(Happy to get feedback on this! It draws from and expounds ideas in this article: http://jetpress.org/v26.2/torres.htm)

Consider a seemingly simple question: if the means were available, who exactly would destroy the world? There is surprisingly little discussion of this question within the nascent field of existential risk studies. But it’s an absolutely crucial issue: what sort of agent would either intentionally or accidentally cause an existential catastrophe?

The first step forward is to distinguish between two senses of an existential risk. Nick Bostrom originally defined the term as: “One where an adverse outcome would either annihilate Earth-originating intelligent life or permanently and drastically curtail its potential.” It follows that there are two distinct scenarios, one endurable and the other terminal, that could realize an existential risk. We can call the former an extinction risk and the latter a stagnation risk. The importance of this distinction with respect to both advanced technologies and destructive agents has been previously underappreciated.

So, the question asked above is actually two questions in disguise. Let’s consider each in turn.

Terror: Extinction Risks

First, the categories of agents who might intentionally cause an extinction catastrophe are fewer and smaller than one might think. They include:

(1) Idiosyncratic actors. These are malicious agents who are motivated by idiosyncratic beliefs and/or desires. There are instances of deranged individuals who have simply wanted to kill as many people as possible and then die, such as some school shooters. Idiosyncratic actors are especially worrisome because this category could have a large number of members (token agents). Indeed, the psychologist Martha Stout estimates that about 4 percent of the human population suffers from sociopathy, resulting in about 296 million sociopaths. While not all sociopaths are violent, a disproportionate number of criminals and dictators have (or very likely have) had the condition.

(2) Future ecoterrorists. As the effects of climate change and biodiversity loss (resulting in the sixth mass extinction) become increasingly conspicuous, and as destructive technologies become more powerful, some terrorism scholars have speculated that ecoterrorists could become a major agential risk in the future. The fact is that the climate is changing and the biosphere is wilting, and human activity is almost entirely responsible. It follows that some radical environmentalists in the future could attempt to use technology to cause human extinction, thereby “solving” the environmental crisis. So, we have some reason to believe that this category could become populated with a growing number of token agents in the coming decades.

(3) Negative utilitarians. Those who hold this view believe that the ultimate aim of moral conduct is to minimize misery, or “disutility.” Although some negative utilitarians like David Pearce see existential risks as highly undesirable, others would welcome annihilation because it would entail the elimination of suffering. It follows that if a “strong” negative utilitarian had a button in front of her that, if pressed, would cause human extinction (say, without causing pain), she would very likely press it. Indeed, on her view, doing this would be the morally right action. Fortunately, this version of negative utilitarianism is not a position that many non-academics tend to hold, and even among academic philosophers it is not especially widespread.

(4) Extraterrestrials. Perhaps we are not alone in the universe. Even if the probability of life arising on an Earth-analog is low, the vast number of exoplanets suggests that the probability of life arising somewhere may be quite high. If an alien species were advanced enough to traverse the cosmos and reach Earth, it would very likely have the technological means to destroy humanity. As Stephen Hawking once remarked, “If aliens visit us, the outcome would be much as when Columbus landed in America, which didn’t turn out well for the Native Americans.”

(5) Superintelligence. The reason Homo sapiens is the dominant species on our planet is due almost entirely to our intelligence. It follows that if something were to exceed our intelligence, our fate would become inextricably bound up with its will. This is worrisome because recent research shows that even slight misalignments between our values and those motivating a superintelligence could have existentially catastrophic consequences. But figuring out how to upload human values into a machine poses formidable problems — not to mention the issue of figuring out what our values are in the first place.

Making matters worse, a superintelligence could process information at about 1 million times faster than our brains, meaning that a minute of time for us would equal approximately 2 years in time for the superintelligence. This would immediately give the superintelligence a profound strategic advantage over us. And if it were able to modify its own code, it could potentially bring about an exponential intelligence explosion, resulting in a mind that’s many orders of magnitude smarter than any human. Thus, we may have only one chance to get everything just right: there’s no turning back once an intelligence explosion is ignited.

A superintelligence could cause human extinction for a number of reasons. For example, we might simply be in its way. Few humans worry much if an ant genocide results from building a new house or road. Or the superintelligence could destroy humanity because we happen to be made out of something it could use for other purposes: atoms. Since a superintelligence need not resemble human intelligence in any way — thus, scholars tell us to resist the dual urges of anthropomorphizing and anthropopathizing — it could be motivated by goals that appear to us as utterly irrational, bizarre, or completely inexplicable.

Terror: Stagnation Risks

Now consider the agents who might intentionally try to bring about a scenario that would result in a stagnation catastrophe. This list subsumes most of the list above in that it includes idiosyncratic actors, future ecoterrorists, and superintelligence, but it probably excludes negative utilitarians, since stagnation (as understood above) would likely induce more suffering than the status quo today. The case of extraterrestrials is unclear, given that we can infer almost nothing about an interstellar civilization except that it would be technologically sophisticated.

For example, an idiosyncratic actor could harbor not a death wish for humanity, but a “destruction wish” for civilization. Thus, she or he could strive to destroy civilization without necessarily causing the annihilation of Homo sapiens. Similarly, a future ecoterrorist could hope for humanity to return to the hunter-gatherer lifestyle. This is precisely what motivated Ted Kaczynski: he didn’t want everyone to die, but he did want our technological civilization to crumble. And finally, a superintelligence whose values are misaligned with ours could modify Earth in such a way that our lineage persists, but our prospects for future development are permanently compromised. Other stagnation scenarios could involve the following categories:

(6) Apocalyptic terrorists. History is overflowing with groups that not only believed the world was about to end, but saw themselves as active participants in an apocalyptic narrative that’s unfolding in realtime. Many of these groups have been driven by the conviction that “the world must be destroyed to be saved,” although some have turned their activism inward and advocated mass suicide.

Interestingly, no notable historical group has combined both the genocidal and suicidal urges. This is why apocalypticists pose a greater stagnation terror risk than extinction risk: indeed, many see their group’s survival beyond Armageddon as integral to the end-times, or eschatological, beliefs they accept. There are almost certainly less than about 2 million active apocalyptic believers in the world today, although emerging environmental, demographic, and societal conditions could cause this number to significantly increase in the future, as I’ve outlined in detail elsewhere (see Section 5 of this paper).

(7) States. Like terrorists motivated by political rather than transcendent goals, states tend to place a high value on their continued survival. It follows that states are unlikely to intentionally cause a human extinction event. But rogue states could induce a stagnation catastrophe. For example, if North Korea were to overcome the world’s superpowers through a sudden preemptive attack and implement a one-world government, the result could be an irreversible decline in our quality of life.

So, there are numerous categories of agents that could attempt to bring about an existential catastrophe. And there appear to be fewer agent types who would specifically try to cause human extinction than to merely dismantle civilization.

Error: Extinction and Stagnation Risks

There are some reasons, though, for thinking that error (rather than terror) could constitute the most significant threat in the future. First, almost every agent capable of causing intentional harm would also be capable of causing accidental harm, whether this results in extinction or stagnation. For example, an apocalyptic cult that wants to bring about Armageddon by releasing a deadly biological agent in a major city could, while preparing for this terrorist act, inadvertently contaminate its environment, leading to a global pandemic.

The same goes for idiosyncratic agents, ecoterrorists, negative utilitarians, states, and perhaps even extraterrestrials. (Indeed, the large disease burden of Europeans was a primary reason Native American populations were decimated. By analogy, perhaps an extraterrestrial destroys humanity by introducing a new type of pathogen that quickly wipes us out.) The case of superintelligence is unclear, since the relationship between intelligence and error-proneness has not been adequately studied.

Second, if powerful future technologies become widely accessible, then virtually everyone could become a potential cause of existential catastrophe, even those with absolutely no inclination toward violence. To illustrate the point, imagine a perfectly peaceful world in which not a single individual has malicious intentions. Further imagine that everyone has access to a doomsday button on her or his phone; if pushed, this button would cause an existential catastrophe. Even under ideal societal conditions (everyone is perfectly “moral”), how long could we expect to survive before someone’s finger slips and the doomsday button gets pressed?

