Artificial Society Lectures AS pt1 PDF

Artificial society Lectures Lecture 1. 2. What is artificial intelligence? Ada Lovelace (1815 - 1852) : - English mathematician and writer - Known for her work on Charles Babbage’s proposed mechanical general-purpose computer - The machine had applications beyond pure calculation and published...

Artificial society Lectures Lecture 1. 2. What is artificial intelligence? Ada Lovelace (1815 - 1852) : - English mathematician and writer - Known for her work on Charles Babbage’s proposed mechanical general-purpose computer - The machine had applications beyond pure calculation and published the first algorithm intended to be carried out by such a machine How would you define intelligence? -> The ability to adapt -> The ability to deal with complex structure -> the ability to understand self-awareness -> Is it difficult to define the word intelligence? What is intelligence? No definition of what intelligence is but there is a possible one: 1. The ability to achieve complex objectives ( emotional development, problem-solving, learning, thinking logically, etc.) -> Max Tegmark 2. A richly structured space of diverse information processing capacities -> Margaret Boden John McCarthy ( 1927-2011) coined the notion of artificial intelligence (1955) - We propose that a 2-month, 10-man study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College in Hanover, New Hampshire. - The study is to proceed based on the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it. - An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves. We think that a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer. Rodney Brooks: AI “ started as a field whose goal it was to replicate human-level intelligence in a machine” 1 of 37  Overall: => Difficult to define intelligence => There is no right definition of intelligence => The definition has changed over the years. The way we see intelligence now is not the same as we saw it serval years ago. => There are different types of intelligence ( social, creative, analytical, practical, mathematical, emotional) => Emotional and language intelligence have been ignored for a long time Or definition has changed over the years: The development of AI has changed the definition of intelligence Before being intelligent meant being good at mathematics Now focus on different parts of intelligence Di erent types of intelligence Social intelligence: involves how we make sense of the people around us. Social Intelligence requires a basic understanding of people and a set of skills for successful social interaction with others. Emotional intelligence: the ability to manage both your own emotions and understand the emotions of people around you. There are five key elements to EI: self-awareness, self-regulation, motivation, empathy, and social skills. Analytical intelligence: also referred to as componential intelligence, includes academic tasks, problem-solving abilities, and abstract reasoning. Whenever you have to complete a task that requires you to compare, contrast, evaluate, analyze, or make a logical judgment, you are using analytical intelligence Mathematical intelligence: is the ability of students in terms of numbers and logic, which involves the skills to process words and numbers, to use logic and analyze problems logically, to find formulas and to do investigations scientifically Practical intelligence: as the ability that individuals use to find a more optimal fit between themselves and the demands of the environment through adapting, shaping, or selecting a new environment in the pursuit of personally valued goals (Sternberg, 1985, 1997) Dartmouth summer research project on arti cial intelligence (1956) => Workshop widely considered to be the founding event of artificial intelligence as a field. => It was a brainstorming session between mathematics. => Are those the father of AI? 2 of 37 ff fi Does intelligence require subjectivity and consciousness? Mapping intelligence onto machines -> Rule-based AI ( top-down) VS Data-based AI (bottom-up) -> Humans have vision VS AI; image and pattern recognition -> We have speech VS Natural Language Processing ( NLP) - e.g. large language models -> We can recognise emotions vs. emotion recognition algorithms -> Caveat: high uncertainty ( inaccuracy) with NLP and emotion recognition! Language and emotions have long been ignored as parts of intelligence -> Robots have become containers and the embodiment of AI. -> Human intelligence prevails, it is much more complex and multifaceted than machine intelligence. Example: self-driving cars -> sensors, cameras ( eyes), and microphones ( ears), to learn from experiences. They have some difficulties recognise some of the traffic signs e.g. information commissioner warns firms over emotional analysis technologies What are the failures of chatGPT? - the reference list? - is not good in maths - We have emotions but ChatGPT dont have emotions - Sometimes the wrong information - Lose track of the conversation 3 of 37 Cybernetics? Focus on chess: - chess was well-known, popular, and prestigious - Generally recognised as a complex, creative game that requires strategy & planning - Association with mathematics and computing, and with programming ability - There was a comprehensive and extensive body of historical and theoretical literature on chess Is Deep Blue (IBM) intelligent? Deep Blue was a supercomputer developed by IBM specifically for playing chess and was best known for being the first artificial intelligence construct to ever win a chess match against a reigning world champion, Grandmaster Garry Kasparov, under regular time controls. => The chess computer was not able to adapt =>Deep blue is not intellect Despite the promise… - computer chess did not produce new theories ex about human cognitive processes or theoretical computer science. - failed to deliver on its larger promise as a tool for exploring the underlying