AI-Kindergarten: Building biological-like artificial intelligence

Video in TIB AV-Portal: AI-Kindergarten: Building biological-like artificial intelligence

Formal Metadata

AI-Kindergarten: Building biological-like artificial intelligence
Alternative Title
AI-kindergarten: Building biological-like artificial intelligence
Title of Series
Part Number
Number of Parts
CC Attribution - ShareAlike 3.0 Germany:
You are free to use, adapt and copy, distribute and transmit the work or content in adapted or unchanged form for any legal purpose as long as the work is attributed to the author in the manner specified by the author or licensor and the work or content is shared also in adapted form only under the conditions of this license.
Release Date
Production Place

Content Metadata

Subject Area
Using the theory of practopoiesis, I will define first the problem of building biological-like artificial intelligence (or strong AI) and then explain the optimal efforts needed for a successful solution. This effort will require transferring knowledge from humans to machines in a manner that is more similar to what teachers do in preschools than what programmers and engineers do in labs. Thus, to create a strong AI, we will first have to create and employ a gigantic AI-kindergarten.
Computer animation Meeting/Interview Rule of inference Metropolitan area network
Computer animation Term (mathematics) Order (biology) 1 (number) Online help Mereology Theory
Artificial neural network Interactive television Mereology Mechanism design Mathematics Computer animation Integrated development environment Term (mathematics) output Cuboid Energy level Form (programming) Physical system
Mathematics Computer animation Integrated development environment Term (mathematics) Weight Interactive television Energy level Data structure Mereology Physical system Wave packet
Mechanism design Group action Process (computing) Computer animation Integrated development environment Personal digital assistant Semiconductor memory Phase transition Energy level Theory Product (business)
Computer animation Artificial neural network Hypermedia Moment (mathematics) Energy level output Self-organization Food energy Rule of inference Theory Physical system
Greatest element Artificial neural network Decision theory Content (media) Instance (computer science) Cartesian coordinate system Rule of inference Theory Vector potential Process (computing) Computer animation Semiconductor memory Natural number Order (biology) Selectivity (electronic) Object (grammar)
Functional (mathematics) Mapping Interactive television Set (mathematics) Computational intelligence Category of being Mathematics Word Process (computing) Computer animation Integrated development environment Lecture/Conference Self-organization Energy level Flag
Point (geometry) Category of being Process (computing) Different (Kate Ryan album) Semiconductor memory Multiplication sign Adaptive behavior Species Evolute Theory
Process (computing) Computer animation Adaptive behavior Reflexive space Theory
Group action Key (cryptography) Computational intelligence Proper map Theory Arithmetic mean Database normalization Process (computing) Computer animation Integrated development environment Energy level Data structure Proxy server Error message Physical system Autopoiesis
Process (computing) Computer animation Semiconductor memory Cuboid Computational intelligence Computer programming Physical system
Computer animation Semiconductor memory Order (biology) Universe (mathematics) Mögliche-Welten-Semantik Right angle Formal language
Noise (electronics) Key (cryptography) State of matter Adaptive behavior Range (statistics) Data storage device Bit Total S.A. Number Lecture/Conference Energy level Right angle Compilation album
Medical imaging Goodness of fit Computer animation Visualization (computer graphics) State of matter Variety (linguistics) Different (Kate Ryan album) Demoscene Number
Presentation of a group Channel capacity Different (Kate Ryan album) Artificial neural network Visual system Line (geometry) Physical system Number
Point (geometry) Noise (electronics) Key (cryptography) Multiplication sign Fitness function Planning Order of magnitude Number Estimator Database normalization Computer animation Different (Kate Ryan album) Right angle
Lecture/Conference State of matter Robotics Multiplication sign Adaptive behavior Physical system
Implementation Algorithm Artificial neural network Variety (linguistics) Moment (mathematics) Adaptive behavior Physical law Type theory Computer animation Software Different (Kate Ryan album) Energy level Self-organization output Right angle Data structure
Mechanism design Process (computing) Computer animation Moment (mathematics) Virtual machine Basis <Mathematik> Evolute
Type theory Process (computing) Computer animation Semiconductor memory Multiplication sign Software developer Virtual machine Bit Basis <Mathematik> Computer-assisted translation Evolute Resultant
Process (computing) Meeting/Interview Lecture/Conference Semiconductor memory Civil engineering Software developer Virtual machine Energy level Selectivity (electronic) Flow separation Marginal distribution
Point (geometry) Frequency Computer animation Multiplication sign Virtual machine
Point (geometry) Greatest element Computational intelligence Rule of inference Theory Number Semiconductor memory Different (Kate Ryan album) Energy level Data structure Task (computing) Physical system Key (cryptography) Weight Feedback Computer simulation Bit Evolute System call Entire function Category of being Type theory Process (computing) Computer animation Integrated development environment Self-organization Species Metric system
Functional (mathematics) Goodness of fit Computer animation Energy level Mereology Perspective (visual) Physical system
Addition 4 (number) Computer animation Data structure Evolute Rule of inference Theory Physical system
Programmer (hardware) Computer animation Software Key (cryptography) Energy level Data structure Computer programming Form (programming) Physical system
Arithmetic mean Greatest element Computer animation Key (cryptography) Personal digital assistant Software developer Sampling (statistics) Interactive television Finite-state machine Data structure Physical system Number
Goodness of fit Greatest element Computer animation Real number Feedback Normal (geometry) Right angle Data structure Rule of inference Physical system
Computer animation Multiplication sign Direction (geometry) Connectivity (graph theory) Physical law Sampling (statistics) Set (mathematics) Rule of inference Form (programming) Physical system
NP-hard Point (geometry) Virtual machine Interactive television Water vapor Mereology Word Process (computing) Computer animation Integrated development environment Semiconductor memory Similitude (model) Energy level Self-organization Theorem Figurate number Data structure Condition number
Multiplication sign Surface Direction (geometry) Water vapor Bit Mereology Mechanism design Process (computing) Computer animation Integrated development environment Software Natural number Endliche Modelltheorie Physical system