Statistically speaking, a world populated by only 1 billion people would almost certainly self-destruct within a 10-year period if the probability of any individual accidentally pressing a doomsday button were a mere 0.00001 percent per decade. Or, alternatively: if only 500 people in the world were to gain access to a doomsday button, and if each of these individuals had a 1 percent chance of accidentally pushing the button per decade, humanity would have a meager 0.6 percent chance of surviving beyond 10 years. Thus, even if the likelihood of mistakes is infinitesimally small, planetary doom will be virtually guaranteed for sufficiently large populations.

The Two Worlds Thought Experiment

The good news is that a focus on agential risks, as I’ve called them, and not just the technological tools that agents might use to cause a catastrophe, suggests additional ways to mitigate existential risk. Consider the following thought-experiment: a possible world A contains thousands of advanced weapons that, if in the wrong hands, could cause the population of A to go extinct. In contrast, a possible world B contains only a single advanced “weapon of total destruction” (WTD). Which world is more dangerous? The answer is obviously world A.

But it would be foolishly premature to end the analysis here. Imagine further that A is populated by compassionate, peace-loving individuals, whereas B is overrun by war-mongering psychopaths. Now which world appears more likely to experience an existential catastrophe? The correct answer is, I would argue, world B.

In other words: agents matter as much as, or perhaps even more than, WTDs. One simply can’t evaluate the degree of risk in a situation without taking into account the various agents who could become coupled to potentially destructive artifacts. And this leads to the crucial point: as soon as agents enter the picture, we have another variable that could be manipulated through targeted interventions to reduce the overall probability of an existential catastrophe.

The options here are numerous and growing. One possibility would involve using “moral bioenhancement” techniques to reduce the threat of terror, given that acts of terror are immoral. But a morally enhanced individual might not be less likely to make a mistake. Thus, we could attempt to use cognitive enhancements to lower the probability of catastrophic errors, on the (tentative) assumption that greater intelligence correlates with fewer blunders.

Furthermore, implementing stricter regulations on CO2 emissions could decrease the probability of extreme ecoterrorism and/or apocalyptic terrorism, since environmental degradation is a “trigger” for both.

Another possibility, most relevant to idiosyncratic agents, is to reduce the prevalence of bullying (including cyberbullying). This is motivated by studies showing that many school shooters have been bullied, and that without this stimulus such individuals would have been less likely to carry out violent rampages. Advanced mind-reading or surveillance technologies could also enable law enforcement to identify perpetrators before mass casualty crimes are committed.

As for superintelligence, efforts to solve the “control problem” and create a friendly AI are of primary concern among many many researchers today. If successful, a friendly AI could itself constitute a powerful mitigation strategy for virtually all the categories listed above.

(Note: these strategies should be explicitly distinguished from proposals that target the relevant tools rather than agents. For example, Bostrom’s idea of “differential technological development” aims to neutralize the bad uses of technology by strategically ordering the development of different kinds of technology. Similarly, the idea of police “blue goo” to counter “grey goo” is a technology-based strategy. Space colonization is also a tool intervention because it would effectively reduce the power (or capacity) of technologies to affect the entire human or posthuman population.)

Agent-Tool Couplings

Devising novel interventions and understanding how to maximize the efficacy of known strategies requires a careful look at the unique properties of the agents mentioned above. Without an understanding of such properties, this important task will be otiose. We should also prioritize different agential risks based on the likely membership (token agents) of each category. For example, the number of idiosyncratic agents might exceed the number of ecoterrorists in the future, since ecoterrorism is focused on a single issue, whereas idiosyncratic agents could be motivated by a wide range of potential grievances.[1] We should also take seriously the formidable threat posed by error, which could be nontrivially greater than that posed by terror, as the back-of-the-envelope calculations above show.

Such considerations, in combination with technology-based risk mitigation strategies, could lead to a comprehensive, systematic framework for strategically intervening on both sides of the agent-tool coupling. But this will require the field of existential risk studies to become less technocentric than it currently is.

[1] Although, on the other hand, the stimulus of environmental degradation would be experienced by virtually everyone in society, whereas the stimuli that motivate idiosyncratic agents might be situationally unique. It’s precisely issues like these that deserve further scholarly research.

These are my notes and observations after attending the Safety in Artificial Intelligence (SafArtInt) conference, which was co-hosted by the White House Office of Science and Technology Policy and Carnegie Mellon University on June 27 and 28. This isn't an organized summary of the content of the conference; rather, it's a selection of points which are relevant to the control problem. As a result, it suffers from selection bias: it looks like superintelligence and control-problem-relevant issues were discussed frequently, when in reality those issues were discussed less and I didn't write much about the more mundane parts.

SafArtInt has been the third out of a planned series of four conferences. The purpose of the conference series was twofold: the OSTP wanted to get other parts of the government moving on AI issues, and they also wanted to inform public opinion.

The other three conferences are about near term legal, social, and economic issues of AI. SafArtInt was about near term safety and reliability in AI systems. It was effectively the brainchild of Dr. Ed Felten, the deputy U.S. chief technology officer for the White House, who came up with the idea for it last year. CMU is a top computer science university and many of their own researchers attended, as well as some students. There were also researchers from other universities, some people from private sector AI including both Silicon Valley and government contracting, government researchers and policymakers from groups such as DARPA and NASA, a few people from the military/DoD, and a few control problem researchers. As far as I could tell, everyone except a few university researchers were from the U.S., although I did not meet many people. There were about 70-100 people watching the presentations at any given time, and I had conversations with about twelve of the people who were not affiliated with existential risk organizations, as well as of course all of those who were affiliated. The conference was split with a few presentations on the 27th and the majority of presentations on the 28th. Not everyone was there for both days.

Felten believes that neither "robot apocalypses" nor "mass unemployment" are likely. It soon became apparent that the majority of others present at the conference felt the same way with regard to superintelligence. The general intention among researchers and policymakers at the conference could be summarized as follows: we need to make sure that the AI systems we develop in the near future will not be responsible for any accidents, because if accidents do happen then they will spark public fears about AI, which would lead to a dearth of funding for AI research and an inability to realize the corresponding social and economic benefits. Of course, that doesn't change the fact that they strongly care about safety in its own right and have significant pragmatic needs for robust and reliable AI systems.

Most of the talks were about verification and reliability in modern day AI systems. So they were concerned with AI systems that would give poor results or be unreliable in the narrow domains where they are being applied in the near future. They mostly focused on "safety-critical" systems, where failure of an AI program would result in serious negative consequences: automated vehicles were a common topic of interest, as well as the use of AI in healthcare systems. A recurring theme was that we have to be more rigorous in demonstrating safety and do actual hazard analyses on AI systems, and another was that we need the AI safety field to succeed in ways that the cybersecurity field has failed. Another general belief was that long term AI safety, such as concerns about the ability of humans to control AIs, was not a serious issue.

On average, the presentations were moderately technical. They were mostly focused on machine learning systems, although there was significant discussion of cybersecurity techniques.

The first talk was given by Eric Horvitz of Microsoft. He discussed some approaches for pushing into new directions in AI safety. Instead of merely trying to reduce the errors spotted according to one model, we should look out for "unknown unknowns" by stacking models and looking at problems which appear on any of them, a theme which would be presented by other researchers as well in later presentations. He discussed optimization under uncertain parameters, sensitivity analysis to uncertain parameters, and 'wireheading' or short-circuiting of reinforcement learning systems (which he believes can be guarded against by using 'reflective analysis'). Finally, he brought up the concerns about superintelligence, which sparked amused reactions in the audience. He said that scientists should address concerns about superintelligence, which he aptly described as the 'elephant in the room', noting that it was the reason that some people were at the conference. He said that scientists will have to engage with public concerns, while also noting that there were experts who were worried about superintelligence and that there would have to be engagement with the experts' concerns. He did not comment on whether he believed that these concerns were reasonable or not.

An issue which came up in the Q&A afterwards was that we need to deal with mis-structured utility functions in AI, because it is often the case that the specific tradeoffs and utilities which humans claim to value often lead to results which the humans don't like. So we need to have structural uncertainty about our utility models. The difficulty of finding good objective functions for AIs would eventually be discussed in many other presentations as well.