mechanisms of human intelligence - Important concept: AI winter - Computers and humans are essentially playing an entirely different game - Humans think ahead merely by 1 or 2 moves but computers rely instead on perception, pattern recognition… = an even more complex cognitive activity! COG (MIT) Learning through the environment. This robot has eyes, ears, lot of scenes learns the same way as a child not a lot of present-day AI involved extremely difficult to learn in the environment Using computer and robotic technology we seek to better understand and emulate human intelligence. The humanoid robot Cog, shown above, is the primary robotic platform used by the group to explore human intelligence. Several other robotic platforms are also being used by the group to efficiently and concurrently explore more specialized aspects of human intelligence 4 of 37 - humanoid robot designed to interact with humans to learn from these experiences, similar to how children learn - Not rule-based - ears ( microphones), eyes ( cameras), hands ( touch, motion): the robot had to respond to and interact with the environment - Now at the MIT museum Paradigm shifts Cognitive approach - Rule-based AI Interaction with the physical environment ( contextual and behavioristics approach) Behaviour-based approach: the coupling of perception and action gives rise to intelligence Self-learning: AI = techniques such as machine learning, deep learning and neural networks that aim to make computers learn from experiences - data-based AI e.g. the pause giant AI experiments: an open letter Elon Musk signed for the pause and a few months later he developed his own AI chatbot Pause giant AI experiments: an open letter: We call on all AI labs to immediately pause for at least 6 months the training of AI systems more powerful than GPT-4 AI is a game changer because of: 1. New technological developments and applications 2. Increased storage capacity, speed and computing power, which, according to Moore’s law, doubles every 18 to 25 months 3. Big data: huge data sets to train algorithms on 1. Data-based AI 2. Self-learning algorithms The AI ski trip or the hype of AI Movavec’s paradox: What is difficult for AI is easy for humans and the other way around => e.g. calculation 5 of 37 Is AI always preferable? - optimisation - Speed - Collect, analyse and process an. Enormous amount of data - Memory - consistency - More objective The construction of collective intellect or imagination. Internetworked data would then provide the technical infrastructure for the collective brain or hyper cortex of living communities” “We pass from the Cartesian cogito to cogitamus” Our emotions are a disadvantage to being objective? Alpha Go ( deep mind) intelligent? Is a Chinese complex board game 2017 Learn by itself, with the supervision of humans YouTube makes us stay on YouTube Philosophical questions concerning AI? Philosophical questions regarding AI - where is the boundary between human and machine? - To what extent can computers be creative? - Can machines think? - Can AI truly be intelligent? - Are humans ( thinking) machines? - Align with ancient-old philosophical questions? - How does the brain work? - What is consciousness? Turing test: Or the imitation games When the human can not tell that he is talking to a robot, the machine passes the test 6 of 37 -> when you can trick a human -> if the human can not say If it is a human or a bot that is doing the conversation. Strong AI: AI has (self)-consciousness, free will, agency, and intentionality -> John Searle’s Chinese room thought experiment Superintelligence: “ An intellect that is much smarter than the best human brains In practically every field, including scientific creativity, general wisdom and social skills.” (Nick Bostrom ) Singularity: “ A threshold crossing where our computers have become more adept at designing themselves than we are. When this occurs, intelligent artefacts will go it alone, designing and building new intelligent artefacts hit smarter-than-human intelligence. The crossing of this threshold is referred to as the singularity. The singularity is the point at which machines become sufficiently intelligent to start teaching themselves how to design machines” Weak AI: Is a type of artificial intelligence that is limited to a specific or narrow area. e.g. Siri, Alexa, Google Home,… Ethical concerns: - hidden morality: biases -> how do the algorithms do? Is it real? New scams are created - Privacy - Ethics different from a decision tree - The prospect of reducing ethics to a logically consistent principle or set of laws is suspect, given the complex intuitions people have about right and wrong. - The project of building AMAs highlights the need for a richer understanding of human morality. Important to test the product before putting it on the market! Need to have a lot of diversity !!!! 7 of 37 Lecture 2. the future of work and politics Take a philosophical view of work => us as human beings, we are capable of doing things, but technological artefacts also What is work? Work => concept is quite obvious to us but when we start to pick things apart, we find that it is not easy to define what work is. Even the period which we call “rest” is increasingly structured by work. Work structures our entire life. The structure of work is something even larger, work has an enormous impact on our social and political life as well. Do you think that generative AI will cause signi cant disruption to the labour market? -> Generative AI is expected to have a significant impact on the labour market -> Manual labour has mostly been replaced with robots -> But some people do very fine manual labour that a robot cannot do -> The creative work cannot be done by an AI/ robot “These jobs are going son, and they Ain’t comin’ back, to your hometown” ( my hometown, born in the USA, Bruce Springsteen, 1985) => was talking about the industrial works and the different