AI Researchers On AI Risk

by Scott Alexander 4y22nd May 2015No comments

8


I first became interested in AI risk back around 2007. At the time, most people’s response to the topic was “Haha, come back when anyone believes this besides random Internet crackpots.”

Over the next few years, a series of extremely bright and influential figures including Bill Gates, Stephen Hawking, and Elon Musk publically announced they were concerned about AI risk, along with hundreds of other intellectuals, from Oxford philosophers to MIT cosmologists to Silicon Valley tech investors. So we came back.

Then the response changed to “Sure, a couple of random academics and businesspeople might believe this stuff, but never real experts in the field who know what’s going on.”

Thus pieces like Popular Science’s Bill Gates Fears AI, But AI Researchers Know Better:

When you talk to A.I. researchers—again, genuine A.I. researchers, people who grapple with making systems that work at all, much less work too well—they are not worried about superintelligence sneaking up on them, now or in the future. Contrary to the spooky stories that Musk seems intent on telling, A.I. researchers aren’t frantically installed firewalled summoning chambers and self-destruct countdowns.

And Fusion.net’s The Case Against Killer Robots From A Guy Actually Building AI:

Andrew Ng builds artificial intelligence systems for a living. He taught AI at Stanford, built AI at Google, and then moved to the Chinese search engine giant, Baidu, to continue his work at the forefront of applying artificial intelligence to real-world problems. So when he hears people like Elon Musk or Stephen Hawking—people who are not intimately familiar with today’s technologies—talking about the wild potential for artificial intelligence to, say, wipe out the human race, you can practically hear him facepalming.

And now Ramez Naam of Marginal Revolution is trying the same thing with What Do AI Researchers Think Of The Risk Of AI?:

Elon Musk, Stephen Hawking, and Bill Gates have recently expressed concern that development of AI could lead to a ‘killer AI’ scenario, and potentially to the extinction of humanity. None of them are AI researchers or have worked substantially with AI that I know of. What do actual AI researchers think of the risks of AI?

It quotes the same couple of cherry-picked AI researchers as all the other stories – Andrew Ng, Yann LeCun, etc – then stops without mentioning whether there are alternate opinions.

There are. AI researchers, including some of the leaders in the field, have been instrumental in raising issues about AI risk and superintelligence from the very beginning. I want to start by listing some of these people, as kind of a counter-list to Naam’s, then go into why I don’t think this is a “controversy” in the classical sense that dueling lists of luminaries might lead you to expect.

The criteria for my list: I’m only mentioning the most prestigious researchers, either full professors at good schools with lots of highly-cited papers, or else very-well respected scientists in industry working at big companies with good track records. They have to be involved in AI and machine learning. They have to have multiple strong statements supporting some kind of view about a near-term singularity and/or extreme risk from superintelligent AI. Some will have written papers or books about it; others will have just gone on the record saying they think it’s important and worthy of further study.

If anyone disagrees with the inclusion of a figure here, or knows someone important I forgot, let me know and I’ll make the appropriate changes:

* * * * * * * * * *

Stuart Russell (wiki) is Professor of Computer Science at Berkeley, winner of the IJCAI Computers And Thought Award, Fellow of the Association for Computing Machinery, Fellow of the American Academy for the Advancement of Science, Director of the Center for Intelligent Systems, Blaise Pascal Chair in Paris, etc, etc. He is the co-author of Artificial Intelligence: A Modern Approach, the classic textbook in the field used by 1200 universities around the world. On his website, he writes:

The field [of AI] has operated for over 50 years on one simple assumption: the more intelligent, the better. To this must be conjoined an overriding concern for the benefit of humanity. The argument is very simple:

1. AI is likely to succeed.
2. Unconstrained success brings huge risks and huge benefits.
3. What can we do now to improve the chances of reaping the benefits and avoiding the risks?

Some organizations are already considering these questions, including the Future of Humanity Institute at Oxford, the Centre for the Study of Existential Risk at Cambridge, the Machine Intelligence Research Institute in Berkeley, and the Future of Life Institute at Harvard/MIT. I serve on the Advisory Boards of CSER and FLI.

