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 fifth section in the reading guide: Forms of superintelligence. This corresponds to Chapter 3, on different ways in which an intelligence can be super.
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: Chapter 3 (p52-61)
Summary
- A speed superintelligence could do what a human does, but faster. This would make the outside world seem very slow to it. It might cope with this partially by being very tiny, or virtual. (p53)
- A collective superintelligence is composed of smaller intellects, interacting in some way. It is especially good at tasks that can be broken into parts and completed in parallel. It can be improved by adding more smaller intellects, or by organizing them better. (p54)
- A quality superintelligence can carry out intellectual tasks that humans just can't in practice, without necessarily being better or faster at the things humans can do. This can be understood by analogy with the difference between other animals and humans, or the difference between humans with and without certain cognitive capabilities. (p56-7)
- These different kinds of superintelligence are especially good at different kinds of tasks. We might say they have different 'direct reach'. Ultimately they could all lead to one another, so can indirectly carry out the same tasks. We might say their 'indirect reach' is the same. (p58-9)
- We don't know how smart it is possible for a biological or a synthetic intelligence to be. Nonetheless we can be confident that synthetic entities can be much more intelligent than biological entities.
- Digital intelligences would have better hardware: they would be made of components ten million times faster than neurons; the components could communicate about two million times faster than neurons can; they could use many more components while our brains are constrained to our skulls; it looks like better memory should be feasible; and they could be built to be more reliable, long-lasting, flexible, and well suited to their environment.
- Digital intelligences would have better software: they could be cheaply and non-destructively 'edited'; they could be duplicated arbitrarily; they could have well aligned goals as a result of this duplication; they could share memories (at least for some forms of AI); and they could have powerful dedicated software (like our vision system) for domains where we have to rely on slow general reasoning.
Notes
- This chapter is about different kinds of superintelligent entities that could exist. I like to think about the closely related question, 'what kinds of better can intelligence be?' You can be a better baker if you can bake a cake faster, or bake more cakes, or bake better cakes. Similarly, a system can become more intelligent if it can do the same intelligent things faster, or if it does things that are qualitatively more intelligent. (Collective intelligence seems somewhat different, in that it appears to be a means to be faster or able to do better things, though it may have benefits in dimensions I'm not thinking of.) I think the chapter is getting at different ways intelligence can be better rather than 'forms' in general, which might vary on many other dimensions (e.g. emulation vs AI, goal directed vs. reflexive, nice vs. nasty).
- Some of the hardware and software advantages mentioned would be pretty transformative on their own. If you haven't before, consider taking a moment to think about what the world would be like if people could be cheaply and perfectly replicated, with their skills intact. Or if people could live arbitrarily long by replacing worn components.
- The main differences between increasing intelligence of a system via speed and via collectiveness seem to be: (1) the 'collective' route requires that you can break up the task into parallelizable subtasks, (2) it generally has larger costs from communication between those subparts, and (3) it can't produce a single unit as fast as a comparable 'speed-based' system. This suggests that anything a collective intelligence can do, a comparable speed intelligence can do at least as well. One counterexample to this I can think of is that often groups include people with a diversity of knowledge and approaches, and so the group can do a lot more productive thinking than a single person could. It seems wrong to count this as a virtue of collective intelligence in general however, since you could also have a single fast system with varied approaches at different times.
- For each task, we can think of curves for how performance increases as we increase intelligence in these different ways. For instance, take the task of finding a fact on the internet quickly. It seems to me that a person who ran at 10x speed would get the figure 10x faster. Ten times as many people working in parallel would do it only a bit faster than one, depending on the variance of their individual performance, and whether they found some clever way to complement each other. It's not obvious how to multiply qualitative intelligence by a particular factor, especially as there are different ways to improve the quality of a system. It also seems non-obvious to me how search speed would scale with a particular measure such as IQ.
- How much more intelligent do human systems get as we add more humans? I can't find much of an answer, but people have investigated the effect of things like team size, city size, and scientific collaboration on various measures of productivity.
