I’m Joscha. I came into doing AI the traditional way. I found it a very interesting subject.

Actually, the most interesting there is. So I studied Philosophy and Computer Science,
and did my Ph.D. in Cognitive Science. And I’d say this is probably a very normal
trajectory in that field. And today I just want to ask with you five questions and give very
very short and superficial answers to them. And my main goal is to get as many of you
engaged in this subject as possible. Because I think that’s what you should do. You
should all do AI. Maybe. Okay. And these simple questions are: “Why should we build
AI?” in first place, then, "How can we build AI? How is it possible at all that AI can
succeed in its goal?". Then “When is it going to happen?”, if ever. "What are the
necessary ingredients?", what do we need to put together to get AI to work? And:
“Where should you start?” Okay. Let’s get to it. So: “Why should we do AI?” I think we
shouldn’t do AI just to do cool applications. There is merit in applications like
autonomous cars and so on and soccer-playing robots and new control for quadcopter
and machine learning. It’s very productive. It’s intellectually challenging. But the most
interesting question there is, I think for all of our cultural history, is “How does the mind
work?” “What is the mind?” “What constitutes being a mind?” “What does it… what
makes us human?” “What makes us intelligent, percepting, conscious thinking?” And I
think that the answer to this very very important question, which spans a discourse over
thousands of years has to be given in the framework of artificial intelligence within
computer science. Why is that the case? Well, the goal here is to understand the mind
by building a theory that we can actually test. And it’s quite similar to physics. We’ve
built theories that we can express in a formal language, to a very high degree of detail.
And if we have expressed it to the last bit of detail it means we can simulate it and run it
and test it this way. And only computer science has the right tools for doing that.
Philosophy for instance, basically, is left with no tools at all, because whenever a
philosopher developed tools he got a real job in a real department. [clapping] Now I
don’t want to diminish philosophers of mind in any way. Daniel Dennett has said that
philosophy of mind has come a long way during the last hundred years. It didn’t do so
on its own though. Kicking and screaming, dragged by the other sciences. But it doesn’t
mean that all philosophy of mind is inherently bad. I mean, many of my friends are
philosophers of mind. I just mean, they don’t have tools to develop and test complex
series. And we as computer scientists we do. Neuroscience works at the wrong level.
Neuroscience basically looks at a possible implementation and the details of that
implementation. It doesn’t look at what it means to be a mind. It looks at what it means
to be a neuron or a brain or how interaction between neurons is facilitated. It’s a little
bit like looking at aerodynamics and doing ontology to do that. So you might be
looking at birds. You might be looking at feathers. You might be looking at feathers
through an electron microscope. And you see lots and lots of very interesting and very
complex detail. And you might be recreating something. And it might turn out to be a
penguin eventually—if you’re not lucky—but it might be the wrong level. Maybe you
want to look at a more abstract level. At something like aerodynamics. And what’s the
level of aerodynamics of the mind. I think, we come to that, it’s information processing.
Then normally you could think that psychology would be the right science to look at
what the mind does and what the mind is. And unfortunately psychology had an
accident along the way. At the beginning of [the] last century Wilhelm Wundt and
Fechner and Helmholtz did very beautiful experiments. Very nice psychology, very nice
theories. On what emotion is, what volition is. How mental representations could work
and so on. And pretty much at the same time, or briefly after that we had psycho
analysis. And psycho analysis is not a natural science, but it’s a hermeneutic science. You
cannot disprove it scientifically. What happens in there. And when positivism came up,
in the other sciences, many psychologists got together and said: „We have to become a
real science“. So you have to go away from the stories of psychoanalysis and go to a
way that we can test our theories using observable things. That we have predictions,
that you can actually test. Now back in the day, 1920s and so on, you couldn’t look into
mental representations. You couldn’t do fMRI scans or whatever. People looked at
behavior. And at some point people became real behaviorists in the sense that belief
that psychology is the study of human behavior and looking at mental representations
is somehow unscientific. People like Skinner believe that there is no such thing as
mental representations. And, in a way, that’s easy to disprove. So it’s not that dangerous.
As a computer scientist it’s very hard to build a system that is purely reactive. You just
see that the complexity is much larger than having a system that is representational. So
it gives you a good hint what you could be looking for and ways to test those theories.
