Artificial Intelligence

Made-Up Minds

Posted by Peter on June 20, 2016
Posted in: Artificial Intelligence, Book Reviews, The Human Mind. Tagged: .

Book review: Made-Up Minds: A Constructivist Approach to Artificial Intelligence, by Gary L. Drescher.

It’s odd to call a book boring when it uses the pun “ontology recapitulates phylogeny”[1]. to describe a surprising feature of its model. About 80% of the book is dull enough that I barely forced myself to read it, yet the occasional good idea persuaded me not to give up.

Drescher gives a detailed model of how Piaget-style learning in infants could enable them to learn complex concepts starting with minimal innate knowledge.
Continue Reading

AGI Timelines

Posted by Peter on June 8, 2016
Posted in: Artificial Intelligence, Economics. Tagged: , , , .

One of most important assumptions in The Age of Ems is that non-em AGI will take a long time to develop.

1.

Scott Alexander at SlateStarCodex complains that Robin rejects survey data that uses validated techniques, and instead uses informal surveys whose results better fit Robin’s biases [1]. Robin clearly explains one reason why he does that: to get the outside view of experts.

Whose approach to avoiding bias is better?

  • Minimizing sampling error and carefully documenting one’s sampling technique are two of the most widely used criteria to distinguish science from wishful thinking.
  • Errors due to ignoring the outside view have been documented to be large, yet forecasters are reluctant to use the outside view.

So I rechecked advice from forecasting experts such as Philip Tetlock and Nate Silver, and the clear answer I got was … that was the wrong question.

Tetlock and Silver mostly focus on attitudes that are better captured by the advice to be a fox, not a hedgehog.

The strongest predictor of rising into the ranks of superforecasters is perpetual beta, the degree to which one is committed to belief updating and self-improvement.

Tetlock’s commandment number 3 says “Strike the right balance between inside and outside views”. Neither Tetlock or Silver offer hope that either more rigorous sampling of experts or dogmatically choosing the outside view over the inside view help us win a forecasting contest.

So instead of asking who is right, we should be glad to have two approaches to ponder, and should want more. (Robin only uses one approach for quantifying the time to non-em AGI, but is more fox-like when giving qualitative arguments against fast AGI progress).

2.

What Robin downplays is that there’s no consensus of the experts on whom he relies, not even about whether progress is steady, accelerating, or decelerating.

Robin uses the median expert estimate of progress in various AI subfields. This makes sense if AI progress depends on success in many subfields. It makes less sense if success in one subfield can make the other subfields obsolete. If “subfield” means a guess about what strategy best leads to intelligence, then I expect the median subfield to be rendered obsolete by a small number of good subfields [2]. If “subfield” refers to a subset of tasks that AI needs to solve (e.g. vision, or natural language processing), then it seems reasonable to look at the median (and I can imagine that slower subfields matter more). Robin appears to use both meanings of “subfield”, with fairly similar results for each, so it’s somewhat plausible that the median is informative.

3.

Scott also complains that Robin downplays the importance of research spending while citing only a paper dealing with government funding of agricultural research. But Robin also cites another paper (Ulku 2004), which covers total R&D expenditures in 30 countries (versus 16 countries in the paper that Scott cites) [3].

4.

Robin claims that AI progress will slow (relative to economic growth) due to slowing hardware progress and reduced dependence on innovation. Even if I accept Robin’s claims about these factors, I have trouble believing that AI progress will slow.

I expect higher em IQ will be one factor that speeds up AI progress. Garrett Jones suggests that a 40 IQ point increase in intelligence causes a 50% increase in a country’s productivity. I presume that AI researcher productivity is more sensitive to IQ than is, say, truck driver productivity. So it seems fairly plausible to imagine that increased em IQ will cause more than a factor of two increase in the rate of AI progress. (Robin downplays the effects of IQ in contexts where a factor of two wouldn’t much affect his analysis; he appears to ignore them in this context).

I expect that other advantages of ems will contribute additional speedups – maybe ems who work on AI will run relatively fast, maybe good training/testing data will be relatively cheap to create, or maybe knowledge from experimenting on ems will better guide AI research.

5.

Robin’s arguments against an intelligence explosion are weaker than they appear. I mostly agree with those arguments, but I want to discourage people from having strong confidence in them.