Group action Different (Kate Ryan album) Robotics Multiplication sign Feedback Virtual machine Right angle Bit Proper map Wave packet
Process (computing) Computer animation Order (biology) Energy level Bit Wave packet
Mathematics Process (computing) Computer animation Integrated development environment Robotics Interactive television Virtual machine Energy level Normal (geometry) Computer simulation Bit Theory
Process (computing) Computer animation Inheritance (object-oriented programming) Strategy game Personal digital assistant Robotics Right angle Pattern language Computational intelligence Mereology Subset
Process (computing) Computer animation Semiconductor memory Virtual machine Energy level Video game Normal (geometry) Metric system Event horizon Wave packet
Process (computing) Computer animation Multiplication sign Connectivity (graph theory) Evolute Food energy Resultant
Point (geometry) Process (computing) Computer animation Information Semiconductor memory Moment (mathematics) Mereology
Process (computing) Computer animation Causality Semiconductor memory Term (mathematics) Different (Kate Ryan album) Multiplication sign Energy level Evolute
Pairwise comparison Process (computing) Computer animation Software developer Species Office suite Figurate number Evolute Event horizon Sequence
Point (geometry) Software developer Multiplication sign Feedback Similarity (geometry) Computational intelligence Evolute Rule of inference Formal language Revision control Mathematics Computer animation Video game Energy level Software testing Analytic continuation Spacetime
Functional (mathematics) Channel capacity Real number Virtual machine Sound effect Insertion loss Product (business) Goodness of fit Mathematics Process (computing) Computer animation Integrated development environment Semiconductor memory Video game Endliche Modelltheorie Physical system
Lecture/Conference Software developer Business model Number
Point (geometry) Force Personal digital assistant Keyboard shortcut Price index Mereology Computational intelligence
Computer animation Term (mathematics) View (database) Software developer Source code Virtual machine Series (mathematics) Product (business)
Domain name Word Latent heat Computer animation Multiplication sign Weight Cuboid Bit Species Event horizon Writing Physical system
Covering space Domain name Block (periodic table) 1 (number) Database Surface of revolution Number Personal computer Latent heat Computer animation Internetworking Calculation Binary multiplier
Three-dimensional space Code Image resolution Physical law Sampling (statistics) Mathematical analysis Bit Mereology Evolute Word Computer animation Different (Kate Ryan album) Order (biology) Quicksort
NP-hard Type theory Process (computing) Computer animation Bit rate Multiplication sign Decision theory Feedback Video game Resultant
Computer animation Evolute Resultant
Computer animation Evolute Theory
Film editing Computer animation Meeting/Interview Closed set Network topology Evolute Number
Computer animation
a man and a and
hello with and so on my goal today is to tell you about how to create artificial intelligence and by that I don't mean
artificial intelligence that you have in your telephones right now I mean the artificial intelligence that we would really like to have the 1 you see in
movies for example Robert that you know that you would have at home and you could send to make up the rules or to to carry new work to help you with your homework but not do the homework for just
help you with all and to do that I will have to basically so part I will tell you about the new brain-mind theory has been recently
developed I will tell you how this this theory applied to artificial intelligence or a tie I will tell you why we need something like a kindergarten in in which we will play with the rabbits in order to develop such intelligence and then I tell you something about the safety term portended AI save it doesn't when it wasn't because intelligence that it doesn't start rolling the world would be 1 still ask to be
the ones who ruled the world and so to begin
and understanding about what intelligence is about we have to think in terms of agents that interact with the world so the goal is basically to have an agent which is
and here is gray box that has a knowledge of what to do when there is interaction with the with the external world there's some sensory inputs and then the agent has to know what to do if it knows what to do it's intelligent the more it the more intelligent this that is the basic idea not the question the main question is how to get the knowledge where it where
is this knowledge gonna come from and what we have so far today is a solution of this form you have another
mechanism another system that actually also interact with the environment gets an input from the environment and then makes changes within the agent and basically teachers this level of agent on what to do normally in neuro science we call this neural plasticity where is this part would be the neural network that interacts in artificial intelligence you would call this part up of but policy it whereas the lower part would be called of something like a learning mechanism now he this 1 we can describe as you want it has only 1 level of interaction with the environment and
the other 1 you can we can describe this T 2 has 2 levels of interaction with
the environment here is an example of an artificial of the dark part is the top level that interacts with the environment has
knowledge and what to do hopefully and this orange part red-orange part is the part that learns that teachers than that reported to it makes a comparison on between what has been done and what needs to be done makes changes to the synaptic weights and teachers this
the to structure is how we think today about the brain and it's also how we think today but
artificial intelligence we think in terms of 1 leveled it performs the the behavior in another level the training is the system to to move from the
case however with the new theory tells us that we have actually have to have 3 levels we have to have a d 3 system and let me walk you
through the idea how it works in our own minds how our own minds work which which 3 levels at the lowest level we have genes genes our our basically learning mechanism of plasticity mechanisms these genes create our long-term memory this big blob that you see our our our our long-term memory however that theory because if this is not
yet finished product of an agent that can interact with the environment we use this long the memory only to adopt once again at 1 new I so the long-term memory is not the knowledge on what to do when but the knowledge and how to adapt how to learn and then we'll learn quickly with various phase groups fast process and we we create something really high up high
up which is called a working knowledge and that's knowledge that momentarily is use for this particular situation and only then can be produced the so for example if an agent has to or human has to interact with the chair has to recognize that year the traditional