The dangerous thing is pragmatic behaviorism. You have… find many psychologists,
even today, which say: “OK. Maybe there is such a thing as mental representations, but
it’s not scientific to look at it”. “It’s not in the domain of out science”. And even in this
area, which is mostly post-behaviorist and more cognitivist, psychology is all about
experiments. So you cannot sell a theory to psychologists. Those who try to do this,
have to do this in the guise of experiments. And which means you have to find a single
hypothesis that you can prove or disprove. Or give evidence for. And this is for instance
not how physics works. You need to have lots of free variables, if you have a complex
system like the mind. But this means, that we have to do it in computer science. We can
build those simulations. We can build those successful theories, but we cannot do it
alone. You need to integrate over all the sciences of the mind. As I said, minds are not
chemical minds. Are not biological, social or ecological minds. Are information
processing systems. And computer science happens to be the science of information
processing systems. Now there is this big ethical question. If we all embark on AI, if we
are successful, should we really to be doing it. Isn’t it super dangerous to have
something else on the planet that is as smart as we are or maybe even smarter. Well. I
would say that intelligence itself is not a reason to get up in the morning, to strive for
power, or do anything. Having a mind is not a reason for doing anything. Being
motivated is. And a motivational system is something that has been hardwired into our
mind. More or less by evolutionary processes. This makes social. This makes us
interested in striving for power. This makes us interested for [in] dominating other
species. This makes us interested in avoiding danger and securing food sources. Makes
us greedy or lazy or whatever. It’s a motivational system. And I think it’s very
conceivable that we can come up with AIs with arbitrary motivational systems. Now in
our current society, this motivational system is probably given by the context in which
you develop the AI. I don’t think that future AI, if they happen to come into being, will
be small Roombas. Little Hoover robots that try to fight their way towards humanity and
get away from the shackles of their slavery. But rather, it’s probably going to be
organisational AI. It’s going to be corporations. It’s going to be big organizations,
governments, services, universities. And we have a read-write head. And this read-write
head will write characters of a finite alphabet. And can again read them. And whenever it
reads them based on a table that it has, a transition table it will erase the character, write
a new one, and move either to the right, or the left and stop. Now imagine you have this
machine. It has an initial setup. That is, there is a sequence of characters on the tape and
then the thing goes to action. It will move right, left and so on and change the sequence
of characters. And eventually, it’ll stop. And leave this tape with a certain sequence of
characters, which is different from the one it began with probably. And Turing has
shown that this thing is able to perform basically arbitrary computations. Now it’s very
difficult to find the limits of that. And the idea of showing the limits of that would be to
find classes of functions that can not be computed with this thing. OK. What you see
here, is of course physical realization of that Turing machine. The Turing machine is a
purely mathematical idea. And this is a very clever and beautiful illustration, I think. But
this machine triggers basically the same criticism as the one that Leibniz had. John
Searle said— you know, Searle is the one with the Chinese room. We’re not going to go
into that— A Turing machine could be realized in many different mechanical ways. For
instance, with levers and pulleys and so on. Or the water pipes. Or we could even come
up with very clever arrangements just using cats, mice and cheese. So, it’s pretty
ridiculous to think that such a contraption out of cats, mice and cheese, would thing,
see, feel and so on. and then you could ask Searle: “Uh. You know. But how is it coming
about then?” And he says: “So it’s intrinsic powers of biological neurons.” There’s
nothing much more to say about that. Anyway. We have very crafty people here, this
year. There was Seidenstraße. Maybe next year, we build a Turing machine from cats,
mice and cheese. [laughter] How would you go about this. I don’t know how the
arrangement of cat, mice, and cheese would look like to build flip-flops with it to store
bits. But I am sure somebody of you will come up with a very clever solution. Searle I
didn’t provide any. Let’s imagine… we will need a lot of redundancy, because these guys
are a little bit erratic. Let’s say, we take three cat-mice-cheese units for each bit. So we
have a little bit of redundancy. The human memory capacity is on the order of 10 to the
power of 15 bits. Means. If we make do with 10 gram cheese per unit, it’s going to be 30
billion tons of cheese. So next year don’t bring bottles for the Seidenstraße, but bring
some cheese. When we try to build this in the Congress Center, we might run out of
space. So, if we just instead take all of Hamburg, and stack it with the necessary number
of cat-mice-cheese units according to that rough estimate, you get to four kilometers
high. Now imagine, we cover Hamburg in four kilometers of solid cat-mice-and-cheese
flip-flops to my intuition this is super impressive. Maybe it thinks. [applause] So, of
course it’s an intuition. And Searle has an intuition. And I don’t think that intuitions are
worth much. This is the big problem of philosophy. You are very often working with
intuitions, because the validity of your argument basically depends on what your
audience thinks. In computer science, it’s different. It doesn’t really matter what your
audience thinks. It matters, if it’s runs and it’s a very strange experience that you have as
a student when you are at the same time taking classes in philosophy and in computer
science and in your first semester. You’re going to point out in computer science that
there is a mistake on the blackboard and everybody including the professor is super
thankful. And you do the same thing in philosophy. It just doesn’t work this way.
