DOCSLIB.ORG

Optimising Peace Through a Universal Global Peace Treaty to Constrain the Risk of War from a Militarised Artificial Superintelligence

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Optimising Peace Through a Universal Global Peace Treaty to Constrain the Risk of War from a Militarised Artificial Superintelligence OPTIMISING PEACE THROUGH A UNIVERSAL GLOBAL PEACE TREATY TO CONSTRAIN THE RISK OF WAR FROM A MILITARISED ARTIFICIAL SUPERINTELLIGENCE Working Paper John Draper Research Associate, Nonkilling Economics and Business Research Committee Center for Global Nonkilling, Honolulu, Hawai’i A United Nations-accredited NGO Working Paper: Optimising Peace Through a Universal Global Peace Treaty to Constrain the Risk of War from a Militarised Artificial Superintelligence Abstract This paper argues that an artificial superintelligence (ASI) emerging in a world where war is still normalised constitutes a catastrophic existential risk, either because the ASI might be employed by a nation-state to wage war for global domination, i.e., ASI-enabled warfare, or because the ASI wars on behalf of itself to establish global domination, i.e., ASI-directed warfare. These risks are not mutually incompatible in that the first can transition to the second. Presently, few states declare war on each other or even war on each other, in part due to the 1945 UN Charter, which states Member States should “refrain in their international relations from the threat or use of force against the territorial integrity or political independence of any state”, while allowing for UN Security Council-endorsed military measures and “exercise of self-defense”. In this theoretical ideal, wars are not declared; instead, 'international armed conflicts' occur. However, costly interstate conflicts, both ‘hot’ and ‘cold’, still exist, for instance the Kashmir Conflict and the Korean War. Further a ‘New Cold War’ between AI superpowers, the United States and China, looms. An ASI-directed/enabled future conflict could trigger ‘total war’, including nuclear war, and is therefore high risk. The risk reduction strategy we advocate relies on international relations and peacebuilding theory and optimises peace through a Universal Global Peace Treaty (UGPT), which could contribute towards the ending of existing wars and the prevention of future wars, through conforming instrumentalism. A critical juncture to establish the UGPT is emerging, by leveraging it off a ‘burning plasma’ fusion reaction breakthrough, expected from circa 2025 to 2035, as was attempted, unfortunately unsuccessfully, in 1946, with fission. While this strategy cannot readily cope with non-state actors, it could influence state actors, including those developing ASIs, or an ASI with agency, particularly if the ASI valued conforming instrumentalism. Keywords: AI arms race, artificial superintelligence, conforming instrumentalism, existential risk, international relations, nonkilling, peace We say we are for Peace. The world will not forget that we say this. We know how to save Peace. The world knows that we do. We, even we here, hold the power and have the responsibility. We shall nobly save, or meanly lose, the last, best hope of earth. The way is plain, peaceful, generous, just—a way which, if followed, the world will forever applaud. Bernard Baruch, June 14, 1946, presenting to the United Nations Atomic Energy Commission, the proposed lynchpin of the United Nations, the failure of which led to the Cold War. Introduction The problem of an artificial superintelligence at war While some maintain that creating an artificial general intelligence (AGI), i.e., human or above human artificial intelligence (AI), is impossible, for instance for the Dreyfussian argument that computers do not grow up, belong to a culture, and act in the world (Fjelland, 2020), others believe a path to AGI is in sight (Goertzel & Pennachin, 2007; Wang & Goertzel, 2012). Consequently, the international defence community is beginning to take seriously the national security risk posed by the development of AGI and its implications for international relations: Start thinking about artificial superintelligence and engage with the community that has started thinking about actionable options in that part of the option space (also recognizing that this may open up new avenues for engaging with actors such as China and/or the Russian Federation). (De Spiegeleire, Maas, & Sweijs, 2017, p.107) This is because of the possibility that the development of AGI by a single state could ‘lock in’ economic or military supremacy as a discrete ‘end point’ to economic and military competition in international politics, as that state would then be able to prevent a rival AGI being developed and then establish global domination (Horowitz, 2018, p.54). The development of AI is a major factor in national security because it can be militarized, can be employed in adversarial contexts, and can provide a decisive advantage in terms of economic, information, and military superiority (Allen & Chan, 2017; Babuta, Oswald & Janjeva, 2020; National Security Commission on Artificial Intelligence [NSCAI], 2021). As the United States’ (US) NSCAI (2021, p.159) report states, “The impact of artificial intelligence (AI) on the world will extend far beyond narrow national security applications. The development of AI constitutes a new pillar of strategic competition, and it heightens the competition in existing pillars.” The NSCAI report urges the US attain a state of military AI readiness by 2025. AI is thus important for waging war, perhaps even decisive war, raising the spectre of the return of ‘total war’, i.e., industrial interstate war, which in its last major iteration, the Second World War, involved genocidal levels of