The most suspicious of those arguments is that gains in software algorithmic efficiency “remain surprisingly close to the rate at which hardware costs have fallen. This suggests that algorithmic gains have been enabled by hardware gains”. He cites only (Grace 2013) in support of this. That paper doesn’t comment on whether hardware changes enable software changes. The evidence seems equally consistent with that or with the hypothesis that both are independently caused by some underlying factor. I’d say there’s less than a 50% chance that Robin is correct about this claim.

Robin lists 14 other reasons for doubting there will be an intelligence explosion: two claims about AI history (no citations), eight claims about human intelligence (one citation), and four about what causes progress in research (with the two citations mentioned earlier). Most of those 14 claims are probably true, but it’s tricky to evaluate their relevance.

Conclusion

I’d say there’s maybe a 15% chance that Robin is basically right about the timing of non-em AI given his assumptions about ems. His book is still pretty valuable if an em-dominated world lasts for even one subjective decade before something stranger happens. And “something stranger happens” doesn’t necessarily mean his analysis becomes obsolete.

Footnotes

[1] – I can’t find any SlateStarCodex complaint about Bostrom doing something in Superintelligence that’s similar to what Scott accuses Robin of, when Bostrom’s survey of experts shows an expected time of decades for human-level AI to become superintelligent. Bostrom wants to focus on a much faster takeoff scenario, and disagrees with the experts, without identifying reasons for thinking his approach reduces biases.

[2] – One example is that genetic algorithms are looking fairly obsolete compared to neural nets, now that they’re being compared on bigger problems than when genetic algorithms were trendy.

Robin wants to avoid biases from recent AI fads by looking at subfields as they were defined 20 years ago. Some recent changes in AI are fads, but some are increased wisdom. I expect many subfields to be dead ends, given how immature AI was 20 years ago (and may still be today).

[3] – Scott quotes from one of three places that Robin mentions this subject (an example of redundancy that is quite rare in the book), and that’s the one place out of three where Robin neglects to cite (Ulku 2004). Age of Em is the kind of book where it’s easy to overlook something important like that if you don’t read it more carefully than you’d read a normal book.

I tried comparing (Ulku 2004) to the OECD paper that Scott cites, and failed to figure out whether they disagree. The OECD paper is probably consistent with Robin’s “less than proportionate increases” claim that Scott quotes. But Scott’s doubts are partly about Robin’s bolder prediction that AI progress will slow down, and academic papers don’t help much in evaluating that prediction.

If you’re tempted to evaluate how well the Ulku paper supports Robin’s views, beware that this quote is one of its easier to understand parts:

In addition, while our analysis lends support for endogenous growth theories in that it confirms a significant relationship between R&D stock and innovation, and between innovation and per capita GDP, it lacks the evidence for constant returns to innovation in terms of R&D stock. This implies that R&D models are not able to explain sustainable economic growth, i.e. they are not fully endogenous.

The Age of Em

Posted by Peter on May 24, 2016
Posted in: Artificial Intelligence, Book Reviews, Economics, Miscellaneous, Virtual Worlds. Tagged: , , , , .

Book review: The Age of Em: Work, Love and Life when Robots Rule the Earth, by Robin Hanson.

This book analyzes a possible future era when software emulations of humans (ems) dominate the world economy. It is too conservative to tackle longer-term prospects for eras when more unusual intelligent beings may dominate the world.

Hanson repeatedly tackles questions that scare away mainstream academics, and gives relatively ordinary answers (guided as much as possible by relatively standard, but often obscure, parts of the academic literature).

Assumptions

Hanson’s scenario relies on a few moderately controversial assumptions. The assumptions which I find most uncertain are related to human-level intelligence being hard to understand (because it requires complex systems), enough so that ems will experience many subjective centuries before artificial intelligence is built from scratch. For similar reasons, ems are opaque enough that it will be quite a while before they can be re-engineered to be dramatically different.

Hanson is willing to allow that ems can be tweaked somewhat quickly to produce moderate enhancements (at most doubling IQ) before reaching diminishing returns. He gives somewhat plausible reasons for believing this will only have small effects on his analysis. But few skeptics will be convinced.

Some will focus on potential trillions of dollars worth of benefits that higher IQs might produce, but that wealth would not much change Hanson’s analysis.

