Getting AI Right: Introductory Notes on AI & Society

This dialogue is from an early scene in the 2014 film Ex Machina, which Nathan has invited Caleb to determine whether succeeded creating artificial intelligence.1 The achievement of powerful general intelligence long held a grip on our imagination not only for its exciting as well worrisome possibilities, but also suggestion new, uncharted era humanity. In opening his 2021 BBC Reith Lectures, titled “Living with Artificial Intelligence,” Stuart Russell states that “the eventual emergence general-purpose [will be] biggest event human history.”2Over last decade, rapid succession impressive results brought wider public attention possibilities intelligence. machine vision, researchers demonstrated systems could recognize objects as, if better than, humans some situations. Then came games. Complex games strategy have been associated superior intelligence, and so when AI beat best players at chess, Atari games, Go, shogi, StarCraft, Dota, world took notice. It was just Als (although astounding it first happened), escalating progression how they did it: initially by learning expert play, then self-play, teaching themselves principles ground up, eventually yielding single learn, win several structurally different hinting possibility generally intelligent systems.3Speech recognition natural language processing seen headline-grabbing advances. Most recently large models capable generating human-like outputs. Progress particular significance given role always played notions reasoning, understanding. While advances mentioned thus far may seem abstract, those driverless cars robots more tangible their embodied often biomorphic forms. Demonstrations such exhibiting increasingly complex autonomous behaviors physical captured attention.Also headlines various branches science related techniques used tools advance research materials environmental sciences high energy physics astronomy.4 A few highlights, spectacular fifty-year-old protein-folding problem AlphaFold, suggest soon help tackle science's hardest problems, health life sciences.5While tend feature demonstrations future come, technologies are already here pervade daily lives than many realize. Examples include recommendation systems, search, translators - now covering one hundred languages facial recognition, speech text (and back), digital assistants, chatbots customer service, fraud detection, decision support management scientific research, name few. all these examples others, AI-related become components other software hardware methods incorporating messy real-world inputs into inferences, predictions, and, cases, actions. As director Future Humanity Institute University Oxford, Nick Bostrom noted back 2006, “A lot cutting-edge filtered applications, without being called because once something becomes useful enough common it's labeled anymore.”6As scope, use, usefulness grown individual users, fields, companies types organizations, governments, too concerns worked (such bias systems), or misused (as deepfakes), resulted harms (in predicting crime, example), accidents fatalities self-driving cars).7Dædalus devoted volume topic 1988, contributions founders field, among others. Much issue concerned questions making progress, turning point, foundations, mathematical, technical, philosophical-with much disagreement. However, there perhaps rediscovery, alternative path toward connectionist approach notion neural nets-and burgeoning optimism this approach's potential. Since 1960s, had relegated fringes favor symbolic formalism representing world, knowledge it, machines can reason about it. Yet no essay mood time Hilary Putnam's “Much Ado About Not Very Much.” Putnam questioned Dædalus itself: “Why whole Dædalus? Why don't we wait until achieves issue?” He concluded:This indeed since 1988 be does rehash same debates; else happened since, mostly result success rediscovered reimagined, discussed volume. aims capture where AI's development growing uses impact society. themes herein colored my own involvement AI. Besides television, films, books I grew up with, interest began earnest 1989 when, undergraduate Zimbabwe, undertook project model train network.9 went do robotics Oxford. Over years, involved academia labs developing studying economy, tracking working others business, policy, labor grappling opportunities challenges society.10The authors twenty-five essays range scientists technologists frontier developments social forefront analyzing impacts organized ten sections. Half sections focused development, half intersections aspects addition diversity topics, expertise, vantage points, bring views benefits, am grateful accepting invitation write essays.Before proceeding further, say what mean increasing pervasiveness led conflation confusion exactly counts helped current trend-among humanities, startups, established companies, even governments-to associate anything involving learning, data science, algorithms, robots, automation sorts simply reflect