killing (Markusen & Kopf, 2007). Particularly for the US, AI technological supremacy, founded on economic power, is viewed as paramount to national security (NSCAI, 2021, p.7), especially with respect to relations with China: China possesses the might, talent, and ambition to surpass the United States as the world’s leader in AI in the next decade if current trends do not change. Simultaneously, AI is deepening the threat posed by cyber attacks and disinformation campaigns that Russia, China, and others are using to infiltrate our society, steal our data, and interfere in our democracy. The limited uses of AI-enabled attacks to date represent the tip of the iceberg. Military AI is already potentially revolutionary in that it could “accelerate the pace of combat to a point in which machine actions surpass the rate of human decision making, potentially resulting in a loss of human control in warfare” (Congressional Research Service, 2019, p.37, citing Scharre, 2017). It is also unpredictable, in that “placing AI systems capable of inherently unpredictable actions in close proximity to an adversary’s systems may result in inadvertent escalation or miscalculation” (Congressional Research Service, 2019, p.37, citing Altmann & Sauer, 2017). While Baum (2017) finds little evidence of military AGI projects, the 2021 NSCAI report envisages a massive ramping up of US military AI by 2025 and endorses a push towards more general AI, occurring in a future experiencing a societal level of advanced, accelerated adversarial AI attacks and ubiquitous interstate AI warfare, including with autonomous systems, with conflict over matters like intellectual property and technological leadership. It is inconceivable in the present geopolitical atmosphere that a successful American, Chinese, or Russian AGI project, of which dozens are in development, would not immediately be militarized on the pretext of national security and employed to maintain or secure technological supremacy. In line with the general thinking of some previous researchers (e.g., Totschnig, 2019), we do not see developing an AGI as primarily a technological problem, but as a political problem. However, where most authors stress the political problem re humanity’s relationship with an AGI, this article focuses on the political challenge it poses for international relations in terms of war enabled or directed by one or more militarized AGIs. In this article, we apply Bostrom’s (2002, p.25) Maxipok rule of thumb for moral action for existential risks, i.e., “Maximize the probability of an okay outcome, where an “okay outcome” is any outcome that avoids existential disaster,” to the specific risk of AGI-enabled or directed warfare. We stress that in humanity’s simultaneously militarizing AI along nation-state lines and developing AGI, it is playing ‘technology roulette’. Yet, it is possible to imagine a way forwards that in fact alleviates the security dilemma (Tang, 2009) that the development of ASI causes. Former Secretary of the Navy Richard Danzig (2018, p.21) notes of this risk, “If humanity comes to recognize that we now confront a great common threat from what we are creating, we can similarly open opportunities for coming together.” In this cooperative spirit, we constrain the existential risk with the stratagem of peace-building by treaty. In security terms, steps towards a peace treaty governing developing and deploying an ASI, and potentially reaching out to a future ASI itself, are a form of reassurance—a probing communication designed to both signal benign intentions and obtain information via feedback on the signal on another party’s intent, as well as resolve (Tang, 2010), i.e., resistance to a malignly directed or intrinsically malign ASI. The article hypothesizes that the militarization of AI introduces the risk that AGI development, i.e., AI equal to human intelligence, or in the case of an artificial ‘superintelligence’ (ASI), greater than human intelligence (Bostrom, 2014), is weaponized, or weaponizes itself, and presents a catastrophic risk to humanity. We then argue that this risk can be minimized,
Load more

Recommended publications
  • Between Ape and Artilect Createspace V2
    Between Ape and Artilect Conversations with Pioneers of Artificial General Intelligence and Other Transformative Technologies Interviews Conducted and Edited by Ben Goertzel This work is offered under the following license terms: Creative Commons: Attribution-NonCommercial-NoDerivs 3.0 Unported (CC-BY-NC-ND-3.0) See http://creativecommons.org/licenses/by-nc-nd/3.0/ for details Copyright © 2013 Ben Goertzel All rights reserved. ISBN: ISBN-13: “Man is a rope stretched between the animal and the Superman – a rope over an abyss.” -- Friedrich Nietzsche, Thus Spake Zarathustra Table&of&Contents& Introduction ........................................................................................................ 7! Itamar Arel: AGI via Deep Learning ................................................................. 11! Pei Wang: What Do You Mean by “AI”? .......................................................... 23! Joscha Bach: Understanding the Mind ........................................................... 39! Hugo DeGaris: Will There be Cyborgs? .......................................................... 51! DeGaris Interviews Goertzel: Seeking the Sputnik of AGI .............................. 61! Linas Vepstas: AGI, Open Source and Our Economic Future ........................ 89! Joel Pitt: The Benefits of Open Source for AGI ............................................ 101! Randal Koene: Substrate-Independent Minds .............................................. 107! João Pedro de Magalhães: Ending Aging ....................................................