jobs. The relationship between political, and material relationships What is work: 1. Thinking philosophically about work. - To perform work or fulfil duties for wages or salary - perform repeated operations. There is this data of work which involves doing something regularly, repeatedly and to get a salary. - To perform or carry through a task requiring sustained effort or continuous repeated operations. - To exert oneself physically or mentally, especially in sustained effort for a purpose or under compulsion or necessity - To function or operate according to the plan or design To produce the desired effect or result - To exert an influence or tendency - Putting an effort for a purpose or under compulsion or necessity. The notion of effort is. Extremely important in the motivational part of work. - It is morally good to expend your effort on your duty. 8 of 37 fi - To function or operate according to the plan or design: doing this regular activity according to a plan or design - In the modern era, we are constantly looking to improve efficiency !! “An exertion of mind or body undergone partly or wholly with a view to some other good than the pleasure derived directly from work”~ Alfred Marshall “Work is a modern concept stemming from the Industrial Revolution. It refers to a performance, intended for others, possessing use-value for them and hence entitling the person carrying it out to a certain reward or compensation. It is performed in a public, not private space and is intended for others as social not private individuals and has a recognised social validation or value” ~ ( Andre Gorz) The distribution between making things and making efforts.~(Hannah Arendt) Gorz and Arendt are both seeing how the structure of work is changing and they are concerned about the fact that this structure leads to the change of political and social organisation. The work of our hands as distinguished from the labour of our bodies... fabricates the sheer unending variety of things whose sum total constitutes the human artifice -> The effort needs to be put into cooperation Wages against housework -> the household labour of a woman The moral value of effort -> It is good to work hard The social value of effort -> -> All modern thinking realises on a preset opposition that people have to work together and cooperate EGE opinion (de nition of work) : “Work is a practice by which people ( besides seeking to ensure that their own tangible and intangible needs are met) contribute something to their families communities or societies. This makes work a key aspect of our personal and social identities, it is both personal and connects us to the communities we live in. It follows from this that work includes more than only paid work and employment. Contributions that people make to the functioning and flourishing of our society => group of experts that is trying to expand the definition of work by claiming that work can also be made unpaid to help our society ( such as caring for children or the elderly) form a very significant part of work in our society even if they are unpaid History of work: James Scott The beginning of the formation of a stat-like political structure - the story of 9 of 37 fi the relationship between the formation of the state and agriculture. A linear story about the production of certain types of artefacts (mainly grain). What is the state? -some form of government - Geographical borders - Division of labour Early state: A kind of machine for the extraction of value from a human person as a potential producer of value through labour. Most of the value which has been extracted was grain. It was a real problem to maintain the population to produce value. => As soon as people could, they moved away. What emerges out of this technology is a certain idea of the human person ( individual subject entity) as a potential producer of value through labour. The state is still using this to measure how many taxes you need to pay. Ernest Gellner: Talks about the emergency of the nation-state This period of commercialisation requires a transformation of the kind of people that were living and functioning within this political entity. People need to be specialised. A person with a common education is needed to be able to maintain order between many people. Due to a change in the socio-technical context, a political change is emerging and a new type of person emerges as well. Work, technology and politics Ernest Mass democratic leverage results from the capacity to withdraw labour/ energy from the economy Democracy is about conflict labour was the weapon of the masses The future of work question: The crisis of work or the future of work problem Despite high levels of employment Rising inequality and precarity: shift of GDP from labour to capital Loss of meaning at work Concerns about the automation of roles, tasks and jobs EU Automation Anxiety: - how technologies are going to impact the labour market and our social structure and eventually political structure. - Global opinion is not uniform - EU opinion can be quite contradictory 10 of 37 - Automation brings many benefits as soon as the position of the human is not completely replaced The future of work problem: Has existed since we started to talk about “work” within the modern industrial period. The big question: does this process of replacement reduce the number of available jobs? People are unhappy at work -> the problem has to do with the quality of the job but also with the quality. Even those in the most privileged form of employment seem unsatisfied. Luddites: a group of early 19th-century English workmen destroying laboursaving machines as a protest. => Right decision? - yes, took 100 years for living standards to recover(the level of comfort in which people live, which usually depends on how much money they have.) - No, productivity and innovation create more jobs than they destroy Crisis of work - despite