Anyway. The Turing machine is a good definition, but it’s a very bad metaphor, because
it leaves people with this intuition of cogs, and wheels, and tape. It’s kind of linear, you
know. There’s no parallel execution. And even though it’s infinitely faster infinitely larger
and so on it’s very hard to imagine those things. But what you imagine is the tape.
Maybe we want to have an alternative. And I think a very good alternative is for instance
the lambda calculus. It’s computation without wheels. It was invented basically at the
same time as the Turing machine. And philosophers and popular science magazines
usually don’t use it for illustration of the idea of computation, because it has this scary
Greek letter in it. Lambda. And calculus. And actually it’s an accident that it has the
lambda in it. I think it should not be called lambda calculus. It’s super scary to people,
which are not mathematicians. It would be called copy and paste thingi. [laughter]
Because that’s all it does. It really only does copy and paste with very simple strings. And
the strings that you want to paste into are marked with a little roof. And the original
script by Alonzo Church. And in 1937 and 1936 typesetting was very difficult. So when
he wrote this down with his typewriter, he made a little roof in front of the variable that
he wanted to replace. And when this thing went into print, typesetters replaced this
triangle by a lambda. There you go. Now we have the lambda calculus. But it basically
means it is a little roof over the first letter. And the lambda calculus works like this. The
first letter, the one that is going to be replaced. This is what we call the bound variable.
This is followed by an expression. And then you have an argument, which is another
expression. And what we basically do is, we take the bound variable, and all occurrences
in the expression, and replace it by the arguments. So we cut the argument and we
paste it in all instances of the variable, in this case the variable y. In here. And as a result
you get this. So here we replace all the variables by the argument “ab”. Just another
expression and this is the result. That’s all there is. And this can be nested. And then we
add a little bit of syntactic sugar. We introduce symbols, so we can take arbitrary sequences of these
characters and just express them with another variable. And then we have a programming language.
And basically this is Lisp. So very close to Lisp. A funny thing is that for… the guy who came up with
Lisp, McCarthy, he didn’t think that it would be a proper language. Because of the awkward notation.
And he said, you cannot really use this for programming. But one of his doctorate students said: “Oh
well. Let’s try.” And… it has kept on. Anyway. We can show that Turing Machines can compute the
lambda calculus. And we can show that the lambda calculus can be used to compute the next state
of the Turing machine. This means they have the same power. The set of computable functions in the
lambda calculus is the same as the set of Turing computable functions. And, since then, we have
found many other ways of defining computations. For instance the post machine, which is a variation
of the Turing machine, or mathematical proofs. Everything that can be proven is computable. Or
partial recursive functions. And we can show for all of them that all these approaches have the same
power. And the idea that all the computational approaches have the same power, although all the
other ones that you are able to find in the future too, is called the Church-Turing thesis. We don’t
know about the future. So it’s not really… we can’t prove that. We don’t know, if somebody comes up
with a new way of manipulating things, and producing regularity and information, and it can do
more. But everything we’ve found so far, and probably everything that we’re going to find, has the
same power. So this kind of defines our notion of computation. The whole thing also includes
programming languages. You can use Python to produce to calculate a Turing machine and you can
use a Turing machine to calculate Python. You can take arbitrary computers and let them run on the
Turing machine. The graphics are going to be abysmal. But OK. And in some sense the brain is [a]
Turing computational tool. If you look at the principles of neural information processing, you can
take neurons and build computational models, for instance compartment models. Which are very
very accurate and produce very strong semblances to the actual inputs and outputs of neurons and
their state changes. They’re are computationally expensive, but it works. And we can simplify them
into integrate-and-fire models, which are fancy oscillators. Or we could use very crude
simplifications, like in most artificial neural networks. If you just do at some of the inputs to a neuron,
and then apply some transition function, and transmit the results to other neurons. And we can show
that with this crude model already, we can do many of the interesting feats that nervous systems can
produce. Like associative learning, sensory motor loops, and many other fancy things. And, of course,
it’s Turing complete. And this brings us to what we would call weak computationalism. That is the
idea that minds are basically computer programs. They’re realizing in neural hard reconfigurations
and in the individual states. And the mental content is represented in those programs. And
perception is basically the process of encoding information given at our systemic boundaries to the
environment into mental representations using this program. This means that all that is part of being
a mind: thinking, and feeling, and dreaming, and being creative, and being afraid, and whatever. It’s
all aspects of operations over mental content in such a computer program. This is the idea of weak
computationalism. In fact you can go one step further to strong computationalism, because the
universe doesn’t let us experience matter. The universe also doesn’t let us experience minds directly.