Others will prefer an inside view analysis which focuses on the chance that higher IQs will better enable us to handle risks of superintelligent software. Hanson’s analysis implies we should treat that as an unlikely scenario, but doesn’t say what we should do about modest probabilities of huge risks.

Another way that Hanson’s assumptions could be partly wrong is if tweaking the intelligence of emulated Bonobos produces super-human entities. That seems to only require small changes to his assumptions about how tweakable human-like brains are. But such a scenario is likely harder to analyze than Hanson’s scenario, and it probably makes more sense to understand Hanson’s scenario first.

Wealth

Wages in this scenario are somewhat close to subsistence levels. Ems have some ability to restrain wage competition, but less than they want. Does that mean wages are 50% above subsistence levels, or 1%? Hanson hints at the former. The difference feels important to me. I’m concerned that sound-bite versions of book will obscure the difference.

Hanson claims that “wealth per em will fall greatly”. It would be possible to construct a measure by which ems are less wealthy than humans are today. But I expect it will be at least as plausible to use a measure under which ems are rich compared to humans of today, but have high living expenses. I don’t believe there’s any objective unit of value that will falsify one of those perspectives [1].

Style / Organization

The style is more like a reference book than a story or an attempt to persuade us of one big conclusion. Most chapters (except for a few at the start and end) can be read in any order. If the section on physics causes you to doubt whether the book matters, skip to chapter 12 (labor), and return to the physics section later.

The style is very concise. Hanson rarely repeats a point, so understanding him requires more careful attention than with most authors.

It’s odd that the future of democracy gets less than twice as much space as the future of swearing. I’d have preferred that Hanson cut out a few of his less important predictions, to make room for occasional restatements of important ideas.

Many little-known results that are mentioned in the book are relevant to the present, such as: how the pitch of our voice affects how people perceive us, how vacations affect productivity, and how bacteria can affect fluid viscosity.

I was often tempted to say that Hanson sounds overconfident, but he is clearly better than most authors at admitting appropriate degrees of uncertainty. If he devoted much more space to caveats, I’d probably get annoyed at the repetition. So it’s hard to say whether he could have done any better.

Conclusion

Even if we should expect a much less than 50% chance of Hanson’s scenario becoming real, it seems quite valuable to think about how comfortable we should be with it and how we could improve on it.

Footnote

[1] – The difference matters only in one paragraph, where Hanson discusses whether ems deserve charity more than do humans living today. Hanson sounds like he’s claiming ems deserve our charity because they’re poor. Most ems in this scenario are comfortable enough for this to seem wrong.

Hanson might also be hinting that our charity would be effective at increasing the number of happy ems, and that basic utilitarianism says that’s preferable to what we can do by donating to today’s poor. That argument deserves more respect and more detailed analysis.

Probably Approximately Correct

Posted by Peter on May 24, 2016
Posted in: Artificial Intelligence, Book Reviews. Tagged: .

Book review: Probably Approximately Correct: Nature’s Algorithms for Learning and Prospering in a Complex World, by Leslie Valiant.

This book provides some nonstandard perspectives on machine learning and evolution, but doesn’t convince me there’s much advantage to using those perspectives. I’m unsure how much of that is due to his mediocre writing style. He often seems close to saying something important, but never gets there.

He provides a rigorous meaning for the concept of learnability. I suppose that’s important for something, but I can’t recall what.

He does an ok job of explaining how evolution is a form of learning, but Eric Baum’s book What is Thought? explains that idea much better.

The last few chapters, where he drifts farther from his areas of expertise, are worse. Much of what he says there only seems half-right at best.

One example is his suggestion that AI researchers ought to put a lot of thought into how teaching materials are presented (similar to how schools are careful to order a curriculum, from simple to complex concepts). I doubt that that reflects a reasonable model of human learning: children develop an important fraction of their intelligence before school age, with little guidance for the order in which they should learn concepts (cf. Piaget’s theory of cognitive development); and unschooled children seem to choose their own curriculum.

My impression of recent AI progress suggests that a better organized “curriculum” is even farther from being cost-effective there – progress seems to be coming more from better ways of incorporating unsupervised learning.

I’m left wondering why anyone thinks the book is worth reading.

How much Diversity of AGI-Risk Organizations is Optimal?

