 and vice-versa. Ladies and gentlemen, please be reserved in the national organizations and acknowledging trust at the department and make sure that you sit down, please. And make sure that the University of Siena of the so called the University of Siena is going to take the floor. Ladies and gentlemen, to open this ceremony of the University of Siena at the opening of this ceremony I have the opportunity to extend my greetings and heartfelt thanks to the author present as it's her and to all those who have voted on the organization of this ceremony greetings and heartfelt thanks not merely for family who are the distinguished guests Professor Jann Lecun Professor Lecun University in Cannons who will go into the area. We are pleased to be honored by your participation in this ceremony at the University of Siena. A few days ago, I have started to think about this speech and I have started to like it yesterday I have asked people to challenge me to give me some help and I have asked for time to write to use the loudest in 30 seconds I have asked but believe me I don't like it but my speech is extremely meaningful personally because I would like to start telling you something about the University of Siena the University of Siena is one of the oldest universities in the world with nearly 80 centuries of history the first talk when days back to December when you see 1140 we belong the group of the oldest university the place where we are they one century after the establishment of the University of Siena it was in the 13th century 1340 our university is inseparably linked to its city and its territories yet it has always been a part of an international network of relationships its first students in the 13th century who are also foreign students of German origin the University of Siena is a decision characterised by a market wide variety of indigenous students operating in numerous areas of science and language from experimental science of economics and social sciences from medical sciences to literacy philosophical history the University of Siena whose numerous excellence envires digital knowledge from humanities to history from life sciences to medicine from super sciences to engineering these excelencies connect us to the cultural context the business world and the product district of Germany and in Germany in the specific field of engineering we have a history and a growth especially the people computer engineering in performance and outcomes of the business intelligence this achievement has not only impacted our study programmes master of science and PhD programmes but also research activities and their dissemination of knowledge thereby creating numerous relationships between the business world these relationships in the Tasmanian region have led to the creation of the company network called SEA with the contribution of the Montenegro-Fasci-Siena Foundation on the one these relationships in the same region will allow us to acquire a super computer through the creation of a specific joint venture between the University of Siena and the network of cameras with the total investment of over a million volumes the Honorary degree we have the pleasure of comparing this morning perfectly aligned with their research field that characterised the University of Siena and the people of artificial intelligence the Honorary degree constitutes recognition of the journey the immigration of experience and knowledge and many achievements made available to the institutional scientific community the Honorary degree is an act of recognition and a form of gratitude for the contribution made to science and knowledge this decision was made by the University of Siena when this year we decided to compare the degree in artificial intelligence as an automation engineering to prove this to be done at home fulfilled by the research of the University of Siena last month or later we had made this decision because like all the universities in this country our institution needs a place of science, knowledge and culture and it must remain so in the face of the attempt to legitimise self-scientists and non-scientists today the University of Siena a region which is an online association with a new graduate who thankfully became part of our community finally let me invite our guests and all the participants to enjoy our saluting you all for attending these moments in our University thank you so much Professor Valerio Vignoli Head of Department of Information Engineering and Mathematics is going to take the floor and read the motivation for the University of Siena to confer a normally degree in artificial intelligence and automation engineering Professor Jantler Kuhn A student of mine who leads distinguished research and the pleasure for me to be here today to introduce the award of the LARIA of knowledge artificial intelligence to commission this award was strongly desired by the Department of Environmental Engineering for the outstanding scientific and provided solution. A researcher who has given and is the significant contribution in several research teams, starting from machine learning to computer vision, mobile audience, professional neuroscience, just to be in addition to his academic achievements, his engineers, he have received the attention being by a professor who knows practical applications. On this point, I just want to mention the field of optical character recognition that we have all experienced. The same thing provided by the professor, evidenced by the number of recognitions he has received over the course of time. Among these recognitions, one stands out. They do not work. He received the award in 2018, together with his colleagues George Mulvaney and Joe Green, for the theoretical and the engineering practicals in the field of team development work. During the award, which is now also known as the Nobel Prize in good science, it recognizes him as one of the most influential researchers within the market vision. In a few minutes, I will turn to the fact that we outline the reasons why we are gathered here today to compare this. Before leaving the floor, I would like to express my sincere hope that this event, and all with a remarkable achievement of Professor Wilma, will serve as the inspiration for the future generations and the research of my department and of our students. Professor Manco Gori of the Department of Information Engineering and Mathematics is invited to reach the training for the land action. First of all, I would like to thank the partners present, and all