high levels of employment - Rising inequality and precarity ( shift of GDP from labour to capital) - Loss of meaning at work - Concerns about automation of roles, tasks and jobs - Evidence of polarization ( economic, regional, …) - Lingering concerns about technical unemployment becoming structural Very coarse outline of the debate : Position 1: The automatisation of work is going to lead to structural technological unemployment. Position 2: no it won't Position 3: who knows, but it will change the way we work and subsequently live. There is already displacement and polarization 11 of 37 Lecture 3 What computers can do How people used to think of intelligent computers. => Computers slavishly follow instructions They were trying to make the impression of mimicking intelligence -> ELIZA ELIZA -> just a programming code and it gives answers to questions Artificial intelligence report by James Lighthill => concluded that AI is complicated but also it might be useful for the future in this time AI was not a fashion. AI is more of a game -> in order to make predictions by using certain rules. The history of AI 1941: Computers were built and had a program was a German electromechanical computer designed by Konrad Zuse in 1938 and completed in 1941. It was the world's first working programmable, fully automatic digital computer. 1944: Computers had a program were the first electronic digital machines with programmability, albeit limited in modern terms. => passed the Turing test Turing asked this question -> “Can machines think?” Hubris of early AI research “... every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it. “ ~ attributed to John McCarthy (1958) “Within a generation, the problem of creating ‘ artificial intelligence’ will be substantially solved.” ~ Marvin Minsky (1967) Eliza: Eliza 1966 by Joseoh Weizenbaum Chatbot presenting itself as a psychotherapist Very influential in the field of human-machine interaction SHDRLU: - SHDRLU by Terry Winograd (1968-1970) - Understanding of natural language to do action in a physical world 12 of 37 Shakey: - Shakey Robot by Standford Research Institute - representation and planning on a mobile robot - connected to a computer via a radio link. - The computer could process the incoming data and send commands to the circuits controlling the robot's motors. - Shakey was the first mobile robot with enough artificial intelligence to navigate on its own through a set of rooms. - Among its component technologies were a TV camera, a range finder, radio communications, and a set of drive wheels controlled by stepping motors. Grey Walter tortoises: - Elmer and Elsie - grey Walter (1948-1949)) - Simple robots that had animal-like behaviour Perceptron - Frank Rosenblat in 1958 built a machine to recognise images “... the embryo of an electronic computer that [....] will be able to walk, see, write, reproduce itself and be conscious of its existence.” Perceptron is a single-layer neural network linear or a Machine Learning algorithm used for supervised learning of various binary classifiers. It works as an artificial neuron to perform computations by learning elements and processing them for detecting the business intelligence and capabilities of the input data. 13 of 37 Backpropagation - Backpropagation ( automatic differentiation) is the workhorse of training of training neural networks - Seppo linnainmaa ( findland, 1970) Backpropagation is a process involved in training a neural network. It involves taking the error rate of forward propagation and feeding this loss backwards through the neural network layers to fine-tune the weights. Backpropagation is the essence of neural net training. The Lighthill report - prof James Lighthill - Lucasian professor of mathematics at the University of Cambridge - Illustrious predecessors: Isaac Newton, Stephen Hawking, And Charles Babbage - in 1972btasked by the British government to write a report on artificial intelligence Lighthill's report Identifies 3 categories of artificial intelligence: - Advanced automation - Bridge robot’s - Central nervous system and brains Men build robots to compensate for the lack of female capability to give birth to children Men-building robots form a relationship which may be called pseudo-maternal Robot research is an entertainment industry Perceptron is a single-layer neural network linear or a Machine Learning algorithm used for supervised learning of various binary classifiers. It works as an artificial neuron to perform computations by learning elements and processing them for detecting the business intelligence and capabilities of the input data. Predictions of Lighthill about how the robots/ AI will develop Disappointments: - Speech recognition - Translation - Recognition - Automatic theorem proving -... 14 of 37 Lighthill main criticism the real world is complex Robots mostly operate in a toy world:( simple room ) they can only deal with restricted, easy problems Increasing complexity leads to a combinatorial explosion Combinatorial explosion - the need for computer power grows exponentially with the size of the problem being solved - How many combinations can you make with How did AI do after the Lighthill report: - Lighthill’s report ( and similar reports) killed AI research funding for over 10 years and tarnished its reputation for 20 more years. - Disappointment in machine translation in 1966 - Marvin Minksy and Seymour Papert’s book “ perceptrons” that ridiculed neural net research. - AI only slowly recovered in 1990 The AI -e ect - it occurs when onlookers discount the behaviour of an AI program by arguing that it is not real intelligence - Every time they figure out a piece of it, it stops being magical. It's just a computation. By the 21st century, AI will be everywhere - every minute of your day, some AI algorithm decides for you - The drama of it all is that we dont think of it as AI. - Ai is a slow burner - magic solutions do not exist 