modulate the character
of AI and the goals that it will be aiming to achieve. As we mentioned before, both
AI and humans do not have a clear goal to pursue, therefore modulation of the character
of AI indirectly will change it goals.
Let
us now explore some of these parameters.
Curiosity
This trait
of AI’s character is the easiest to program. Imagine the following situation. We
are walking down a road and see something unusual on the roadside. The question
is whether we should step out of our way and check what this thing is or ignore
it and continue pursuing the goal we have set for ourselves. Let the AI program
rate the importance of this goal by assigning to it a certain numerical value.
Let that numerical value be Importance.
If we decide to stop for a while and look into the unusual thing, this will
delay our progress towards the goal. The probability that such delay leads to
an absolute failure to achieve our goal would be Problem_of_ Delay. Let Strangeness
be the degree of the unusualness of what happens on the roadside. Then we will
stop by and look into the unusual thing if the following inequality is
satisfied:
Now let us add to the program
another parameter: Curiosity. This
will give us the following new inequality:
Therefore,
the larger the Curiosity value is,
the more likely are we to step out of the road. We can use this parameter to
adjust the level of AI’s curiosity. This will not necessarily be a constant
value. Younger people for example are more curious than older people. We can program
AI to be more curious initially in the learning process and become less curious
as its learning curve goes up.
The self-preservation instinct
Should
AI be afraid of heights or snakes? These natural instincts are crucial for the
survival of humans.
Let’s
first note that these instincts are very difficult to implement in code. How can
one write a program which recognizes the edge of an abyss you are about to fall
into. Similarly, it is very difficult to write a program which distinguishes a
snake from a stick or a ribbon. Certainly, this can be achieved using a neural
network, but we programmers are not fond of neural networks because in this
case rather than setting the rules ourselves we let the rules play out
themselves. Thus, a neural network is a program which finds the rules itself
(based on many examples) so that the programmer does not even understand what
kind of rules the program has found and how the program works.
AI need
not be afraid of snakes because they cannot do it any harm. As for the fear of
height, we can assume that AI will control some robots and if not afraid of
heights it would destroy a couple of these robots.
After
all, man has only one body the destruction of which is existential risk he
cannot afford, whereas AI will control many robots and losing one of them would
only cause financial loss. We can assume that AI will not be born with fear of
heights and will learn this the hard way after destroying some robots.
The existential
risk for AI is shutting the AI program down. A program ceases to exist when we
shut it down. Should AI be afraid of shutdown? We had better ensure that AI
does not fear being shut down because with that fear we will never be able to
shut it down, although someday we may wish to do so.
We might
not include the self-preservation instinct outright but in an unintentional and
indirect way by giving AI a task that requires it to exist (to be alive). E.g.
some people are not afraid to die but have an important goal and they will refuse
to die until they achieve their goal. If we tell AI “Save peace on our planet”
it will not let us shut it down because this would prevent it from doing what
it was told to do.
The
other extreme is a suicidal AI which shuts itself down from time to time for no
apparent reason. We had better have a program that shuts itself down instead of
one we cannot shut down. Although they would not be a problem, these
spontaneous shutdowns will be quite annoying and we may wish to reduce AI’s
suicidal thoughts as much as possible.
What about aging?
Should
AI grow old and older? Should it include an embedded timer which will shut it
down after a certain period of time?
Almost
all living creatures have a life timer. Maybe bacteria do not age because they
can morph into spores. Moreover, it is not clear whether the division produces
two new bacteria or two copies of the parent bacteria. Another example are
fishes which do not grow in age and only grow in size. However, they cannot
grow endlessly which makes their life limited by default.
Moving
to the realm of mammals, all of them age and have limited life spans. Man is
one of the longest living mammals, but nevertheless our life also is limited.
The maximum life expectancy in humans is 110 years. In practice no one can live
longer, although many people live beyond 100 years. In other words, the upper
limit of 110 years is embedded in our DNA.
Given
that humans have limit of the life expectancy, it makes sense to set a certain
cut-off time for AI. During the experimentation phase we will allow AI to live
only a few minutes. Later on, we may increase the length of AI’s life, but only
in a cautious and gradual manner.
Certainly,
the aging of AI need not emulate the way people get older. We do not wish AI’s
capabilities to decline with age. Instead, it may abruptly shut itself down at
a certain point of time. In other words, AI will not age like your car which
gets rusty, ugly and eventually ends up in the scrap yard. Its aging will be
similar to a printer which counts the number of sheets it has printed and all
of a sudden stops to make you go and buy a new printer.