the research involved in this work of a new, symbolic learning history and learning in the practical sciences with the other parallel work. We are going to recognize today as a parallel work for the companies which has been recognized by the Cien International Foundation, and I would also like to thank the companies from the University of Cien. As you can see in the picture, the prestigious Grand Institute for Mathematics and Science is also a sheet scientist. And our research activity has been providing a fruitful contribution to the foundations of the overview of the AI and has provided an impressive impact on companies and society. As previously mentioned by Professor Vignoni, who is one of the most distinguished achievements in this three number that is received by the other achievements. So then I'm just saying, oh, it's a confusion that I'm really doing the same course in such a complex society, because whether it's pioneering in establishing foundations on neural networks and AI, and the second mark is, at least for me, very important, maybe because of my age and maybe because I was studying my energy in the mirror, and to which professors can be so important on the pitch. So let's start briefly with the impact on companies and society. Well, I'd like to point out that has it since the technologies on large range models have appeared in the web, I think the advice of who through us work with information has been actively transported. Maybe most perhaps even what needs to be can be nowadays provided by intelligent agents which are however still exposed to new states. Regardless of the position by the matric, on the impact of these technologies on society, we believe that the non-imperial generative of almost 20 years of land and tradition has been having a profound impact on society, education, and human relations. But the explosion we need to see in artificial intelligence, I think is when we own a recent education that's childhood, the impact in other fields in each of these robotics of communication has been changing up a lot. The adoption of artificial intelligence or climate recognition and especially the impact in medicine and biology can dramatically contribute to the improvement of the quality of climate. Now, if you can design this what contributed most to this development of AI on the teardrops of 100, I think the only non-screw professional development spending through it all and creating at their labs by focusing on new techniques and learning algorithm and providing, at the same time, I think in extraordinary edits or engineering skills. A remarkable example which was quickly also mentioned by Pennsylvania's type of research can serve treatment and on treatment less. So, the technology which was based on was licensed by specialist providers or by assistants such as national education system and it was massively used in the United States. Now, interestingly the technology behind these systems were people like me who teach machine learning from years of the university you know, connected immediately and really might be popular at least in the character recognition database, which is now associated with this game widely used in machine learning courses who could learn the world. So, it is a fantastic plane I think primarily with the consumption and the use of large language models and overall, fortunately, the AI could be have a look at the large language models made for example the Bama model and overall, I would say throughout his career he has built a very refined reputation for converting ideas into technologies. That is probably one of the most single way of distinguishing professoring. I mean, now quickly you are about his pioneering role in the foundation of deep learning and artificial intelligence. While in large his most impactful contributions can be in conjunction with the extortion of the environment as well as one of the major protagonists behind this technological impact. He has been one of those deep convolutional neural networks. A class of neural networks that has become widely the most popular eventually in real-world applications. On 2015 together we judge the game to be eventually published on the page of essay and paper which has been having impressive influence in artificial intelligence and cognitive science and among others it has even brought a lot of applications in the field of engineering, medicine, and biology. As yet the paper has received as many as 66,000 citations Scholar Google citations, that is speaking, we are talking about 35 citations per page. The paper discusses the importance of the internal representations that have developed in neural networks which is in fact one of the secrets for boosting performance that we have experimented in the last decade. Now I like also to mention in my opinion that Tesla can play on the foundations of neural networks during the so-called first overview So the best connection is way is typically marked in the 80s and in order to better grasp the crucial pioneering role that is spread in AI one should be in mind that in the 80s the connectionist approach was quite much For example one that is not as successful in artificial intelligence by African twins from the Massachusetts Institute of Technology which at that time was sort of vital for the DC research and vaccine position in the global field of academic life of scientists who have decided to follow the arc of neural networks was not easy at that time since the field was regarded as failure of AI which was quite failure also in computer science This might have been the relevance of the field that was mostly perceived in Europe and surely in England On October 1980 I met the young man at the International and Operates Connectionist in Perspective One of the first events which I did at home was the first connectionism after an entire period following the publication of the seminal book and it was immediately revealed to me that despite his young age young was destined to be an protagonist in a new field of research that I had to follow in law with He presented the paper entitled Generalization and Neutral Science Strategies where the notion of sedentary innovation that must be used for the commercial land where already here we need to be successful in the experiment at that time When considering the contribution to science I believe that in addition to the widely accepted parametric behaviors one should be able to identify the new distinct achievements There is no doubt that Perspective would have set the primary role