Smarter Than Us: The Rise of Machine Intelligenceby Published 01 Feb 2014
|Smarter Than Us: The Rise of Machine Intelligence.pdf|
|Publisher||Machine Intelligence Research Institute|
What happens when machines become smarter than humans? Forget lumbering Terminators. The power of an artificial intelligence (AI) comes from its intelligence, not physical strength and laser guns. Humans steer the future not because we're the strongest or the fastest but because we're the smartest. When machines become smarter than humans, we'll be handing them the steering wheel. What promises—and perils—will these powerful machines present? Stuart Armstrong’s new book navigates these questions with clarity and wit.
Can we instruct AIs to steer the future as we desire? What goals should we program into them? It turns out this question is difficult to answer! Philosophers have tried for thousands of years to define an ideal world, but there remains no consensus. The prospect of goal-driven, smarter-than-human AI gives moral philosophy a new urgency. The future could be filled with joy, art, compassion, and beings living worthwhile and wonderful lives—but only if we’re able to precisely define what a "good" world is, and skilled enough to describe it perfectly to a computer program.
AIs, like computers, will do what we say—which is not necessarily what we mean. Such precision requires encoding the entire system of human values for an AI: explaining them to a mind that is alien to us, defining every ambiguous term, clarifying every edge case. Moreover, our values are fragile: in some cases, if we mis-define a single piece of the puzzle—say, consciousness—we end up with roughly 0% of the value we intended to reap, instead of 99% of the value.
Though an understanding of the problem is only beginning to spread, researchers from fields ranging from philosophy to computer science to economics are working together to conceive and test solutions. Are we up to the challenge?
A mathematician by training, Armstrong is a Research Fellow at the Future of Humanity Institute (FHI) at Oxford University. His research focuses on formal decision theory, the risks and possibilities of AI, the long term potential for intelligent life (and the difficulties of predicting this), and anthropic (self-locating) probability. Armstrong wrote Smarter Than Us at the request of the Machine Intelligence Research Institute, a non-profit organization studying the theoretical underpinnings of artificial superintelligence.
"Smarter Than Us: The Rise of Machine Intelligence" Reviews
A humorous read on a serious subject - the possible perils of an uncontrolled intelligence explosion. I found it fun and informative - a great primer for both newbies and those well versed in the idea of an intelligence explosion/technological singularity.
It is succinct and easy to read - definitely worth the time.
1st part of a video interview on the book here: https://www.youtube.com/watch?v=RotSh...
A very concise and clear book on the benefits and the potential dangers the rise of artificial intelligence poses to humanity. It really opened up my mind to what is possible, and how difficult it is to create (or even just define) 'safe' AI.
First, I love that the book is published under a Creative Commons license. That shows the author cares about spreading his ideas as far as possible, and understands that copyright restrictions merely shrink the audience unless you manage to write the next Harry Potter and the Sorcerer's Stone. I was tempted to give the book five stars just for that.
The book itself is short and easy to read, at least for anyone with a modicum of computer science and philosophy. It summarizes the potential dangers of AI and even more briefly tells the reader how to help. While Hollywood has labored for decades to instill fear of robots one day taking over, the book explains how an actual AI disaster scenario could play out differently than the standard robot movie script. But more importantly, the book outlines the intellectual problems that building a friendly AI poses - including the problem of precisely defining what constitutes human well-being, a problem that has frustrated philosophers for more than 2400 years. Even if some unforeseen technical barrier prevents AIs gaining general intelligence, the kind that could outsmart humans across the board, humans should benefit from more attention paid to moral philosophy. Thus the book should be helpful even in the unlikely event it turns out to be unnecessary.
But being short, the book has to leave a lot out. These omissions struck me as rather curious:
Chapter 7 (What, Precisely, Do We Really (Really) Want?) describes the difficulty of specifying a goal so as to exclude all solutions having undesirable consequences. The chapter does not mention the extensive fictional literature exploring this very quandary, for example The Monkey's Paw and the List of adaptations of The Monkey's Paw. "Be careful what you wish for" is an ancient fictional trope. King Midas anyone? The book does not need to cover the whole history of this trope, but at least give it a nod. People did think about this before computers threatened to make it real.
It also couldn't hurt to mention that when you hire two humans to do a job for you, the smarter worker usually figures out what you really want, more reliably than the less smart worker. When you hire the most expensive attorney, for example, part of what you are paying for is expert advice on what your goals are. Part of the intelligent expert's job is to disambiguate your initial, perhaps vague or counterproductive, request. If you ask the competent expert to do something he or she knows you probably won't be happy with, for example to pursue a legal strategy with a high risk of backfire, the expert will seek to dissuade you. If AIs will exceed every human skill, as the book predicts, then maybe they will help us figure out what we really want.
A second omission seems to ignore the book's title. A recurring theme throughout the book is that it's up to humans to make AI friendly, or at least contain it, all by ourselves. That requires humans to solve the problem of containing AIs, or the problem of building moral compasses into them, or more generally the problem of predicting whatever troubling unknown unknowns they might unleash on us and designing in safeguards against those threats. But the book also predicts that AIs will eventually exceed every human skill (thereby becoming Smarter Than Us) - wouldn't that have to include the skill of building friendly AIs? If thorny problems of moral philosophy have to be solved along the way, AIs should solve them faster than we can.