Just as nuclear fusion researchers consider the problem of containment of fusion reactions as one of the primary problems of their field, it seems inevitable that issues of control and safety will become central to AI as the field matures. The research questions are beginning to be formulated and range from highly technical (foundational issues of rationality and utility, provable properties of agents, etc.) to broadly philosophical.

He makes a similar point on edge.org, writing:

As Steve Omohundro, Nick Bostrom, and others have explained, the combination of value misalignment with increasingly capable decision-making systems can lead to problems—perhaps even species-ending problems if the machines are more capable than humans. Some have argued that there is no conceivable risk to humanity for centuries to come, perhaps forgetting that the interval of time between Rutherford’s confident assertion that atomic energy would never be feasibly extracted and Szilárd’s invention of the neutron-induced nuclear chain reaction was less than twenty-four hours.

He has also tried to serve as an ambassador about these issues to other academics in the field, writing:

What I’m finding is that senior people in the field who have never publicly evinced any concern before are privately thinking that we do need to take this issue very seriously, and the sooner we take it seriously the better.

David McAllester (wiki) is professor and Chief Academic Officer at the U Chicago-affilitated Toyota Technological Institute, and formerly served on the faculty of MIT and Cornell. He is a fellow of the American Association of Artificial Intelligence, has authored over a hundred publications, has done research in machine learning, programming language theory, automated reasoning, AI planning, and computational linguistics, and was a major influence on the algorithms for famous chess computer Deep Blue. According to an article in the Pittsburgh Tribune Review:

Chicago professor David McAllester believes it is inevitable that fully automated intelligent machines will be able to design and build smarter, better versions of themselves, an event known as the Singularity. The Singularity would enable machines to become infinitely intelligent, and would pose an ‘incredibly dangerous scenario’, he says.

On his personal blog Machine Thoughts, he writes:

Most computer science academics dismiss any talk of real success in artificial intelligence. I think that a more rational position is that no one can really predict when human level AI will be achieved. John McCarthy once told me that when people ask him when human level AI will be achieved he says between five and five hundred years from now. McCarthy was a smart man. Given the uncertainties surrounding AI, it seems prudent to consider the issue of friendly AI…

The early stages of artificial general intelligence (AGI) will be safe. However, the early stages of AGI will provide an excellent test bed for the servant mission or other approaches to friendly AI. An experimental approach has also been promoted by Ben Goertzel in a nice blog post on friendly AI. If there is a coming era of safe (not too intelligent) AGI then we will have time to think further about later more dangerous eras.

He attended the AAAI Panel On Long-Term AI Futures, where he chaired the panel on Long-Term Control and was described as saying:

McAllester chatted with me about the upcoming ‘Singularity’, the event where computers out think humans. He wouldn’t commit to a date for the singularity but said it could happen in the next couple of decades and will definitely happen eventually. Here are some of McAllester’s views on the Singularity. There will be two milestones: Operational Sentience, when we can easily converse with computers, and the AI Chain Reaction, when a computer can bootstrap itself to a better self and repeat. We’ll notice the first milestone in automated help systems that will genuinely be helpful. Later on computers will actually be fun to talk to. The point where computer can do anything humans can do will require the second milestone.

Hans Moravec (wiki) is a former professor at the Robotics Institute of Carnegie Mellon University, namesake of Moravec’s Paradox, and founder of the SeeGrid Corporation for industrial robotic visual systems. His Sensor Fusion in Certainty Grids for Mobile Robots has been cited over a thousand times, and he was invited to write the Encyclopedia Britannica article on robotics back when encyclopedia articles were written by the world expert in a field rather than by hundreds of anonymous Internet commenters.