in both fabrication and compulsion of neural mentors that have changed the date of A.I. in the last decades Professor Dukui began setting the game against the wind of discourage the adoption of huge networks because of parametric editing which was one of the research that I believed some understanding coherence patterns in promotion of the connectionism approach aimed to each specificity then pioneering hope of the impact of creation and almost in contrast with deep neural mentors of huge dimensions the clear sign of his long third history and strong foundational capability however his research path is also a mainly characterized by his extraordinary capability in the energy of large-scale systems which is in fact one of the major reasons why the field of artificial intelligence is so important about this he recently proposed a path towards autonomous machine intelligence where machines learn more than humans and are driven by intrinsic objectives rather than by hardwired foreign external supervision or external reward these studies arise from his claim that scaling up on neural architectures like in current large language models is not enough for capturing many interesting and subtle aspects of human intelligence this path on autonomous machine intelligence should really intersect the research interest of many participants to the workshop on neural symbolic learning and reasoning which is about style so like other top level scientists in the history of science I think professor began his career sailing against the wind against the winds but ultimately strongly contributed plant in the current exciting territory of the AI in conclusion I want to you young who expresses my gratitude for having accepted our invitation to offer our volume this part of the moment we are enthusiastic and eagerly looking forward to follow your information into the stars young fun to finish the days and deliver his dissertation entitled the present and future of artificial intelligence I was this time human it's the emotion so much the fact that you are more each other for being by your side and this is me this is a year again this is a real pleasure to be here it's a full site and I'm going to tell you a little bit about the present and future of AI if I don't see it and that has kind of dominated the future of AI okay so today AI means machinery and if there is something that it seems as long as press that's creating as a part of my career it's the idea that AI is it's very hard to just write it from scratch write it forward and it is intelligent what we can do is imitate to some extent the inspiration from biology and industries who make themselves intelligent and serve biology and every biological system that has brain is capable of learning in fact even a lot of organisms can be able to predict that but I don't see the stability of that I hope that as he was making that I decided to design the AI system I decided that I should design the cells and it started working on machinery at the time in the early years actually that machinery was not it actually existed as a field and was used as I say in fact most of the efforts of machinery in the time frame were due to the start of the domain due to the start of the state of the signal and it filters the field to the engineering but then it's here as kind of a close to the AI and so this was a very low time for engineers until the late 80s when there was a normal but really today AI is machinery and it's really three times the machinery that existed that are being used practice or just research and so it's really important for the system to simulate the system of the situation and the way the machinery works is by trying to try something and then get an indication from the environment of the teacher to whether the identity was real bad if the number of items that are possible is very large this process is extremely slow people thought and first that it was with millions of money because it's extremely efficient so it's useful for games because in games we can get millions of games against itself essentially but in real life primary what's broken is too slow the second type of money which by far was the main one course of AI and so recently for the last 10 years it has it's still supervised money since when it has been used for training the machine to recognize for example, or the next day or for such recognition and there you can tell the machine whether the correct answer is which machine to produce an answer the answer is correct the correct answer would be to find your internal structure so that next time the same training example is an answer that is closer to the one we want that's much more efficient because you're getting more efficient in the system and then there's a third movement which basically has become dominant for the past few years you see examples of this in geometry of AI but there's other bases as well that still self-supervive in the team so something I'd be like competing for something like 80 years or so just the idea that the system is not trained in tasks other than capturing the structure of the data that it observes for example like predicting the future so we show the system the piece of the video and ask the system what is the result of the next or show the system the piece of text and ask it to predict the words that are coming next this is the way the government resembles the train so in order to have a task or anything, the system learns to represent the structure of the data by just predicting in and it's really thinking about the world but it's not productive as we need first I really wasn't to say advice to everybody but it's easier to understand because it's so much more so let's say I need to train the system to distinguish images of the forest or images of the caravan which connects millions of images of cars and then show the image of the car of the system ask it to predict an output if the output is car from the meeting if it says an airplane and it says something else which has the knobs of the system so then the output is completely gone and this is a brilliant example indeed to believe possibly and in the end the machine learning such that the system essentially converts the set of the system airplane and is able to recognize the process of the airplane even for examples and this works really well for images but it also distracts the transmission in the long-term mathematically the way we represent this problem is the problem of mathematical optimization so we can the distance we basically