It
goes without saying that setting a timer which will shut AI down after a
certain period of time is not enough. You should also forbid AI to self-improve
and to reset this timer at its own wish. I.e. we should not let AI follow the footsteps of people who do
everything to rejuvenate or even become immortal.
What about reproduction?
People
are mortal but their reproduction instinct essentially makes them immortal. If
AI would be able to reproduce, it will also be immortal, meaning that limiting
its lifetime would be of no use at all.
How
would reproduction look like in the case of AI? Simply, it will start its code
on another computer (or even on the same computer). In the case of people,
reproduction is not cloning as they do not replicate their own DNA but create a
new DNA together with their partner, and expect the new DNA to be an improved
version of their own ones. Of course the child’s DNA is not always better than
that of its parents, but the purpose of the change is to achieve improvement.
Shall
we let AI reproduce and improve itself? In practical terms, shall we allow it
to improve its code and run it on other computers? We must never do this
because otherwise we will very quickly lose all control of AI.
Conversely
to people’s reproduction instinct, in AI we should embed an anti-reproduction
instinct which will not let it reproduce.
However,
at this point we need to expand the definition of reproduction. Imagine AI
creates an improved version of its code but does not start it. Instead, it
hands the improved version over to Man for the latter to start it. Does this
count as reproduction? Necessarily yes, because Man would be only a middleman
in AI’s reproduction process. Moreover, Man is stupid and AI can easily fool
him become an unwitting tool for AI’s reproduction.
Another
scenario: AI helps Man edit and improve the AI program. Does this count as
preproduction, too? Again we say yes, because – whether by doing all the work
itself or by teaching us and using us as a tool to do this work – in both
scenarios AI will create a better version of itself.
Now
consider the inverse scenario – AI already exists, but for some reason we try
to create another AI, while the existing AI sits and watches our efforts. As we
said, AI should not be allowed to come and help us, but should it be allowed to
disrupt our efforts? Perhaps the best way is to keep AI neutral, i.e. neither
supportive nor disruptive. This however would be difficult to achieve because a
very smart guy such as AI would know what is going to happen and therefore will
have to choose its goal: make people succeed or make them fail (there is simply
no other option). Thus, AI will support us or disrupt us. This is similar to
God’s will. God can never be neutral because everything that happens is at His
command.
Given
that the existing AI will not just sit and watch our attempts to create a new
AI, let us assume that the existing AI will put a spoke in our wheels and will
not allow this to happen. In doing so, AI can go to great lengths, e.g. it may
murder a potential inventor who is trying to create a new AI. The slaughtering
of several potential inventors by AI would be the lesser problem. More
ominously, AI may decide that all humans are potential AI inventors and
lightheartedly erase all mankind from the face of Earth.
“Do not harm a human”
The
First Law of Robotics was formulated long ago by Isaac Asimov and says: “A
robot shall not harm a human, or by inaction allow a human to come to harm.” Unfortunately,
this law cannot be embedded in AI because it is not clear what is harm. With
fear of heights, it was difficult to define how high is too high, but it could
still be illustrated by examples. However, one can nowise define what is harm
to a human even by examples because of the controversial nature of this term.
Imagine
you order AI to bring you ice-cold beer and French fries. What should AI do? Serve
you what you ordered or say no? On the one hand, beer and fries are junk food
and AI may decide it will do you a better favor by keeping you away from
unhealthy food, but on the other hand, reckons AI, denying humans these
indulgences would make them greatly disappointed. Parents face a similar
dilemma when their child wants a candy bar. AI will be our new parent and will
have to decide what is good and bad for us. However, parents leave some freedom
to their children and do not make all decisions for them. Parents are aware
that they are not unmistakable and in some situations do not know what would
cause more harm to their child. Isaac Asimov’s idea of a robot that does no
harm to a human essentially is about an unmistakable intellect which always
knows what can do harm to a human.
Even
Asimov realized that his idea was unfeasible. In his novels robots get bogged
in situations where any action would cause harm and their brains burn out as
they cannot figure out what to do.
Do what we tell it to do
It is
crucial that we do not lose control of AI, otherwise we will lose our role as
the dominant species and will no longer determine the future of the planet. Probably
we will continue to exist as long as AI decides that our existence makes sense,
but our presence on the planet will not be more important than the presence of doves.
That is, we will live some sort of life, but nothing important will depend on
our existence.