The next talk was given by Andrew Moore of Carnegie Mellon University, who claimed that his talk represented the consensus of computer scientists at the school. He claimed that the stakes of AI safety were very high - namely, that AI has the capability to save many people's lives in the near future, but if there are any accidents involving AI then public fears could lead to freezes in AI research and development. He highlighted the public's irrational tendencies wherein a single accident could cause people to overlook and ignore hundreds of invisible lives saved. He specifically mentioned a 12-24 month timeframe for these issues.

Moore said that verification of AI system safety will be difficult due to the combinatorial explosion of AI behaviors. He talked about meta-machine-learning as a solution to this, something which is being investigated under the direction of Lawrence Schuette at the Office of Naval Research. Moore also said that military AI systems require high verification standards and that development timelines for these systems are long. He talked about two different approaches to AI safety, stochastic testing and theorem proving - the process of doing the latter often leads to the discovery of unsafe edge cases.

He also discussed AI ethics, giving an example 'trolley problem' where AI cars would have to choose whether to hit a deer in order to provide a slightly higher probability of survival for the human driver. He said that we would need hash-defined constants to tell vehicle AIs how many deer a human is worth. He also said that we would need to find compromises in death-pleasantry tradeoffs, for instance where the safety of self-driving cars depends on the speed and routes on which they are driven. He compared the issue to civil engineering where engineers have to operate with an assumption about how much money they would spend to save a human life.

He concluded by saying that we need policymakers, company executives, scientists, and startups to all be involved in AI safety. He said that the research community stands to gain or lose together, and that there is a shared responsibility among researchers and developers to avoid triggering another AI winter through unsafe AI designs.

The next presentation was by Richard Mallah of the Future of Life Institute, who was there to represent "Medium Term AI Safety". He pointed out the explicit/implicit distinction between different modeling techniques in AI systems, as well as the explicit/implicit distinction between different AI actuation techniques. He talked about the difficulty of value specification and the concept of instrumental subgoals as an important issue in the case of complex AIs which are beyond human understanding. He said that even a slight misalignment of AI values with regard to human values along one parameter could lead to a strongly negative outcome, because machine learning parameters don't strictly correspond to the things that humans care about.

Mallah stated that open-world discovery leads to self-discovery, which can lead to reward hacking or a loss of control. He underscored the importance of causal accounting, which is distinguishing causation from correlation in AI systems. He said that we should extend machine learning verification to self-modification. Finally, he talked about introducing non-self-centered ontology to AI systems and bounding their behavior.

The audience was generally quiet and respectful during Richard's talk. I sensed that at least a few of them labelled him as part of the 'superintelligence out-group' and dismissed him accordingly, but I did not learn what most people's thoughts or reactions were. In the next panel featuring three speakers, he wasn't the recipient of any questions regarding his presentation or ideas.

Tom Mitchell from CMU gave the next talk. He talked about both making AI systems safer, and using AI to make other systems safer. He said that risks to humanity from other kinds of issues besides AI were the "big deals of 2016" and that we should make sure that the potential of AIs to solve these problems is realized. He wanted to focus on the detection and remediation of all failures in AI systems. He said that it is a novel issue that learning systems defy standard pre-testing ("as Richard mentioned") and also brought up the purposeful use of AI for dangerous things.

Some interesting points were raised in the panel. Andrew did not have a direct response to the implications of AI ethics being determined by the predominantly white people of the US/UK where most AIs are being developed. He said that ethics in AIs will have to be decided by society, regulators, manufacturers, and human rights organizations in conjunction. He also said that our cost functions for AIs will have to get more and more complicated as AIs get better, and he said that he wants to separate unintended failures from superintelligence type scenarios. On trolley problems in self driving cars and similar issues, he said "it's got to be complicated and messy."

Dario Amodei of Google Deepbrain, who co-authored the paper on concrete problems in AI safety, gave the next talk. He said that the public focus is too much on AGI/ASI and wants more focus on concrete/empirical approaches. He discussed the same problems that pose issues in advanced general AI, including flawed objective functions and reward hacking. He said that he sees long term concerns about AGI/ASI as "extreme versions of accident risk" and that he thinks it's too early to work directly on them, but he believes that if you want to deal with them then the best way to do it is to start with safety in current systems. Mostly he summarized the Google paper in his talk.

In her presentation, Claire Le Goues of CMU said "before we talk about Skynet we should focus on problems that we already have." She mostly talked about analogies between software bugs and AI safety, the similarities and differences between the two and what we can learn from software debugging to help with AI safety.

Robert Rahmer of IARPA discussed CAUSE, a cyberintelligence forecasting program which promises to help predict cyber attacks. It is a program which is still being put together.

In the panel of the above three, autonomous weapons were discussed, but no clear policy stances were presented.

John Launchbury gave a talk on DARPA research and the big picture of AI development. He pointed out that DARPA work leads to commercial applications and that progress in AI comes from sustained government investment. He classified AI capabilities into "describing," "predicting," and "explaining" in order of increasing difficulty, and he pointed out that old fashioned "describing" still plays a large role in AI verification. He said that "explaining" AIs would need transparent decisionmaking and probabilistic programming (the latter would also be discussed by others at the conference).

The next talk came from Jason Gaverick Matheny, the director of IARPA. Matheny talked about four requirements in current and future AI systems: verification, validation, security, and control. He wanted "auditability" in AI systems as a weaker form of explainability. He talked about the importance of "corner cases" for national intelligence purposes, the low probability, high stakes situations where we have limited data - these are situations where we have significant need for analysis but where the traditional machine learning approach doesn't work because of its overwhelming focus on data. Another aspect of national defense is that it has a slower decision tempo, longer timelines, and longer-viewing optics about future events.

He said that assessing local progress in machine learning development would be important for global security and that we therefore need benchmarks to measure progress in AIs. He ended with a concrete invitation for research proposals from anyone (educated or not), for both large scale research and for smaller studies ("seedlings") that could take us "from disbelief to doubt".

The difference in timescales between different groups was something I noticed later on, after hearing someone from the DoD describe their agency as having a longer timeframe than the Homeland Security Agency, and someone from the White House describe their work as being crisis reactionary.

The next presentation was from Andrew Grotto, senior director of cybersecurity policy at the National Security Council. He drew a close parallel from the issue of genetically modified crops in Europe in the 1990's to modern day artificial intelligence. He pointed out that Europe utterly failed to achieve widespread cultivation of GMO crops as a result of public backlash. He said that the widespread economic and health benefits of GMO crops were ignored by the public, who instead focused on a few health incidents which undermined trust in the government and crop producers. He had three key points: that risk frameworks matter, that you should never assume that the benefits of new technology will be widely perceived by the public, and that we're all in this together with regard to funding, research progress and public perception.

In the Q&A between Launchbury, Matheny, and Grotto after Grotto's presentation, it was mentioned that the economic interests of farmers worried about displacement also played a role in populist rejection of GMOs, and that a similar dynamic could play out with regard to automation causing structural unemployment. Grotto was also asked what to do about bad publicity which seeks to sink progress in order to avoid risks. He said that meetings like SafArtInt and open public dialogue were good.

One person asked what Launchbury wanted to do about AI arms races with multiple countries trying to "get there" and whether he thinks we should go "slow and secure" or "fast and risky" in AI development, a question which provoked laughter in the audience. He said we should go "fast and secure" and wasn't concerned. He said that secure designs for the Internet once existed, but the one which took off was the one which was open and flexible.

Another person asked how we could avoid discounting outliers in our models, referencing Matheny's point that we need to include corner cases. Matheny affirmed that data quality is a limiting factor to many of our machine learning capabilities. At IARPA, we generally try to include outliers until they are sure that they are erroneous, said Matheny.

Another presentation came from Tom Dietterich, president of the Association for the Advancement of Artificial Intelligence. He said that we have not focused enough on safety, reliability and robustness in AI and that this must change. Much like Eric Horvitz, he drew a distinction between robustness against errors within the scope of a model and robustness against unmodeled phenomena. On the latter issue, he talked about solutions such as expanding the scope of models, employing multiple parallel models, and doing creative searches for flaws - the latter doesn't enable verification that a system is safe, but it nevertheless helps discover many potential problems. He talked about knowledge-level redundancy as a method of avoiding misspecification - for instance, systems could identify objects by an "ownership facet" as well as by a "goal facet" to produce a combined concept with less likelihood of overlooking key features. He said that this would require wider experiences and more data.