He is also the author of Robot: Mere Machine to Transcendent Mind, which Amazon describes as:

In this compelling book, Hans Moravec predicts machines will attain human levels of intelligence by the year 2040, and that by 2050, they will surpass us. But even though Moravec predicts the end of the domination by human beings, his is not a bleak vision. Far from railing against a future in which machines rule the world, Moravec embraces it, taking the startling view that intelligent robots will actually be our evolutionary heirs.” Moravec goes further and states that by the end of this process “the immensities of cyberspace will be teeming with unhuman superminds, engaged in affairs that are to human concerns as ours are to those of bacteria”.

Shane Legg is co-founder of DeepMind Technologies (wiki), an AI startup that was bought for Google in 2014 for about $500 million. He earned his PhD at the Dalle Molle Institute for Artificial Intelligence in Switzerland and also worked at the Gatsby Computational Neuroscience Unit in London. His dissertation Machine Superintelligence concludes:

If there is ever to be something approaching absolute power, a superintelligent machine would come close. By definition, it would be capable of achieving a vast range of goals in a wide range of environments. If we carefully prepare for this possibility in advance, not only might we avert disaster, we might bring about an age of prosperity unlike anything seen before.

In a later interview, he states:

AI is now where the internet was in 1988. Demand for machine learning skills is quite strong in specialist applications (search companies like Google, hedge funds and bio-informatics) and is growing every year. I expect this to become noticeable in the mainstream around the middle of the next decade. I expect a boom in AI around 2020 followed by a decade of rapid progress, possibly after a market correction. Human level AI will be passed in the mid 2020’s, though many people won’t accept that this has happened. After this point the risks associated with advanced AI will start to become practically important…I don’t know about a “singularity”, but I do expect things to get really crazy at some point after human level AGI has been created. That is, some time from 2025 to 2040.

He and his co-founders Demis Hassabis and Mustafa Suleyman have signed the Future of Life Institute petition on AI risks, and one of their conditions for joining Google was that the company agree to set up an AI Ethics Board to investigate these issues.

Steve Omohundro (wiki) is a former Professor of Computer Science at University of Illinois, founder of the Vision and Learning Group and the Center for Complex Systems Research, and inventor of various important advances in machine learning and machine vision. His work includes lip-reading robots, the StarLisp parallel programming language, and geometric learning algorithms. He currently runs Self-Aware Systems, “a think-tank working to ensure that intelligent technologies are beneficial for humanity”. His paper Basic AI Drives helped launch the field of machine ethics by pointing out that superintelligent systems will converge upon certain potentially dangerous goals. He writes:

We have shown that all advanced AI systems are likely to exhibit a number of basic drives. It is essential that we understand these drives in order to build technology that enables a positive future for humanity. Yudkowsky has called for the creation of ‘friendly AI’. To do this, we must develop the science underlying ‘utility engineering’, which will enable us to design utility functions that will give rise to the consequences we desire…The rapid pace of technological progress suggests that these issues may become of critical importance soon.”

See also his section here on “Rational AI For The Greater Good”.

Murray Shanahan (site) earned his PhD in Computer Science from Cambridge and is now Professor of Cognitive Robotics at Imperial College London. He has published papers in areas including robotics, logic, dynamic systems, computational neuroscience, and philosophy of mind. He is currently writing a book The Technological Singularity which will be published in August; Amazon’s blurb says:

Shanahan describes technological advances in AI, both biologically inspired and engineered from scratch. Once human-level AI — theoretically possible, but difficult to accomplish — has been achieved, he explains, the transition to superintelligent AI could be very rapid. Shanahan considers what the existence of superintelligent machines could mean for such matters as personhood, responsibility, rights, and identity. Some superhuman AI agents might be created to benefit humankind; some might go rogue. (Is Siri the template, or HAL?) The singularity presents both an existential threat to humanity and an existential opportunity for humanity to transcend its limitations. Shanahan makes it clear that we need to imagine both possibilities if we want to bring about the better outcome.