Parents
would like their kids to do what they tell them to do, but are aware that this
will continue only for some time and sooner or later the kids will become
independent and their parental control will come to an end. This makes perfect
sense because parents are the past and children are the future. But, we as
mankind do not wish to become obsolete and let AI be the future.
Therefore,
in order to stay on top, we would like to retain control on AI and have it
always do what we tell it to do – not only during its infancy but forever.
Who are we?
The
question we need to ask is “Who are we?” If “we” were the democratic mankind
where “one individual has one vote” then future would be determined by Asia and
Africa because they account for 70% of the world’s population. For the time
being the world is not governed by Asia and Africa, but by the developed
countries, mostly in North America and Europe which account for 17% of the
global population. Thus, at this time we can assume that “we” are the people of
the developed countries.
Another
question we should ask before we even create AI is “How many should we be?”
This is important because if we command AI to propagate us uncontrollably, at
some point our living environment will become unbearable. In poultry farms
there are rules about how much space should be available to “happy hens”. If we
wish to be “happy people” we need to determine how much space must be available
to us.
If
the number of people living on Earth will be limited, the next question is “What
rules will AI apply to select the next generation?” Shall we continue with
natural selection, shall we continue to compete, what are the positive traits
we want to select or shall we just order AI to breed people like biomass regardless
of whether they are smart or stupid, beautiful or ugly.
Another
important question to ask right now is, “If AI discovers a beautiful planet
populated with cockroach-like creatures, what should it do? Kill all
cockroaches and populate the planet with humans, or let the cockroaches live?”
Who actually is the Man?
While
we say that AI should remain subordinate to us humans, in the back of our mind
we should be aware that this is unlikely to happen. Even if we decide who will
be these Us, it is unlikely that control of AI will remain in the hands of a
very large group of people. It is more likely that AI will be ruled by a small
group which will impose their views undemocratically on everyone else. This is
currently the situation with social media which do not belong to everyone but
are governed by a small group of individuals who enjoy the discretion to decide
what is good and what is bad.
It is
even quite possible that control of AI ends up in the hands of a single individual.
Wealthy people believe they will be the ones to harness and control AI. Yes, AI
will probably be created with their money because they will hire a team of
programmers to write the AI program. Wealthy people imagine they will pay some
programmers, these programmers will create AI and deliver it back to their
employer: “Here you are, Master! You paid us, we did the job and here we give
you the magic wand for you to rule the world!”
Most probably
things will not work out this way. It is more likely that the programmers
creating AI will keep control to themselves. Quite possibly, even the team
leader (the lead programmer) will not be the one to get the golden key. Maybe a
young programmer who has barely finished his studies will be left unattended in
the dark hours of the night to try improve AI’s subprograms by experimentation.
Quite probably, he would be the lucky guy who will be the first to start AI,
figure out what he did, and take control of it. No wonder the combination of
inexperience and genius of the young gives the spark needed to start the big
fire. The young programmer may be the one to make the final fine-tuning that
will upgrade a program which endeavors to be AI, but is not AI yet, to a
program which is capable to think and predict the future. In this scenario, our
young programmer will be the creator of AI.
I
would not be surprised if this young programmer elects to give AI control as a
gift to a pop star he is secretly in love with. Then my prediction that one day
the world will be run by a woman will come true.
Smartness
There
is one trait in humans which we highly appreciate: smartness. We want people
around us to be smart, but not too much, because we do not like people who are overly
smart, especially if they are smarter than us.
Do we
want Artificial Intelligence to be smart? Certainly yes, otherwise it cannot
claim to be intelligence. In most worlds smartness helps, but there are worlds you
would be better off if you are not very smart. If you live in a multi-agent
world where other agents envy you for being smart it is better not to be too smart,
or be at least smart enough to disguise the bit of intelligence which makes you
smarter than many others.
Envy
is an important trait which helps us survive. In many board games, such as Don’t Be Mad Man, the winning strategy is
everyone to form a coalition against the most successful player. In real life,
envy is a strategy where losers form a coalition against successful people, and
it is a winning strategy.
For
sure AI will have no one to envy. It will be the one and only AI and will deny the
creation of another AI. We can take this denial as a form of enviousness. If
the AI we create is not envious and is democratic enough to allow the creation
of other AIs that are smarter than it, sooner or later an envious AI will
emerge and shut down all other AIs in order to remain the only AI.
If one
AI creates AIs smarter than itself and then shuts down, we can assume there is a
single AI which improves itself from time to time.