There were many other speakers who brought up a similar set of issues: the user of cybersecurity techniques to verify machine learning systems, the failures of cybersecurity as a field, opportunities for probabilistic programming, and the need for better success in AI verification. Inverse reinforcement learning was extensively discussed as a way of assigning values. Jeanette Wing of Microsoft talked about the need for AIs to reason about the continuous and the discrete in parallel, as well as the need for them to reason about uncertainty (with potential meta levels all the way up). One point which was made by Sarah Loos of Google was that proving the safety of an AI system can be computationally very expensive, especially given the combinatorial explosion of AI behaviors.

In one of the panels, the idea of government actions to ensure AI safety was discussed. No one was willing to say that the government should regulate AI designs. Instead they stated that the government should be involved in softer ways, such as guiding and working with AI developers, and setting standards for certification.

Pictures: https://imgur.com/a/49eb7

In between these presentations I had time to speak to individuals and listen in on various conversations. A high ranking person from the Department of Defense stated that the real benefit of autonomous systems would be in terms of logistical systems rather than weaponized applications. A government AI contractor drew the connection between Mallah's presentation and the recent press revolving around superintelligence, and said he was glad that the government wasn't worried about it.

I talked to some insiders about the status of organizations such as MIRI, and found that the current crop of AI safety groups could use additional donations to become more established and expand their programs. There may be some issues with the organizations being sidelined; after all, the Google Deepbrain paper was essentially similar to a lot of work by MIRI, just expressed in somewhat different language, and was more widely received in mainstream AI circles.

In terms of careers, I found that there is significant opportunity for a wide range of people to contribute to improving government policy on this issue. Working at a group such as the Office of Science and Technology Policy does not necessarily require advanced technical education, as you can just as easily enter straight out of a liberal arts undergraduate program and build a successful career as long as you are technically literate. (At the same time, the level of skepticism about long term AI safety at the conference hinted to me that the signalling value of a PhD in computer science would be significant.) In addition, there are large government budgets in the seven or eight figure range available for qualifying research projects. I've come to believe that it would not be difficult to find or create AI research programs that are relevant to long term AI safety while also being practical and likely to be funded by skeptical policymakers and officials.

I also realized that there is a significant need for people who are interested in long term AI safety to have basic social and business skills. Since there is so much need for persuasion and compromise in government policy, there is a lot of value to be had in being communicative, engaging, approachable, appealing, socially savvy, and well-dressed. This is not to say that everyone involved in long term AI safety is missing those skills, of course.

I was surprised by the refusal of almost everyone at the conference to take long term AI safety seriously, as I had previously held the belief that it was more of a mixed debate given the existence of expert computer scientists who were involved in the issue. I sensed that the recent wave of popular press and public interest in dangerous AI has made researchers and policymakers substantially less likely to take the issue seriously. None of them seemed to be familiar with actual arguments or research on the control problem, so their opinions didn't significantly change my outlook on the technical issues. I strongly suspect that the majority of them had their first or possibly only exposure to the idea of the control problem after seeing badly written op-eds and news editorials featuring comments from the likes of Elon Musk and Stephen Hawking, which would naturally make them strongly predisposed to not take the issue seriously. In the run-up to the conference, websites and press releases didn't say anything about whether this conference would be about long or short term AI safety, and they didn't make any reference to the idea of superintelligence.

I sympathize with the concerns and strategy given by people such as Andrew Moore and Andrew Grotto, which make perfect sense if (and only if) you assume that worries about long term AI safety are completely unfounded. For the community that is interested in long term AI safety, I would recommend that we avoid competitive dynamics by (a) demonstrating that we are equally strong opponents of bad press, inaccurate news, and irrational public opinion which promotes generic uninformed fears over AI, (b) explaining that we are not interested in removing funding for AI research (even if you think that slowing down AI development is a good thing, restricting funding yields only limited benefits in terms of changing overall timelines, whereas those who are not concerned about long term AI safety would see a restriction of funding as a direct threat to their interests and projects, so it makes sense to cooperate here in exchange for other concessions), and (c) showing that we are scientifically literate and focused on the technical concerns. I do not believe that there is necessarily a need for the two "sides" on this to be competing against each other, so it was disappointing to see an implication of opposition at the conference.

Anyway, Ed Felten announced a request for information from the general public, seeking popular and scientific input on the government's policies and attitudes towards AI: https://www.whitehouse.gov/webform/rfi-preparing-future-artificial-intelligence

Overall, I learned quite a bit and benefited from the experience, and I hope the insight I've gained can be used to improve the attitudes and approaches of the long term AI safety community.

http://www.ted.com/talks/nick_bostrom_what_happens_when_our_computers_get_smarter_than_we_are

Artificial intelligence is getting smarter by leaps and bounds — within this century, research suggests, a computer AI could be as "smart" as a human being. And then, says Nick Bostrom, it will overtake us: "Machine intelligence is the last invention that humanity will ever need to make." A philosopher and technologist, Bostrom asks us to think hard about the world we're building right now, driven by thinking machines. Will our smart machines help to preserve humanity and our values — or will they have values of their own?

I realize this might go into a post in a media thread, rather than its own topic, but it seems big enough, and likely-to-prompt-discussion enough, to have its own thread.

I liked the talk, although it was less polished than TED talks often are. What was missing I think was any indication of how to solve the problem. He could be seen as just an ivory tower philosopher speculating on something that might be a problem one day, because apart from mentioning in the beginning that he works with mathematicians and IT guys, he really does not give an impression that this problem is already being actively worked on.

The argument that AIs provided with a reward channel will observe their controllers and learn to manipulate them is a valid one. Unfortunately, it's often framed in a way that feels counterintuitive or extreme, especially to AI designers. It typically starts with the standard reinforcement learning scenario, then posits that the AI becomes superintelligent and either manipulates the controller with super-social powers, or breaks out and gains control of its reward channel, killing or threatening its controllers.

And that is a fair argument. But conceptually, it leaps from a standard reinforcement learning scenario, to a science-fiction-sounding scenario. It might help to have intermediate scenarios: to show that even lower intelligence AIs might start exhibiting the same sort of behaviour, long before it gets to superintelligence.

So consider the following scenario. Some complex, trainable AI is tasked with writing automated news stories for a student newspaper. It trawls the web and composes its stories, then gets reward and feedback from the editors. Assume there are two editors for this newspaper, and they work on alternate days. The two editors have somewhat different ideas as to what constitutes a good story, so their feedbacks are different. After a while, the AI finds that it gets higher reward by using a certain style on Monday, Wednesday and Friday, and another style on Tuesday and Thursdays - this is a simple consequence of its reward mechanism.

But the rota isn't perfect. Sometime the Monday editor will edit a story so late on Monday that it's Tuesday, and sometimes the Tuesday editor will be up early and edit a story at the same time. This pattern is more common during the holidays, when the editors go to different time zones, often travel, and become more irregular in their editing.

So if the AI complex and skilled enough, then, simply through simple feedback, it will start building up a picture of its editors. It will figure out when they are likely to stick to a schedule, and when they will be more irregular. It will figure out the difference between holidays and non-holidays. Given time, it may be able to track the editors moods and it will certainly pick up on any major change in their lives - such as romantic relationships and breakups, which will radically change whether and how it should present stories with a romantic focus.

It will also likely learn the correlation between stories and feedbacks - maybe presenting a story define roughly as "positive" will increase subsequent reward for the rest of the day, on all stories. Or maybe this will only work on a certain editor, or only early in the term. Or only before lunch.

Thus the simple trainable AI with a particular focus - write automated news stories - will be trained, through feedback, to learn about its editors/controllers, to distinguish them, to get to know them, and, in effect, to manipulate them.

This may be a useful "bridging example" between standard RL agents and the superintelligent machines.

An Article on Motherboard reports about  Alien Minds by Susan Schneider who claiThe Dominant Life Form In the Cosmos Is Probably Superintelligent Robots. The article is crosslinked to other posts about superintelligence and at the end discusses the question why these alien robots leave us along. The arguments puts forth on this don't convince me though. 