- The things we might think of as collective intelligences - e.g. companies, governments, academic fields - seem notable to me for being slow-moving, relative to their components. If someone were to steal some chewing gum from Target, Target can respond in the sense that an employee can try to stop them. And this is no slower than an individual human acting to stop their chewing gum from being taken. However it also doesn't involve any extra problem-solving from the organization - to the extent that the organization's intelligence goes into the issue, it has to have already done the thinking ahead of time. Target was probably much smarter than an individual human about setting up the procedures and the incentives to have a person there ready to respond quickly and effectively, but that might have happened over months or years.
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.
- Produce improved measures of (substrate-independent) general intelligence. Build on the ideas of Legg, Yudkowsky, Goertzel, Hernandez-Orallo & Dowe, etc. Differentiate intelligence quality from speed.
- List some feasible but non-realized cognitive talents for humans, and explore what could be achieved if they were given to some humans.
- List and examine some types of problems better solved by a speed superintelligence than by a collective superintelligence, and vice versa. Also, what are the returns on “more brains applied to the problem” (collective intelligence) for various problems? If there were merely a huge number of human-level agents added to the economy, how much would it speed up economic growth, technological progress, or other relevant metrics? If there were a large number of researchers added to the field of AI, how would it change progress?
- How does intelligence quality improve performance on economically relevant tasks?
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 'intelligence explosion kinetics', a topic at the center of much contemporary debate over the arrival of machine intelligence. To prepare, read Chapter 4, The kinetics of an intelligence explosion (p62-77). The discussion will go live at 6pm Pacific time next Monday 20 October. Sign up to be notified here.
I agree not only with this sentence, but with this entire post. Which of the many, many degrees of freedom of a neuron, are "housekeeping" and don't contribute to "information management and processing" (quotes mine, not SteveG's) is far from obvious, and it seems likely to me that, even with a liberal allocation of the total degrees of freedom of a neuron to some sub-partitiioned equivalence class of "mere" (see following remarks for my reason for quotes) housekeeping, there are likely to be many, many remaining nodes in the directed graph of that neuron's phase space that participate in the instantiation and evolution of an informational state of the sort we are interested in (non-housekeeping).
And, this is not even to mention adjacent neuroglia, etc, that are in that neuron's total phase space, actively participating in the relevant (more than substrate-maintenance) set of causal loops -- as I argued in my post that WBE is not well-defined, a while back.
Back to what SteveG said about the currently unknown level of detail that matters (to the kind of information processing we are concerned with ... more later about this very, very important point); for now: we must not be too temporally-centric, i.e. thinking that the dynamically evolving information processing topology that a neuron makes relevant contributions to, is bounded, temporally, with a window beginning with: dendritic and membrane level "inputs" (receptor occupation, prevailing ionic environment, etc), and ending with: one depolarization -- exocytosis and/or the reuptake and clean-up shortly thereafter.
The gene expression-suppression and the protein turnover within that neuron should, arguably, also be thought of as part of the total information processing action of the cell... leaving this out is not describing the information processing act completely. Rather, it is arbitrarily cutting off our "observation" right before and after a particular depolarization and its immediate sequelae.
The internal modifications of genes and proteins that are going to effect future, information processing (no less than training of ANNs effects future behavior of of the ANN witin that ANNs information ecology) should be thought of, perhaps, as a persistent type of data structure itself. LTP of the whole ecology of the brain may occur on many levels beyond canonical synaptic remodeling.
We don't know yet which ones we can ignore -- e ven after agreeing on some others that are likely substrate maintenance only.
Another way of putting this or an entwined issue is: What are the temporal bounds of an information processing "act"? In a typical Harvard architecture substrate design, natural candidates would be, say, the time window of a changed PSW (processor status word), or PC pointer, etc.
But at a different level of description, it could be the updating of a Dynaset, a concluded SIMD instruction on a memory block representing a video frame, or anything in between.
It depends, ie, on both the "application" and aspects of platform archiceture.
I think it productive, at least, to stretch our horizons a bit (not least because of the time dilation of artificial systems relative to biological ones -- but again, this very statement itself has unexamined assumptions about the window -- spatial and temporal -- of a processed / processable information "packet" in both systems, bio and synthetic) and remain open about assumptions about what must be actively and isomorphically simulated, and what may be treated like "sparse brain" at any given moment.
I have more to say about this, but it fans out into several issues that I should put in multiple posts.