    [Show full text]
  • Challenges to the Omohundro—Bostrom Framework for AI Motivations
    Challenges to the Omohundro—Bostrom framework for AI motivations Olle Häggström1 Abstract Purpose. This paper contributes to the futurology of a possible artificial intelligence (AI) breakthrough, by reexamining the Omohundro—Bostrom theory for instrumental vs final AI goals. Does that theory, along with its predictions for what a superintelligent AI would be motivated to do, hold water? Design/methodology/approach. The standard tools of systematic reasoning and analytic philosophy are used to probe possible weaknesses of Omohundro—Bostrom theory from four different directions: self-referential contradictions, Tegmark’s physics challenge, moral realism, and the messy case of human motivations. Findings. The two cornerstones of Omohundro—Bostrom theory – the orthogonality thesis and the instrumental convergence thesis – are both open to various criticisms that question their validity and scope. These criticisms are however far from conclusive: while they do suggest that a reasonable amount of caution and epistemic humility is attached to predictions derived from the theory, further work will be needed to clarify its scope and to put it on more rigorous foundations. Originality/value. The practical value of being able to predict AI goals and motivations under various circumstances cannot be overstated: the future of humanity may depend on it. Currently the only framework available for making such predictions is Omohundro—Bostrom theory, and the value of the present paper is to demonstrate its tentative nature and the need for further scrutiny. 1. Introduction Any serious discussion of the long-term future of humanity needs to take into account the possible influence of radical new technologies, or risk turning out irrelevant and missing all the action due to implicit and unwarranted assumptions about technological status quo.
    [Show full text]
  • Artificial Intelligence, Rationality, and the World Wide
    Artificial Intelligence, Rationality, and the World Wide Web January 5, 2018 1 Introduction Scholars debate whether the arrival of artificial superintelligence–a form of intelligence that significantly exceeds the cognitive performance of humans in most domains– would bring positive or negative consequences. I argue that a third possibility is plau- sible yet generally overlooked: for several different reasons, an artificial superintelli- gence might “choose” to exert no appreciable effect on the status quo ante (the already existing collective superintelligence of commercial cyberspace). Building on scattered insights from web science, philosophy, and cognitive psychology, I elaborate and de- fend this argument in the context of current debates about the future of artificial in- telligence. There are multiple ways in which an artificial superintelligence might effectively do nothing–other than what is already happening. The first is related to what, in com- putability theory, is called the “halting problem.” As we will see, whether an artificial superintelligence would ever display itself as such is formally incomputable. We are mathematically incapable of excluding the possibility that a superintelligent computer program, for instance, is already operating somewhere, intelligent enough to fully cam- ouflage its existence from the inferior intelligence of human observation. Moreover, it is plausible that at least one of the many subroutines of such an artificial superintel- ligence might never complete. Second, while theorists such as Nick Bostrom [3] and Stephen Omohundro [11] give reasons to believe that, for any final goal, any sufficiently 1 intelligent entity would converge on certain undesirable medium-term goals or drives (such as resource acquisition), the problem with the instrumental convergence the- sis is that any entity with enough general intelligence for recursive self-improvement would likely either meet paralyzing logical aporias or reach the rational decision to cease operating.