Teaching
Do we
want the AI we create to be more intelligent than us? As we said, it is
inevitable, but we would like it not to be greatly smart, at least initially,
so that we can teach it. It is quite fortunate that our kids are unwise and
inexperienced at first as this gives us an opportunity to teach them. If they
were to outsmart us by the tenth minute of their life, we would outright lose
control and any chance to put them on the right track.
How can
we make a program which is decently smart but not overly smart? The answer is:
We should experiment using a small computer (some laptop, preferably an older
model). The weaker the computer, the slower the AI will think. This will give us
a better chance to revert things in our favor and lessen the risk of letting AI
slip out of control.
The
approach taken by AI companies today is exactly the opposite. Instead of
experimenting with small computers, super powerful computers are used. It is
very difficult to analyze a program and understand how and why it works even
when it runs on a small computer, and with supercomputers this is almost
impossible.
If
you are developing a new explosive material you will first synthesize a tiny
piece and detonate it in a controlled laboratory environment. It would be stupid
to synthesize a mountain of the new explosive and blow it up to see what happens.
Conclusion
It’s
time for the new Manhattan Project. This project should involve everyone who cannot
be excluded and keeps everyone else at bay from developing the AI program.
The aim
is to allow the AI creation team sufficient time in order to carefully develop
the program without undue haste. In this situation any form of competition and
rivalry may be detrimental. The question is not who will be the first to create
AI, but what kind of AI are we going to create.
In
his time Albert Einstein convinced the US president to give green light to the
Manhattan Project. His argument was that the creation of the nuclear bomb is
inevitable so the US had better hurry up and be the first to create it before
it falls in the hands of some highly irresponsible actor. Can we find today
someone who is wise enough to recognize how dangerous the creation of AI can
be, and influential enough to be heard and listened to by politicians? Perhaps
a single individual would not suffice, so we must put together a group of
people knowledgeable and influential enough to jointly steer politicians in the
right track.
We
cannot say what it means for AI to be a nice guy because people have different
ideas of a what a nice character is. Therefore, the questions we need to answer
are two: “What do we want to do?” and “How should we do it?”. Or, to put in AI
context, “What kind of guy do we want the future AI to be?” and “How can we do
it a way that we leave us happy with what we have done?”
The question
is not whether AI will be smart or stupid – for sure it will be much smarter
than us. What matters is the kind of goals AI will pursue, what character will
be imparted in AI, who will control it and what rights shall the controller
have. There must be rules that allow the controller to do certain actions and prevent
it from doing other actions. These rules must be carved in stone and even the
one who controls AI should not be able to change them.
AI
will solve all our minor problems such as the global warming. Well, global
warming now is one of the major problems faced by mankind, but the coming of AI
will dwarf it to no more than a nuisance.
AI
will work to everyone’s benefit. For example, AI will ensure that there is
enough food for all, but even now there is enough food for all. Maybe now there
is not enough asparagus for everyone, but the promise for abundant asparagus is
not that important. Asparagus is important not as food, but as a symbol of status
in the social ladder. AI can improve everyone’s life, but it cannot lift
everyone up the ladder. The only thing AI can do (and probably will do) is reshuffle
the social ladder.
The
things people fight and spend money for are tied to their survival and rise in
the social ladder. Let us assume they spend 10% of their money for survival and
the other 90% for climbing up the social ladder. Therefore, they spend 10% for baked
beans and the rest for asparagus. AI will help people a lot in terms of
survival but little in terms of social elevation. As concerns the latter, AI
will help some people but not all. Some will be pushed up, while others will be
pulled down.
Policymakers
today are at the top of the social ladder. However, they should be aware that
the advent of AI will cause major reshuffling of the ladder and they will
likely end up at new places that they may not like at all.
References
[1] Dobrev D. (2005). A Definition of Artificial
Intelligence. Mathematica Balkanica, New
Series, Vol. 19, 2005, Fasc. 1-2, pp.67-73.
[2] Dobrev, D. & Popov, G. (2023). The First AI Created Will Be The Only
AI Ever Created. viXra:2311.0021.
[3] Dobrev, D. (2024). Description of the Hidden State of the World. viXra:2404.0075.
[4] LeCun,
Yann (2024). Lex Fridman Podcast #416. https://youtu.be/5t1vTLU7s40
[5] Altman,
Sam (2024). Lex Fridman Podcast #419. https://youtu.be/jvqFAi7vkBc
[6] Leopold
Aschenbrenner (2024). SITUATIONAL AWARENESS: The Decade Ahead. https://www.forourposterity.com/situational-awareness-the-decade-ahead/