One collection of issues deals with: is "intelligence" a process (or processes) actively in play; is it a capacity to spawn effective, active processes; is it a state of being, like occurrently knowing occupying a subject's specious present, like one of Whitehead's "occasions of experience?"
Should we get right down to, and at last stop finessing around the elephant in the room: the question of whether consciousness is relevant to intelligence , and if so, when should we head-on start looking aggressively and rigorously at retiring the Turing Test, and supplanting it with one that enfolds consciousness and intelligence together, in their proper ratio? (This ratio is to be determined, of course, since we haven't even allowed ourselves to formally address the issue with both our eyes -- intelligenge and consciousness --open. Maybe looking through both issues, confers insight -- like depth vision, to push the metaphor of using two eyes. )
Look, if interested, for my post late tomorrow, Sunday, about the three types of information (at least) in the brain. I will title it as such, for anyone looking for it.
Personally, I think this week is the best thus far, in its parity with my own interests and ongoing research topics. Especially the 4 "For In-depth Ideas" points at the top, posted by Katja. All 4 are exactly what I am most interested in, and working most actively on. But of course that is just me; everyone will have their own favorites.
It is my personal agony (to be melodramatic about it) that I had some external distractions this week, so I am getting a late start on what might have been my best week.
But I will add what I can, Sunday evening (at least about the three types of information, and hopefully other posts. I will come back here even after the "kinetics" topic begins, so those persons in here who are interested in Katja's 4 In-depth issues, might wish to look back here later next week, as well as Sunday night or Monday morning, if you are interested in those issues as much as I am.
I am also an enthusiast for plumbing the depths of the quality idea, as well as, again, point number one on Katja's "In-depth Research" idea list for this week, which is essentially the issue of whether we can replace the Turing Test with -- now my own characterization follows, not Katja's, so "blame me" (or applaud if you agree) -- something much more satisfactory, with updated conceptual nuance representative of cognitive sciences and progressive AI as they are (esp the former) in 2015, not 1950.
By that I refer to theories, less preemptively suffocated by the legacy of logical positivism, which has been abandoned in the study of cognition and consciousness by mainstream cognitive science researchers; physicists doing competent research on consciousness; neuroscience and physics-literate philosophers; and even "hard-nosed" neurologists (both clinical and theoretical) who are doing down and detailed, bench level neuroscience.
As an aside, a brief look around confers the impression that some people on this web site still seem to think that being "critical thinkers" is somehow to be identified with holding (albeit perhaps semi-consciously) the scientific ontology of the 19th century, and subscribing to philosophy-of-science of the 1950's.
Here's the news, for those folks: the universe is made of information, not Rutherford-style atoms, or particles obeying Newtonian mechanics. Ask a physicist: naive realism is dead. So are many brands of hard "materialism" in philosophy and cognitive science.
Living in the 50's is not being "critical", is is being uninformed. Admitting that consciousness exists, and trying to ferret out its function, is not new-agey, it is realistic. Accepting reality is pretty much a necessary condition of being "less wrong."
And I think it ought to be one of the core tasks we never stray too far from, in our study of, and our pursuit of the creation of, HLAI (and above.)
Okay, late Saturday evening, and I was loosening my tie a bit... and, well, now I'll to get back to what contemporary bench-science neurologists have to say, to shock some of us (it surprised me) out of our default "obvious* paradigms, even our ideas about what the cortex does.
I'll try to post a link or two in the next day or two, to illustrate the latter. I recently read one by neurologists (research and clinical) who study children born en-cephalic (basically, just a spinal column and medulla, with an empty cavity full of CS fluid, in the rest of their cranium.) You won't believe what the team in this one paper presents, about consciousness in these kids. Large database of patients over years of study. And these neurologists are at the top of their game. It will have you rethinking some ideas we all thought were obvious, about what the cortex does. But let me introduce that paper properly, when I post the link, in a future message.
Before that, I want to talk about the three kinds of information in the brain -- maybe two, maybe 4, but important categorical differences (thermodynamic vs. semantic-referential, for starters), and what it means to those of us interested in minds and their platform-independent substrates, etc. I'll try to have something about that up, here, Sunday night sometime.
No, information ontology isn't a done deal.