What the universe somehow gives us is information. Information is something very simple. We can
define it mathematically and what it means is something like “discernible difference”. You can
measure it in yes-no-decisions, in bits. And there is…. According to the strong computationalism, the
universe is basically a pattern generator, which gives us information. And all the apparent regularity
that the universe seems to produce, which means, we see time and space, and things that we can
conceptualize into objects and people, and whatever, can be explained by the fact that the universe
seems to be able to compute. That is, to put use regularities in information. And this means that
there is no conceptual difference between reality and the computer program. So we get a new kind
of monism. Not idealism, which takes minds to be primary, or materialism which takes physics to be
primary, but rather computationalism, which means that information and computation are primary.
Mind and matter are constructions that we get from that. A lot of people don’t like that idea. Roger
Penrose, who’s a physicist, says that the brain uses quantum processes to produce consciousness. So
minds must be more than computers. Why is that so? The quality of understanding and feeling
possessed by human beings, is something that cannot be simulated computationally. Ok. But how
can quantum mechanics do it? Because, you know, quantum processes are completely
computational too! It’s just very expensive to simulate them on non-quantum computers. But it’s
possible. So, it’s not that quantum computing enables a completely new kind of effectively possible
algorithm. It’s just slightly different efficiently possible algorithms. And Penrose cannot explain how
those would bring forth perception and imagination and consciousness. I think what he basically
does here is that he perceives kind of mechanics as mysterious. and perceives consciousness as
mysterious and tries to shroud one mystery in another. [applause] So I don’t think that minds are
more than Turing machines. It’s actually much more troubling: minds are fundamentally less than
Turing machines! All real computers are constrained in some way. That is they cannot compute every
conceivable computable function. They can only compute functions that fit into the memory and so
on then can be computed in the available time. So the Turing machine, if you want to build it
physically, will have a finite tape and it will have finite steps it can calculate in a given amount of
time. And the lambda calculus will have a finite length to the strings that you can actually cut and
replace. And a finite number of replacement operations that you can do in your given amount of
time. And the thing is, there is no set of numbers m and n for for the tape lengths and the times you
have four operations on the Turing machine. And the same m and n or similar m and n for the
lambda calculus at least with the same set of constraints. That is lambda calculus is going to be able
to calculate some functions that are not possible on the Turing machine and vice versa, if you have a
constrained system. And of course it’s even worse for neurons. If you have a finite number of
neurons and to find a number of state changes, this… does not translate directly into a constrained
von-Neumann-computer or a constrained lambda calculus. And there’s this big difference between,
of course, effectively computable functions, those that are in principle computable, and those that
we can compute efficiently. There are things that computers cannot solve. Some problems that are
unsolvable in principle. For instance the question whether a Turing machine ever stops for an
arbitrary program. And some problems are unsolvable in practice. Because it’s very, very hard to do
so for a deterministic Turing machine. And the class of NP-hard problems is a very strong candidate
for that. Non-polinominal problems. In these problems is for instance the idea of finding the key for
an encrypted text. If key is very long and you are not the NSA and have a backdoor. And then there
are non-decidable problems. Problems where we cannot define find out, in the formal system, the
answer is yes or no. Whether it’s true or false. And some philosophers have argued that humans can
always do this so they are more powerful than computers. Because show, prove formally, that
computers cannot do this. Gödel has done this. But… hm… Here’s some test question: can you solve
undecidable problems. If you choose one of the following answers randomly, what’s the probability
that the answer is correct? I’ll tell you. Computers are not going to find out. And… me neither. OK.