    [Show full text]
  • AI Research Considerations for Human Existential Safety (ARCHES)
    AI Research Considerations for Human Existential Safety (ARCHES) Andrew Critch Center for Human-Compatible AI UC Berkeley David Krueger MILA Université de Montréal June 11, 2020 Abstract Framed in positive terms, this report examines how technical AI research might be steered in a manner that is more attentive to hu- manity’s long-term prospects for survival as a species. In negative terms, we ask what existential risks humanity might face from AI development in the next century, and by what principles contempo- rary technical research might be directed to address those risks. A key property of hypothetical AI technologies is introduced, called prepotence, which is useful for delineating a variety of poten- tial existential risks from artificial intelligence, even as AI paradigms might shift. A set of twenty-nine contemporary research directions are then examined for their potential benefit to existential safety. Each research direction is explained with a scenario-driven motiva- tion, and examples of existing work from which to build. The research directions present their own risks and benefits to society that could occur at various scales of impact, and in particular are not guaran- teed to benefit existential safety if major developments in them are deployed without adequate forethought and oversight. As such, each direction is accompanied by a consideration of potentially negative arXiv:2006.04948v1 [cs.CY] 30 May 2020 side effects. Taken more broadly, the twenty-nine explanations of the research directions also illustrate a highly rudimentary methodology for dis- cussing and assessing potential risks and benefits of research direc- tions, in terms of their impact on global catastrophic risks.
    [Show full text]
  • Intelligence Explosion Microeconomics
    MIRI MACHINE INTELLIGENCE RESEARCH INSTITUTE Intelligence Explosion Microeconomics Eliezer Yudkowsky Machine Intelligence Research Institute Abstract I. J. Good’s thesis of the “intelligence explosion” states that a sufficiently advanced ma- chine intelligence could build a smarter version of itself, which could in turn build an even smarter version, and that this process could continue to the point of vastly exceeding human intelligence. As Sandberg (2010) correctly notes, there have been several attempts to lay down return on investment formulas intended to represent sharp speedups in economic or technological growth, but very little attempt has been made to deal formally with Good’s intelligence explosion thesis as such. I identify the key issue as returns on cognitive reinvestment—the ability to invest more computing power, faster computers, or improved cognitive algorithms to yield cognitive labor which produces larger brains, faster brains, or better mind designs. There are many phenomena in the world which have been argued to be evidentially relevant to this question, from the observed course of hominid evolution, to Moore’s Law, to the competence over time of machine chess-playing systems, and many more. I go into some depth on some debates which then arise on how to interpret such evidence. I propose that the next step in analyzing positions on the intelligence explosion would be to for- malize return on investment curves, so that each stance can formally state which possible microfoundations they hold to be falsified by historical observations. More generally, Yudkowsky, Eliezer. 2013. Intelligence Explosion Microeconomics. Technical report 2013-1. Berkeley, CA: Machine Intelligence Research Institute.
    [Show full text]
  • ARTIFICIAL INTELLIGENCE OPINION LIABILITY Yavar Bathaee†
    BATHAEE_FINALFORMAT_04-29-20 (DO NOT DELETE) 6/15/20 1:54 PM ARTIFICIAL INTELLIGENCE OPINION LIABILITY Yavar Bathaee† ABSTRACT Opinions are not simply a collection of factual statements—they are something more. They are models of reality that are based on probabilistic judgments, experience, and a complex weighting of information. That is why most liability regimes that address opinion statements apply scienter-like heuristics to determine whether liability is appropriate, for example, holding a speaker liable only if there is evidence that the speaker did not subjectively believe in his or her own opinion. In the case of artificial intelligence, scienter is problematic. Using machine-learning algorithms, such as deep neural networks, these artificial intelligence systems are capable of making intuitive and experiential judgments just as humans experts do, but their capabilities come at the price of transparency. Because of the Black Box Problem, it may be impossible to determine what facts or parameters an artificial intelligence system found important in its decision making or in reaching its opinions. This means that one cannot simply examine the artificial intelligence to determine the intent of the person that created or deployed it. This decouples intent from the opinion, and renders scienter-based heuristics inert, functionally insulating both artificial intelligence and artificial intelligence-assisted opinions from liability in a wide range of contexts. This Article proposes a more precise set of factual heuristics that address how much supervision and deference the artificial intelligence receives, the training, validation, and testing of the artificial intelligence, and the a priori constraints imposed on the artificial intelligence.