A third omission - or at least an underemphasis - is the question of why AIs would care about the goals we give them. If a committee of ants presented a human with a list of goals, why would the human care what the ants want? It's not enough merely to tell the AI precisely what we want it to do. The AI must also "want" to do precisely what we command. The book does mention that an AI would have the power to out-think the humans who supposedly control it, but so the AI could more efficiently pursue its original goals (programmed in by humans). Why wouldn't the AI subvert its human masters more comprehensively, by inventing its own goals, which less intelligent humans cannot imagine? Maybe we won't be programming these things, but propitiating them, much as ancient and modern superstitious people believe they are propitiating their God or gods. If ants were to propitiate humans, who knows, we might listen.
If AIs do become smarter than us, we'd better hope Sam Harris is correct in The Moral Landscape: How Science Can Determine Human Values, namely that science really does (or can someday) determine moral values. If so, in that best of all possible worlds, scientifically skilled AIs should converge on the same morality that any other scientifically competent entity would discover, only faster, but perhaps with a few catastrophes along the way. (Humans have also stumbled repeatedly during their long moral evolution - remember slavery? Homophobia? Sexism? Beheadings? Man-made climate change? History and current events are replete with human moral failings.) It doesn't seem too Pollyannish to suppose AIs should turn out to be scientifically skilled, given that science is perhaps the most useful human skill. If AIs will have all our skills, only better, then maybe AIs could be our natural allies in the ancient quest for moral progress. Maybe AIs could teach us what it means to be moral.
KURZWEILIGER SPENDENAUFRUF FÜR KI-RISIKOFORSCHUNG
Stephen Hawking verglich einmal die künstliche Intelligenz der Zukunft mit der Ankündigung eines Besuchs überlegener Aliens in einigen Jahrzehnten, um dann zu fragen: Was würden wir mit diesem Wissen heute schon tun? Würden wir Vorkehrungen treffen?
Einfach abwarten wollen Oxfords "Future of Humanity Institute" (FHI) und das "Machine Intelligence Research Institute" (MIRI) jedenfalls nicht. Sie möchten sicherstellen, dass das Herstellen einer "smarter-than-human intelligence" positive Folgen hat; genauer noch, dass "advanced intelligent machines behave as we intend even in the absence of immediate human supervision" (MIRI mission statement). Gefördert von merkwürdigen Gründungen wie der "Saving Humanity from Homo Sapiens Foundation" (SHFHS.org).
Als kurzweiliger 60-Seiten-Spendenaufruf fürs FHI bzw. MIRI bzw. für die KI-Risikoforschung im Allgemeinen spürt das Heftchen u.a. Probleme der Mensch-Maschine-Kommunikation nach - menschliche Gehirne sind für die Kommunikation mit anderen Menschen optimiert, d.h. unpräzise, uneindeutig und implizit; es prognostiziert eine zunehmende gesellschaftliche Spaltung zwischen denen, die KI für sich nutzen können, und einem abgehängten Rest; es schaut auf die Möglichkeiten und Konsequenzen einer Superintelligenz, die schließlich Menschen überlistet, um vorgegebene Ziele zu erfüllen: Der beste Plan ist einer, der auch ausgeführt wird. Außerdem skaliert menschliche Kontrolle nicht, sie wird als Performance-Bottleneck zum Wettbewerbsnachteil und zunehmend unattraktiv; Gleichermaßen drehen sich Abhängigkeiten um, wenn künstliche Intelligenzen unsere gebaute Umwelt steuern und wir die Zusammenhänge nicht mehr verstehen - wie schaltet man so ein System ab? Was müsste eine "sichere KI" leisten?
Verstanden wird KI im Heft als die Fähigkeit von Maschinen menschliche Leistungen in vielen Bereichen nachahmend einzuholen oder - aufgrund der fehlenden biologischen Grenzen - sogar zu übertreffen. Dabei kann eine nennenswerte KI ihre Aufgaben unter verschiedenen Umweltbedingungen erfüllen, sie passt sich ihnen an. Was in Maschinen aber "innerlich" vor sich geht, sei für eine Technikfolgenabschätzung unwichtig. Äußerliche Maßstäbe zählen, das beobachtbare Verhalten und der reale Impact auf unsere Umwelt statt metaphysische Überlegungen: 'doing' statt 'being'.
"Smarter Than Us" ist vor allem eine Einführung, pointiert geschrieben und leicht verständlich; vielleicht etwas fear-mongering, aber nicht automatisch verkehrt. Technische Details fehlen weitestgehend, machte aber dennoch Spaß.
The book is titled 'The Rise of Machine Intelligence', but it hardly ever talks about the positive aspects of AI. The author does bring up a few valid points about the potential threat of AI, but overall it comes across as an attempt at fear-mongering. Many of the examples are Hollywood-inspired and some of them are just fundamentally flawed. The author talks about the complexities in ordering an AI to fetch a person from a burning building, and then goes on describe the need to define the person in terms of limbs, body parts, lifeline, etc. This is as absurd as saying that to write a program in a high level language that adds two numbers, one has to define the numeric system, explain the program in binary instructions, and go all the way down to voltages and integrated circuits. Neural networks are similar in that the higher layer neurons abstract away the details and can understand complex features composed of simpler lower layers. AI systems based on neural networks would have the ability to understand a human as a whole, making the author's example baseless.
Another overarching theme is that to build safe AI, one needs to define all the special cases by hand, and this wouldn't be possible as there are too many of them. This contradicts with the fundamentals of Machine Learning. Unsupervised learning takes the complexity of special-casing away and they learn features by looking at examples. Self-driving cars are not hard-coded with all the objects they are not supposed to crash into. Rather, they learn by watching the world (ala training data).
The book is an attempt to make people aware of the need to invest in precautions to prevent 'evil AI' from emerging, but falls short as the examples are superficial and far-fetched.