I mostly appear from 3:30 to 5:30.
Added 16Dec: I’m told this video has had over 100,000 views so far.
I mostly appear from 3:30 to 5:30.
Added 16Dec: I’m told this video has had over 100,000 views so far.
A recent burst of econo–blog posts on the subject of a future robot based economy mostly seem to treat the subject as if those few bloggers were the only people ever to consider the subject. But in fact, people have been considering the subject for centuries. I myself have written dozens of posts just here on this blog.
So let me offer a quick robot econ primer, i.e. important points widely known among folks who have long discussed the subject, but often not quickly rediscovered by dilettantes new to the subject:
There’s lots more I could add, but this should be enough for now.
My ex-co-blogger Eliezer Yudkowsky last June:
I worry about conversations that go into “But X is like Y, which does Z, so X should do reinterpreted-Z”. Usually, in my experience, that goes into what I call “reference class tennis” or “I’m taking my reference class and going home”. The trouble is that there’s an unlimited number of possible analogies and reference classes, and everyone has a different one. I was just browsing old LW posts today (to find a URL of a quick summary of why group-selection arguments don’t work in mammals) and ran across a quotation from Perry Metzger to the effect that so long as the laws of physics apply, there will always be evolution, hence nature red in tooth and claw will continue into the future – to him, the obvious analogy for the advent of AI was “nature red in tooth and claw”, and people who see things this way tend to want to cling to that analogy even if you delve into some basic evolutionary biology with math to show how much it isn’t like intelligent design. For Robin Hanson, the one true analogy is to the industrial revolution and farming revolutions, meaning that there will be lots of AIs in a highly competitive economic situation with standards of living tending toward the bare minimum, and this is so absolutely inevitable and consonant with The Way Things Should Be as to not be worth fighting at all. That’s his one true analogy and I’ve never been able to persuade him otherwise. For Kurzweil, the fact that many different things proceed at a Moore’s Law rate to the benefit of humanity means that all these things are destined to continue and converge into the future, also to the benefit of humanity. For him, “things that go by Moore’s Law” is his favorite reference class.
I can have a back-and-forth conversation with Nick Bostrom, who looks much more favorably on Oracle AI in general than I do, because we’re not playing reference class tennis with “But surely that will be just like all the previous X-in-my-favorite-reference-class”, nor saying, “But surely this is the inevitable trend of technology”; instead we lay out particular, “Suppose we do this?” and try to discuss how it will work, not with any added language about how surely anyone will do it that way, or how it’s got to be like Z because all previous Y were like Z, etcetera. (more)
When we shared this blog, Eliezer and I had a long debate here on his “AI foom” claims. Later, we debated in person once. (See also slides 34,35 of this 3yr-old talk.) I don’t accept the above as characterizing my position well. I’ve written up a summaries before, but let me try again, this time trying to more directly address the above critique.
Eliezer basically claims that the ability of an AI to change its own mental architecture is such a potent advantage as to make it likely that a cheap unnoticed and initially low ability AI (a mere “small project machine in a basement”) could without warning over a short time (e.g., a weekend) become so powerful as to be able to take over the world.
As this would be a sudden big sustainable increase in the overall growth rate in the broad capacity of the world economy, I do find it useful to compare to compare this hypothesized future event to the other pasts events that produce similar outcomes, namely a big sudden sustainable global broad capacity rate increase. The last three were the transitions to humans, farming, and industry.
I don’t claim there is some hidden natural law requiring such events to have the same causal factors or structure, or to appear at particular times. But I do think these events suggest a useful if weak data-driven prior on the kinds of factors likely to induce such events, on the rate at which they occur, and on their accompanying inequality in gains. In particular, they tell us that such events are very rare, that over the last three events gains have been spread increasingly equally, and that these three events seem mainly due to better ways to share innovations.
Eliezer sees the essence of his scenario as being a change in the “basic” architecture of the world’s best optimization process, and he sees the main prior examples of this as the origin of natural selection and the arrival of humans. He also sees his scenario as differing enough from the other studied growth scenarios as to make analogies to them of little use.