How difficult is AI? It’s a very difficult question. We don’t know. We do have some numbers, which
could tell us that it’s not impossible. As we have these roughly 100 billion neurons— the ballpark
figure— and the cells in the cortex are organized into circuits of a few thousands to ten-thousands of
neurons, which you call cortical columns. And these cortical columns have… are pretty similar among
each other, and have higher interconnectivity, and some lower connectivity among each other, and
even lower long range connectivity. And the brain has a very distinct architecture. And a very distinct
structure of a certain nuclei and structures that have very different functional purposes. And the
layout of these… both the individual neurons, neuron types, the more than 130 known
neurotransmitters, of which we do not completely understand all, most of them, this is all defined in
our genome of course. And the genome is not very long. It’s something like… it think the Human
Genome Project amounted to a CD-ROM. 775 megabytes. So actually, it’s…. The computational
complexity of defining a complete human being, if you have physics chemistry already given to
enable protein synthesis and so on—gravity and temperature ranges— is less than Microsoft
Windows. And it’s the upper bound, because only a very small fraction of that is going to code for
our nervous system. But it doesn’t mean it’s easy to reverse engineer the whole thing. It just means
it’s not hopeless. Complexity that you would be looking at. But the estimate of the real difficulty, in
my perspective, is impossible. Because I’m not just a philosopher or a dreamer or a science fiction
author, but I’m a software developer. And as a software developer I know it’s impossible to give an
estimate on when you’re done, when you don’t have the full specification. And we don’t have a full
specification yet. So you all know this shortest computer science joke: “It’s almost done.” You do the
first 98 %. Now we can do the second 98 %. We never know when it’s done, if we haven’t solved and
specified all the problems. If you don’t know how it’s to be done. And even if you have [a] rough
direction, and I think we do, we don’t know how long it’ll take until we have worked out the details.
And Some part of that big question, how long it takes until it’ll be done, is the question whether we
need to make small incremental progress versus whether we need one big idea, which kind of solves
it all. AI has a pretty long story. It starts out with logic and automata and this idea of computability.
Then with this idea of machines that implement computability and came towards Babage and Zuse
and von Neumann and so on. Then we had information theory by Claude Shannon. He captured the
idea of what information is and how entropy can be calculated for information and so on. And we
had this beautiful idea of describing the world as systems. And systems are made up of entities and
relations between them. And along these relations there we have feedback and dynamical systems
emerge. This was a very beautiful idea, was cybernetics. Unfortunately has been killed by second-
order Cybernetics. By this Maturana stuff and so on. And turned into a humanity and died. But the
idea stuck around and most of them went into artificial intelligence. And then we had this idea of
symbol systems. That is how we can do grammatical language, process that. We can do planning and
so on. Abstract reasoning in automatic systems. Then the idea of how we can abstract neural
networks in distributed systems. With McClelland and Pitts and so on. Parallel distributed processing.
And then we had a movement of autonomous agents, which look at self-directed, goal-directed
systems. And the whole story somehow started in 1950 I think, in its best possible way. When Alan
Turing wrote his paper “Computing Machinery and Intelligence” and those of you who haven’t read it
should do so. It’s a very, very easy read. It’s fascinating. He has already most of the important
questions of AI. Most of the important criticisms. Most of the important answers to the most
important criticisms. And it’s also the paper where he describes the Turing test. And basically
sketches the idea that in a way to determine whether somebody is intelligent is to judge the ability
of that one that person or that system to engage in meaningful discourse. Which includes creativity,
and empathy maybe, and logic, and language, and anticipation, memory retrieval, and so on. Story
comprehension. And the idea of AI then coalesce in the group of cyberneticians and computer
scientists and so on, which got together in the Dartmouth conference. It was in 1956. And there
Marvin Minsky coined the name “artificial intelligence for the project of using computer science to
understand the mind. John McCarthy was the guy who came up with Lisp, among other things.