theory would tell you there is a system that takes inputs and then raises some fly off here is that you what this theory tells you what we do when we see it yeah we actually reprogram the entire neural network of the brain such as to prepare it to interact with that but it which are not with the energy with that but it which that we see right now the sensory input is gone and this acquired this 3 levels
of organization and they work with different speeds the highest 1 works the fact this is the sensory motor rules the lowest 1 is the smaller 1 as the learning and the media 1 is this new 1 that is fairly fast but still it's all worth this is to meet you fast right and that whenever it goes red this media 1 this is that the moment where new new thought comes about new reconstruction of the system becomes with that we can build the whole
theory of mind explanation is called a 3 Traverso theory of the mind at the bottom we have learning rule psychological drives are instance that comes the learning process which creates a long-term memories but these long-term memories as I said this is not red network you still have to apply those in order to create a ready natural to interact in this process of application of those as our our decisions that we make are of the process of recognizable object in
detention and so on then we create the contents of working memory potential selection us is assessed things and only that can interact with and what now this has
implications of implications what does it
mean what is the spot what does it mean to
think of traditional way traditional fear of what is it mean to think basically we makes thinking equivalent to mathematical mapping function so you have some
some set a set B and a mapping between a and B from sensors for muscles that would be to traditional Weber have just the 1 process of thinking but if you have 2 levels of organization which we think then the thought is something else In the flag becomes a change in a computational properties of you know it's
it's something much as the computation it's a change of properties that occurs through interaction with the environment so in other words the
thought is an adaptation so much like you can think the theory of emotion species and that change the properties so thought is pretty much conceptually the same process just occurs much faster that's 1 difference much much faster of course and the other difference is that it is much more informed such that random guessing like evolution has to just randomly pretty much get things In our thought process we don't get we have a huge amount of knowledge we know how to adapt that's our of the memory tells us how to adapt and use practical situation but the point is that we have to add up all the time so that's this process
here in the middle and by the way this process in the mother said that fast adaptation process is the reason why we are conscious and our whatever you if you if you ask me questions always when do I know that this process happens to me and my
answer is this is the only thing you know everything that you experience is this adaptation process
you cannot experience the learning process from the very bottom you cannot experience the fast reflexes on the top but with the experience conscious process is this year this medium adaptation process there
is a recent paper that explains these things in much more detail that you can see
in the papers so reason that it actually is is if you look here about you see came from the future so I have this how I know this things I have contacted the future
now the theory called police this is derived from Greek Procter
means proxies actions of police systems creation so basically the
fundamental idea behind is that it's not the computation that we have to pay attention to but would have to pay attention
to is the process of creating new capabilities for acting for interacting with the environment so this is this adaptive process here that you see with the error is the process of for users of creation mind and this is attributed to the famous theory of
autopoiesis which started out with such a ideas OK now how does it
apply to realize what is the key to apply this theory to pay up if the if the if the theory tells us OK AI has to have if you wanna have intelligent AI we have to have 3 levels of redundancy then there are 2 basic questions 1st what is it by 2 ways it better then just to ordain just 1 we will answer that question and then the next question is more difficult to answer is how do you create 1 how do you provided this system with the proper structure knowledge and so on and that that question we will also discuss now so 1st the question
of how do we get the
white who why is the advantage of this process where is the advantage of this process and to
understand that let this 1st consider what is the advantage of having the to resist
1 what would be required if you wanted to make the intelligence system that is only the 1 that has everything here in this great box on the left so it doesn't need to learn
anything he knows everything that is there to be no 1 ever needed so maybe the 1st initial designer self of our artificial intelligence and may
have naively believed that everything could be already preprogrammed in a computer that it which is that problem program them put the roles and and you just work hard enough finally will program huge program of a computer that knows everything doesn't learn anything else well that turned out very quickly that it was not possible now he think about how big memory this computer would need to
have helping a memory with a biological system you need to have in order to be born on the planet Earth and already know everything there is needed to be known this would mean all possible worlds all possible histories of possible cultures all possible languages all combine to help bigger brain you would need but roughly this size and
here's the galaxy so it's like the size of the election probably or the entire universe to know all that right there without having good needing to learn anything new now this guy looks very confused and of course his because he knows everything about how to live on any of these planets in this galaxy but he just can't get there is too
big for so the question is OK this is the 1 so the question is then can if he
chooses them which has 1 more level of adaptation to reduce this need for huge such a huge brain to
smaller right to get less than 2 kilograms of biological might do have enough there you do
that FIL up and I did actually
some compilation look approximately to
accomplish this thing so but if you can take if you take the not total number of neurons in the brain number of synapse is in the range number of bits that this synapse considering the right you can estimate how much of the knowledge can bring store and then you can assume that everything is maximally optimize there's is no noise and so on you can cut but how many different things a brain of that size could possibly do if it is the key to break it does out it is about 10 to the 15 of states that's the 1 with 15 0 salt-like a big number but is it big enough that question is it big enough for we something data then you have to cope with another number and in the other number is how many states do