15 of 37 ff Speech recognition: superhuman? Breakthrough in speech recognition: - In 2012 - Translated English to other languages. - The translated voice sounds exactly like the original voice. - Great improvement for this time Transcribing speech Now there are automated subtitling of videos and meeting - Ex. If you are doing a Zoom meeting you can add the subtitles and follow the whole conversation with it. This can be practical if there are people who are speaking very low, are shay and dont speak loud, have a strange accent,... and so on ASR for atypical population automated speech recognition (ASR) is still problematic, especially for atypical populations. - ASR is not yet optimal for - Different ages - Languages - Dialects e.g. A kid is asking Alexa, to put his favourite song but Alexa does not understand the command and is pouting something completely wrong and inappropriate content. There experiment: They asked different children to read a book. Methodology: -children’s speech in a school setting in England. -11 children, average age M= 4.9, SD = 0.3; 5F/ 6M -Three children of utterances: - words: ( one, two, three) - simple sentences ( the horse is in the stable) - spontaneous speech 16 of 37 - three recording devices: - NAO ( V5.0, running Naoqi V2.1.4) - Studio-grade microphone (Rode NT1-A) - Portable audio recorded (Zoom H1) ASR performance for child speech Fast forward to 2023: openAIs Whisper 1550M model Text-to -speech Before 2000 “Stephen Hawking” parametric speech synthesis After 2015 The voice is more neural speech synthesis Looks more natural -> more human Speech recognition is really impressive Neural TTS 17 of 37 Machine translation - in the last 15 years, some amazing revolutions have taken place - First powered by statistical machine translation, now replaced by neural machine translation - In recent years, transformers revolutionised neural MT. - The translation became better and better. - The end of the translation fails Recognising type and handwriting - type is virtually at 100% performance - Handwriting depends, all powered by new seq-2-seq neural networks 18 of 37 OCR performance Google Photo -> Google picture can scan the picture and translate it on the picture. This is an AI that does that. Say gen -> Is an application that translates what you are saying -> Matches the lips with what the translation is saying. -> The translation is excellent -> It looks like it was really you that is speaking that language Combinatorial explosion In the 1970s we expected “ the curse of dimensionality” to form an insurmountable obstacle to AI’s progress “... a general obstacle to the construction of a [...] system on a large knowledge base which results from the explosive growth of any combinatorial expression” ( Lighthill) “... chess-playing programs [...] results are discouraging” ( Lighthill) Neural networks: Neurons -> the connection which neurons make between each other ( synapses) 19 of 37 Humans wanted to create AI machines, that could resemble humans, by taking inspiration from this process Perceptron -> This technology is an electronic brain. Ex. If you give some analogue pictures to it, the perceptron will recognise what gender the person is. This machine makes predictions, says if it is right or wrong and gives either penalties or rewards. The structure of the process of the Perceptron: An input layer ( identifying the picture) -> the hidden layer can be one or more ( processing of these neurons, trying to find the best combination) -> output layer (answer) Transformation in computer science: Do not need to code to train a computer -> biggest transformation in computer science -> just by giving computers data they analyse it and learn Deep Neural Networks Deep NN diagnosis of melanoma Now you can take a picture of your skin and it will say if you have cancer or not Limitations of Deep Neural Networks: - requires a large amount of training examples - Task-specific learning 20 of 37  Large language models 2 important concepts - next word prediction The princess lived in a beautiful castle - distributional semantics You can get an idea of the meaning of words by looking Transformers - new seq-2-seq architecture with 2 improvements over last year's technology - Attention: keep track of which data is useful to disambiguate incoming data - Parallel training: speed up training enormously over sequential training - Breakthrough in Natural Language Processing Large language models: chatGPT - an unprecedented leap in quality in language generation text comprehension and conversational AI leading to more robust NLI systems - rich understanding of language but also the physical and social world - adaptation through one shot/ few shots promoting - Democratised access to LLM’s - Heightened awareness of AI, increased attention to responsible AI Di usion model - trained on hundreds of millions of images and contextual text - A professor teaching in the beautiful city of Maastricht 21 of 37 ff Reinforcement learning - trail and error learning with rewards - Can deal with delayed reward - The idea has been known since 1950 but has had limited success - Some successes that only got academics exited - TDGammon, a backgammon playing AI in 1992 Deep reinforcement learning Deep Mind’s Atari player Ai learns to play at an expert level through practice Go => 10^80 atoms in the observable universe! Reinforcement learning played Go better than the reigning human champion in 2016 The machine plays against a person and after some time it learns to predict his moves. Alphafold Reinforcement learning reduced the cooling cost of Google data centres by 30% 22 of 37 Controlling plasma in a fusion reactor a fusion reactor needs plasma in which hydrogen fuses into helium Plasma is not stable - > 150 million degrees AI to the rescue Where is my robot? We do have robots ( vacuum cleaners, company robots, cooking robots,...) Millions of robots in factories ( that do