However, since most global bio or econ growth processes can be thought of as optimization processes, this comes down to his judgement on what counts as a “basic” structure change, and on how different such scenarios are from other scenarios. And in my judgement the right place to get and hone our intuitions about such things is our academic literature on global growth processes.
Economists have a big literature on processes by which large economies grow, increasing our overall capacities to achieve all the things we value. There are of course many other growth literatures, and some of these deal in growths of capacities, but these usually deal with far more limited systems. Of these many growth literatures it is the economic growth literature that is closest to dealing with the broad capability growth posited in a fast growing AI scenario.
It is this rich literature that seems to me the right place to find and hone our categories for thinking about growing broadly capable systems. One should review many formal theoretical models, and many less formal applications of such models to particular empirical contexts, collecting “data” points of what is thought to increase or decrease growth of what in what contexts, and collecting useful categories for organizing such data points.
With such useful categories in hand one can then go into a new scenario such as AI foom and have a reasonable basis for saying how similar that new scenario seems to old scenarios, which old scenarios it seems most like if any, and which parts of that new scenario are central vs. peripheral. Yes of course if this new area became mature it could also influence how we think about other scenarios.
But until we actually see substantial AI self-growth, most of the conceptual influence should go the other way. Relying instead primarily on newly made up categories and similarity maps between them, concepts and maps which have not been vetted or honed in dealing with real problems, seems to me a mistake. Yes of course a new problem may require one to introduce some new concepts to describe it, but that is hardly the same as largely ignoring old concepts.
So, I fully grant that the ability of AIs to intentionally change mind designs would be a new factor in the world, and it could make a difference for AI ability to self-improve. But while the history of growth over the last few million years has seen many dozens of factors come and go, or increase and decrease in importance, it has only seen three events in which overall growth rates greatly increased suddenly and sustainably. So the mere addition of one more factor seems unlikely to generate foom, unless our relevant categories for growth causing factors suggest that this factor is unusually likely to have such an effect.
This is the sense in which I long ago warned against over-reliance on “unvetted” abstractions. I wasn’t at all trying to claim there is one true analogy and all others are false. Instead, I argue for preferring to rely on abstractions, including categories and similarity maps, that have been found useful by a substantial intellectual community working on related problems. On the subject of an AI growth foom, most of those abstractions should come from the field of economic growth.
This weekend I was in a AAAI (Association for the Advancement of Artificial Intelligence) Fall Symposium on Machine Aggregation of Human Judgment. It was my job to give a short summary about our symposium to the eight co-located symposia. Here is what I said.
In most of AI, data is input, and judgements are output. But here humans turn data into judgements, and then machines and institutions combine those judgements. This work is often inspired by a “wisdom of crowds” idea that we often rely too much on arrogant over-rated experts instead of the under-rated insight of everyone else. Boo elites; rah ordinary folks!
Many of the symposium folks are part of the IARPA ACE project, which is structured as a competition between four teams, each of which must collect several hundred participants to answer the same real-time intelligence questions, with roughly a hundred active questions at any one time. Each team uses a different approach. The two most common ways are to ask many people for estimates, and then average them somehow, or to have people trade in speculative betting markets. ACE is now in its second of four years. So, what have we learned?
First, we’ve learned that it helps to transform probability estimates into log-odds before averaging them. Weights can then correct well for predictable over- or under-confidence. We’ve also learned better ways to elicit estimates. For example, instead of asking for a 90% confidence interval on a number, it is better to ask for an interval, and then for a probability. It works even better to ask about an interval someone else picked. Also, instead of asking people directly for their confidence, it is better to ask them how much their opinion would change if they knew what others know.
Our DAGGRE team is trying to improve accuracy by breaking down questions into a set of related correlated questions. ACE has also learned how to make people better at estimating, both by training them in basic probability theory, and by having them work together in teams.