Posted by Peter on December 6, 2015
Posted in: Artificial Intelligence. Tagged: , .

This post is partly a response to arguments for only donating to one charity and to an 80,000 Hours post arguing against diminishing returns. But I’ll focus mostly on AGI-risk charities.

Diversifying Donations?

The rule that I should only donate to one charity is a good presumption to start with. Most objections to it are due to motivations that diverge from pure utilitarian altruism. I don’t pretend that altruism is my only motive for donating, so I’m not too concerned that I only do a rough approximation of following that rule.

Still, I want to follow the rule more closely than most people do. So when I direct less than 90% of my donations to tax-deductible nonprofits, I feel a need to point to diminishing returns [1] to donations to justify that.

With AGI risk organizations, I expect the value of diversity to sometimes override the normal presumption even for purely altruistic utilitarians (with caveats about having the time needed to evaluate multiple organizations, and having more than a few thousand dollars to donate; those caveats will exclude many people from this advice, so this post is mainly oriented toward EAs who are earning to give or wealthier people).

Diminishing Returns?

Before explaining that, I’ll reply to the 80,000 Hours post about diminishing returns.

The 80,000 Hours post focuses on charities that mostly market causes to a wide audience. The economies of scale associated with brand recognition and social proof seem more plausible than any economies of scale available to research organizations.

The shortage of existential risk research seems more dangerous than any shortage of charities which are devoted to marketing causes, so I’m focusing on the most important existential risk.

I expect diminishing returns to be common after an organization grows beyond two or three people. One reason is that the founders of most organizations exert more influence than subsequent employees over important policy decisions [2], so at productive organizations founders are more valuable.

For research organizations that need the smartest people, the limited number of such people implies that only small organizations can have a large fraction of employees be highly qualified.

I expect donations to very young organizations to be more valuable than other donations (which implies diminishing returns to size on average):

  • It takes time to produce evidence that the organization is accomplishing something valuable, and donors quite sensibly prefer organizations that have provided such evidence.
  • Even when donors try to compensate for that by evaluating the charity’s mission statement or leader’s competence, it takes some time to adequately communicate those features (e.g. it’s rare for a charity to set up an impressive web site on day one).
  • It’s common for a charity to have suboptimal competence at fundraising until it grows large enough to hire someone with fundraising expertise.
  • Some charities are mainly funded by a few grants in the millions of dollars, and I’ve heard reports that those often take many months between being awarded and reaching the charities’ bank (not to mention delays in awarding the grants). This sometimes means months when a charity has trouble hiring anyone who demands an immediate salary.
  • Donors could in principle overcome these causes of bias, but as far as I can tell, few care about doing so. EA’s come a little closer to doing this than others, but my observations suggest that EA’s are almost as lazy about analyzing new charities as non EA’s.
  • Therefore, I expect young charities to be underfunded.

Why AGI risk research needs diversity

I see more danger of researchers pursuing useless approaches for existential risks in general, and AGI risks in particular (due partly to the inherent lack of feedback), than with other causes.

The most obvious way to reduce that danger is to encourage a wide variety of people and organizations to independently research risk mitigation strategies.

I worry about AGI-risk researchers focusing all their effort on a class of scenarios which rely on a false assumption.

The AI foom debate seems superficially like the main area where a false assumption might cause AGI research to end up mostly wasted. But there are enough influential people on both sides of this issue that I expect research to not ignore one side of that debate for long.

I worry more about assumptions that no prominent people question.

I’ll describe how such an assumption might look in hindsight via an analogy to some leading developers of software intended to accomplish what the web ended up accomplishing [3].

Xanadu stood out as the leading developer of global hypertext software in the 1980s to about the same extent that MIRI stands out as the leading AGI-risk research organization. One reason [4] that Xanadu accomplished little was the assumption that they needed to make money. Part of why that seemed obvious in the 1980s was that there were no ISPs delivering an internet-like platform to ordinary people, and hardware costs were a big obstacle to anyone who wanted to provide that functionality. The hardware costs declined at a predictable enough rate that Drexler was able to predict in Engines of Creation (published in 1986) that ordinary people would get web-like functionality within a decade.

A more disturbing reason for assuming that web functionality needed to make a profit was the ideology surrounding private property. People who opposed private ownership of home, farms, factories, etc. were causing major problems. Most of us automatically treated ownership of software as working the same way as physical property.