measure our wellness systems performing by measuring the distance of the time the system uses with the action of the water in the distance over millions of examples and then a single value the goal of the class what we need to do is the system so that we find the value of those parameters that he might set us that the system most of the time serves the correct and we do this that's so essentially imagine yourself somewhere the Italian house and there is fog you're lost in the mountains and you're trying to look for the valley this fog you can't see anything around so what you do is you turn around and you figure out which way it goes down down and you take a step down and you keep taking steps down the hill until you reach the valley that is a very slight scaling which of course isn't like it so in the case here everybody here we make the entrance so the question now he goes in the direction of the system and that's where the background is in so what do we do with that what are the parameters we're going to adjust that's what we've done with the inspiration for biology this is an inspiration that goes back to the 1950s in fact that was in the 40s the idea that intelligence is a version of the phenomenon in the brain interaction between very very large number of very simple parameters in the neurons the neurons in the brain if you simplify the function of the basically a very simple function basically with the average with some of the variables and then compare the average to the threshold it's larger than the threshold then the neuron gets activated it's it's deactivated and then actually it's very simple it's going to accept products and editions so take a moment and look at all the inputs computer systems with weights weights that can also be learned then compare the average to the threshold the spotless threshold the average to the threshold is below the threshold at the same moment it's very simple the only operations we need for this is the multiplication addition if you can do billions of billions of them per second you can learn very much more than that and that's basically what it will mean so it's called incrementing because you build the systems by stacking multiple layers and those systems may have billions, billions hundreds of billions starting to get into the trillions of those adjustable parameters which are the weights that connect to those products so the way this works is you put an input to the output of the network telling the average it's supposed to reduce and then figure out in which way the average to adjust the parameters consists of the weights that connect to the network so that the output is constant and that's not the multiplication algorithm it can work basically the transcription of something invented by the order it's basically derivative of the function it's very simple so nothing very sophisticated so it's a mathematical experiment powerful and basically probably it signals the networks to indicate how much to adjust the parameter in the system so that the other distance is constant so the next question is why do you connect with each other so then the system can usefully solve the problem like they say image recognition in this case image recognition and that's really when you have a function on the network so this is something important in the 1980s and we started working on it then and basically the architecture of how an artificial man is is inspired by the structure of the visual cortex in mammals mammals in AI as vision and noise the cortex is represented in a small area in the visual field and it detects very simple on the and that detection is replicated in a very good view that's automatically that's the operation which is required in the networks and those are actually successful and you can apply them to anything else because we just didn't have any data that was available in the computer it's not very powerful in this one and so we could recognize digits we could recognize multiple digits this is an example of what would system apply and then the field of lost interest is going to take the mid-19th and for almost 10 years most people dismissed in all that as the useless, interesting but not really something that should be studied seriously and it's because partly the theory to explain the function was very difficult very complex objects and even in theory it's only sort of a cement representation so people dismissed it because they could not write a theory for it for instance, other machine learning which is good to understand but in each time in 2012 I think each of them started implementing commercial nets on the GDU so the same processors that I used roughly which were starting to be very common and those systems are those are designed to perform multiplication in addition to present history within so they were really more creative for the past and they were really cheap because they were designed for gamers essentially so they implemented the commercial nets pretty good results on the standard benchmark in the internet in 2012 and this kind of revolutionized the field of revolution and anything like writing assistance systems for cars in Europe nowadays it was called an ADES Atlantic Emergency Rating System which means detect obstacles and cause the cars to break those commercial nets 80% of the market is not a popular item Israeli company and then various other players so the same it reduces collisions to represent the stage lights it's made in all respects in all respects by my colleagues a few years ago that shows that those systems can not only detect and classify them but also detect and identify them in the level the Asian thing also the project and the region in the image is the category of the project it's called instant segmentation or panoramic computer vision so the performance of the system is not to say that computer vision is a good solution but really the performance is incredible this is open source very good this has caused a lot of performance in things like medical imaging this is a support for my colleagues in the department where they use self supervised learning in combination with other things too and commercial nets tumors in networks other work in collaboration between English department and data which basically uses deep learning for restoring the quality of MRI which is basically no systems and images with one quarter of the time that it takes normally to collect the data instead of having to lie down in a a magnetic or inness that's because of the properties that we can use to