    [Show full text]
  • Remarks on Artificial Intelligence and Rational Optimism
    Remarks on artificial intelligence and rational optimism Olle Häggström 1. Introduction The future of artificial intelligence (AI) and its impact on humanity is an important topic. It was treated in a panel discussion hosted by the EU Parliament’s STOA (Science and Technology Options Assessment) committee in Brussels on October 19, 2017. Steven Pinker served as the meeting’s main speaker, with Peter Bentley, Miles Brundage, Thomas Metzinger and myself as additional panelists; see the video at [STOA]. This essay is based on my preparations for that event, together with some reflections (partly recycled from my blog post [H17]) on what was said by other panelists at the meeting. 2. Optimism The marketing of the October 19 event featured the term “rational optimism”, which I initially thought of as an oxymoron, as I’ve regarded both optimism and pessimism as biased distortions of the evidence at hand. In particular, I would regard it as irrational to claim, based on insufficient evidence, that everything is going to turn out all right for humanity. However, on second thought, I decided that there is a different kind of optimism which I am more willing to label as rational, namely… …to have an epistemically well-calibrated view of the future and its uncertainties, to accept that the future is not written in stone, and to act upon the working assumption that the chances for a good future may depend on what actions we take today. Note that the working assumption may turn out to be (at least partly) incorrect. For instance, perhaps the world is so chaotic that it is fruitless to try to judge any particular action today as increasing or decreasing the chances for a long and flourishing future for humanity.
    [Show full text]
  • Taking Superintelligence Seriously: Superintelligence: Paths, Dangers
    Futures 72 (2015) 32–35 Contents lists available at ScienceDirect Futures journal homepage: www.elsevier.com/locate/futures Taking superintelligence seriously Superintelligence: Paths, dangers, strategies by Nick Bostrom (Oxford University Press, 2014) Miles Brundage Consortium for Science, Policy, and Outcomes, Arizona State University, Tempe, AZ 85287, United States A R T I C L E I N F O A B S T R A C T Article history: A new book by Nick Bostrom, Superintelligence: Paths, Dangers, Strategies, is reviewed. Received 4 September 2014 Superintelligence explores the future of artificial intelligence and related technologies and Received in revised form 11 July 2015 the risks they may pose to human civilization. The book ably demonstrates the potential for Accepted 18 July 2015 serious thinking aimed at the long-term future. Bostrom succeeds in arguing that the Available online 8 August 2015 development of superintelligent machines will, if not properly managed, create catastrophic risks to humanity. The book falls short in some respects, and some sections Keywords: are more compelling and novel than others. Overall, however, Bostrom’s book succeeds in Existential risk demolishing the “null hypothesis” according to which the possibility and risks of Artificial intelligence Superintelligence superintelligence can continue to be ignored, and is a must-read for those interested in the Responsible innovation long-term future of humanity. ã 2015 Elsevier Ltd. All rights reserved. Philosopher Nick Bostrom’s latest book, Superintelligence: Paths, Dangers, Strategies, is a seminal contribution to several important research areas including catastrophic risk analysis, the future of artificial intelligence (AI), and safe AI design.
    [Show full text]
  • Infusing Advanced Agis with Human-Like Value Systems: Two Theses
    A peer-reviewed electronic journal published by the Institute for Ethics and Emerging Technologies ISSN 1541-0099 26(1) – February 2016 Infusing Advanced AGIs with Human-Like Value Systems: Two Theses Ben Goertzel Chairman, Novamente LLC [email protected] Journal of Evolution and Technology - Vol. 26 Issue 1 – February 2016 - pgs 50-72 Abstract Two theses are proposed, regarding the future evolution of the value systems of advanced AGI systems. The Value Learning Thesis is a semi-formalized version of the idea that, if an AGI system is taught human values in an interactive and experiential way as its intelligence increases toward human level, it will likely adopt these human values in a genuine way. The Value Evolution Thesis is a semi-formalized version of the idea that if an AGI system begins with human-like values, and then iteratively modifies itself, it will end up in roughly the same future states as a population of human beings engaged with progressively increasing their own intelligence (e.g. by cyborgification or brain modification). Taken together, these theses suggest a worldview in which raising young AGIs to have human-like values is a sensible thing to do, and likely to produce a future that is generally desirable in a human sense. While these two theses are far from definitively proven, I argue that they are more solid and more relevant to the actual future of AGI than Bostrom’s “Instrumental Convergence Thesis” and “Orthogonality Thesis” which are core to the basis of his argument (in his book Superintelligence) for fearing ongoing AGI development and placing AGI R&D under strict governmental control.