But the biggest thing we’ve learned is that people are unequal – the best way to get good crowd wisdom is to have a good crowd. Contributions that most improve accuracy are more extreme, more recent, by those who contribute more often, and come with more confidence. In our DAGGRE system, most value comes from a few dozen of our thousands of participants. True, these elites might not be the same folks you’d have picked via resumes, and tracking success may give better incentives. But still, what we’ve most learned about the wisdom of crowds is that it is best to have an elite “crowd.”
Miller discusses several possible paths to super-intelligence, but never says which paths he thinks likely, nor when any might happen. However, he is confident that one will happen eventually, he calls Kurzweil’s 2045 forecast “robust”, and he offers readers personal advice as if something will happen in their lifetimes.
I get a lot of coverage in chapter 13, which discusses whole brain emulations. (And Katja is mentioned on pp.213-214.) While Miller focuses mostly on what emulations imply for humans, he does note that many ems could die from poverty or obsolescence. He make no overall judgement on the scenario, however, other than to once use the word “dystopian.”
While Miller’s discussion of emulations is entirely of the scenario of a large economy containing many emulations, his discussion of non-emulation AI is entirely of the scenario of a single “ultra AI”. He never considers a single ultra emulation, nor an economy of many AIs. Nor does he explain these choices.
On ultra AIs, Miller considers only an “intelligence explosion” scenario where a human level AI turns itself into an ultra AI “in a period of weeks, days, or even hours.” His arguments for this extremely short timescale are:
I’ve said before that I don’t see how these imply a weeks timescale for one human level AI to make itself more powerful than the entire rest of the world put together. Miller explains my skepticism:
As Hanson told me, the implausibility of some James Bond villains illustrates a reason to be skeptical of an intelligence explosion. A few of these villains had their own private islands on which they created new powerful weapons. But weapons development is a time and resource intensive task, making it extremely unlikely that the villains small team of followers could out-innovate all of the weapons developers in the rest of the world by producing spectacularly destructive instruments that no other military force possessed. Thinking that a few henchmen, even if led by an evil genius, would do a better job at weapons development than a major defense contractor is as silly as believing that the professor on Gilligan’s Island really could have created his own coconut based technology. …
Think of an innovation race between a single AI and the entirety of mankind. For an intelligence explosion to occur, the AI has to not only win the race, but finish before humanity completes its next stride. A sufficiently smart AI could certainly do this, but an AI only a bit brighter than von Neumann would not have the slightest chance of achieving this margin of victory. (pp.215-216)
As you can tell from this quotation, Miller’s book often reads like the economics textbook he wrote. He is usually content to be a tutor, explaining common positions and intuitions behind common arguments. He does, however, explain some of his personal contributions to this field, such as his argument that preventing the destruction of the world can be a public good undersupplied by private firms, and that development might slow down just before an anticipated explosion, if investors think non-investors will gain or lose just as much as investors from the change.
I’m not sure this book has much of a chance to get very popular. The competition is fierce, Miller isn’t already famous, and while his writing quality is good, it isn’t at the popular blockbuster popular book level. But I wish his book all the success it can muster.
From ’85 to ’93 I was an AI researcher, first at Lockheed AI Center, then at the NASA Ames AI group. In ’91 I presented at IJCAI, the main international AI conference, on a probability related paper. Back then this was radical – one questioner at my talk asked “How can this be AI, since it uses math?” Probability specialists created their own AI conference UAI, to have a place to publish.
Today probability math is well accepted in AI. The long AI battle between the neats and scruffs was won handily by the neats – math and theory are very accepted today. UAI is still around though, and a week ago I presented another probability related paper there (slides, audio), on our combo prediction market algorithm. And listening to all the others talks at the conference let me reflect on the state of the field, and its progress in the last 21 years.