hype AI, acknowledgment wave wide-ranging use usefulness. think both true; like this. period referred winter, during progress live expectations, reticence most call AI.Two definitions typically ability artificially beings, usually human, do. For example, is:The abilities invoked visual perception, capacity reason, solve discover meaning, generalize, learn experience. Definitions type considered limiting human-centricity benchmarks set (more later). second try free define agent system, whatever origin, makeup, method, as:This definition suggests pursuit goals, self-generated, learned.13 That employed throughout yields insights own.These definitional distinctions notwithstanding, term chagrin come cognitive computer scientist Marvin Minsky “suitcase word.”14 packed variously, depending who you ask, approaches achieving including based logic, probability, information control theory, networks, inference, planning methods, instantiations software, hardware, case perceive, move, manipulate objects.Three cut through discussions volume: 1) Where development? 2) What pose society? 3) How really us?Notions date way antiquity.15 Philosophers, too, them Hobbes, Leibnitz, Descartes, dreaming time; Daniel Dennett Descartes anticipated Turing Test.16 idea computation-based traces Alan Turing's invention universal 1930s, ideas contemporaries mid-twentieth century. But birth know attributed famed Dartmouth summer workshop 1956. proposal two-month John McCarthy, Minsky, Nathaniel Rochester, Claude Shannon whereby “An attempt will made find make language, form abstractions concepts, kinds problems reserved humans, improve themselves.”17In respective volume, “From So Simple Beginning: Species Intelligence” “If We Succeed,” complementary ways, Nigel Shadbolt chart key periods excitement aforementioned winters. spring underway 1990s, breakthroughs appearing over years so: Jeffrey Dean describes title “golden decade,” pace wide sectors society, areas research.18 characterized achieve experience, deep reinforcement together probability ways learn.19A brief history here: 1950s, were two dominant visions One vision computers create logic representation there, akin mind. espoused Allen Newell Hebert Simon, along Closely “heuristic search” supposed essentially exploring space answers. inspired brain, rather mind, sought learning. became known approach, units perceptrons connected connection neurons brain. At time, Frank Rosenblatt. initial visions, dominate, decades, successes, so-called systems.Not benefit championing advocates plentiful funding, suggested weight intellectual tradition-exemplified Boole, Frege, Russell, Church, others-that symbols formalize axiomatize reasoning. late 1980s grow again largely work David Rumelhart, Geoffrey Hinton, James McClelland, philosophical Hubert Dreyfus Dreyfus's “Making Mind Versus Modeling Brain: Intelligence Back Branchpoint.”20 then, statistical back-propagation, training (supervised unsupervised) characterize approach.Kevin Scott, “I Do Think Means You Means: Intelligence, Cognitive Work & Scale,” reminds us Ray Solomonoff linking theory compress potentially generalize emerging realization built come. benefited boon availability algorithms thanks growth Internet applications services. explosion new instruments observation platforms data-generating breakthroughs, astronomy genomics. Equally important co-evolution used, especially chip architectures suited parallel computations data- compute-intensive networks approaches, discusses.Several delve subfields AI.21 essay, “Searching Computer Vision North Stars,” Fei-Fei Li Ranjay Krishna creation standard sets ImageNet benchmarking performance. “Human Language Understanding Reasoning” “The Curious Case Commonsense Chris Manning Yejin Choi discuss eras processing, recent comprising hundreds billions parameters transformer self-supervised vast amounts data.22 resulting pretrained take prompts trained specifically generate outputs, images, code, more, Mira Murati discusses illustrates “Language Coding Creativity.” Some started refer foundational trained, adaptable tasks outputs.23 despite unexpected performance, still shortcomings limitations highlighted elsewhere, developers.24In Machines Our Future,” Daniela Rus robotic underlying technologies, integrated design enables operate world. She highlights “industrial” conceptualizing draw biological systems. robotics, generally, tension copy inspiration organisms behavior. Elsewhere, researcher Demis Hassabis colleagues explored neuroscience inspire each other, although direction Alexis Baria Keith Cross suggested.25Despite limitations, conceptually hard AI.26 distinguish hand problematic shortcomings, perform intended safely, produces biased toxic outputs lead harm, impinges privacy, generates false characteristics lack explainability, loss trust. These rightly regulatory bodies, researchers, whom increased focus technical ethics issues.27 flurry efforts develop responsible bodies industry academia, Partnership aim share practices.28 