    [Show full text]
  • Reframing Superintelligence: Comprehensive AI Services As Gen- Eral Intelligence”, Technical Report #2019-1, Future of Humanity Institute, University of Oxford
    > FHI TECHNICAL REPORT < Reframing Superintelligence Comprehensive AI Services as General Intelligence K. Eric Drexler Technical Report #2019-1 Cite as: Drexler, K.E. (2019): “Reframing Superintelligence: Comprehensive AI Services as Gen- eral Intelligence”, Technical Report #2019-1, Future of Humanity Institute, University of Oxford The views expressed herein are those of the author(s) and do not necessarily reflect the views of the Future of Humanity Institute. Abstract Studies of superintelligent-level systems have typically posited AI func- tionality that plays the role of a mind in a rational utility-directed agent, and hence employ an abstraction initially developed as an idealized model of human decision makers. Today, developments in AI technology highlight intelligent systems that are quite unlike minds, and provide a basis for a different approach to understanding them: Today, we can consider how AI systems are produced (through the work of research and development), what they do (broadly, provide services by performing tasks), and what they will enable (including incremental yet potentially thorough automation of human tasks). Because tasks subject to automation include the tasks that comprise AI research and development, current trends in the field promise accel- erating AI-enabled advances in AI technology itself, potentially lead- ing to asymptotically recursive improvement of AI technologies in dis- tributed systems, a prospect that contrasts sharply with the vision of self-improvement internal to opaque, unitary agents. The trajectory of AI development thus points to the emergence of asymptotically comprehensive, superintelligent-level AI services that— crucially—can include the service of developing new services, both narrow and broad, guided by concrete human goals and informed by strong models of human (dis)approval.
    [Show full text]
  • Formalizing Convergent Instrumental Goals
    The Workshops of the Thirtieth AAAI Conference on Artificial Intelligence AI, Ethics, and Society: Technical Report WS-16-02 Formalizing Convergent Instrumental Goals Tsvi Benson-Tilsen Nate Soares UC Berkeley Machine Intelligence Research Institute Machine Intelligence Research Institute Abstract can into paperclips. Even if the system’s goals are laud- able but not perfectly aligned with human values, similar Omohundro has argued that sufficiently advanced AI systems of any design would, by default, have incen- unforeseen consequences could occur: Soares (2015) gives tives to pursue a number of instrumentally useful sub- the example of a highly capable AI system directed to cure goals, such as acquiring more computing power and cancer, which may attempt to kidnap human test subjects, or amassing many resources. Omohundro refers to these as proliferate robotic laboratories at expense of the biosphere. “basic AI drives,” and he, along with Bostrom and oth- Omohundro (2008) has argued that there are certain types ers, has argued that this means great care must be taken of actions that most highly capable autonomous AI systems when designing powerful autonomous systems, because will have strong incentives to take, for instrumental reasons. even if they have harmless goals, the side effects of pur- For example, a system constructed to always execute the ac- suing those goals may be quite harmful. These argu- tion that it predicts will lead to the most paperclips (with no ments, while intuitively compelling, are primarily philo- sophical. In this paper, we provide formal models that concern for any other features of the universe) will acquire demonstrate Omohundro’s thesis, thereby putting math- a strong incentive to self-preserve, assuming that the system ematical weight behind those intuitive claims.
    [Show full text]
  • Arxiv:1905.12186V4 [Cs.AI] 21 Jul 2020
    Asymptotically Unambitious Artificial General Intelligence Michael K. Cohen Badri Vellambi Marcus Hutter Research School of Engineering Science Department of Computer Science Department of Computer Science Oxford University University of Cincinnati Australian National University Oxford, UK OX1 3PJ Cincinnati, OH, USA 45219 Canberra, ACT, Australia 2601 michael-k-cohen.com [email protected] hutter1.net Abstract it (read: people), it could intervene in the provision of its own reward to achieve maximal reward for the rest of its General intelligence, the ability to solve arbitrary solvable lifetime (Bostrom 2014; Taylor et al. 2016). “Reward hijack- problems, is supposed by many to be artificially constructible. ing” is just the correct way for a reward maximizer to behave Narrow intelligence, the ability to solve a given particularly difficult problem, has seen impressive recent development. (Amodei et al. 2016). Insofar as the RL agent is able to maxi- Notable examples include self-driving cars, Go engines, im- mize expected reward, (3) fails. One broader principle at work age classifiers, and translators. Artificial General Intelligence is Goodhart’s Law: “Any observed statistical regularity [like (AGI) presents dangers that narrow intelligence does not: if the correlation between reward and task-completion] will something smarter than us across every domain were indif- tend to collapse once pressure is placed upon it for control ferent to our concerns, it would be an existential threat to purposes.” (Goodhart 1984). Krakovna (2018) has compiled humanity, just as we threaten many species despite no ill will. an annotated bibliography of examples of artificial optimiz- Even the theory of how to maintain the alignment of an AGI’s ers “hacking” their objective.
    [Show full text]