Overall I can’t complain much about emphasis. I saw roughly the right mix of theory vs. application, of general vs. specific results, etc. I doubt the field would progress more than a factor of two faster if such parameters were exactly optimized. The most impressive demo I saw was Video In Sentences Out, an end-to-end integrated system for writing text summaries of simple videos. Their final test stats:
Human judges rated each video-sentence pair to assess whether the sentence was true of the video and whether it described a salient event depicted in that video. 26.7% (601/2247) of the video-sentence pairs were deemed to be true and 7.9% (178/2247) of the video-sentence pairs were deemed to be salient.
This is actually pretty impressive, once you understand just how hard the problem is. Yes, we have a long way to go, but are making steady progress.
So how far have we come in last twenty years, compared to how far we have to go to reach human level abilities? I’d guess that relative to the starting point of our abilities of twenty years ago, we’ve come about 5-10% of the distance toward human level abilities. At least in probability-related areas, which I’ve known best. I’d also say there hasn’t been noticeable acceleration over that time. Over a thirty year period, it is even fair to say there has been deceleration, since Pearl’s classic ’88 book was such a big advance.Robin Hanson
I asked a few other folks at UAI who had been in the field for twenty years to estimate the same things, and they roughly agreed – about 5-10% of the distance has been covered in that time, without noticeable acceleration. It would be useful to survey senior experts in other areas of AI, to get related estimates for their areas. If this 5-10% estimate is typical, as I suspect it is, then an outside view calculation suggests we probably have at least a century to go, and maybe a great many centuries, at current rates of progress.
Added 21Oct: At the recent Singularity Summit, I asked speaker Melanie Mitchell to estimate how far we’ve come in her field of analogical reasoning in the last twenty years. She estimated 5 percent of the way to human level abilities, with no noticeable acceleration.
Added 11Dec: At the Artificial General Intelligence conference, Murray Shanahan says that looking at his twenty years experience in the knowledge representation field, he estimates we have come 10% of the way, with no noticeable acceleration.
Added 4Oct’13: At an NSF workshop on social computing, Wendy Hall said that in her twenty years in computer-assisted training, we’ve moved less than 1% of the way to human level abilities. Claire Cardie said that in her twenty years in natural language processing, we’ve come 20% of the way. Boi Faltings says that in his field of solving constraint satisfaction problems, they were past human level abilities twenty years ago, and are even further past that today.
Let me clarify that I mean to ask people about progress in a field of AI as it was conceived twenty years ago. Looking backward one can define areas in which we’ve made great progress. But to avoid selection biases, I want my survey to focus on areas as they were defined back then.
Added 21May’14: At a private event, after Aaron Dollar talked on robotics, he told me that in twenty years we’ve come less than 1% of the distance to human level abilities in his subfield of robotic grasping manipulation. But he has seen noticeable acceleration over that time.
Added 28Aug’14: After coming to a talk of mine, Peter Norvig told me that he agrees with both Claire Cardie and Boi Faltings, that on speech recognition and machine translation we’ve gone from not usable to usable in 20 years, though we still have far to go on deeper question answering, and for retrieving a fact or page that is relevant to a search query we’ve far surpassed human ability in recall and do pretty well on precision.
Added 14Sep’14: At a closed academic workshop, Timothy Meese, who researches early vision processing in humans, told me he estimates about 5% progress in his field in the last 20 years, with a noticeable deceleration.
Added 4Jan’15: At a closed meeting, Francesca Rossi, expert in constraint reasoning, gave an estimate of 10%, with deceleration. Margret Boden, author of Artificial Intelligence and Natural Man (1977), estimated 5%, but for no particular subfield.
Added 6July’15: David Kelley, expert in big data analysis, says 5% in last twenty years, sees acceleration only in last 2-3 years, not before that.
People are often interested in robot ethics. I have argued before that this is strange. I offered two potential explanations:
A more obvious explanation now: people are just more interested in ethics when the subject is far away, for instance in the future. This is the prediction of construal level theory. It says thinking about something far away makes you think more abstractly, and in terms of goals and ideals rather than low level constraints. Ethics is all this.
So a further prediction would be that when we come to use robots a lot, expertise from robot ethicists will be in as little demand as expertise from washing machine ethicists is now.