A summary and review of Bostrom's Superintelligence is in the December issue of Reason magazine, and is now posted online at Reason.com.

This is part of a weekly reading group on Nick Bostrom's book, Superintelligence. For more information about the group, and an index of posts so far see the announcement post. For the schedule of future topics, see MIRI's reading guide.

Welcome. This week we discuss the 11th section in the reading guide: The treacherous turn. This corresponds to Chapter 8.

This post summarizes the section, and offers a few relevant notes, and ideas for further investigation. Some of my own thoughts and questions for discussion are in the comments.

There is no need to proceed in order through this post, or to look at everything. Feel free to jump straight to the discussion. Where applicable and I remember, page numbers indicate the rough part of the chapter that is most related (not necessarily that the chapter is being cited for the specific claim).

Reading: “Existential catastrophe…” and “The treacherous turn” from Chapter 8

Summary

1. The possibility of a first mover advantage + orthogonality thesis + convergent instrumental values suggests doom for humanity (p115-6)
   1. First mover advantage implies the AI is in a position to do what it wants
   2. Orthogonality thesis implies that what it wants could be all sorts of things
   3. Instrumental convergence thesis implies that regardless of its wants, it will try to acquire resources and eliminate threats
   4. Humans have resources and may be threats
   5. Therefore an AI in a position to do what it wants is likely to want to take our resources and eliminate us. i.e. doom for humanity.
2. One kind of response: why wouldn't the makers of the AI be extremely careful not to develop and release dangerous AIs, or relatedly, why wouldn't someone else shut the whole thing down? (p116)
3. It is hard to observe whether an AI is dangerous via its behavior at a time when you could turn it off, because AIs have convergent instrumental reasons to pretend to be safe, even if they are not. If they expect their minds to be surveilled, even observing their thoughts may not help. (p117)
4. The treacherous turn: while weak, an AI behaves cooperatively. When the AI is strong enough to be unstoppable it pursues its own values. (p119)
5. We might expect AIs to be more safe as they get smarter initially - when most of the risks come from crashing self-driving cars or mis-firing drones - then to get much less safe as they get too smart. (p117)
6. One can imagine a scenario where there is little social impetus for safety (p117-8): alarmists will have been wrong for a long time, smarter AI will have been safer for a long time, large industries will be invested, an exciting new technique will be hard to set aside, useless safety rituals will be available, and the AI will look cooperative enough in its sandbox.
7. The conception of deception: that moment when the AI realizes that it should conceal its thoughts (footnote 2, p282)

Another view

Danaher:

This is all superficially plausible. It is indeed conceivable that an intelligent system — capable of strategic planning — could take such treacherous turns. And a sufficiently time-indifferent AI could play a “long game” with us, i.e. it could conceal its true intentions and abilities for a very long time. Nevertheless, accepting this has some pretty profound epistemic costs. It seems to suggest that no amount of empirical evidence could ever rule out the possibility of a future AI taking a treacherous turn. In fact, its even worse than that. If we take it seriously, then it is possible that we have already created an existentially threatening AI. It’s just that it is concealing its true intentions and powers from us for the time being.

I don’t quite know what to make of this. Bostrom is a pretty rational, bayesian guy. I tend to think he would say that if all the evidence suggests that our AI is non-threatening (and if there is a lot of that evidence), then we should heavily discount the probability of a treacherous turn. But he doesn’t seem to add that qualification in the chapter. He seems to think the threat of an existential catastrophe from a superintelligent AI is pretty serious. So I’m not sure whether he embraces the epistemic costs I just mentioned or not.

Notes

1. Danaher also made a nice diagram of the case for doom, and relationship with the treacherous turn:

2. History

According to Luke Muehlhauser's timeline of AI risk ideas, the treacherous turn idea for AIs has been around at least 1977, when a fictional worm did it:

1977: Self-improving AI could stealthily take over the internet; convergent instrumental goals in AI; the treacherous turn. Though the concept of a self-propagating computer worm was introduced by John Brunner's The Shockwave Rider (1975), Thomas J. Ryan's novel The Adolescence of P-1 (1977) tells the story of an intelligent worm that at first is merely able to learn to hack novel computer systems and use them to propagate itself, but later (1) has novel insights on how to improve its own intelligence, (2) develops convergent instrumental subgoals (see Bostrom 2012) for self-preservation and resource acquisition, and (3) learns the ability to fake its own death so that it can grow its powers in secret and later engage in a "treacherous turn" (see Bostrom forthcoming) against humans.

3. The role of the premises

Bostrom's argument for doom has one premise that says AI could care about almost anything, then another that says regardless of what an AI cares about, it will do basically the same terrible things anyway. (p115) Do these sound a bit strange together to you? Why do we need the first, if final values don't tend to change instrumental goals anyway?

It seems the immediate reason is that an AI with values we like would not have the convergent goal of taking all our stuff and killing us. That is, the values we want an AI to have are some of those rare values that don't lead to destructive instrumental goals. Why is this? Because we (and thus the AI) care about the activites the resources would be grabbed from. If the resources were currently being used for anything we didn't care about, then our values would also suggest grabbing resources, and look similar to all of the other values. The difference that makes our values special here is just that most resources are already being used for them somewhat. 

4. Signaling

It is hard to tell apart a safe and an unsafe AI, because both would like to look safe. This is a very common problem in human interactions. For instance, it can be nontrivial to tell a genuine lover from a gold digger, a businessman from a conman, and an expert from a crank. All of them want to look like the desirable sort. Particularly similar to the AI case is that of hiring a new employee for a trial period. You will sometimes find that the employee's values are much better aligned during the trial period, and then they undergo a 'treacherous turn' once they have been hired more thoroughly.

'Costly signaling' is a general purpose solution to this problem, which works some of the time. The basic idea is this. Everyone has instrumental reasons to look like the good kind of person, but perhaps their reasons aren't exactly as strong as one other's, or the strength of their desire is harder to act on for one group than the other, so we can construct a set of options that will lead the different types of people to choose differently, even though they know this will set them apart. For instance, while an honest businessman and a conman would both like to say 'I'm an honest businessman', and to behave honestly if you watch them for five minutes, a conman would benefit less from expensive shopfront as an honest businessman, because his reputation is less valuable, so a brand is a signal of being honest.

Similarly, we could perhaps design situations where most harmful AIs would have a reasonable incentive to reveal themselves. This is difficult because so much is at stake, compared to the cost of the signal. But I haven't thought enough about it to rule it out, and I haven't seen anyone else do so either.

5. When would the 'conception of deception' take place?

6. Surveillance of the mind

Bostrom is not optimistic that AI researchers could protect much against risks by watching AI mental processes directly for signs of deception, rapid progress, or undesirable scheming. He says if the AI is smart it will realize this might happen, and hide the offending thoughts.

This seems an open question to me, for several reasons:

1. Making your mental contents look innocuous while maintaining their semantic content sounds potentially very hard
2. Especially for a creature which has only just become smart enough to realize it should treacherously turn
3. From the AI's perspective, even if it is smart, surveillance could seem fairly unlikely, especially if we deceive it about its surroundings

In-depth investigations

If you are particularly interested in these topics, and want to do further research, these are a few plausible directions, some inspired by Luke Muehlhauser's list, which contains many suggestions related to parts of Superintelligence. These projects could be attempted at various levels of depth.

1. How transparent are AI minds likely to be? Should we expect to be able to detect deception? What are the answers to these questions for different specific architectures and methods? This might be relevant.
2. Are there other good ways to filter AIs with certain desirable goals from others? e.g. by offering them choices that would filter them.

How to proceed

This has been a collection of notes on the chapter.  The most important part of the reading group though is discussion, which is in the comments section. I pose some questions for you there, and I invite you to add your own. Please remember that this group contains a variety of levels of expertise: if a line of discussion seems too basic or too incomprehensible, look around for one that suits you better!

Next week, we will talk about 'malignant failure modes' (as opposed presumably to worse failure modes). To prepare, read “Malignant failure modes” from Chapter 8. The discussion will go live at 6pm Pacific time next Monday December 1. Sign up to be notified here.