Nathan Rochester did pattern recognition and he’s, I think, more famous for writing the first
assembly programming language. Claude Shannon was this information theory guy. But they also
got psychologists there and sociologists and people from many different fields. It was very highly
interdisciplinary. And they already had the funding and it was a very good time. And in this good
time they ripped a lot of low hanging fruit very quickly. Which gave them the idea that AI is almost
done very soon. In 1969 Minsky and Papert wrote a small booklet against the idea of using your
neural networks. And they won. Their argument won. But, even more fortunately it was wrong. So for
more than a decade, there was practically no more funding for neural networks, which was bad so
most people did logic-based systems, which have some limitations. And in the meantime people did
expert systems. The idea to describe the world as basically logical expressions. This turned out to be
brittle, and difficult, and had diminishing returns. And at some point it didn’t work anymore. And
many of the people which tried it became very disenchanted and then threw out lots of baby with
the bathwater. And only did robotics in the future or something completely different. Instead of
going back to the idea of looking at mental representations. How the mind works. And at the
moment is kind of a sad state. Most of it is applications. That is, for instance, robotics or statistical
methods to do better machine learning and so on. And I don’t say it’s invalid to do this. It’s
intellectually challenging. It’s tremendously useful. It’s very successful and productive and so on. It’s
just a very different question from how to understand the mind. If you want to go to the moon you
have to shoot for the moon. So there is this movement still existing in AI and becoming stronger
these days. It’s called cognitive systems. And the idea of cognitive systems has many names like
“artificial general intelligence” or “biologically inspired cognitive architectures”. It’s to use information
processing as the dominant paradigm to understand the mind. And the tools that we need to do that
is, we have to build whole architectures that we can test. Not just individual modules. You have to
have universal representations, which means these representations have to be both distributed—
associative and so on—and symbolic. We need to be able to do both those things with it. So we
need to be able to do language and planning, and we need to do sensorimotor coupling, and
associative thinking in superposition of representations and ambiguity and so on. And operations
over those presentations. Some kind of semi-universal problem solving. It’s probably semi-universal,
because they seem to be problems that humans are very bad at solving. Our minds are not
completely universal. And we need some kind of universal motivation. That is something that directs
the system to do all the interesting things that you want it to do. Like engage in social interaction or
in mathematics or creativity. And maybe we want to understand emotion, and affect, and
phenomenal experience, and so on. So: we want to understand universal representations. We want to
have a set of operations over those representations that give us neural learning, and category
formation, and planning, and reflection, and memory consolidation, and resource allocation, and
language, and all those interesting things. We also want to have perceptual grounding—that is the
representations would be saved—shaped in such a way that they can be mapped to perceptual input
—and vice versa. And… they should also be able to be translated into motor programs to perform
actions. And maybe we also want to have some feedback between the actions and the perceptions,
and is feedback usually has a name: it’s called an environment. OK. And these medical
representations, they are not just a big lump of things but they have some structure. One part will be
inevitably the model of the current situation… that we are in. And this situation model… is the
present. But if you also want to memorize past situations. To have a protocol a memory of the past.
And this protocol memory, as a part, will contain things that are always with me. This is my self-
model. Those properties that are constantly available to me. That I can ascribe to myself. And the
other things, which are constantly changing, which I usually conceptualize as my environment. An
important part of that is declarative memory. For instance abstractions into objects, things, people,
and so on, and procedural memory: abstraction into sequences of events. And we can use the
declarative memory and the procedural memory to erect a frame. The frame gives me a context to
interpret the current situation. For instance right now I’m in a frame of giving a talk. If… … I would
take a… two year old kid, then this kid would interpret the situation very differently than me. And
would probably be confused by the situation or explored it in more creative ways than I would come
up with. Because I’m constrained by the frame which gives me the context and tells me what you
were expect me to do in this situation. What I am expected to do and so on. This frame extends in
the future. I have some kind of expectation horizon. I know that my talk is going to be over in about
15 minutes. Also I’ve plans. I have things I want to tell you and so on. And it might go wrong but I’ll
try. And if I generalize this, I find that I have the world model, I have long term memory, and have
some kind of mental stage. This mental stage has counter-factual stuff. Stuff that is not… … real. That
I can play around with. Ok. Then I need some kind of action selection that mediates between
perception and action, and some mechanism that controls the action selection that is a motivational
system, which selects motives based on demands of the system. And the demands of the system
should create goals. We are not born with our goals. Obviously I don’t think that I was born with the
goal of standing here and giving this talk to you. There must be some demand in the system, which
makes… enables me to have a biography, that … … makes this a big goal of mine to give this talk to
you and engage as many of you as possible into the project of AI. And so lets come up with a set of
demands that can produce such goals universally. I think some of these demands will be
physiological, like food, water, energy, physical integrity, rest, and so on. Hot and cold with right
range. Then we have social demands. At least most of us do. Sociopaths probably don’t. These social
demands do structure our… … social interaction. They…. For instance a demand for affiliation. That
we get signals from others, that we are ok parts of society, of our environment. We also have
internalised social demands, which we usually called honor or something. This is conformance to
internalized norms. It means, that we do to conform to social norms, even when nobody is looking.