Marcus Hutter (wiki) is a professor in the Research School of Computer Science at Australian National University. He has previously worked with the Dalle Molle Institute for Artificial Intelligence and National ICT Australia, and done work on reinforcement learning, Bayesian sequence prediction, complexity theory, Solomonoff induction, computer vision, and genomic profiling. He has also written extensively on the Singularity. In Can Intelligence Explode?, he writes:

This century may witness a technological explosion of a degree deserving the name singularity. The default scenario is a society of interacting intelligent agents in a virtual world, simulated on computers with hyperbolically increasing computational resources. This is inevitably accompanied by a speed explosion when measured in physical time units, but not necessarily by an intelligence explosion…if the virtual world is inhabited by interacting free agents, evolutionary pressures should breed agents of increasing intelligence that compete about computational resources. The end-point of this intelligence evolution/acceleration (whether it deserves the name singularity or not) could be a society of these maximally intelligent individuals. Some aspect of this singularitarian society might be theoretically studied with current scientific tools. Way before the singularity, even when setting up a virtual society in our imagine, there are likely some immediate difference, for example that the value of an individual life suddenly drops, with drastic consequences.

Jurgen Schmidhuber (wiki) is Professor of Artificial Intelligence at the University of Lugano and former Professor of Cognitive Robotics at the Technische Universitat Munchen. He makes some of the most advanced neural networks in the world, has done further work in evolutionary robotics and complexity theory, and is a fellow of the European Academy of Sciences and Arts. In Singularity Hypotheses, Schmidhuber argues that “if future trends continue, we will face an intelligence explosion within the next few decades”. When asked directly about AI risk on a Reddit AMA thread, he answered:

Stuart Russell’s concerns [about AI risk] seem reasonable. So can we do anything to shape the impacts of artificial intelligence? In an answer hidden deep in a related thread I just pointed out: At first glance, recursive self-improvement through Gödel Machines seems to offer a way of shaping future superintelligences. The self-modifications of Gödel Machines are theoretically optimal in a certain sense. A Gödel Machine will execute only those changes of its own code that are provably good, according to its initial utility function. That is, in the beginning you have a chance of setting it on the “right” path. Others, however, may equip their own Gödel Machines with different utility functions. They will compete. In the resulting ecology of agents, some utility functions will be more compatible with our physical universe than others, and find a niche to survive. More on this in a paper from 2012.

Richard Sutton (wiki) is professor and iCORE chair of computer science at University of Alberta. He is a fellow of the Association for the Advancement of Artificial Intelligence, co-author of the most-used textbook on reinforcement learning, and discoverer of temporal difference learning, one of the most important methods in the field.

In his talk at the Future of Life Institute’s Future of AI Conference, Sutton states that there is “certainly a significant chance within all of our expected lifetimes” that human-level AI will be created, then goes on to say the AIs “will not be under our control”, “will compete and cooperate with us”, and that “if we make superintelligent slaves, then we will have superintelligent adversaries”. He concludes that “We need to set up mechanisms (social, legal, political, cultural) to ensure that this works out well” but that “inevitably, conventional humans will be less important.” He has also mentioned these issues at a presentation to the Gadsby Institute in London and in (of all things) a Glenn Beck book: “Richard Sutton, one of the biggest names in AI, predicts an intelligence explosion near the middle of the century”.

Andrew Davison (site) is Professor of Robot Vision at Imperial College London, leader of the Robot Vision Research Group and Dyson Robotics Laboratory, and inventor of the computerized localization-mapping system MonoSLAM. On his website, he writes:

At the risk of going out on a limb in the proper scientific circles to which I hope I belong(!), since 2006 I have begun to take very seriously the idea of the technological singularity: that exponentially increasing technology might lead to super-human AI and other developments that will change the world utterly in the surprisingly near future (i.e. perhaps the next 20–30 years). As well as from reading books like Kurzweil’s ‘The Singularity is Near’ (which I find sensational but on the whole extremely compelling), this view comes from my own overview of incredible recent progress of science and technology in general and specificially in the fields of computer vision and robotics within which I am personally working. Modern inference, learning and estimation methods based on Bayesian probability theory (see Probability Theory: The Logic of Science or free online version, highly recommended), combined with the exponentially increasing capabilities of cheaply available computer processors, are becoming capable of amazing human-like and super-human feats, particularly in the computer vision domain.