automated work) Sometimes, the robot takes a lot of time to do a task e.g. this robot takes 30 minutes for each towel that it needs to fold it Building autonomous social robots is hard - moravec’s paradox: in AI/ robotics, hard problems are easy and easy problems are hard Paradox: “ It is comparatively easy to make computers exhibit adult-level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility” (Hans Moravec, 1988) Still, robot research matters We still need to seek new findings in robots and give them a specific task. e.g. a robot/camera that can recognise the licence plate of a car that is full of distractions Social robots: A social robot is an autonomous robot that interacts and communicates with humans or other autonomous physical agents by following social behaviours and rules attached to its role. “ Almost anything interesting that happens with robots involves people.” ~ rod brook e.g Pepper, Anki, woo wee keep on 23 of 37 Our human brain Pareidolia: seeing human-like features in non-human visual stimuli Located in our fusiform face area ( FSA) -> Our brain is researching for faces even on non-human being Fritz Heider and Marianne Simmel ( 1944) Animated shapes on a 2D screen give the impression that they interact and that they have goals, beliefs and emotions. Social interaction is brain-complete Social-multi-modal interaction engages all brain areas and all cognitive faculties. Memory, motor cognition, language, visual perception, reasoning, face recognition/ detection, and object recognition,... are there few cognitive faculties that are not implicated in social interaction Large language models for embodied interaction Tanzanian Swahili: 0.0080% of chatGPT training data Where are we heading The red line is what Lighthill predicted. But the green line is more what is happening nowadays 24 of 37 Take-home messages: - AI has made unexpected progress in the last 10 years, and expectation-shattering progress in 2023 - We have broken through Moravec’s paradox by getting a grip on the combinatorial explosion - AI has not yet delivered on the high expectations set 70 years ago - Bridging the digital to analogue world has proven harder than expected. - Artificial general intelligence ( AGI) - the holy grail for many - is not on the horizon yet. - AI is continuously evolving - Computer science transformation: focused on programming to learn and make intelligent decisions. - Breakthrough of deep learning: now we have certain limitations, but we can use deep learning because of the large amount of data - Limitations of deep learning: if you do not have significant resources, it is quite a challenge 25 of 37 Lecture 4. Biases of technology -> Technology is often not designed for different genders, ethnicities Technology is smart -> how much do we know about intelligence? How can we project the notion of smart on machines if we are not sure what the concept really means? AI is a philosophical discipline -> technology helps us learn what intelligence is. Phenomenology -> the study of experience our immediate experience of the world Hermeneutics -> study of interpretation, how do you interpret the world Example of self-driving cars: Why dont we have self-driving cars yet? Is it technologically not possible? Driving is rational/ intuitive/ and rule-based You have a body which allows you to participate in life In a self-driving car where is the body? Self-driving cars are unpredictable They follow rules but when something changes, they get confused and are not able to make a new decision. They dont have experience We drive with intuition Enbody -> be an expression of or give a tangible or visible form to (an idea, quality, or feeling). "a national team that embodies competitive spirit and skill" Artificial intelligence -> is an algorithm Alan Turing “ As soon as it works, no one calls it AI anymore” Alan Turing asked the question: “Can a machine think?” The test uses various standards such as deception and empathy -> the ability to communicate like a human The standards are often invisible, if we count one thing, we may not count other things Language is the key, and symbolism is universal by which computers and humans can be related It is very difficult to develop something entirely new with machines, machines can do only what they are programmed to do. 26 of 37 Relationalism and representationalism Turing has never talked about these philosophical traditions because he did not know that he was working on it Turing thinks that we think like this: - human thinking: input -> interpretation -> output - Humans make sense of their bodies by referring to computers - Our mind is like a bank book and we write new information inside of that book - The world consists of objects and things which follow rules Rationalism -> Everything can be translated into rules, the world consists of facts that represent the discrete and explicit, simple and independent fact that exists and that we know exactly what it is. Representationalism is the information processing model of the intelligence -> idea that we have a representation in our mind, it creates a duality regarding the fact that the world is just a representation created in our mind. The information processing model of intelligence -> based on rationalism. Works with a fit set of data on which it draws upon to better understand the world. The common-space problem -> How do we behave in a particular situation? We make instantaneous decisions that are not based on rules. It is a problem that frustrates AI researchers. Hubert Dreyfus Machine intelligence will never replace human intelligence because they cannot think as we do. Computers work with representations of the world, we do not work with representations of the world but we work with the world itself He understands that knowledge is situated and embodied -> phenomenology and hermeneutics vision 27 of 37 Lecture 5: AI utopias and promises What are the different futures