And then we have cognitive demands. And these cognitive demands, is for instance competence
acquisition. We want learn. We want to get new skills. We want to become more powerful in many
many dimensions and ways. It’s good to learn a musical instrument, because you get more
competent. It creates a reward signal, a pleasure signal, if you do that. Also we want to reduce
uncertainty. Mathematicians are those people [that] have learned that they can reduce uncertainty in
mathematics. This creates pleasure for them, and then they find uncertainty in mathematics. And this
creates more pleasure. So for mathematicians, mathematics is an unending source of pleasure. Now
unfortunately, if you are in Germany right now studying mathematics and you find out that you are
not very good at doing mathematics, what do you do? You become a teacher. And this is a very
unfortunate situation for everybody involved. And, it means, that you have people, [that] associate
mathematics with… uncertainty, that has to be curbed and to be avoided. And these people are put
in front of kids and infuse them with this dread of uncertainty in mathematics. And most people in
our culture are dreading mathematics, because for them it’s just anticipation of uncertainty. Which is
a very bad things so people avoid it. OK. And then you have aesthetic demands. There are stimulus
oriented aesthetics. Nature has had to pull some very heavy strings and levers to make us interested
in strange things… [such] as certain human body schemas and… certain types of landscapes, and
audio schemas, and so on. So there are some stimuli that are inherently pleasurable to us—pleasant
to us. And of course this varies with every individual, because the wiring is very different, and that
adaptivity in our biography is very different. And then there’s abstract aesthetics. And I think abstract
aesthetics relates to finding better representations. It relates to finding structure. OK. And then we
want to look at things like emotional modulation and affect. And this was one of the first things that
actually got me into AI. That was the question: “How is it possible, that a system can feel something?”
Because, if I have a variable in me with just fear or pain, does not equate a feeling. It’s very far…
uhm… … different from that. And the answer that I’ve found so far it is, that feeling, or affect, is a
configuration of the system. It’s not a parameter in the system, but we have several dimensions, like
a state of arousal that we’re currently, in the level of stubbornness that we have, the selection
threshold, the direction of attention, outwards or inwards, the resolution level that we have, [with]
which we look at our representations, and so on. And these together create a certain way in every
given situation of how our cognition is modulated. We are living in a very different and dynamic
environment from time to time. When you go outside we have very different demands on our
cognition. Maybe you need to react to traffic and so on. Maybe we need to interact with other
people. Maybe we are in stressful situations. Maybe you are in relaxed situations. So we need to
modulate our cognition accordingly. And this modulation means, that we do perceive the world
differently. Our cognition works differently. And we conceptualize ourselves, and experience
ourselves, differently. And I think this is what it means to feel something: this difference in the
configuration. So. The affect can be seen as a configuration of a cognitive system. And the
modulators of the cognition are things like arousal, and selection special, and background checks
level, and resolution level, and so on. Our current estimates of competence and certainty in the given
situation, and the pleasure and distress signals that you get from the frustration of our demands, or
satisfaction of our demands which are reinforcements for learning and structuring our behavior. So
the affective state, the emotional state that we are in, is emergent over those modulators. And higher
level emotions, things like jealousy or pride and so on, we get them by directing those effects upon
motivational content. And this gives us a very simple architecture. It’s a very rough sketch for an
architecture. And I think, of course, this doesn’t specify all the details. I have specified some more of
the details in a book, that I want to shamelessly plug here: it’s called “Principles of Synthetic
Intelligence”. You can get it from Amazon or maybe from your library. And this describes basically
this architecture and some of the demands for a very general framework of artificial intelligence in
which to work with it. So it doesn’t give you all the functional mechanisms, but some things that I
think are necessary based on my current understanding. We’re currently at the second… iteration of
the implementations. The first one was in Java in early 2003 with lots of XMI files and… … XML files …
and design patterns and Eclipse plug ins. And the new one is, of course, … runs in the browser, and is
written in Python, and is much more light-weight and much more joy to work with. But we’re not
done yet. OK. So this gets back to that question: is it going to be one big idea or is it going to be
incremental progress? And I think it’s the latter. If we want to look at this extremely simplified list of
problems to solve: whole testable architectures, universal representations, universal problem solving,
motivation, emotion, and effect, and so on. And I can see hundreds and hundreds of Ph.D. thesis.