Another shortcoming significant diversity-especially respect gender race people researching documented years.29 gap right, consequently, society broadly.On hand, things yet that, solved, powerful, capable, Lecture, pioneers Yoshua Bengio, Yann LeCun, Hinton stock stands difficulties out-of-distribution generalization.30 highlight reasoning commonsense understanding, surprising performance models. computational linguists Emily Bender Alexander Koller challenged resembling meaning.31 “Multi-Agent Systems: Technical Ethical Challenges Functioning Mixed Group,” Kobi Gal Barbara Grosz multi-agent highlighting conceptual difficulties-such agents, belief intentionality-as ethical cooperative competitive settings, agents machines. Allan Dafoe provide overview open AI.32 Indeed, sense adequate theories sociotechnical embedding scope societal expands.And proving elsewhere illustrate-including embedded capabilities evaluate steer experiments going beyond heuristics-based experimental termed “self-driving laboratories”33-getting understand mathematics theorize novel concepts remain grand AI.34 Indeed discoveries enable game-changing humanities greatest opportunities, motivation build systems.Beyond particulars subfield list continue limit includes one-shot cross-domain generalizations, causal grounding, complexities timescales memory, meta-cognition.35 Consideration raises question approaches-mostly building larger multimodal models, learning-are sufficient, entirely needed addition, neuroscience-inspired semantic representations On kind additional might needed, community divided, believe approaches36 further evolution compute reach limits.37The debate sufficiency closely achieved, so, when. (AGI) defined distinction sometimes narrow AI: is, developed fine-tuned specific playing chess. AGI, least humans-that applicable any situation conceptions, evolve itself, goals preferences. Though whether, how, AGI achieved matter debate, agree would profound implications-beneficial worrisome-for humanity, depicted popular books38 films 2001: Space Odyssey Terminator Matrix Machina Her. Whether imminent not, agreement should prepare safety control, alignment compatibility governance multiple varieties emerge, factor considerations AGI.Most investment, commercial activity today variety numerous forms: terms speciation hardly potential economic gains economy.39 organizations primary goal. Among well-known DeepMind OpenAI, generality, though AGI.Perhaps widely jobs work. new. 1964, wake era's automation, jobs, President Lyndon Johnson empaneled National Commission Technology, Automation, Economic Progress.40 commission's conclusions prosperity basic fact technology destroys work.” studies effect, in, reached similar gained lost. sectoral occupational transitions, skill wage effects-not existence broadly-that present challenges.41 “Automation, Work,” Laura Tyson Zysman implications workers. Michael Spence goes Augmentation, Value Creation Distribution Income Wealth,” distributional issues income wealth within between countries, created, countries. Trap: Promise Peril Human-Like Erik Brynjolfsson runs risk substitutes for, complements, labor. concludes regard, outcomes work, depend incentives governments.42Still, concern remains conclusion created lost draws patterns past look of. arguments why break technology-driven change include: first, unlike past, technological happening faster markets (including workers) systems' adapt slow mismatched; second, now, mechanized routine tasks, forward, taking nonroutine creative tacit knowledge, indication, socioempathic out question.43 words, “There think, create. Moreover, increase rapidly until-in visible future-the handle coextensive mind applied.” Herbert Simon 1957.44Acknowledging elicits responses: First, emerge value done sake, doing humans. response material abundance, need labor, scale abundance sufficient everyone's needs. And happens, humanity face challenge Keynes framed: “For man faced real, permanent problem-how freedom pressing cares, occupy leisure, compound won him, wisely agreeably well.”45 close majority Keynes's challenge, AI- automation-related effects must addressed near future, inequality effects, education, skilling, alongside machines-issues Zysman, Spence, volume.Jobs aspect economy impacted provides directional estimate huge bounty fully realized: global GDP $750 trillion, times today's GDP. before get realized productivity competitiveness ensure intense competition countries deployment, national level, while United States ahead, acknowledged China fast becoming major player evidenced infrastructure, ecosystems, reports.46 Such likely market structure Eric Schmidt, elsewhere.47 dynamics requiring collective action safety) competitors, Amanda Askell, Miles Brundage, Gillian Hadfield highlighted.48Nations reasons want security-in surveillance, signals cyber operations, defense battle-space superiority, weapons, disinformation forms sociopolitical warfare-is clear. “AI, Great Power Competition Security,” Schmi