People who are too young to remember attitudes toward free / open source software before about 1997 will have some trouble believing how reluctant people were to imagine valuable software being free. [5] Attitudes changed unusually fast due to the demise of communism and the availability of affordable internet access.

A few people (such as RMS) overcame the focus on cold war issues, but were too eccentric to convert many followers. We should pay attention to people with similarly eccentric AGI-risk views.

If I had to guess what faulty assumption AGI-risk researchers are making, I’d say something like faulty guesses about the nature of intelligence or the architecture of feasible AGIs. But the assumptions that look suspicious to me are ones that some moderately prominent people have questioned.

Vague intuitions along these lines have led me to delay some of my potential existential-risk donations in hopes that I’ll discover (or help create?) some newly created existential-risk projects which produce more value per dollar.

Conclusions

How does this affect my current giving pattern?

My favorite charity is CFAR (around 75 or 80% of my donations), which improves the effectiveness of people who might start new AGI-risk organizations or AGI-development organizations. I’ve had varied impressions about whether additional donations to CFAR have had diminishing returns. They seem to have been getting just barely enough money to hire employees they consider important.

FLI is a decent example of a possibly valuable organization that CFAR played some hard-to-quantify role in starting. It bears a superficial resemblance to an optimal incubator for additional AGI-risk research groups. But FLI seems too focused on mainstream researchers to have much hope of finding the eccentric ideas that I’m most concerned about AGI-researchers overlooking.

Ideally I’d be donating to one or two new AGI-risk startups per year. Conditions seem almost right for this. New AGI-risk organizations are being created at a good rate, mostly getting a few large grants that are probably encouraging them to focus on relatively mainstream views [6].

CSER and FLI sort of fit this category briefly last year before getting large grants, and I donated moderate amounts to them. I presume I didn’t give enough to them for diminishing returns to be important, but their windows of unusual need were short enough that I might well have come close to that.

I’m a little surprised that the increasing interest in this area doesn’t seem to be catalyzing the formation of more low-budget groups pursuing more unusual strategies. Please let me know of any that I’m overlooking.

See my favorite charities web page (recently updated) for more thoughts about specific charities.

[1] – Diminishing returns are the main way that donating to multiple charities at one time can be reconciled with utilitarian altruism.

[2] – I don’t know whether it ought to work this way, but I expect this pattern to continue.

[3] – they intended to accomplish a much more ambitious set of goals.

[4] – probably not the main reason.

[5] – presumably the people who were sympathetic to communism weren’t attracted to small software projects (too busy with politics?) or rejected working on software due to the expectation that it required working for evil capitalists.

[6] – The short-term effects are probably good, increasing the diversity of approaches compared to what would be the case if MIRI were the only AGI-risk organization, and reducing the risk that AGI researchers would become polarized into tribes that disagree about whether AGI is dangerous. But a field dominated by a few funders tends to focus on fewer ideas than one with many funders.

Uploaded Bonobos

Posted by Peter on November 6, 2015
Posted in: Artificial Intelligence, The Flynn Effect. Tagged: , , , .

I’d like to see more discussion of uploaded ape risks.

There is substantial disagreement over how fast an uploaded mind (em) would improve its abilities or the abilities of its progeny. I’d like to start by analyzing a scenario where it takes between one and ten years for an uploaded bonobo to achieve human-level cognitive abilities. This scenario seems plausible, although I’ve selected it more to illustrate a risk that can be mitigated than because of arguments about how likely it is.

I claim we should anticipate at least a 20% chance a human-level bonobo-derived em would improve at least as quickly as a human that uploaded later.

Considerations that weigh in favor of this are: that bonobo minds seem to be about as general-purpose as humans, including near-human language ability; and the likely ease of ems interfacing with other software will enable them to learn new skills faster than biological minds will.

The most concrete evidence that weighs against this is the modest correlation between IQ and brain size. It’s somewhat plausible that it’s hard to usefully add many neurons to an existing mind, and that bonobo brain size represents an important cognitive constraint.

I’m not happy about analyzing what happens when another species develops more powerful cognitive abilities than humans, so I’d prefer to have some humans upload before the bonobos become superhuman.