we need or the number is how many states can actually human brain and mind produce and
recognize so that's the question of how many for example different visual scenes I can understand how many different sentences I could possibly understand because this number against the the total variety of the things that my brain can do and the question is is it about 10 to the 15 there is a big not this number is difficult to estimate but it you give an example home good that it could be a big imagine I tell you the of red banana is riding on a pink costs you most people can understand it can make an image of red non obtained horse riding but how many different
combinations of such where things are are meaningfully understandable to humans
process but imagine now that you can not only that you could take this same see an trolley in many different ways painted it in many different ways of in the defeat our visual system can understand in our cognitive system can also understand that this is a red line on the writing and of the courts and today's artificial intelligence systems are completely lost present at that now what is the number of common it's it's difficult to estimate this not work but I found a way by using the known capacity of working memory on and this number that they get is
about 10 to the 20 this is 5 orders of magnitude bigger 5 this is 5 zeros difference where this is about 100 thousand times the brain we have and is a very uh very optimistic and pessimistic estimates are made sure
that the number that I don't I'm not subjective right so this is probably what we need is probably much bigger than 10 to the planning and what the brain can really create the fit was key tool agent is less than 10 to the 15 and because there must be some redundancies noises and so on so we have a problem their so had to be solved this problem so what is well let me just
point out that's of it that if
we have a T to system then we can have a smaller brains then the whole galaxy but still a big brain is still doesn't state in our had this 1 in this illustration is has a size of of a planet if it is to do so if we get from the size of a galaxy to the
size of a planet with 1 step before adaptation in the next step before adaptation In the d 3 system can bring this to the can bring to this the brain of our size and if we can do that then we can build in
principle robots that as intelligent at us and in the same time as the as I stated require as many resources as brains require and 2 no OK now the
next question How do we build so what what I
have shown you so far you may agree or not but we live in for a moment that we need a type of the 3 organizations and we need something that has this 3 levels of organization this all of this can be a global possibly flexible enough intelligent enough to create to create this variety of behavior that we need this means that today's AI algorithms that tries to stack everything in this much smaller adaptive structure cannot possibly work so if global would like to produce which deep learning networks for the use of a something that's human-like intelligent they would have to increase the size of the resources increase and increase in agrees to the size of this huge Robertson that I have that you have seen and answer have to have more flexibility pocket more flexibility is not hard to build technologically sees drew relatively easy to implement right adjusted different types of software written law of sensory inputs and so on but what is really hard to get at
is that our our brain are mine worsened the basis of genes we have G is tells us what to do and this is not just
1 learning mechanism this this is hundreds if not thousands of principles of how to build our brains from the moment we're born and and this has been accumulated through a long process of evolution that the billions of years and that's where the knowledge it's not a question is how do we get it into a ID thinking get genes something that corresponds to genes which would then and all you could call machines you know if they could get machine genome right then it would be very easy you just take a Robert with started as a little baby nourish it grow with and 2 years later it's an adult Robert knows everything about the future but the problem is how to get enough and not spend 1 billion years in in evolution but spend a few years that's what
AI kindergartens about is about tracking this problem the now
we the 1 of the checks that they acted garden uses is the fact that evolution has already done that job once and the the result of this job is already in origins store so we have to find a way to transfer the knowledge from arginines into the machine to know if you could crack that then we we don't have to repeat the evolution we just have to somehow figure out how to get knowledge from origins to the to our and you cannot just copy basis with the memories you know even if a true we have we know that you know but we don't understand it we cannot just copy basically have to do something completely different and and this completed different thing on how to cat how to the development time of some knowledge we have already used this technology quite a bit we already know this type of knowledge and it's called education and so
what I was culture knows where civilization knows is accumulated knowledge over at least several thousands of years if not longer and yet we are capable of transferring this knowledge to our kids with the 20 30 years they pick up a good chunk of it and they can operate they don't have to repeat the thousands of years of the development of human civilization there is a new method of doing it quickly of using this knowledge into the quicker and again in the education we don't copy and paste like and with the members that the knowledge we have to do some work but the work is is not thousands of years it's only furious now what we wanna what we need to do with AI engineering is a similar process but not at the level of long-term memory but at the level of the genes thank the same trick but at the level of we have to test the knowledge from margins into the machines you know not the we had any experience in doing it it should be dual and it's called selective breeding of
animals we basically through selective breeding of animals we turned what walls
Nice domicile dogs and so we have been doing it and it going of course genetically-engineered nowadays more and more but the point is that it's global also in a short period of time relative which of the how long would it have taken to turn into into DOD's if we did not interfere proba forever if you never but with our interference relatively quickly we could we could we could turn them in an optimal and also also a bunch of other
animals we could turn into our to our needs so if can we find a way to combine on the the fast
learning that we see education with gene on 1 inflation we can't we can turn became create machine gene on size of the machines do exactly what we need them as robots and that the important thing is that for that to do that it's not