doing? utopian digital society? - very green - Light brighter Dystopian digital society - a lot of dark colours - A lot of robots The utopian pictures are a source of inspiration -> see the green buildings in Milano Look kind of the same as the utopian images Starting point of the tradition: Utopia -> Thomas more in 1516 A sketch of an ideal society in his book The term utopia means “a place that does not actually exist. The island is an interesting place: 54 cities, strictly democratic, no private property, welfare state, and religious tolerance. Utopia: - de Optimo Reipublicae statu deque nova insula utopia - Thomas More 1516 - Sketch of an ideal society - Commentary on the present - Book 1: Dialogue of Counsel - adventures of the traveller Raphael Hythlodaeu - Utopia Book 2: Discourse on Utopia -description of an imaginary country The island contains 54 cities - strictly democratic - No private properties - Welfare state - Religious tolerance Interpretation is not st 28 of 37 straightforward Is it ideal or not? -> There are no failures -> There are no errors There were a lot of critiques: On private properties On the different religious Other examples: Francis Bacon's (1626) new Atlantis -> an island governed by scientific knowledge Brig quite modern ideas Jonathan Swift (1726) Gulliver Travels -> Voyage to Lilliput; satire Samuel Butler (1872) Erewhon -> Intelligent machines taking over Edward M foster (1928) The Machine Stops -> underground living people fed by big machine Goerge Orwell (1949) nineteen eighty-four -> Big Brother, doublethink, newspeak Ray Kurzweil (2005) the singularity is near -> The 6 epochs of evolution Francis Bacon (1561- 1626) Political and scientific career Transformer of science Prime minister “Knowledge is power” He was concerned about how society deals with knowledge. At a university at the time, people were engaged in a lot of debates. He thought that this was not the way to go -> more with experience/ experiments. He was not enthusiastic about that, he wanted to find a new way to produce knowledge. Experiments were getting more fashionable. The experimental way is the way to go. Unlike what they were learning at university, with the experimental way they would have had knowledge which improve mankind. Bacon's first book The Proficience and Advancement of Learning (How Was University Education in the Modern Time 17the Century) pseudo-science -> a collection of beliefs Decadent debates -> not about important issues Decadent learning -> reading a lot of historical literature The seven free arts: You could find the seven free arts in a curriculum at the university 29 of 37 - Trivium: - Grammar - Rhetorics - Dialectics - Quadrivium: - Arithmetics - Geometrics - Astronomy - Music Novum organum (1620): Experimental knowledge - Useful knowledge through observations - Knowledge will improve the fate of mankind First sentences of Bacon’s book: Man... understands as much as his observations on the order of nature permit him […] The unassisted hand, and the understanding left to itself, possess but little power. Effects are produced by the means of instruments […] Knowledge and human power are synonymous. New Atlantis (1626) - Ideal community of scientists ->still a picture of the modern universities - Supported by the government - Division of labour - merchants -> collect books and instruments - Pioneers -> design new experiments - Compilers -> present data in tables - Benefactors -> think about practical consequences and benefits - Inoculaters -> perform experiments - Interpreters -> interpret the results They are not innocent they really guide the way of thinking 30 of 37 AI utopian. Example: the singularity is near Kurzweil 2005 Premise: - Singularity -> non-biological intelligence grows exponentially so taking over at some point Some consequences: - production automated - Fighting diseases with nanobots - Cybernetic bodies - All needs to be fulfilled - Eternal life in VR worlds The 6 epochs Physics and chemistry -> Information in atomic structures. The universe got more structure after the Big Bang Biology -> the information that you have in the DNA Brains -> information is not only in DNA but also in neuro patterns Technology -> information in hardware and software Merger of technology and human intelligence -> the methods of biology( including human intelligence) are integrated into the ( exponentially expanding) human technology base. Ex chatGPT we did not expect that to do as much as what he is doing The universe wakes up -> patterns of matter-energy in the universe become saturated with intelligence processes and knowledge Omnipresent utopia-dystopia Boston Dynamics => They are creating things that people feel are fascinating but also frightening. They make use of the sense of ‘oh wow what is going on?” to do marketing. This kind of project stories are around us, Is AI a hype? How much attention is there giving -> We can see that AI started in 2007 and it has gone up and up -> However in jobs AI wi slowly going down -> There is a bit less hype 31 of 37 than a few years ago -> Generative AI is at the peak of expectation -> Do not invest too much init After there will be a delusion Edge AI in 2021 was at the top now in 2023 it's already falling down A series of hype cycles -> in 1960 -> 1980-1990 -> 2010 -2020 Different statements - descriptive What is the case? - Normative - How things should be - performative - Statements that do something They do change the world Role of promises in technology? -starting points ( innovation studies) - technological developments are always collective accomplishments, including engineers, firms, governments - actors have to operate under conditions of uncertainty and competition - actors depend on what others do, or what they think others will do - actors use promises as short-cuts to make decisions 32 of 37  - Promises help to legitimate decisions - When funding is needed - inside or