It is hard to even start thinking about all of the implications of this, positive or negative, and here I will just try to state facts and not offer much in the way of opinions (though I should say that I am definitely not in the super-optimistic camp). I strongly think that this is something that scientists and the general public should all be talking about. I’ll make a list here of some ‘singularity indicators’ I come across and try to update it regularly. These are little bits of technology or news that I come across which generally serve to reinforce my view that technology is progressing in an extraordinary, faster and faster way that will have consequences few people are yet really thinking about.

Alan Turing and I. J. Good (wiki, wiki) are men who need no introduction. Turing invented the mathematical foundations of computing and shares his name with Turing machines, Turing completeness, and the Turing Test. Good worked with Turing at Bletchley Park, helped build some of the first computers, and invented various landmark algorithms like the Fast Fourier Transform. In his paper “Can Digital Machines Think?”, Turing writes:

Let us now assume, for the sake of argument, that these machines are a genuine possibility, and look at the consequences of constructing them. To do so would of course meet with great opposition, unless we have advanced greatly in religious tolerance since the days of Galileo. There would be great opposition from the intellectuals who were afraid of being put out of a job. It is probable though that the intellectuals would be mistaken about this. There would be plenty to do in trying to keep one’s intelligence up to the standards set by the machines, for it seems probable that once the machine thinking method had started, it would not take long to outstrip our feeble powers…At some stage therefore we should have to expect the machines to take control.

During his time at the Atlas Computer Laboratory in the 60s, Good expanded on this idea in Speculations Concerning The First Ultraintelligent Machine, which argued:

Let an ultraintelligent machine be defined as a machine that can far surpass all the intellectual activities of any man however clever. Since the design of machines is one of these intellectual activities, an ultraintelligent machine could design even better machines; there would then unquestionably be an ‘intelligence explosion,’ and the intelligence of man would be left far behind. Thus the first ultraintelligent machine is the last invention that man need ever make

* * * * * * * * * *

I worry this list will make it look like there is some sort of big “controversy” in the field between “believers” and “skeptics” with both sides lambasting the other. This has not been my impression.

When I read the articles about skeptics, I see them making two points over and over again. First, we are nowhere near human-level intelligence right now, let alone superintelligence, and there’s no obvious path to get there from here. Second, if you start demanding bans on AI research then you are an idiot.

I agree whole-heartedly with both points. So do the leaders of the AI risk movement.

A survey of AI researchers (Muller & Bostrom, 2014) finds that on average they expect a 50% chance of human-level AI by 2040 and 90% chance of human-level AI by 2075. On average, 75% believe that superintelligence (“machine intelligence that greatly surpasses the performance of every human in most professions”) will follow within thirty years of human-level AI. There are some reasons to worry about sampling bias based on eg people who take the idea of human-level AI seriously being more likely to respond (though see the attempts made to control for such in the survey) but taken seriously it suggests that most AI researchers think there’s a good chance this is something we’ll have to worry about within a generation or two.

But outgoing MIRI director Luke Muehlhauser and Future of Humanity Institute director Nick Bostrom are both on record saying they have significantly later timelines for AI development than the scientists in the survey. If you look at Stuart Armstrong’s AI Timeline Prediction Data there doesn’t seem to be any general law that the estimates from AI risk believers are any earlier than those from AI risk skeptics. In fact, the latest estimate on the entire table is from Armstrong himself; Armstrong nevertheless currently works at the Future of Humanity Institute raising awareness of AI risk and researching superintelligence goal alignment.

The difference between skeptics and believers isn’t about when human-level AI will arrive, it’s about when we should start preparing.

Which brings us to the second non-disagreement. The “skeptic” position seems to be that, although we should probably get a couple of bright people to start working on preliminary aspects of the problem, we shouldn’t panic or start trying to ban AI research.

The “believers”, meanwhile, insist that although we shouldn’t panic or start trying to ban AI research, we should probably get a couple of bright people to start working on preliminary aspects of the problem.