restore the quality of images there's a lot of applications of commercial nets and other deep learning techniques in neuroscience and science in general but in particular, commercial nets are the dominant model that neuroscience intelligence uses to explain how the initial system works there's similar effort to use the deep learning system in a language to try to explain how the human brain in Microsoft's language really as well seems to indicate that the current models that we have in British understanding not reflect the brain understanding of this as I said also applications of science particularly in physics in astrophysics in history material science in particular just one example of a project that was started by my colleagues at META and the idea there is that need to solve climate change and when possible the approach to solve climate change is to solve the energy storage problem we're able to store the energy of the energy could cover a small desert with small amounts just reduce energy instead of marketing separating energy from oxygen and the problem with this is that processes that are known to do this are either scalable or the efficient or the efficient but they're not both scalable and efficient and for them they're more efficient and that's not scalable it's too expensive and the one that was scalable was not efficient so the goal of this is to use AI to discover new coordinates that might facilitate the electrolysis of fire but at the same time are scalable and the data is open so the data is generated by doing the energy simulation very large scale distributing the data and asking scientists around the world can you train an AI system to extrapolate the data and basically invent new compounds that could help with that so this is an ongoing project very risky and as I said it will definitely protect lives of the environment it's useful and the energy simulation any switch here so I talked about these three highlights of learning and self-supervised learning and I'm going to focus on self-supervised learning now because this is really the future but it's still the first time as well and it's really the first version based on AI or machine learning in the last five years or so five to seven years and if you want to find some sort of an energy or metaphor or learning or machine learning with this digital traffic theory you could think of self-supervised learning this is really where most of them should take place and then to learn in particular still you cannot use a supervised code that would be like the ICs of the cake small pieces of the cake important but small and then you can start learning it's just kind of fine tuning I didn't like the cherry on the cake in terms of variety of importance and so one has to accept the fact that machine learning really sucks today compared to humans and animals or machines today really go around very well or very fast as we're saying here there's models they're the same they could use bad switches and when we compare this with humans and animals humans and animals can learn new tasks really quickly with very few trials we understand how the world works humans and animals humans and animals doesn't matter what we both sense okay so self-improvement is really what it has to do with the world and the expected of it let's say we're going to make a video show the initial segment of the machine we ask it to the developers what's going to happen next and then reveal what happens next and we hope the machine will adjust itself to do a good job by predicting what happens next when I say what happens next is that this is already predicting the future and predicting the left half or the last it doesn't really matter in fact the way humans see the context of the mathematical language process is the following you take a piece of text you correct it by you read some of the words the principles words by Marker then you take this to the text and you write it through a very original method in particular words that I've seen in the process of doing so the system writes an internal representation of text that integrates meaning, syntax grammar, just everything and so you can use the internal representation of the text written by the system as an input to its subsequent tasks like translation, case speech deduction summarization eating you want and this works amazingly well it's called the noise input it's an old idea if you combine this with a particular architecture of human interest which you can think of as a different fabric of commercial nets there's two big fabrics commercial nets and transporers and very well put they're combined actually then you can you can trans distance to translate very easily by basically taking two sentences that are approximations of each other and investing in different parts in two sentences you can train those systems to use that on so this idea that then talk to a data system is very old a scallion should be working on this for a while to make this work so let's get from here let's get from Google Geoscience Technology here is what we're going to do today a data system is an old topic it's been a long time for us in recent years it was a self-supervised marketing to test in particular cases where data system is actually supporting us all it's a system put together by my colleagues a multi-volume a data system that actually understands images make stories so this is years before so another example here is a simple moment behind the system together by some of my colleagues that they can translate two languages could be in any direction and it's the magic of it is that it's trained you don't need to have your own time to train that system between all the 4,000 and 40,000 different countries in fact this system has been trained with only data that only works as 2400 and yet this system can translate from any language those are one of the main questions that is several of those not just Italian there are some Indians where they're yet so in total languages which it's right here all the several hundred 700 languages so let's look at one of the thousands of languages so this works pretty well so for speech recognition we can operate some of the on the order of 10 minutes for transcribed speech in a language and that allows to fulfill the desire that most of those important expressions which is you can now have a speech recognition system that works with very rare languages for which you don't need data in fact my colleagues throughout the system they can recognize speech in 11 hundred languages