enough to have a 3 agent simulated on a computer which is relatively easy task but all we need is actually the teeth for and why do we need to before because as I told you we the individual's biological individuals we are to organize the this 3 levels for him as we have the the structure however how did we come around here is through evolution and evolution process is also process of changing ourselves from changing your genes it takes also feedback from the environment and it's this for level for organization so this specie it although individual is the 3 this piece is a whole is for each of the entire biosphere employment Earth is the key for organisms in we need if we wanna make artificial intelligence that is like us we have to be exactly the same structure we the key for structure so out a 5 kindergarten is going ahead of India at for structure but before I tell you
how it will but I will give you a little bit of theory off what it takes to develop deficient out so here's a little bit more formal definition of a few you remember the structure I told you where we have a key to this is a classical AI that you have perhaps a uniform trying to recognize your voice your comments on successful sometimes not now what is the key of understand how this the human each year and this creates an understanding that 3 different types of aspects of the structure so I said at the bottom we have this learning rules and this learning rules human engineer understands very well we created me understand how they work millions 200 black and that kind of thing that that kind of knowledge recall you can call clear knowledge because it's understandable to human beings but then when you high this learning rules these artificial systems are the learn things and fill up some memory with 8 of these metrices of huge amounts of the numbers that could represent some probability that could represent synaptic weights and so on but the point is that this is no more understandable to human the human engineer not
look at it and understand what's going on we just have to use it as such this is referred to as dark knowledge start for the perspective of the of the engineer of the
creator but then there's a good news at the end at the level of behavior of the system we still can understand so although we don't understand this middle part of what the system has learned how it
performs its functions Was it before once we ran it and plight was it performs its function we can understand it at the level of intuition we could go well that's good that's bad that's interesting that's what I want and that's what I don't want to and that level we can call torque knowledge now this is 42 system if we
have a d 3 system with additional of 4th structure that we have to have the same structure at the at the end at the talk of the behavior of a what is something that we have to became judge within its behavior we like or not but is this what you want or not is this what I would do in this situation are not in the middle we have a lot of dark knowledge that we cannot understand and then at the very bottom we have to have understandable things the bottom requires of some because otherwise we cannot even begin to be begin to build front to the very bottom we have this
simple learning rules and the theory of evolution is pretty simple we understand it so in principle we can do it now you remember that do you want the idea of of a developing 1st maybe maybe artificial intelligence that you wanted humans understand
everything and program everything this will be this structure where you would have a clear knowledge and producing intuitive knowledge but that we have today in a form of softer most of our software is developed like this it's actually
developed by human programmers CW understood what they were doing why they were doing and we as the user understands it and OPEC level at the top as it at the level of behavior it makes sense to us or not we like it or not is useful or not and and and so on now when you have you wanna build a key to system really you need you need this additional additional here that's understandable then you can build a dark not something that is not understand not the
requirement is at the bottom that you have something programmable understandable and then the 2nd requirement very important for development efficient origins is to have learning samples and to have to have proper on examples so what we do in normal traditional in AI you have if you wanna for example digit assistant to recommend recognize handwriting idiot examples of correct handwriting so you give users some numbers and you at what these numbers mean and and the new system understood from the from the the interesting thing about this is that they have at the 0 structured means they don't interact with the world at all ah they're just sitting there passively to be uh waiting to be used now look at is to create T 1 agent that is intelligent we need a key to structure With the 0 samples you know
what happens with the 3 structure if you wanna create the infrastructure the 3 agent
we need the system so what we would see the same kind of rules at the bottom but what I need as a role as a as a as a learning examples and Islamic examples what we need is the is the 2 structure we need actual behavior it's not enough to give a just pictures we additional feedback on the behavior we have to tell it that's good that's bad behavior or normal don't do this this was nice good boy that's exactly what happens in the real kindergarten right you have to interact with kids who have to be told and that was
good that was that give them examples do something yourself as a feature so would it was useful so you have to build the 3 agent in the for system with 2
2 2 examples so we can generalize that and here is the law of AI creation this applies to treating any form of AI in the level of of
complex to create and and agent really key and thus fun creation system and the set of k m samples where M is
less than and if you follow this rule you can create any AI 1 the now
we follow this rule and create update I can about and this 1 has to to create we we would have 3 components the 1st component is the simplest 1 is direct engineering direct engineering means just whatever we can find a short where we happen to understand what's not waste time just build it
directly Anderson but many things we can directly understand figure it out hard words scientists figure out somewhere we just build a directly without worry but it's also all this all I can get us that's far we have to go into the into the dark knowledge and for
that we need clearer in Hubei so let's let's see what the playroom theorem is basically a process of transferring knowledge from human to machine but similar principles by which we transfer knowledge from adult human to to child or water to persons interrupting the court impact of this is that you have to transfer knowledge from the lower level of organization and take it from the top you take it from a deeper level from your long-term memory and how does it work you have 1 agent here this see laughter that interacts with the environment has its own structure the green part is the 1 you have another agent that also has its own structure but the point that the necessary condition is that they share the environment there's also a few other conditions and the need to share the balls and so on but I'm not gonna go into details and they have to share the environment if they shared the environment if they work on the same problem they start interacting with each other they become environment for each other and as they