A few people worry that uploading a mouse brain will generate enough understanding of intelligence to quickly produce human-level AGI. I doubt that biological intelligence is simple / intelligible enough for that to work. So I focus more on small tweaks: the kind of social pressures which caused the Flynn Effect in humans, selective breeding (in the sense of making many copies of the smartest ems, with small changes to some copies), and faster software/hardware.

The risks seem dependent on the environment in which the ems live and on the incentives that might drive their owners to improve em abilities. The most obvious motives for uploading bonobos (research into problems affecting humans, and into human uploading) create only weak incentives to improve the ems. But there are many other possibilities: military use, interesting NPCs, or financial companies looking for interesting patterns in large databases. No single one of those looks especially likely, but with many ways for things to go wrong, the risks add up.

What could cause a long window between bonobo uploading and human uploading? Ethical and legal barriers to human uploading, motivated by risks to the humans being uploaded and by concerns about human ems driving human wages down.

What could we do about this risk?

Political activism may mitigate the risks of hostility to human uploading, but if done carelessly it could create a backlash which worsens the problem.

Conceivably safety regulations could restrict em ownership/use to people with little incentive to improve the ems, but rules that looked promising would still leave me worried about risks such as irresponsible people hacking into computers that run ems and stealing copies.

A more sophisticated approach is to improve the incentives to upload humans. I expect the timing of the first human uploads to be fairly sensitive to whether we have legal rules which enable us to predict who will own em labor. But just writing clear rules isn’t enough – how can we ensure political support for them at a time when we should expect disputes over whether they’re people?

We could also find ways to delay ape uploading. But most ways of doing that would also delay human uploading, which creates tradeoffs that I’m not too happy with (partly due to my desire to upload before aging damages me too much).

If a delay between bonobo and human uploading is dangerous, then we should also ask about dangers from other uploaded species. My intuition says the risks are much lower, since it seems like there are few technical obstacles to uploading a bonobo brain shortly after uploading mice or other small vertebrates.

But I get the impression that many people associated with MIRI worry about risks of uploaded mice, and I don’t have strong evidence that I’m wiser than they are. I encourage people to develop better analyses of this issue.

Artificial Superintelligence: A Futuristic Approach

Posted by Peter on August 17, 2015
Posted in: Artificial Intelligence, Book Reviews. Tagged: .

Book review: Artificial Superintelligence: A Futuristic Approach, by Roman V. Yampolskiy.

This strange book has some entertainment value, and might even enlighten you a bit about the risks of AI. It presents many ideas, with occasional attempts to distinguish the important ones from the jokes.

I had hoped for an analysis that reflected a strong understanding of which software approaches were most likely to work. Yampolskiy knows something about computer science, but doesn’t strike me as someone with experience at writing useful code. His claim that “to increase their speed [AIs] will attempt to minimize the size of their source code” sounds like a misconception that wouldn’t occur to an experienced programmer. And his chapter “How to Prove You Invented Superintelligence So No One Else Can Steal It” seems like a cute game that someone might play with if he cared more about passing a theoretical computer science class than about, say, making money on the stock market, or making sure the superintelligence didn’t destroy the world.

I’m still puzzling over some of his novel suggestions for reducing AI risks. How would “convincing robots to worship humans as gods” differ from the proposed Friendly AI? Would such robots notice (and resolve in possibly undesirable ways) contradictions in their models of human nature?

Other suggestions are easy to reject, such as hoping AIs will need us for our psychokinetic abilities (abilities that Yampolskiy says are shown by peer-reviewed experiments associated with the Global Consciousness Project).

The style is also weird. Some chapters were previously published as separate papers, and weren’t adapted to fit together. It was annoying to occasionally see sentences that seemed identical to ones in a prior chapter.

The author even has strange ideas about what needs footnoting. E.g. when discussing the physical limits to intelligence, he cites (Einstein 1905).

Only read this if you’ve read other authors on this subject first.

Superintelligence

Posted by Peter on July 28, 2014
Posted in: Artificial Intelligence, Book Reviews. Tagged: .

Book review: Superintelligence: Paths, Dangers, Strategies, by Nick Bostrom.

This book is substantially more thoughtful than previous books on AGI risk, and substantially better organized than the previous thoughtful writings on the subject.

