When asked their opinions about “human-level artificial intelligence” — aka “artificial general intelligence” (AGI)¹ — many experts understandably reply that these terms haven’t yet been precisely defined, and it’s hard to talk about something that hasn’t been defined.² In this post, I want to briefly outline an imprecise but useful “working definition” for intelligence we tend to use at MIRI. In a future post I will write about some useful working definitions for artificial general intelligence.

Imprecise definitions can be useful

Precise definitions are important, but I concur with Bertrand Russell that

[You cannot] start with anything precise. You have to achieve such precision… as you go along.

Physicist Milan Ćirković agrees, and gives an example:

The formalization of knowledge — which includes giving precise definitions — usually comes at the end of the original research in a given field, not at the very beginning. A particularly illuminating example is the concept of number, which was properly defined in the modern sense only after the development of axiomatic set theory in the… twentieth century.³

For a more AI-relevant example, consider the concept of a “self-driving car,” which has been given a variety of vague definitions since the 1930s. Would a car guided by a buried cable qualify? What about a modified 1955 Studebaker that could use sound waves to detect obstacles and automatically engage the brakes if necessary, but could only steer “on its own” if each turn was preprogrammed? Does that count as a “self-driving car”?

What about the “VaMoRs” of the 1980s that could avoid obstacles and steer around turns using computer vision, but weren’t advanced enough to be ready for public roads? How about the 1995 Navlab car that drove across the USA and was fully autonomous for 98.2% of the trip, or the robotic cars which finished the 132-mile off-road course of the 2005 DARPA Grand Challenge, supplied only with the GPS coordinates of the route? What about the winning cars of the 2007 DARPA Grand Challenge, which finished an urban race while obeying all traffic laws and avoiding collisions with other cars? Does Google’s driverless car qualify, given that it has logged more than 500,000 autonomous miles without a single accident under computer control, but still struggles with difficult merges and snow-covered roads?⁴

Our lack of a precise definition for “self-driving car” doesn’t seem to have hindered progress on self-driving cars very much.⁵ And I’m glad we didn’t wait to seriously discuss self-driving cars until we had a precise definition for the term.

Similarly, I don’t think we should wait for a precise definition of AGI before discussing the topic seriously. On the other hand, the term is useless if it carries no information. So let’s work our way toward a stipulative, operational definition for AGI. We’ll start by developing an operational definition for intelligence.

Read more »

1. I use the HLAI and AGI interchangeably, but lately I’ve been using AGI almost exclusively, because I’ve learned that many people in the AI community react negatively to any mention of “human-level” AI but have no objection to the concept of narrow vs. general intelligence. See also Ben Goertzel’s comments here. ↩
2. Asked when he thought HLAI would be created, Pat Hayes (a past president of AAAI) replied: “I do not consider this question to be answerable, as I do not accept this (common) notion of ‘human-level intelligence’ as meaningful.” Asked the same question, AI scientist William Uther replied: “You ask a lot about ‘human level AGI’. I do not think this term is well defined,” while AI scientist Alan Bundy replied: “I don’t think the concept of ‘human-level machine intelligence’ is well formed.” ↩
3. Sawyer (1943) gives another example: “Mathematicians first used the sign √-1, without in the least knowing what it could mean, because it shortened work and led to correct results. People naturally tried to find out why this happened and what √-1 really meant. After two hundreds years they succeeded.” Dennett (2013) makes a related comment: “Define your terms, sir! No, I won’t. That would be premature… My [approach] is an instance of nibbling on a tough problem instead of trying to eat (and digest) the whole thing from the outset… In Elbow Room, I compared my method to the sculptor’s method of roughing out the form in a block of marble, approaching the final surfaces cautiously, modestly, working by successive approximation.” ↩
4. With self-driving cars, researchers did use many precise external performance measures (e.g. accident rates, speed, portion of the time they could run unassisted, frequency of getting stuck) to evaluate progress, as well as internal performance metrics (speed of search, bounded loss guarantees, etc.). Researchers could see that these bits of progress were in the right direction, even if their relative contribution long-term was unclear. And so it is with AI in general. AI researchers use many precise external and internal performance measures to evaluate progress, but it is difficult to know the relative contribution of these bits of progress toward the final goal of AGI. ↩
5. Heck, we’ve had pornography for millennia and still haven’t been able to define it precisely. Encyclopedia entries for “pornography” often simply quote Justice Potter Stewart: “I shall not today attempt further to define the kinds of material I understand to be [pornography]… but I know it when I see it.” ↩

MIRI was founded in 2000 on the premise that creating¹ Friendly AI might be a particularly efficient way to do as much good as possible.

Some developments since then include:

• The field of “effective altruism” — trying not just to do good but to do as much good as possible² — has seen more publicity and better research than ever before, in particular through the work of GiveWell, the Center for Effective Altruism, the philosopher Peter Singer, and the community at Less Wrong.³
• In his recent PhD dissertation, Nick Beckstead has clarified the assumptions behind the claim that shaping the far future (e.g. via Friendly AI) is overwhelmingly important.
• Due to research performed by MIRI, the Future of Humanity Institute (FHI), and others, our strategic situation with regard to machine superintelligence is more clearly understood, and FHI’s Nick Bostrom has organized much of this work in a forthcoming book.⁴
• MIRI’s Eliezer Yudkowsky has begun to describe in more detail which open research problems constitute “Friendly AI research,” in his view.