it's a single system languages which are transcribed into text so that's pretty amazing similar techniques this self supervised meeting instead of being applied to text and being applied to DNA sequences so we can replace the words by you know amino acids essentially and those systems are able to representations of the structure that captures a lot of the complexity and so now you can use those systems to predict the combination of proteins which is a major advance so two major moves working on this one a deep mind the release system called ATHLEPOL2 which can be used by millions of researchers around the world another system produced by some of my colleagues or some of my colleagues called ESF4 and ESF42 which can be used to predict again the operation of proteins can be used for designing new proteins so direct design is both for the revolution of the next years because we can now we need we can now design a sequence amino acids so that it will take a particular shape and then bind to a particular location in the cell for example and then have it as a core function because it's likely to direction this like we need to put together this database of 650 million researchers it's used again in my thousands of researchers some of its sources are okay so this is the present so we'll be talking about geometry di we're at least in the public the research has been years and here is how the work we train that mass to the last word and we train some gigantic momentum to the last word that we do a text is thousands of words and once the system has been trained spend this system with millions of parameters or just billions when the system has been trained you can use it to just produce text because it knows how to predict the next word but for those of you who suspect and then as you predict the next word and then you check that next word into the input and then as you predict the next word and then you check that input and then you go in small auto-reversed prediction it's an old offset from the statistics and then single processing it's the same idea and most of the sets are amazing in many ways in their ability to really understand the content of text and suppressments but in many ways they make various mistakes and that's because they don't have any experience they only train text and most of human knowledge actually has something to do with language it has to do with our experience in the physical world and so those systems are extremely limited in their intelligence and you know because it's a product it's made it more than just a system in the use of gears when it was before that before that so they don't have anything like intelligence to get there to work you know to train that something on the order of learning again, some more units that's basically the entire internet the entire working internet to read this quantity of text for a person reading 8 hours a day that average key required to reach a thousand years babies, young children in normal language a tiny amount of exposure in the language and they are much smarter than any other systems it kind of has a more example for the text that's relevant but in terms of raw intelligence they are much more smarter than any other systems because we're already here human-level intelligence today to realize it and these systems can do this today so like these and the two of them they train a large number of people and say did you know that the other group dropped a graph of a nice year but listen to it and here is what we saw and then they let the system create a graph spreading so the system buys a critique of a nice graph which we can read and now this hallucination I'm just kidding I'm not interested at all I'm really into jazz so it's like the same thing something it didn't work because it's just not an update of me today of jazz readings that made me extremely sad so those are the best evidence they're useful, exciting there's a lot of applications for that that will reach the next years in nowhere here human intelligence reach that gap so the body ball and the fluency because this we have some kind of intelligence anywhere we go and so in fact I wrote this philosophy paper last year and actually it was this morning this morning people in the race and race have critiques of they kind of reason they kind of plan it's a system that can be able to begin to function in the morning so that's a big fraction I think that critiques in the next years in the eye it's the ability to reason and plan so basically I said that it's the most important of the week but basically people in the race and race don't want to do this so it's not especially intelligence also it's not finishing research it's basically my vision for where I'm searching for the next years so the kind of people that you get to do when you're sharing the work if you do this while you're still overlapping all the attention we can set which is so it's based around this idea that some sort of machine detector basically the center piece is the idea that we have a machine where we have some kind of reason and I call this a project in Germany how it is because when I say this machine intelligence people get scared to say what I'm going to do what you're talking about is it's scary so I'm going to project in Germany a system as a model that allows it to predict how it will be able to evolve so you can understand and particularly is able to predict what's going to happen as a consequence of these two actions if you can predict what's going to happen as a consequence of your actions then you can plan so if I want to get a class out there then you can borrow in the class and then forming the class here because I know it's stable and we know that it's here because the strength of the glass is still so I have this model that allows you to plan this in such actions and what makes the result in the result cycle which is so that's the idea we're going to do when you make the objective sometimes because of the objective it sets itself and then it empowers a sequence of actions that according to its own will satisfy the subjective so it creates steps and actually we can take the second action we can predict the resulting outcomes of actions and then check whether these things will satisfy the objective and perhaps I can just bother checking is the guarantee that the system of all these things are unsafe will not crazy things that will bring the users sometimes like this and one thing that if it was an active schedule will describe that big thing so let's imagine for example