become environment for each other there's a very nature of these adaptive systems of is such that they tend to become models of each other because they tend to become similar to each other that is just a very but I can't I don't have time now to go into the does explain why but there's just a tendency of a system that so they tend to share their knowledge the more they interact with the water we tend to share the knowledge to how they make sure
that the knowledge process in 1 direction is simple free is the learning mechanisms on 1 side and then you have enormous surfaces it has to make sure that 1 that 1 side does molar there's different ways how you can do it in katrina technological systems research just shut off some part of a software for example so it doesn't really learn any more in the in the human it's a little bit more tricky but it's
doable as well and that's what we do all the time how do you make sure when you're minute teacher in the school that kids learn from you but you don't learn from that don't become their them become childish for the lecture about that this possible right you just have to of you just have to worry about where you have to have enough knowledge and then you can do it up the so basically what we need to do
all the chance of this knowledge to to machines we have to transit different bits and pieces of knowledge of proper behavior and this means that the robbers have to be to try some things and they have to get the feedback of what is wrong what is good they have 2 of maybe observe our behavior and also try to learn from from these examples and so this ends up basically that we will have as as this technology develops from simple to more complex at the end we will have something very similar to real kindergarten where where human trainers and educators are playing in the training robots and this is 1
artistic illustration of of what of of such a player look alike in the future and this job that this will lady here's taking use maybe gonna be the job of the future of training Roberts like a teacher like a it's like what the problem is that today this is what people to do to develop now let's talk about the 3rd
party and as the incubator the the player we get all the bits of knowledge and it's not yet transferred to the level of genes and but you get many bits of knowledge will have to probably call
like thousands of neurons of of such stupid and bits of knowledge in order to cover properly or on knowledge knowledge that's the that's stored in our own genes and then you have to have
process to eat agreed all those bits and 1 big agent that nor wasn't all that takes all the bonds and the way we do
at will be again so the same principle which is called enormous stresses theory not too surprise where you will have now not having a more human robot interaction but you will have robert robert interaction but in such a way that 1 Robert is frozen has lower levels of a musician frozenset doesn't change in a more than the other Robert
doesn't have it frozen is flexible and can even change of phone machine G norm and then hand they have to shared environment in Iraq have the same syllables work together on some things and it's kind of this whole thing is computer simulated for example could be also made brought real Roberts interrupting
but problem most cases of the the
simplest way to simulate everything on a computer and up this knowledge can be transferred from small snippets of Roberts on the last that came from a from the parent into bigger and bigger and bigger a robot that has more and more knowledge here on the right and we we just accumulate that knowledge
to this to this process up not
the the way how to integrate this snippet there could be several several other ways it could be um some more optimal ways you don't you don't necessarily want to do everything everything immediately stuck on on 1 on 1 big AI 1 big Robert but you wanna maybe maker of small subsets and then bigger ones and and so on and here is 1 such strategy of how to how to build a which is a part of a pattern that have recently submitted but I'm not going to bother you with details about that but
what what is important to understand that this process in incubator of integrating knowledge is something very similar very equivalent to process of dreaming I what when we sleep and drink that's what we do we also integrate knowledge that we have accumulated through the day not this training is going to be a little different because when mainstream which aim at the level of foreign memories of long-term memory but what we don't have to do with the machines will have to dream up here it will have to dream at the level of gene on what is it you need to dream at a level of gene a usually have like a whole life don't live just a few and dream few events like a normal dreams like few things happened at half an hour 1 hour story you have you can have a whole life dreamed about and they all interact they all come to each other dreams they all live together in the so there's it remind you of something it's very similar to what you have seen in the movie metrics just the other way around what in the
more we have the machines done to us what we would have to do to develop and i is the same thing as the opposite we have to do that to the machines we have 2 kind have them in play role the interacting with real humans and then go into this incubated the dreaming machine where they will have to integrate this knowledge among themselves and they will not know what the reality is necessary but it's also that good but
it is not the and is to understand this process is gonna make that make it develop things much faster so we can roughly say
that each of these 3 components that we have is but thousand time accelerated remember what we need to accelerate
this billions of years of evolution into just years we want and nearly into once so directly engineering is a thousand times roughly accelerator because some things you don't have to really evolving just figure them out you do them directly a param is another roughly thousand time accelerator because
we don't have to repeated the experiments that evolution has died the results of these experiments are already stored energy so we have to just pick them up we have our own intuition of what is the right thing to do we just have to tell robbers and this is right that this is our intuition and that this is the right thing to do it that that way we we save a lot of that we don't have to repeat all the time the and then the incubator also
does something very important that additionally accelerates really but this process makes really faster but I would find don't moment moment just 1 only 1 1 another point the interesting point and this is the following the Tyrone the relationship between player an incubator is