This is part of a weekly reading group on Nick Bostrom's book, Superintelligence. For more information about the group, see the announcement post. For the schedule of future topics, see MIRI's reading guide.

Welcome to the Superintelligence reading group. This week we discuss the first section in the reading guide, Past developments and present capabilities. This section considers the behavior of the economy over very long time scales, and the recent history of artificial intelligence (henceforth, 'AI'). These two areas are excellent background if you want to think about large economic transitions caused by AI.

This post summarizes the section, and offers a few relevant notes, thoughts, and ideas for further investigation. My own thoughts and questions for discussion are in the comments.

There is no need to proceed in order through this post. Feel free to jump straight to the discussion. Where applicable, page numbers indicate the rough part of the chapter that is most related (not necessarily that the chapter is being cited for the specific claim).

Reading: Foreword, and Growth modes through State of the art from Chapter 1 (p1-18)

Summary

Economic growth:

1. Economic growth has become radically faster over the course of human history. (p1-2)
2. This growth has been uneven rather than continuous, perhaps corresponding to the farming and industrial revolutions. (p1-2)
3. Thus history suggests large changes in the growth rate of the economy are plausible. (p2)
4. This makes it more plausible that human-level AI will arrive and produce unprecedented levels of economic productivity.
5. Predictions of much faster growth rates might also suggest the arrival of machine intelligence, because it is hard to imagine humans - slow as they are - sustaining such a rapidly growing economy. (p2-3)
6. Thus economic history suggests that rapid growth caused by AI is more plausible than you might otherwise think.

The history of AI:

1. Human-level AI has been predicted since the 1940s. (p3-4)
2. Early predictions were often optimistic about when human-level AI would come, but rarely considered whether it would pose a risk. (p4-5)
3. AI research has been through several cycles of relative popularity and unpopularity. (p5-11)
4. By around the 1990s, 'Good Old-Fashioned Artificial Intelligence' (GOFAI) techniques based on symbol manipulation gave way to new methods such as artificial neural networks and genetic algorithms. These are widely considered more promising, in part because they are less brittle and can learn from experience more usefully. Researchers have also lately developed a better understanding of the underlying mathematical relationships between various modern approaches. (p5-11)
5. AI is very good at playing board games. (12-13)
6. AI is used in many applications today (e.g. hearing aids, route-finders, recommender systems, medical decision support systems, machine translation, face recognition, scheduling, the financial market). (p14-16)
7. In general, tasks we thought were intellectually demanding (e.g. board games) have turned out to be easy to do with AI, while tasks which seem easy to us (e.g. identifying objects) have turned out to be hard. (p14)
8. An 'optimality notion' is the combination of a rule for learning, and a rule for making decisions. Bostrom describes one of these: a kind of ideal Bayesian agent. This is impossible to actually make, but provides a useful measure for judging imperfect agents against. (p10-11)

Notes on a few things

1. What is 'superintelligence'? (p22 spoiler)
   In case you are too curious about what the topic of this book is to wait until week 3, a 'superintelligence' will soon be described as 'any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest'. Vagueness in this definition will be cleared up later. 
2. What is 'AI'?
   In particular, how does 'AI' differ from other computer software? The line is blurry, but basically AI research seeks to replicate the useful 'cognitive' functions of human brains ('cognitive' is perhaps unclear, but for instance it doesn't have to be squishy or prevent your head from imploding). Sometimes AI research tries to copy the methods used by human brains. Other times it tries to carry out the same broad functions as a human brain, perhaps better than a human brain. Russell and Norvig (p2) divide prevailing definitions of AI into four categories: 'thinking humanly', 'thinking rationally', 'acting humanly' and 'acting rationally'. For our purposes however, the distinction is probably not too important.
3. What is 'human-level' AI? 
   We are going to talk about 'human-level' AI a lot, so it would be good to be clear on what that is. Unfortunately the term is used in various ways, and often ambiguously. So we probably can't be that clear on it, but let us at least be clear on how the term is unclear. 
   
   One big ambiguity is whether you are talking about a machine that can carry out tasks as well as a human at any price, or a machine that can carry out tasks as well as a human at the price of a human. These are quite different, especially in their immediate social implications.
   
   Other ambiguities arise in how 'levels' are measured. If AI systems were to replace almost all humans in the economy, but only because they are so much cheaper - though they often do a lower quality job - are they human level? What exactly does the AI need to be human-level at? Anything you can be paid for? Anything a human is good for? Just mental tasks? Even mental tasks like daydreaming? Which or how many humans does the AI need to be the same level as? Note that in a sense most humans have been replaced in their jobs before (almost everyone used to work in farming), so if you use that metric for human-level AI, it was reached long ago, and perhaps farm machinery is human-level AI. This is probably not what we want to point at.
   
   Another thing to be aware of is the diversity of mental skills. If by 'human-level' we mean a machine that is at least as good as a human at each of these skills, then in practice the first 'human-level' machine will be much better than a human on many of those skills. It may not seem 'human-level' so much as 'very super-human'.
   
   We could instead think of human-level as closer to 'competitive with a human' - where the machine has some super-human talents and lacks some skills humans have. This is not usually used, I think because it is hard to define in a meaningful way. There are already machines for which a company is willing to pay more than a human: in this sense a microscope might be 'super-human'. There is no reason for a machine which is equal in value to a human to have the traits we are interested in talking about here, such as agency, superior cognitive abilities or the tendency to drive humans out of work and shape the future. Thus we talk about AI which is at least as good as a human, but you should beware that the predictions made about such an entity may apply before the entity is technically 'human-level'.
   
   
   Example of how the first 'human-level' AI may surpass humans in many ways.
   
   Because of these ambiguities, AI researchers are sometimes hesitant to use the term. e.g. in these interviews.
4. Growth modes (p1) 
   Robin Hanson wrote the seminal paper on this issue. Here's a figure from it, showing the step changes in growth rates. Note that both axes are logarithmic. Note also that the changes between modes don't happen overnight. According to Robin's model, we are still transitioning into the industrial era (p10 in his paper).
5. What causes these transitions between growth modes? (p1-2)
   One might be happier making predictions about future growth mode changes if one had a unifying explanation for the previous changes. As far as I know, we have no good idea of what was so special about those two periods. There are many suggested causes of the industrial revolution, but nothing uncontroversially stands out as 'twice in history' level of special. You might think the small number of datapoints would make this puzzle too hard. Remember however that there are quite a lot of negative datapoints - you need an explanation that didn't happen at all of the other times in history. 
6. Growth of growth
   It is also interesting to compare world economic growth to the total size of the world economy. For the last few thousand years, the economy seems to have grown faster more or less in proportion to it's size (see figure below). Extrapolating such a trend would lead to an infinite economy in finite time. In fact for the thousand years until 1950 such extrapolation would place an infinite economy in the late 20th Century! The time since 1950 has been strange apparently. 
   
   (Figure from here)
7. Early AI programs mentioned in the book (p5-6)
   You can see them in action: SHRDLU, Shakey, General Problem Solver (not quite in action), ELIZA.
8. Later AI programs mentioned in the book (p6)
   Algorithmically generated Beethoven, algorithmic generation of patentable inventions, artificial comedy (requires download).
9. Modern AI algorithms mentioned (p7-8, 14-15) 
   Here is a neural network doing image recognition. Here is artificial evolution of jumping and of toy cars. Here is a face detection demo that can tell you your attractiveness (apparently not reliably), happiness, age, gender, and which celebrity it mistakes you for.
10. What is maximum likelihood estimation? (p9)
    Bostrom points out that many types of artificial neural network can be viewed as classifiers that perform 'maximum likelihood estimation'. If you haven't come across this term before, the idea is to find the situation that would make your observations most probable. For instance, suppose a person writes to you and tells you that you have won a car. The situation that would have made this scenario most probable is the one where you have won a car, since in that case you are almost guaranteed to be told about it. Note that this doesn't imply that you should think you won a car, if someone tells you that. Being the target of a spam email might only give you a low probability of being told that you have won a car (a spam email may instead advise you of products, or tell you that you have won a boat), but spam emails are so much more common than actually winning cars that most of the time if you get such an email, you will not have won a car. If you would like a better intuition for maximum likelihood estimation, Wolfram Alpha has several demonstrations (requires free download).
11. What are hill climbing algorithms like? (p9)
    The second large class of algorithms Bostrom mentions are hill climbing algorithms. The idea here is fairly straightforward, but if you would like a better basic intuition for what hill climbing looks like, Wolfram Alpha has a demonstration to play with (requires free download).