Bostrom’s discussion of AGI takeoff speed is disappointingly philosophical. Many sources (most recently CFAR) have told me to rely on the outside view to forecast how long something will take. We’ve got lots of weak evidence about the nature of intelligence, how it evolved, and about how various kinds of software improve, providing data for an outside view. Bostrom assigns a vague but implausibly high probability to AI going from human-equivalent to more powerful than humanity as a whole in days, with little thought of this kind of empirical check.

I’ll discuss this more in a separate post which is more about the general AI foom debate than about this book.

Bostrom’s discussion of how takeoff speed influences the chance of a winner-take-all scenario makes it clear that disagreements over takeoff speed are pretty much the only cause of my disagreement with him over the likelihood of a winner-take-all outcome. Other writers aren’t this clear about this. I suspect those who assign substantial probability to a winner-take-all outcome if takeoff is slow will wish he’d analyzed this in more detail.

I’m less optimistic than Bostrom about monitoring AGI progress. He says “it would not be too difficult to identify most capable individuals with a long-standing interest in [AGI] research”. AGI might require enough expertise for that to be true, but if AGI surprises me by only needing modest new insights, I’m concerned by the precedent of Tim Berners-Lee creating a global hypertext system while barely being noticed by the “leading” researchers in that field. Also, the large number of people who mistakenly think they’ve been making progress on AGI may obscure the competent ones.

He seems confused about the long-term trends in AI researcher beliefs about the risks: “The pioneers of artificial intelligence … mostly did not contemplate the possibility of greater-than-human AI” seems implausible; it’s much more likely they expected it but were either overconfident about it producing good results or fatalistic about preventing bad results (“If we’re lucky, they might decide to keep us as pets” – Marvin Minsky, LIFE Nov 20, 1970).

The best parts of the book clarify many issues related to ensuring that an AGI does what we want.

He catalogs more approaches to controlling AGI than I had previously considered, including tripwires, oracles, and genies, and clearly explains many limits to what they can accomplish.

He briefly mentions the risk that the operator of an oracle AI would misuse it for her personal advantage. Why should we have less concern about the designers of other types of AGI giving them goals that favor the designers?

If an oracle AI can’t produce a result that humans can analyze well enough to decide (without trusting the AI) that it’s safe, why would we expect other approaches (e.g. humans writing the equivalent seed AI directly) to be more feasible?

He covers a wide range of ways we can imagine handling AI goals, including strange ideas such as telling an AGI to use the motivations of superintelligences created by other civilizations

He does a very good job of discussing what values we should and shouldn’t install in an AGI: the best decision theory plus a “do what I mean” dynamic, but not a complete morality.

I’m somewhat concerned by his use of “final goal” without careful explanation. People who anthropomorphise goals are likely to confuse at least the first few references to “final goal” as if it worked like a human goal, i.e. something that the AI might want to modify if it conflicted with other goals.

It’s not clear how much of these chapters depend on a winner-take-all scenario. I get the impression that Bostrom doubts we can do much about the risks associated with scenarios where multiple AGIs become superhuman. This seems strange to me. I want people who write about AGI risks to devote more attention to whether we can influence whether multiple AGIs become a singleton, and how they treat lesser intelligences. Designing AGI to reflect values we want seems almost as desirable in scenarios with multiple AGIs as in the winner-take-all scenario (I’m unsure what Bostrom thinks about that). In a world with many AGIs with unfriendly values, what can humans do to bargain for a habitable niche?

He has a chapter on worlds dominated by whole brain emulations (WBE), probably inspired by Robin Hanson’s writings but with more focus on evaluating risks than on predicting the most probable outcomes. Since it looks like we should still expect an em-dominated world to be replaced at some point by AGI(s) that are designed more cleanly and able to self-improve faster, this isn’t really an alternative to the scenarios discussed in the rest of the book.

He treats starting with “familiar and human-like motivations” (in an augmentation route) as an advantage. Judging from our experience with humans who take over large countries, a human-derived intelligence that conquered the world wouldn’t be safe or friendly, although it would be closer to my goals than a smiley-face maximizer. The main advantage I see in a human-derived superintelligence would be a lower risk of it self-improving fast enough for the frontrunner advantage to be large. But that also means it’s more likely to be eclipsed by a design more amenable to self-improvement.

I’m suspicious of the implication (figure 13) that the risks of WBE will be comparable to AGI risks.