outside the organisation/ firm - Government agencies face accountability pressure - Promises to protect the project - May work even when nobody believes the promise - Promises help to guide search activities -when success cannot be calculated, promises are the second-best - They work like heuristics ( rule of thumb) - Actors can be fueled by fear of lagging behind ( missing the boat, unstoppable train) -> Others are doing this so I need to do the same Without this fear there would not be the missing boat -> influences the economy - Promises help to coordinate decisions: - promises are stories with a plot and players - Actors see that others take a position and position themselves as well in the story - This mutual positioning leads to a shared agenda and division of task Self-ful lling prophecies: extreme example of performativity “ The self-fulfilling prophecy is, in the beginning, a false definition of a situation evoking a new behaviour which makes the original false conception come true. [...] such are the perversities of social logic”. Robert k. Merton, 1948 Promises about technology are also self-fulfilling prophecies, a lot of speculation, investments and something will of course come out. With technology, you are not sure whether the definitions are false or not, the only way to check it is to try it out. Whether you believe it or not. Promises do things, they are performative, they are not descriptive, or normative but they are of a different kind, they change reality. 33 of 37 fi Conclusion: promises have consequences: Performativity: statements that do something - descriptive, normative, and performative statements What expectations do: Legimitation, guiding, coordinating Overall dynamics: - promise becomes requirement - Technology did not start with problems or needs but with promises Utopia and promises are not innocent!!!!!!!!!!!!! 34 of 37 Lecture 6. A lot of people are misled by the media, they have the idea that AI is already strong and that they are capable of doing more than what it already can do Ethical concerns 1. Baised technology: Without realising we are dealing with morality on a daily basis -> gossip is a clear example That means that every product made by humans is made with morality -> they become part of technology, algorithms are trained with unfair/ biased data 2. Who teaches self-learning AI systems? Biased moral beings -> in 2016 Microsoft exposed the world to TAY (a chatbot) learned how to tweet as an American teenage girl on Twitter. It says a lot about the naivety of the creator –since Twitter is not a ‘friendly’ community. Within 16 hours Microsoft had to take TAY off the Internet. The tweets were very unethical and naïve. They had to release an open letter in which they admitted that the challenges are as much social as they are technical –they were focusing on the question: Does it work? Without caring about the social part. 3. Privacy The problem of privacy also exists offline The idea of privacy as a right that must be protected If you connect different data sets it is quite easy to deanonymize and give another identity to someone. 4. Morality different from decision tree Comparing cultures there are many similarities in some rules What we do is not easy to put in rules, it is not a decision tree Decision tree -> a tree-like model that acts as a decision support tool, visually displaying decisions and their potential outcomes, consequences, and costs. Even clear rules shared among rules can be context-dependent. It is difficult to teach context to machines Can AI be creative and original? AI painting is a new art market Lady Lovelace -> Computers cannot be creative DeepBach -> AI can compose music The machine just focuses on patterns and it does not know what is odd or not - train the machine with more data 35 of 37 Weak AI and Strong AI -> Everything we have right now is weak AI AI and health: AI is used to detect breast cancer in its early stages. Not as a technology overruling the doctor the doctor but it is used as a tool. Intelligent video monitoring -> Instead of going to a nursing home, elderlies can stay at home and AI is checking on them => privacy concerns From medical to moral decision-making To what extent AI can help us with moral decision-making? Medical decision-making -> AI detects breast cancer at an early stage. When we talk about morality it is more difficult to define -> we are ambiguous human beings we might be not the best ones to make decisions in life. Due to emotional conditions, we might not take all the relevant information into consideration it is necessary to arrive at a well-reasoned moral decision. We lack consistently by not acting on our principles or by favouring our group. Challenges: Empirical (verifiable by observation or experience rather than theory or pure logic) People’s judgements can change when arguments are presented to them. Technical How to teach morals to machines? - context-dependency of morality - Natural language processing (NPL) - No blueprint - Golden rule -> treat others how you want to be treated Ethical Manipulation ( commercial robots, how children conform to what the robots are saying) Moral laziness/ disengagement -> delegating our judgement to a machine Conceptual agent - weak: software or a machine performing a certain task - Strong: to be able to reason and act. Moral agent: - Weak: just following a set of rules - Strong: free will, rationality, consciousness, have moral responsibility 36 of 37  You have to be clear about the definition that you give otherwise people (especially media) will get confused. AI arti cial moral advisor -more effective -Faster -Collect, analyse and process an enormous amount of data -It is consistent -More objective Would this lead to more informed and better moral judgments? Is AI always preferable? Efficiency and consistency are not always better, sometimes what is illegal can also actually make a change in society. 37 of 37 fi