very similar to the relationship that we have in our own minds and our brains between people come close can sleep in the process of sleep and dreaming are all so what we know so far so far from the size of the hippocampus is very important for fast
acquisition of explicit memory and rather they are hippocampus very busy acquiring this information but after that this information needs to be integrated with the rest of the brain mostly with the cortex and how they would do it but during the rest during the sleep during dreaming it basically repite what we have seen and lived through so the day and that way the the the whole thing gets are transferred to other parts of the brain the knowledge concerns of the other part of the brain not clear on
many cubic do very similar thing apparent collects quickly this snippets of small pieces of knowledge what is good thing to do now what is best thing and and so on you call a cause of those and then you have to up those and
integral young there's also a big difference of our brain works at the level of long term memories and the play remains to be to have to go 1 low level and work at the level of genomic DNA and not the all right now the acceleration
process how we're gonna really accelerate this thing that evolution has already has also done 1 more thing for us ready to use so we don't have to experiment in which time and this
is that the actual sequence of evolution it's a sequence of events that happened during the phylogeny of office of the species that means during the evolution process but also during ontogeny during the development of an individual and to be religious figure out what is a good thing to do 1st what can be then what fits
well on top of it the next than what fits fits well on top of X so so what's the sequence we can find out in the in the in the biological literature on comparative psychology comparative neuro science and so on and we have to follow the same
and if we go to follow the signal that we can do the following trick it happened so that the the development of the species and the developer of individual the parallel processes that the quite and not identical but are quite similar so whatever happened later late in the in the development of the species is also happens late in the development of an individual and if you use that we can use the following trick we can train our AI 1st with simple things like taxes like simple
rules avoids damage and so on and then go to more complicated like navigating space using language and so forth on but at each stage we don't have to take a new version of a gene on start from the beginning to make little baby with this gentleman have to live a whole life and make a test whether it will survive and procreate that's what evolution and and that's really
expensive it takes so much time and every change you make you have to read the the whole the whole life really the the whole of this a lot of computation lots of similar so how we think how we can crack palace of of these 2 processes the evolution of development we can actually do something that devotion couldn't be could you take the gene of adults so we could have already adult quotation adults that we could have a I developed a certain level and then you did gist now we don't have to go back to the beginning we can go just from that point and continue and test and develop and get feedback and then development the the next so that's what you see here the city the 1st
steps on the left and then the next thing you what you see is that adding more machines you know says that we'll walk out the old you know we keep it but a lot to be done changes that 1st and then we tested the genes environment walking Altener loss has the effect of walking along the memory here Collett idea taken in mention that because there is a place where we store our semantic knowledge this reminds at this long-term memory doesn't work directly to produce a behavior that is still had required is still very abstract conceptual on knowledge that we need to apply once again in the in the real life situation to produce this working knowledge and then we can interact with the environment and so it gets it gets of frozen effectively as well once we have frozen the mission gene but then when we have done that we can do to the next step and so but a very good thing to doing mean what is to unlock it for a while and have it actually work with its full capacity and this is very good because it's a protest because it very well balanced as the whole system makes the whole system function a row of good and how can we do what was the great idea because also because it's very good
makes a good model of developing a i within within certain societies because you have you have to have a step where you create a new machine but then you have to have a step where you basically applied a solution you sell it out to the sell it for as you need a product to market it may be up and then people use always and while using it and they are improving the the memory of the system and now you take this improved memory but after the use gold digger that 1 and then you change the genome of that punishment this will make the the process lab much more robust and much faster as well and that's a good
also a business model if you like to develop great technology at early stages and continuing through a comparative approach and
at the end we should be able to
create a Robert that's intelligent kind of like us that 1 question is how long would it take
the take 100 years with it take a couple of years so I don't really know of course but we do have some hints it takes 2 to produce a human with such a diploma in hand huddled hot how long
does it take 20 years at least not if it's really hard working smart kid up and we can take this number as an
indicator the that as as a as a reference point for if you want to create the human 20 20 years now if we then we have all the short cuts in the and and and all that stuff if we presumed dead over we have all the resources we need to work on it and we presume that of this becomes a popular thing as popular and as attractive as for example a PC computer what's or intended was also the was where everyone wants to use it most of the part of it everyone wants to help of then comes a lot of resources also investing financial and that case we may think cop OK in that case we can begin thinking from that point on when we have a lot of resources how long it takes 20 years is referenced by then we see how how how we all are as things go 1 up probably we cannot expect much quicker than for years but also if we have resources I don't see any reason why it would take much longer what
maybe 30 years 40 years but a
series of light would take 100 or 200 years but again I don't know so we have to try to to know the
now the last thing I wanna talk about the safety and to say important to that is to
say that it has been a lot of talk lately about AI being potentially revenges AI taking over the world of turning us into paper products now that story the is that you will develop view super-intelligent machines but give it a really attest to create a paper put as many of the proposed as possible and the mission is super intelligent and has a son in the end goal to create a paper and what it does with its superintelligence turns everything in the paper at the endoderm and kills humanity because they're not paper clips in what they prefer the term source and how do we make sure such things than mistakes don't happen and to understand that I
