Computational Thinking - Учебное пособие, страница 9

Data can be encoded as biomolecules – such as DNAstrands, and molecular biology tools can be used to transformthis data to perform, for example, arithmetic or logicoperations.47Challenges of Artificial IntelligenceWords and phrases:Artificial IntelligenceTo percieveInsightCognitive scienceTo overlapFuzzy systemsBiologically inspiredAcute problemCritical inferenceBias– искусственный интеллект– понимать, воспринимать- проницательность- познавательная наука- совмещать– нечеткая система- на основе биотехнологий- острая проблема- критически важный вывод- необъективность,предвзятостьTo underestimate- недооценитьFallacy- заблуждение, ложный выводOft-committed- часто совершаемыйQuantifier- кванторPalatable- привлекательный,удобоваримыйTo diverge- расходиться, разнитьсяFrom scratch- с нулевого уровня, с началаTo hone skills and knowledge - доводить опыт и знания досовершенстваThreshold of criticality- порог критичностиDownright- явный, откровенныйTo pose requirements- накладывать требованияVulnerability- уязвимостьAn endangered species- вымирающий видTo hard-wire- подключить через проводThe modern definition of artificial intelligence (AI) is " thestudy and design of intelligent agents”, where an intelligentagent is a system that perceives its environment and takesactions which maximizes its chances of success.John McCarthy (Sept.4, 1927 – Oct.24 2011), who was oneof the founding pioneers of AI and also coined the very name48of the field, defines it as “the science and engineering ofmaking intelligent machines”.

Other names of the field havebeen proposed, such as computational intelligence, syntheticintelligence or computational rationality.The term artificial intelligence is also used to describe aproperty of machines or programs: the intelligence that thesystem demonstrates. AI research uses tools and insights frommany fields, including computer science, psychology,philosophy, neuroscience, cognitive science, linguistics,operations research, economics, control theory, probability,optimization, and logic.AI research also overlaps with tasks such as robotics, controlsystems, scheduling, data mining, logistics, speechrecognition, facial recognition and many others.Subjects in AI, or computational intelligence, as defined byIEEE Computational Intelligence Society mainly include:neural networks – trainable systems with very strong patternrecognition capabilities; fuzzy systems – techniques forreasoning under uncertainty (widely used in modern industrialand consumer product control systems, capable of workingwith concepts such as 'hot', 'cold', 'warm', and 'boiling');evolutionary computation – applies biologically inspiredconcepts such as populations, mutation and survival of thefittest to generate increasingly better solutions to theproblems.These methods most notably divide into evolutionaryalgorithms (e.g., genetic algorithms) and swarm intelligence(e.g., ant algorithms).

With hybrid intelligent systems,attempts are made to combine these two groups.By far the greatest danger of Artificial Intelligence is thatpeople conclude too early that they understand it. Of course,this problem is not limited to the field of AI. Jacques Monodwrote: “A curious aspect of the theory of evolution is thateverybody thinks he understands it”. The problem seems to beunusually acute in Artificial Intelligence. The field of AI has areputation for making huge promises and then failing todeliver on them.

Most observers conclude that AI is hard; as49indeed it is. But the embarrassment does not stem from thedifficulty. It is difficult to build a star from hydrogen, but thefield of stellar astronomy does not have a terrible reputationfor promising to build stars and then failing.

The criticalinference is not that AI is hard, but that, for some reason, itis very easy for people to think they know far more aboutArtificial Intelligence than they actually do.Anthropomorphic BiasWhen something is universal enough in our everyday lives, wetake it for granted to the point of forgetting it exists.In every known culture, humans experience joy, sadness,disgust, anger, fear, and surprise and indicate these emotionsusing the same facial expressions. We all run the same engineunder our hoods, though we may be painted different colors;a principle which evolutionary psychologists call the psychicunity of humankind.

This observation is both explained andrequired by the mechanics of evolutionary biology. Wehumans forget how alike we are, living in a world that onlyreminds us of our differences. Not surprisingly, human beingsoften “anthropomorphize”— expect humanlike properties ofthat which is not human. When we try to reason about otherminds, each step in the reasoning process may becontaminated by assumptions so ordinary in humanexperience that we take no more notice of them than air orgravity.Anyone seeking to reduce anthropomorphic bias in themselveswould be well-advised to study evolutionary biology forpractice, preferably evolutionary biology with math.

Earlybiologists often anthropomorphized natural selection—theybelieved that evolution would do the same thing they woulddo; they tried to predict the effects of evolution by puttingthemselves “in evolution’s shoes.” The result was a great dealof nonsense. Any two AI designs might be less similar to oneanother than you are to a petunia. The term “ArtificialIntelligence” refers to a vastly greater space of possibilitiesthan does the term “Homo sapiens.” When we talk about“AIs” we are really talking about minds-in-general, or50optimization processes in general.The researchers believe the AI will be friendly, with no strongvisualization of the exact processes involved in producingfriendly behavior, or any detailed understanding of what theymean by friendliness.Underestimating the power of intelligence implies aproportional underestimate of the potential impact ofArtificial Intelligence, which is symmetrical around potentialgood impacts and potential bad impacts.Capability and MotiveThere is a fallacy oft-committed in discussion of ArtificialIntelligence, especially AI of superhuman capability.The following chains of reasoning, considered in isolationwithout supporting argument, all exhibit this fallacy:• A sufficiently powerful Artificial Intelligence couldoverwhelm any human resistance and wipe out humanity.(And the AI would decide to do so.) Therefore, we should notbuild AI.• A sufficiently powerful AI could develop new medicaltechnologies capable of saving millions of human lives.

(Andthe AI would decide to do so.) Therefore, we should build AI.• Once computers become cheap enough, the vast majorityof jobs will be performable by Artificial Intelligence moreeasily than by humans. A sufficiently powerful AI would evenbe better than us at math, engineering, music, art, and all theother jobs we consider meaningful. (And the AI will decide toperform those jobs.)Thus after the invention of AI, humans will have nothing todo, and we’ll starve or watch television.The temptation is to ask what “AIs” will “want,” forgettingthat the space of minds-in-general is much wider than thetiny human dot. One should resist the temptation to spreadquantifiers over all possible minds. Storytellers spinning talesof the distant and exotic land called Future, say how thefuture will be.

They make predictions. They say,“AIs willattack humans with marching robot armies” or “AIs willinvent a cure for cancer.” They do not propose complex51relations between initial conditions and outcomes — thatwould lose the audience. But we need relationalunderstanding to manipulate the future, steer it into a regionpalatable to humankind. If we do not steer, we run the dangerof ending up where we are going.The critical challenge is not to predict that “AIs” will attackhumanity with marching robot armies, or alternatively inventa cure for cancer.

The task is not even to make the predictionfor an arbitrary individual AI design. Rather the task ischoosing into existence some particular powerful optimizationprocess whose beneficial effects can legitimately be asserted.Natural selection isn’t friendly, nor does it hate you, nor willitleaveyoualone.Evolutioncannotbesoanthropomorphized, it does not work like you do. Many pre1960s biologists expected natural selection to do all sorts ofnice things, and rationalized all sorts of elaborate reasons whynatural selection would do it. They were disappointed,because natural selection itself did not start out knowing thatit wanted a humanly-nice result, and then rationalizeelaborate ways to produce nice results using selectionpressures. Thus the events in Nature were outputs of causallydifferent process from what went on in the pre-1960sbiologists’ minds, so that prediction and reality diverged.Technical Failure and Philosophical FailureWe can divide potential failures of attempted Friendly AI intotwo informal fuzzy categories, technical failure andphilosophical failure.