Yann LeCun is probably the most vocal skeptic of AI risk. He was heavily featured in the Popular Science article, was quoted in the Marginal Revolution post, and spoke to KDNuggets and IEEE on “the inevitable singularity questions”, which he describes as “so far out that we can write science fiction about it”. But when asked to clarify his position a little more, he said:

Elon [Musk] is very worried about existential threats to humanity (which is why he is building rockets with the idea of sending humans colonize other planets). Even if the risk of an A.I. uprising is very unlikely and very far in the future, we still need to think about it, design precautionary measures, and establish guidelines. Just like bio-ethics panels were established in the 1970s and 1980s, before genetic engineering was widely used, we need to have A.I.-ethics panels and think about these issues. But, as Yoshua [Bengio] wrote, we have quite a bit of time

Eric Horvitz is another expert often mentioned as a leading voice of skepticism and restraint. His views have been profiled in articles like Out Of Control AI Will Not Kill Us, Believes Microsoft Research Chief and Nothing To Fear From Artificial Intelligence, Says Microsoft’s Eric Horvitz. But here’s what he says in a longer interview with NPR:

KASTE: Horvitz doubts that one of these virtual receptionists could ever lead to something that takes over the world. He says that’s like expecting a kite to evolve into a 747 on its own. So does that mean he thinks the singularity is ridiculous?

Mr. HORVITZ: Well, no. I think there’s been a mix of views, and I have to say that I have mixed feelings myself.

KASTE: In part because of ideas like the singularity, Horvitz and other A.I. scientists have been doing more to look at some of the ethical issues that might arise over the next few years with narrow A.I. systems. They’ve also been asking themselves some more futuristic questions. For instance, how would you go about designing an emergency off switch for a computer that can redesign itself?

Mr. HORVITZ: I do think that the stakes are high enough where even if there was a low, small chance of some of these kinds of scenarios, that it’s worth investing time and effort to be proactive.

Which is pretty much the same position as a lot of the most zealous AI risk proponents. With enemies like these, who needs friends?

A Slate article called Don’t Fear Artificial Intelligence also gets a surprising amount right:

As Musk himself suggests elsewhere in his remarks, the solution to the problem [of AI risk] lies in sober and considered collaboration between scientists and policymakers. However, it is hard to see how talk of “demons” advances this noble goal. In fact, it may actively hinder it.

First, the idea of a Skynet scenario itself has enormous holes. While computer science researchers think Musk’s musings are “not completely crazy,” they are still awfully remote from a world in which AI hype masks less artificially intelligent realities that our nation’s computer scientists grapple with:

Yann LeCun, the head of Facebook’s AI lab, summed it up in a Google+ post back in 2013: “Hype is dangerous to AI. Hype killed AI four times in the last five decades. AI Hype must be stopped.”…LeCun and others are right to fear the consequences of hype. Failure to live up to sci-fi–fueled expectations, after all, often results in harsh cuts to AI research budgets.

AI scientists are all smart people. They have no interest in falling into the usual political traps where they divide into sides that accuse each other of being insane alarmists or ostriches with their heads stuck in the sand. It looks like they’re trying to balance the need to start some preliminary work on a threat that looms way off in the distance versus the risk of engendering so much hype that it starts a giant backlash.

This is not to say that there aren’t very serious differences of opinion in how quickly we need to act. These seem to hinge mostly on whether it’s safe to say “We’ll deal with the problem when we come to it” or whether there will be some kind of “hard takeoff” which will take events out of control so quickly that we’ll want to have done our homework beforehand. I continue to see less evidence than I’d like that most AI researchers with opinions understand the latter possibility, or really any of the technical work in this area. Heck, the Marginal Revolution article quotes an expert as saying that superintelligence isn’t a big risk because “smart computers won’t create their own goals”, even though anyone who has read Bostrom knows that this is exactly the problem.

There is still a lot of work to be done. But cherry-picked articles about how “real AI researchers don’t worry about superintelligence” aren’t it.

[thanks to some people from MIRI and FLI for help with and suggestions on this post]

EDIT: Investigate for possible inclusion: Fredkin, Minsky

8