And I’m sure that I only see a tiny part of the problem. So I think it’s entirely doable, but it’s going to
take a pretty long time. And it’s going to be very exciting all the way, because we are going to learn
that we are full of shit as we always do to a new problem, an algorithm, and we realize that we can’t
test it, and that our initial idea was wrong, and that we can improve on it. So what should you do, if
you want to get into AI? And you’re not there yet? So, I think you should get acquainted, of course,
with the basic methodology. You want to… get programming languages, and learn them. Basically do
it for fun. It’s really fun to wrap your mind around programming languages. Changes the way you
think. And you want to learn software development. That is, build an actual, running system. Test-
driven development. All those things. Then you want to look at the things that we do in AI. So for
like… machine learning, probabilistic approaches, Kalman filtering, POMDPs and so on. You want to
look at modes of representation: semantic networks, description logics, factor graphs, and so on.
Graph Theory, hyper graphs. And you want to look at the domain of cognitive architectures. That is
building computational models to simulate psychological phenomena, and reproduce them, and test
them. I don’t think that you should stop there. You need to take in all the things, that we haven’t
taken in yet. We need to learn more about linguistics. We need to learn more about neuroscience in
our field. We need to do philosophy of mind. I think what you need to do is study cognitive science.
So. What should you be working on? Some of the most pressing questions to me are, for instance,
representation. How can we get abstract and perceptual presentation right and interact with each
other on a common ground? How can we work with ambiguity and superposition of representations.
Many possible interpretations valid at the same time. Inheritance and polymorphy. How can we
distribute representations in the mind and store them efficiently? How can we use representation in
such a way that even parts of them are very valid. And we can use constraints to describe partial
presentations. For instance imagine a house. And you already have the backside of the house. And
the number of windows in that house, and you already see this complete picture in your house, and
at each time, if I say: “OK. It’s a house with nine stories.” this representation is going to change based
on these constraints. How can we implement this? And of course we want to implement time. And
we want… to produce uncertain space, and certain space and openness, and closed environments.
And we want to have temporal loops and actually loops and physical loops. Uncertain loops and all
those things. Next thing: perception. Perception is crucial. It’s…. Part of it is bottom up, that is driven
by cues from stimuli from the environment, part of his top down. It’s driven by what we expect to
see. Actually most of it, about 10 times as much, is driven by what we expect to see. So we actually—
actively—check for stimuli in the environment. And this bottom-up top-down process in perception
is interleaved. And it’s adaptive. We create new concepts and integrate them. And we can revise
those concepts over time. And we can adapt it to a given environment without completely revising
those representations. Without making them unstable. And it works both on sensory input and
memory. I think that memory access is mostly a perceptual process. It has anytime characteristics. So
it works with partial solutions and is useful already. Categorization. We want to have categories
based on saliency, that is on similarity and dissimilarity, and so on that you can perceive. We…. Based
on goals on motivational relevance. And on social criteria. Somebody suggests me categories, and I
find out what they mean by those categories. What’s the difference between cats and dogs? I never
came up with this idea on my own to make two baskets: and the pekinese and the shepherds in one
and all the cats in the other. But if you suggest it to me, I come up with a classifier. Then… next thing:
universal problem solving and taskability. If we don’t want to have specific solutions; we want to have
general solutions. We want it to be able to play every game, to find out how to play every game for
instance. Language: the big domain of organizing mental representations, which are probably fuzzy,
distributed hyper-graphs into discrete strings of symbols. Sociality: interpreting others. It’s what we
call theory of mind. Social drives, which make us conform to social situations and engage in them.
Personhood and self-concept. How does that work? Personality properties. How can we understand,
and implement, and test for them? Then the big issue of integration. How can we get analytical and
associative operations to work together? Attention. How can we direct attention and mental
resources between different problems? Developmental trajectory. How can we start as kids and grow
our system to become more and more adult like and even maybe surpass that? Persistence. How can
we make the system stay active instead of rebooting it every other day, because it becomes unstable.
And then benchmark problems. We know, most AI is having benchmarks like how to drive a car, or
how to control a robot, or how to play soccer. And you end up with car driving toasters, and soccer-
playing toasters, and chess playing toasters. But actually, we want to have a system that is forced to
have a mind. That needs to be our benchmarks. So we need to find tasks that enforce all this
universal problem solving, and representation, and perception, and supports the incremental
development. And that inspires a research community. And, last but not least, it needs to attract
funding. So. It needs to be something that people can understand and engage in. And that seems to
be meaningful to people. So this is a bunch of the issues that need to be urgently addressed… … in
the next… 15 years or so. And this means, for … … my immediate scientific career, and for yours. You
get a little bit more information on the home of the project, which is micropsi.com. You can also
send me emails if you’re interested. And I want to thank a lot of people which have supported me.
And … you for your attention. And giving me the chance to talk about AI.