In-depth investigations

If you are particularly interested in these topics, and want to do further research, these are a few plausible directions:

1. How have investments into AI changed over time? Here's a start, estimating the size of the field.
2. What does progress in AI look like in more detail? What can we infer from it? I wrote about algorithmic improvement curves before. If you are interested in plausible next steps here, ask me.
3. What do economic models tell us about the consequences of human-level AI? Here is some such thinking; Eliezer Yudkowsky has written at length about his request for more.

How to proceed

This has been a collection of notes on the chapter. The most important part of the reading group though is discussion, which is in the comments section. I pose some questions for you there, and I invite you to add your own. Please remember that this group contains a variety of levels of expertise: if a line of discussion seems too basic or too incomprehensible, look around for one that suits you better!

Next week, we will talk about what AI researchers think about human-level AI: when it will arrive, what it will be like, and what the consequences will be. To prepare, read Opinions about the future of machine intelligence from Chapter 1 and also When Will AI Be Created? by Luke Muehlhauser. The discussion will go live at 6pm Pacific time next Monday 22 September. Sign up to be notified here.

I attended Nick Bostrom's talk at UC Berkeley last Friday and got intrigued by these problems again. I wanted to pitch an idea here, with the question: Have any of you seen work along these lines before? Can you recommend any papers or posts? Are you interested in collaborating on this angle in further depth?

The problem I'm thinking about (surely naively, relative to y'all) is: What would you want to program an omnipotent machine to optimize?

For the sake of avoiding some baggage, I'm not going to assume this machine is "superintelligent" or an AGI. Rather, I'm going to call it a supercontroller, just something omnipotently effective at optimizing some function of what it perceives in its environment.

As has been noted in other arguments, a supercontroller that optimizes the number of paperclips in the universe would be a disaster. Maybe any supercontroller that was insensitive to human values would be a disaster. What constitutes a disaster? An end of human history. If we're all killed and our memories wiped out to make more efficient paperclip-making machines, then it's as if we never existed. That is existential risk.

The challenge is: how can one formulate an abstract objective function that would preserve human history and its evolving continuity?

I'd like to propose an answer that depends on the notion of logical depth as proposed by C.H. Bennett and outlined in section 7.7 of Li and Vitanyi's An Introduction to Kolmogorov Complexity and Its Applications which I'm sure many of you have handy. Logical depth is a super fascinating complexity measure that Li and Vitanyi summarize thusly:

Logical depth is the necessary number of steps in the deductive or causal path connecting an object with its plausible origin. Formally, it is the time required by a universal computer to compute the object from its compressed original description.

The mathematics is fascinating and better read in the original Bennett paper than here. Suffice it presently to summarize some of its interesting properties, for the sake of intuition.

• "Plausible origins" here are incompressible, i.e. algorithmically random.
• As a first pass, the depth D(x) of a string x is the least amount of time it takes to output the string from an incompressible program.
• There's a free parameter that has to do with precision that I won't get into here. 
• Both a string of length n that is comprised entirely of 1's, and a string of length n of independent random bits are both shallow. The first is shallow because it can be produced by a constant-sized program in time n. The second is shallow because there exists an incompressible program that is the output string plus a constant sized print function that produces the output in time n.
• An example of a deeper string is the string of length n that for each digit i encodes the answer to the ith enumerated satisfiability problem. Very deep strings can involve diagonalization.
• Like Kolmogorov complexity, there is an absolute and a relative version. Let D(x/w) be the least time it takes to output x from a program that is incompressible relative to w,

From Toby Ord:

Tool assisted speedruns (TAS) are when people take a game and play it frame by frame, effectively providing super reflexes and forethought, where they can spend a day deciding what to do in the next 1/60th of a second if they wish. There are some very extreme examples of this, showing what can be done if you really play a game perfectly. For example, this video shows how to winSuper Mario Bros 3 in 11 minutes. It shows how different optimal play can be from normal play. In particular, on level 8-1, it gains 90 extra lives by a sequence of amazing jumps.

Other TAS runs get more involved and start exploiting subtle glitches in the game. For example, this page talks about speed running NetHack, using a lot of normal tricks, as well as luck manipulation (exploiting the RNG) and exploiting a dangling pointer bug to rewrite parts of memory.

Though there are limits to what AIs could do with sheer speed, it's interesting that great performance can be achieved with speed alone, that this allows different strategies from usual ones, and that it allows the exploitation of otherwise unexploitable glitches and bugs in the setup.

A slightly edited version of "AI risk - executive summary" has been published in "The Conversation", titled "Your essential guide to the rise of the intelligent machines":

The risks posed to human beings by artificial intelligence in no way resemble the popular image of the Terminator. That fictional mechanical monster is distinguished by many features – strength, armour, implacability, indestructability – but Arnie’s character lacks the one characteristic that we in the real world actually need to worry about – extreme intelligence.

Thanks again for those who helped forge the original article. You can use this link, or the Less Wrong one, depending on the audience.

Although it regularly discusses the possibility of superintelligences with the power to transform the universe in the service of some value system - whether that value system is paperclip maximization or some elusive extrapolation of human values - it seems that Less Wrong has never systematically discussed the possibility that we are already within the domain of some superintelligence, and what that would imply. So how about it? What are the possibilities, what are the probabilities, and how should they affect our choices?

Here.

Long story short, it's an attempt to justify the planetarium hypothesis as a solution to the Fermi paradox. The first half is a discussion of how it and things like it are relevant to the intended purview of the blog, and the second half is the meat of the post. You'll probably want to just eat the meat, which I think is relevant to the interests of many LessWrong folk.

The blog is Computational Theology. It's new. I'll be the primary poster, but others are sought. I'll likely introduce the blog and more completely describe it in its own discussion post when more posts are up, hopefully including a few from people besides me, and when the archive will give a more informative indication of what to expect from the blog. Despite theism's suspect reputation here at LessWrong I suspect many of the future posts will be of interest to this audience anyway, especially for those of you who take interest in discussion of the singularity. The blog will even occasionally touch on rationality proper. So you might want to store the fact of the blog's existence somewhere deep in the back of your head. A link to the blog's main page can be found on my LessWrong user page if you forget the url.

I'd appreciate it if comments about the substance of the post were made on the blog post itself, but if you want to discuss the content here on LessWrong then that's okay too. Any meta-level comments about presentation, typos, or the post's relevance to LessWrong, should probably be put as comments on this discussion post. Thanks all!

I did a talk at the 25th Oxford Geek night, in which I had five minutes to present the dangers of AI. The talk is now online. Though it doesn't contain anything people at Less Wrong would find new, I feel it does a reasonable job at pitching some of the arguments in a very brief format.

Just a minor thought connected with the orthogonality thesis: if you claim that any superintelligence will inevitably converge to some true code of morality, then you are also claiming that no measures can be taken by its creators to prevent this convergence. In other words, the superintelligence will be uncontrollable.

Philosopher John Danaher has written an explication and critique of Bostrom's "orthogonality thesis" from "The Superintelligent Will." To quote the conclusion:

 

Summing up, in this post I’ve considered Bostrom’s discussion of the orthogonality thesis. According to this thesis, any level of intelligence is, within certain weak constraints, compatible with any type of final goal. If true, the thesis might provide support for those who think it possible to create a benign superintelligence. But, as I have pointed out, Bostrom’s defence of the orthogonality thesis is lacking in certain respects, particularly in his somewhat opaque and cavalier dismissal of normatively thick theories of rationality.

As it happens, none of this may affect what Bostrom has to say about unfriendly superintelligences. His defence of that argument relies on the convergence thesis, not the orthogonality thesis. If the orthogonality thesis turns out to be false, then all that happens is that the kind of convergence Bostrom alludes to simply occurs at a higher level in the AI’s goal architecture. 