going to New York to Paris and then sitting in my NYU office I do this what I do to plan this is that when I first represent the situation that a very general distraction I say well first I have to take a taxi to go to the airport or really take the subway but I have to go to the airport and catch a plane to Paris and I don't need to know any of the details about this okay so how do I get out of the airport to get out of the building and take the taxi on the street to New York how do I get out of the street to stand up for my chair I think the other area is on the stairs on the street how do I get out of my chair active in the buses in particular so we're planning these actions without thinking about it practically all the way down to the second by the second bus and this time I think no idea how to do this with the AI systems at the moment and the example of this working the missing piece is the ability to run everything that represents the state of the world and I can allow the system to predict this so these actions at multiple levels of abstraction so I'm going towards the end so this leads to a new type of architecture for our construction models which I'm not going to move to but I'm going to allow the system instead of using one to the other by the current system the system will be able to find their answers and be able to guarantee that their answers are safe like in the observance of the diagrams and we're working with this I can't do exactly what I'm going to do it's still very opinion so I'll just give it a little bit so one thing that I'm working with is part of this type through a system and this is using the fact that it's not generative architecture it's a very different architecture and all that time to go into the details of this actually I don't have time I won't forget but maybe just a little bit is the basic idea that I need all the details that I need to have in the video in the future we're training the system to predict an abstract representation of what happens in the video so for example, if I take the video in this room where I'm only counting the gear there's no way that in the production system the details it's half of the room looks to be the clothes everybody is wearing the age and gender when the store looks like when the meeting will look like it's no way that it's predicted but I can predict that it's probably people sitting in chairs as a kind of abstraction I can make that information which is not going to be in details so that's the idea of making predictions in a representation space and have the knowledge to be able to predict it particularly in a general perspective if you're pointing the finger it's an action which is stated so actually so when we're asking people to do now I'm used to thinking about religions in the Irish or something unpopular but I don't know which word you use but I think it's against the mainstream and I'm asking people to believe that the original genetic models if it was jointed with the detectives because although generative rules were in context we use a better genetic model because if you use the genetic you can't expect a better genetic model and then a better but at least we're not thinking so they're very popular so if you do this this architecture actually train the image recognition system and I have a question just to on this part if you take an image you can see it so it's the same idea as the text then you run both images and then you train the representation you train the system to predict the representation the important mass and that was the system to the point that you get the step of the theme and then perhaps again this works in architecture so much better video which to explain in pictures perhaps in an August architecture systems basically and these sense patterns so working with the C's you're working with the systems with the data with the deployables and the taking messages that I'm going to be using in this course my prediction is that you're based by because you're still using essential concepts which you can have those models that you run and you're using general architectures and then you'll have those in the future I've been attending my first two decades it's really important to think there's no question you will have them there's no question that as a computer you will have machines that you can use or really have machines that you can use in very regional things eventually there will be smarter than that in every opinion should we be able to plan like this because we all have intelligent assistants helping us in all their lives and it would be like one of the staff working for us that are smarter than us which is the idea of the situation as the business is taking from a computer the best thing you can do is to have a smart value which sort of kills the better of them so this is going to amplify human intelligence and basically perhaps cause a new business the Italy famous basically started from this experience so he works and uses it and he's he's he's he's so one of this which is able to facilitate AI is the main tool of the society's knowledge and why AI platforms are going to be a repository of all the knowledge that we can all use as a machine so I think in the case of human resources that are similar to the previous one so it's very positive to see very different pictures and sometimes painted in the press with the dangers those dangers are similar to the dangers of AI so this is very difficult to build it's going to be difficult to build AI systems that are safe and operational so the engineering the engineering the engineering school so I think in the future as you know in the future AI systems need to be open and of course they need to be brought to society and this is very important if you see the AI systems of dangerous we need to be able to provide AI systems this is an opinion and we're going to have a discussion about the U.S. which we need to regulate the AI than so many European countries to make these cases because these are going to be very much a favorite taking advantage of conflicts thank you very much thank you artificial intelligence and automation engineering this is very important in the future in the region we're going to have special intelligence and automation engineering in the future in the future in the future in the future in the future in the future in the future in the future in the future in the future in the future in the future in the future in the future in the future in the future in the future in the future in the future