Given these developments, we are in a better position than ever before to assess the value of Friendly AI research as effective altruism.

Still, this is a difficult question. It is challenging enough to evaluate the cost-effectiveness of anti-malaria nets or direct cash transfers. Evaluating the cost-effectiveness of attempts to shape the far future (e.g. via Friendly AI) is even more difficult than that. Hence, this short post sketches an argument that can be given in favor of Friendly AI research as effective altruism, to enable future discussion, and is not intended as a thorough analysis.

Read more »

1. In this post, I talk about the value of humanity in general creating Friendly AI, though MIRI co-founder Eliezer Yudkowsky usually talks about MIRI in particular — or at least, a functional equivalent — creating Friendly AI. This is because I am not as confident as Yudkowsky that it is best for MIRI to attempt to build Friendly AI. When updating MIRI’s bylaws in early 2013, Yudkowsky and I came to a compromise on the language of MIRI’s mission statement, which now reads: “[MIRI] exists to ensure that the creation of smarter-than-human intelligence has a positive impact. Thus, the charitable purpose of [MIRI] is to: (a) perform research relevant to ensuring that smarter-than-human intelligence has a positive impact; (b) raise awareness of this important issue; (c) advise researchers, leaders and laypeople around the world; and (d) as necessary, implement a smarter-than-human intelligence with humane, stable goals” (emphasis added). My own hope is that it will not be necessary for MIRI (or a functional equivalent) to attempt to build Friendly AI itself. But of course I must remain open to the possibility that this will be the wisest course of action as the first creation of AI draws nearer. There is also the question of capability: few people think that a non-profit research organization has much chance of being the first to build AI. I worry, however, that the world’s elites will not find it fashionable to take this problem seriously until the creation of AI is only a few decades away, at which time it will be especially difficult to develop the mathematics of Friendly AI in time, and humanity will be forced to take a gamble on its very survival with powerful AIs we have little reason to trust. ↩
2. One might think of effective altruism as a straightforward application of decision theory to the subject of philanthropy. Philanthropic agents of all kinds (individuals, groups, foundations, etc.) ask themselves: “How can we choose philanthropic acts (e.g. donations) which (in expectation) will do as much good as possible, given what we care about?” The consensus recommendation for all kinds of choices under uncertainty, including philanthropic choices, is to maximize expected utility (Chater & Oaksford 2012; Peterson 2004; Stein 1996; Schmidt 1998:19). Different philanthropic agents value different things, but decision theory suggests that each of them can get the most of what they want if they each maximize their expected utility. Choices which maximize expected utility are in this sense “optimal,” and thus another term for effective altruism is “optimal philanthropy.” Note that effective altruism in this sense is not too dissimilar from earlier approaches to philanthropy, including high-impact philanthropy (making “the biggest difference possible, given the amount of capital invested“), strategic philanthropy, effective philanthropy, and wise philanthropy. Note also that effective altruism does not say that a philanthropic agent should specify complete utility and probability functions over outcomes and then compute the philanthropic act with the highest expected utility — that is impractical for bounded agents. We must keep in mind the distinction between normative, descriptive, and prescriptive models of decision-making (Baron 2007): “normative models tell us how to evaluate… decisions in terms of their departure from an ideal standard. Descriptive models specify what people in a particular culture actually do and how they deviate from the normative models. Prescriptive models are designs or inventions, whose purpose is to bring the results of actual thinking into closer conformity to the normative model.” The prescriptive question — about what bounded philanthropic agents should do to maximize expected utility with their philanthropic choices — tends to be extremely complicated, and is the subject of most of the research performed by the effective altruism community. ↩
3. See, for example: Efficient Charity, Efficient Charity: Do Unto Others, Politics as Charity, Heuristics and Biases in Charity, Public Choice and the Altruist’s Burden, On Charities and Linear Utility, Optimal Philanthropy for Human Beings, Purchase Fuzzies and Utilons Separately, Money: The Unit of Caring, Optimizing Fuzzies and Utilons: The Altruism Chip Jar, Efficient Philanthropy: Local vs. Global Approaches, The Effectiveness of Developing World Aid, Against Cryonics & For Cost-Effective Charity, Bayesian Adjustment Does Not Defeat Existential Risk Charity, How to Save the World, and What is Optimal Philanthropy? ↩
4. I believe Beckstead and Bostrom have done the research community an enormous service in creating a framework, a shared language, for discussing trajectory changes, existential risks, and machine superintelligence. When discussing these topics with my colleagues, it has often been the case that the first hour of conversation is spent merely trying to understand what the other person is saying — how they are using the terms and concepts they employ. Beckstead’s and Bostrom’s recent work should enable clearer and more efficient communication between researchers, and therefore greater research productivity. Though I am not aware of any controlled, experimental studies on the effect of shared language on research productivity, a shared language is widely considered to be of great benefit for any field of research, and I shall provide a few examples of this claim which appear in print. Fuzzi et al. (2006): “The use of inconsistent terms can easily lead to misunderstandings and confusion in the communication between specialists from different [disciplines] of atmospheric and climate research, and may thus potentially inhibit scientific progress.” Hinkel (2008): “Technical languages enable their users, e.g. members of a scientific discipline, to communicate efficiently about a domain of interest.” Madin et al. (2007): “terminological ambiguity slows scientific progress, leads to redundant research efforts, and ultimately impedes advances towards a unified foundation for ecological science.” ↩