Some other predictions, to help check this theory:
More? Which of these are actually true?
There is definitely some conflicting evidence, for instance people feel more compelled to help people in front of them than those in Africa (there was an old OB post on this, but I can’t find it). There are also many other reasons the predictions above may be true. Emerging technologies might prompt more ethical concerns because they are potentially more dangerous for instance. The ethical dimension to killing everyone is naturally prominent. Overall construal level theory still seems to me a promising model for variations in ethical concern.
Added: I’m not confident that there is disproportionate interest compared to other topic areas. I seem to have heard about it too much, but this could be a sampling bias.
Back in July I posted my response to Chalmers’ singularity essay, published in the Journal of Consciousness Studies (JCS) where his paper was published. A paper copy of a JCS issue with thirteen responses recently showed up in my mail, though no JCS electronic copy is yet available. [Added 4Mar: it is now here.] Reading through the responses, the best (besides mine) was by Marcus Hutter.
I didn’t learn much new, but compared to the rest, Hutter is relatively savvy on social issues. He isn’t sure if it is possible to be much more intelligent than a human (as opposed to just thinking faster), but he is sure there is lots of room for improvement overall:
The technological singularity refers to a hypothetical scenario in which technological advances virtually explode. …
When building AIs or tinkering with our virtual selves, we could try out a lot of diﬀerent goals. … But ultimately we will lose control, and the AGIs themselves will build further AGIs. … Some aspects of this might be independent of the initial goal structure and predictable. Probably this initial vorld is a society of cooperating and competing agents. There will be competition over limited (computational) resources, and those virtuals who have the goal to acquire them will naturally be more successful. … The successful virtuals will spread (in various ways), the others perish, and soon their society will consist mainly of virtuals whose goal is to compete over resources, where hostility will only be limited if this is in the virtuals’ best interest. For instance, current society has replaced war mostly by economic competition. … This world will likely neither be heaven nor hell for the virtuals. They will “like” to ﬁght over resources, and the winners will “enjoy” it, while the losers will “hate” it. …
In the human world, local conﬂicts and global war is increasingly replaced by economic competition, which might itself be replaced by even more constructive global collaboration, as long as violaters can quickly and eﬀectively (and non-violently?) be eliminated. It is possible that this requires a powerful single (virtual) world government, to give up individual privacy, and to severely limit individual freedom (cf. ant hills or bee hives).
Hutter noted (as have I) that cheap life is valued less:
Unless a global copy protection mechanism is deliberately installed, … copying virtual structures should be as cheap and eﬀortless as it is for software and data today. The only cost is developing the structures in the ﬁrst place, and the memory to store and the comp to run them. … One consequence … [is] life becoming much more diverse. …
Another consequence should be that life becomes less valuable. … Cheap machines decreased the value of physical labor. … In games, we value our own life and that of our opponents less than real life, … because games can be reset and one can be resurrected. … Why not participate in a dangerous fun activity. … It may be ethically acceptable to freeze, duplicate, slow-down, modify (brain experiments), or even kill (oneself or other) AIs at will, if they are abundant and/or backups are available, just what we are used to doing with software. So laws preventing experimentation with intelligences for moral reasons may not emerge.
Hutter also tried to imagine what such a society would look like from outside:
Imagine an inward explosion, where a ﬁxed amount of matter is transformed into increasingly eﬃcient computers until it becomes computronium. The virtual society like a well-functioning real society will likely evolve and progress, or at least change. Soon the speed of their aﬀairs will make them beyond comprehension for the outsiders. … After a brief period, intelligent interaction between insiders and outsiders becomes impossible. …
Let us now consider outward explosion, where an increasing amount of matter is transformed into computers of ﬁxed eﬃciency. … Outsiders will soon get into resource competition with the expanding computer world, and being inferior to the virtual intelligences, probably only have the option to ﬂee. This might work for a while, but soon … escape becomes impossible, ending or converting the outsiders’ existence.