  • Is that mainly due to “neuromorphic AI” risks? Bostrom’s description of neuromorphic AI is vague, but my intuition is that human intelligence isn’t flexible enough to easily get the intelligence part of WBE without getting something moderately close to human behavior.
  • Is the risk of uploaded chimp(s) important? I have some concerns there, but Bostrom doesn’t mention it.
  • How about the risks of competitive pressures driving out human traits (discussed more fully/verbosely at Slate Star Codex)? If WBE and AGI happen close enough together in time that we can plausibly influence which comes first, I don’t expect the time between the two to be long enough for that competition to have large effects.
  • The risk that many humans won’t have enough resources to survive? That’s scary, but wouldn’t cause the astronomical waste of extinction.

Also, I don’t accept his assertion that AGI before WBE eliminates the risks of WBE. Some scenarios with multiple independently designed AGIs forming a weakly coordinated singleton (which I consider more likely than Bostrom does) appear to leave the last two risks in that list unresolved.

This books represents progress toward clear thinking about AGI risks, but much more work still needs to be done.

Our Mathematical Universe

Posted by Peter on April 14, 2014
Posted in: Artificial Intelligence, Book Reviews, Life, the Universe, and Everything. Tagged: , , .

Book review: Our Mathematical Universe: My Quest for the Ultimate Nature of Reality, by Max Tegmark.

His most important claim is the radical Platonist view that all well-defined mathematical structures exist, therefore most physics is the study of which of those we inhabit. His arguments are more tempting than any others I’ve seen for this view, but I’m left with plenty of doubt.

He points to ways that we can imagine this hypothesis being testable, such as via the fine-tuning of fundamental constants. But he doesn’t provide a good reason to think that those tests will distinguish his hypothesis from other popular approaches, as it’s easy to imagine that we’ll never find situations where they make different predictions.

The most valuable parts of the book involve the claim that the multiverse is spatially infinite. He mostly talks as if that’s likely to be true, but his explanations caused me to lower my probability estimate for that claim.

He gets that infinity by claiming that inflation continues in places for infinite time, and then claiming there are reference frames for which that infinite time is located in a spatial rather than a time direction. I have a vague intuition why that second step might be right (but I’m fairly sure he left something important out of the explanation).

For the infinite time part, I’m stuck with relying on argument from authority, without much evidence that the relevant authorities have much confidence in the claim.

Toward the end of the book he mentions reasons to doubt infinities in physics theories – it’s easy to find examples where we model substances such as air as infinitely divisible, when we know that at some levels of detail atomic theory is more accurate. The eternal inflation theory depends on an infinitely expandable space which we can easily imagine is only an approximation. Plus, when physicists explicitly ask whether the universe will last forever, they don’t seem very confident. I’m also tempted to say that the measure problem (i.e. the absence of a way to say some events are more likely than others if they all happen an infinite number of times) is a reason to doubt infinities, but I don’t have much confidence that reality obeys my desire for it to be comprehensible.

I’m disappointed by his claim that we can get good evidence that we’re not Boltzmann brains. He wants us to test our memories, because if I am a Boltzmann brain I’ll probably have a bunch of absurd memories. But suppose I remember having done that test in the past few minutes. The Boltzmann brain hypothesis suggests it’s much more likely for me to have randomly acquired the memory of having passed the test than for me to actually be have done the test. Maybe there’s a way to turn Tegmark’s argument into something rigorous, but it isn’t obvious.

He gives a surprising argument that the differences between the Everett and Copenhagen interpretations of quantum mechanics don’t matter much, because unrelated reasons involving multiverses lead us to expect results comparable to the Everett interpretation even if the Copenhagen interpretation is correct.

It’s a bit hard to figure out what the book’s target audience is – he hides the few equations he uses in footnotes to make it look easy for laymen to follow, but he also discusses hard concepts such as universes with more than one time dimension with little attempt to prepare laymen for them.

The first few chapters are intended for readers with little knowledge of physics. One theme is a historical trend which he mostly describes as expanding our estimate of how big reality is. But the evidence he provides only tells us that the lower bounds that people give keep increasing. Looking at the upper bound (typically infinity) makes that trend look less interesting.

The book has many interesting digressions such as a description of how to build Douglas Adams’ infinite improbability drive.