What might, however, be significant is whether the higher-level convergence is a convergence towards certain moral beliefs or a convergence toward nihilistic beliefs. If it is the former, then friendliness might be necessitated, not simply possible. If it is the latter, then all bets are off. A nihilistic agent could do pretty anything since, no goals would be rationally entailed.

 

Underground Q&A session with Nick Bostrom (http://www.nickbostrom.com) on existential risks and artificial intelligence with the Oxford Transhumanists (recorded 10 October 2011).

http://www.youtube.com/watch?v=KQeijCRJSog

Below is part 2 of the first draft of my book Ethics and Superintelligence. Your comments and constructive criticisms are much appreciated.

This is not a book for a mainstream audience. Its style is that of contemporary Anglophone philosophy. Compare to, for example, Chalmers' survey article on the singularity.

Bibliographic references and links to earlier parts are provided here.

Part 2 is below...

***

Late in the Industrial Revolution, Samuel Butler (1863) worried about what might happen when machines become more capable than the humans who designed them:

…we are ourselves creating our own successors; we are daily adding to the beauty and delicacy of their physical organisation; we are daily giving them greater power and supplying by all sorts of ingenious contrivances that self-regulating, self-acting power which will be to them what intellect has been to the human race. In the course of ages we shall find ourselves the inferior race.

…the time will come when the machines will hold the real supremacy over the world and its inhabitants…

By the time of the computer, Alan Turing (1950) realized that machines will one day be capable of genuine thought:

I believe that at the end of the century…  one will be able to speak of machines thinking without expecting to be contradicted.

Turing (1951/2004) concluded:

…it seems probable that once the machine thinking method has 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…

All-powerful machines are a staple of science fiction, but one of the first serious arguments that such a scenario is likely came from the statistician I.J. Good (1965):

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.

Vernor Vinge (1993) called this future event the “technological singularity.” Though there are several uses of the term “singularity” in futurist circles (Yudkowky 2007), I will always use the term to refer to Good’s predicted intelligence explosion.

David Chalmers (2010) introduced another terminological convention that I will borrow:

Let us say that AI is artificial intelligence of human level or greater (that is, at least as intelligent as an average human). Let us say that AI+ is artificial intelligence of greater than human level (that is, more intelligent than the most intelligent human). Let us say that AI++ (or superintelligence) is AI of far greater than human level (say, at least as far beyond the most intelligent human as the most intelligent human is beyond a mouse).

With this in place, Chalmers formalized Good’s argument like so:

1.     There will be AI (before long, absent defeaters).

2.     If there is AI, there will be AI+ (soon after, absent defeaters).

3.     If there is AI+, there will be AI++ (soon after, absent defeaters).

4.     Therefore, there will be AI++ (before too long, absent defeaters).

I will defend Chalmers’ argument in greater detail than he has, using “before long” to mean “within 150 years,” using “soon after” to mean “within two decades,” and using “before too long” to mean “within two centuries.” My definitions here are similar to Chalmers’ definitions, but more precise.

Following Chalmers, by “defeaters” I mean “anything that prevents intelligent systems (human or artificial) from manifesting their capacities to create intelligent systems.” Defeaters include “disasters, disinclination, and active prevention.”

Disasters include catastrophic events that would severely impede scientific progress, such as supervolcano eruption, asteroid impact, cosmic rays, climate change, pandemic, nuclear war, biological warfare, an explosion of nanotechnology, and so on. The risk of such disasters and others are assessed in Bostrom & Cirkovic (2008).

Disinclination refers to a lack of interest in developing AI of human-level general intelligence. Given the enormous curiosity of the human species, and the power that human-level AI could bring its creators, I think long-term disinclination is unlikely.

Active prevention of the development of human-level artificial intelligence has already been advocated by Thomas Metzinger (2004), though not because of the risk to humans. Rather, Metzinger is concerned about the risk to artificial agents. Early AIs will inevitably be poorly designed, which could lead to enormous subjective suffering for them that we cannot predict. One might imagine an infant from near Cherynobl whose parts are so malformed by exposure to nuclear radiation during development that its short existence is a living hell. In working toward human-level artificial intelligence, might we be developing millions of internally malformed beings that suffer horrible subjective experiences but are unable to tell us so?

It is difficult to predict the likelihood of the active prevention of AI development, but the failure of humanity to halt the development of ever more powerful nuclear weapons (Norris & Kristensen 2009) – even after tasting their destructive power – does not inspire optimism.

Later, we will return to consider these potential defeaters again. For now, let us consider the premises of Chalmers’ argument.

***

As previously announced, I plan to post the first draft of the book, Ethics and Superintelligence, in tiny parts, to the Less Wrong discussion area. Your comments and constructive criticisms are much appreciated.

This is not a book for a mainstream audience. Its style is that of contemporary Anglophone philosophy. Compare to, for example, Chalmers' survey article on the singularity.

Bibliographic references are provided here.

This "part 1" section is probably the only part of which I will post revision to Less Wrong. Revisions of further parts of the book will probably not appear publicly until the book is published.

Revised part 1 below....

1. The technological singularity is coming soon.

Every year, computers surpass human abilities in new ways. A program written in 1956 was able to prove mathematical theorems, and found a more elegant proof for one of them than Russell and Whitehead had given in Principia Mathematica (MacKenzie 1995). By the late 1990s, “expert systems” had surpassed human ability in a wide range of tasks.[i] In 1997, IBM’s Deep Blue defeated the reigning World Chess Champion Garry Kasparov (Campbell et al. 2002). In 2011, IBM’s Watson beat the best human players at a much more complicated game: Jeopardy! (Someone, 2011). Recently, a robot scientist was programmed with our scientific knowledge about yeast, then posed its own hypotheses, tested them, and assessed the results. It answered a question about yeast that had baffled human scientists for 150 years (King 2011).

Many experts think that human-level general intelligence may be created within this century.[ii] This raises an important question. What will happen when an artificial intelligence (AI) surpasses human ability at designing artificial intelligences?

I.J. Good (1965) speculated that such an AI would be able to improve its own intelligence, leading to a positive feedback loop of improving intelligence – an “intelligence explosion.” Such a machine would rapidly become intelligent enough to take control of the internet, use robots to build itself new hardware, do science on a massive scale, invent new computing technology and energy sources, or achieve similar dominating goals. As such, it could be humanity’s last invention (Bostrom 2003).

Humans would be powerless to stop such a “superintelligence” (Bostrom 1998) from accomplishing its goals. Thus, if such a scenario is at all plausible, then it is critically important to program the goal system of this superintelligence such that it does not cause human extinction when it comes to power.

Success in that project could mean the difference between a utopian solar system of unprecedented harmony and happiness, and a solar system in which all available matter (including human flesh) has been converted into parts for a planet-sized computer built to solve difficult mathematical problems.[iii]

The technical challenges of designing the goal system of such a superintelligence are daunting.[iv] But even if we can solve those problems, the question of which goal system to give the superintelligence remains. It is at least partly a question of philosophy – a question of ethics.

***

In this chapter I argue that a single, powerful superintelligence - one variety of what Bostrom (2006) calls a “singleton" - is likely to arrive within the next 200 years unless a worldwide catastrophe drastically impedes scientific progress.

The singleton will produce very different future worlds depending on which normative theory is used to design its goal system. In chapter two, I survey many popular normative theories, and conclude that none of them offer an attractive basis for designing the motivational system of a machine superintelligence.

Chapter three reformulates and strengthens what is perhaps the most developed plan for the design of the singleton’s goal system ­– Eliezer Yudkowsky’s (2004) “Coherent Extrapolated Volition.” Chapter four considers some outstanding worries about this plan.

In chapter five I argue that we cannot decide how to design the singleton’s goal system without considering meta-ethics, because normative theory depends on meta-ethics. The next chapter argues that we should invest little effort in meta-ethical theories that do not fit well with our emerging reductionist picture of the world, just as we quickly abandon scientific theories that don’t fit the available scientific data. I also identify several meta-ethical positions that I think are good candidates for abandonment.

But the looming problem of the technological singularity requires us to have a positive theory, too. Chapter seven proposes some meta-ethical claims about which I think naturalists should come to agree. In the final chapter, I consider the implications of these meta-ethical claims for the design of the singleton’s motivational system.

***