When foragers were outside of farmer societies, or farmers outside of industrial cities, change was faster on the inside, and the faster change got the harder it was for outsiders to understand. But there was no sharp boundary when understanding became “impossible.” While farmers were greedy for more land, and displaced foragers on farmable (or herd able) land quickly in farming doubling time terms, industry has been much less expansionary. While eventually industry might displace all farming, farming modes of production can continue to use land for many industry doubling times into an industrial revolution.
Similarly, a new faster economic growth mode might well continue to let old farming and industrial modes of production continue for a great many doubling times of the new mode. If land area is not central to the new mode of production, why expect old land uses to be quickly displaced?
I “won” in the sense of gaining more audience votes — the vote was 45-40 (him to me) before, and 32-33 after the debate. That makes me two for two, after my similar “win” over Bryan Caplan (42-10 before, 25-20 after). This probably says little about me, however, since contrarians usually “win” such debates.
Our topic was: Compared to the farming and industrial revolutions, intelligence explosion first-movers will quickly control a much larger fraction of their new world. He was pro, I was con. We also debated this subject here on Overcoming Bias from June to December 2008. Let me now try to summarize my current position.
The key issue is: how chunky and powerful are as-yet-undiscovered insights into the architecture of “thinking” in general (vs. on particular topics)? Assume there are many such insights, each requiring that brains be restructured to take advantage. (Ordinary humans couldn’t use them.) Also assume that the field of AI research reaches a key pivotal level of development. And at that point, imagine some AI research team discovers a powerful insight, and builds an AI with an architecture embodying it. Such an AI might then search for more such insights more efficiently than all other the AI research teams who share their results put together.
This new fast AI might then use its advantage to find another powerful insight, restructure itself to take advantage of it, and so on until it was fantastically good at thinking in general. (Or if the first insight were super-powerful, it might jump to this level in one step.) How good? So good that it could greatly out-compete the entire rest of the world at the key task of learning the vast ocean of specific knowledge and insights useful for functioning in the world. So good that even though it started out knowing almost nothing, after a few weeks it knows more than the entire rest of the world put together.
(Note that the advantages of silicon and self-modifiable code over biological brains do not count as relevant chunky architectural insights — they are available to all competing AI teams.)
In the debate, Eliezer gave six reasons to think very powerful brain architectural insights remain undiscovered:
My responses: Continue reading "Debating Yudkowsky" »
The natural and common human obsession with how much [robot] values differ overall from ours distracts us from worrying effectively. … [Instead:]
1. Reduce the salience of the them-us distinction relative to other distinctions. …
2. Have them and us use the same (or at least similar) institutions to keep peace among themselves and ourselves as we use to keep peace between them and us.
I just wrote a 3000 word new comment on this paper, for a journal. Mostly I complain Chalmers didn’t say much beyond what we should have already known. But my conclusion is less meta:
The most robust and promising route to low cost and mutually beneficial mitigation of these [us vs. superintelligence] conflicts is strong legal enforcement of retirement and bequest contracts. Such contracts could let older generations directly save for their later years, and cheaply pay younger generations to preserve old loyalties. Simple consistent and broad-based enforcement of these and related contracts seem our best chance to entrench the enforcement of such contracts deep in legal practice. Our descendants should be reluctant to violate deeply entrenched practices of contract law for fear that violations would lead to further unraveling of contract practice, which threatens larger social orders built on contract enforcement.
As Chalmers notes in footnote 19, this approach is not guaranteed to work in all possible scenarios. Nevertheless, compare it to the ideal Chalmers favors:
AI systems such that we can prove they will always have certain benign values, and such that we can prove that any systems they will create will also have those values, and so on … represents a sort of ideal that we might aim for (p.35).
Compared to the strong and strict controls and regimentation required to even attempt to prove that values disliked by older generations could never arise in any later generations, enforcing contracts where older generations pay younger generations to preserve specific loyalties seems to me a far easier, safer and more workable approach, with many successful historical analogies on which to build.