will I will walk you through a little bit through the history of other mentions that we've made of you know our
past and how all these inventions basically emulate human capabilities and how we get also superhuman capabilities what is it really means and these all have made may have started as a for example with the invention of hand at hand
axes is very good for opinion most a much better than human bare hands the your finger nite so you get superhuman capabilities already there immediately as soon as you have been mentioned by the by the important thing is that this is a domain specific now the next innovation bone in there really good for throwing spears immediately superhuman capabilities but for the limited domain only domain-specific and the same with a writing system run memorize words superhuman capabilities of new 304 hundred thousands of his original precisely a correctly and so on by the domain-specific pretty much everything weight in the event as time goes on this domains persists super here has a super capability but it is domain-specific and also these buddhist boxes gets smaller and smaller because they have Over time less and less impact on the ability of our species to survive so the
1st thing to have much more important than the later things a little less blocks less important but occasion their little bigger ones that happened like for example this 1 here with this could be the industrial revolution really great thing happened from that follow lots of things afterwards then another 1 maybe is the notion of a personal computer and Internet came and so on and so on or nothing's all this things our domain specific a handheld cover a hand-held
calculator is superhuman capabilities in inability to multiply numbers but it's so that's where it is limited
personal computer superhuman in the search database but that's it not it building such a i through a I kinda got maybe another important step of biggest portal thing
it could help us solve engineering problems all kinds of things up but it will necessarily be domain-specific we cannot possibly create intelligence that's beyond what we have and this is the the the the law of AI creation and code is preventing us from the the to create such
analysis we have to have all this we have to give it all the samples In other words we cannot created and we can you give only the samples of the intelligence that we have we cannot give it anything else than what we have now if you measure here that there is a 3rd dimension of some sort of intelligence that not human from some other planet or whatever we cannot give it to them we can build up because we have no capability to give them samples of that so the the the law of a Ikeda of AI creation tells us that this no way that we can build build intelligent beyond well of course you could do away with all evolution but it'll take billions of years that's a completely different story Matthew a I Kanedera we cannot do it can non-transparent knowledge in order she happened most likely is that we will just give it a part of our knowledge we will never give it everything because we will not you need we cannot give them the knowledge was really hard to give them the knowledge of what is it like to be hungry and and have a good meal afterwards they will never understand that probably because we will never build them really physiology of being hungry and all that there will be no market need for it so there always could be in little bit lost if you
try to if you have a Robert at home do things for you this really good well and then some new
situation happens that they of never knew before Bennett's this this kind of situation where it's something human that robbers don't have they will have a hard time figuring out what's going on because they'd never experienced when like we humans also understand which do not understand each other if we didn't live the same type of life and is home you
know know rate we don't know really what you know how how is it to be an answer now because we haven't been there and so the less announced can talk about themselves stuff that we cannot understand same thing with a i if they don't have the the hunger is they don't have the sexual drive is they don't fall in love really sociologically the whole the whole thing they're never gonna get it completely to they're always going to be autistic a little over and that's really good news because it there's gonna be still lots of jobs for us said a lot of the things we melody and make decisions your being yours of of our kitchen at least so all of but they will do all this rough things hard things that that are kind of boring for us we figured out and instant 1 interesting inside any more challenging that's what a I can prove kindergarten can produce it cannot possibly produce intelligence that is way of doing this kind of the result of to rule the world there will always be obedient obedience there's 1 more thing that's important we have to worry about is that we could possibly training to be an evil and of this type of feedback this is like that of someone selectively breeding
will start to all but to produce even you know walls of more aggressive bigger
no 1 has ever done so if you look at the history of like and we don't do we don't tend to do such stupid things maybe someone has tried to to breed the blue bolts like that and make them mean 1 dangerous but we never heard of him and his wife and so up we're probably not going to be so stupid to train a item to being selfish to be mean that there is this dark part of human we have to admit we go and create wars we go and try to rule the world but that's a result of our own evolution in our own circumstances in which we evolved we had to do to survive our at out of the previous but guys we didn't have this drives didn't so why we actually just 1 over there now we are in a
charge enacted the gun of the evolution of a would have things will happen movement choose like we have chosen the prize for all other technology and everything else it's it's you know
our natural instinct is to choose the the thing that's good for us and so are you money in when you educative can sometimes and when they get to be teenagers and that starts to rebelling you wanted to use the genes and with Roberts we can do that then we can make them actually had delayed obedient around so if we can't turn without
having to theorem much of a theory can turn a wolf into this little Q thank we can also make sure that our our AI is is safe so as
you know mimic conclusions when you to the tree strong AI or biological like the tyrant close to human-level intelligence AI we need 3 3 2 versus not 2 ways we have to but for that we need to create we need for tourists to create 3 to recite and infinite for tourists and to create it quickly we need a I can divide and sub due to cut billions of years of evolution into small number of years and safe
thank you very much
and will