SRG 4: Biological Cognition, BCIs, Organizations

[-]JoshuaFox11y90

What personal factors, if any, cause some people to tend towards one direction or another in some of these key prognostications?

For example, do economists tend more towards multiagent scenarios while computer scientists or ethicists tend more towards singleton prognostications?

Do neuroscientists tend more towards thinking that WBE will come first and AI folks more towards AGI, or the opposite?

Do professional technologists tend to have earlier timelines and others later timelines, or vice versa?

Do tendencies towards the political left or right influence s... (read more)

[-]AlexMennen11y80

Lifelong depression of intelligence due to iodine deficiency remains widespread in many impoverished inland areas of the world--an outrage given that the condition can be prevented by fortifying table salt at a cost of a few cents per person and year.
According to the World Health Organization in 2007, nearly 2 billion individuals have insufficient iodine intake. Severe iodine deficiency hinders neurological development and leads to cretinism, which involves an average loss of about 12.5 IQ points. The condition can be easily and inexpensively prevented th

... (read more)

[-][anonymous]11y160

As of July 30, GiveWell considers the International Council for the Control of Iodine Deficiency Disorders Global Network (ICCIDD) a contender for their 2014 recommendation, according to their ongoing review. They also mention that they're considering the Global Alliance for Improved Nutrition (GAIN), which they've had their eye on for a few years. They describe some remaining uncertainties -- this has been a major philanthropic success for the past couple decades, so why is there a funding gap now, well before the work is finished? Is it some sort of donor fatigue, or are the remaining countries that need iodization harder to work in, or is it something else?

(Also, average gains from intervention seem to be more like 3-4 IQ points.)

7CarlShulman11y

Part of their reason for funding deworming is also improvements in cognitive skills, for which the evidence base just got some boost.

[-]KatjaGrace11y70

Do you have a prefered explanation for the Flynn effect?

The Norwegian military conscripts above were part of a paper suggesting an interesting theory I hadn't heard before: that children are less intelligent as more are added to families, and so intelligence has risen as the size of families has shrunk.

3satt11y

My guess: different causes of the Flynn effect dominated at different times (and maybe in different places, too). For instance, Richard Lynn argued in 1990 that nutrition was the main explanation of the Flynn effect, but Flynn has recently counterargued that nutrition is unlikely to have contributed much since 1950 or so. Another example. Rick Nevin reckons decreasing lead exposure for children has made all the difference, but when I did my own back-of-the-envelope calculations using NHANES data for teenagers from the late 1970s to 2010, it looked like lead probably had a big impact between the late 1970s and early 1990s (maybe 5 IQ points, as average lead levels sank from ~10μg/dL to ~2μg/dL), but not since then, because blood lead concentrations had fallen so low that further improvement (down to ~1μg/dL) made little difference to IQ. Shrinking families would likely be a third factor along these lines, maybe kicking in hardest in mid-century, bolstering IQs after nutrition fell away as a key factor but before declining lead exposure made much difference. Edit, November 27: fixing the Richard Lynn paper link.

4KatjaGrace11y

And these different trends would tend to be consistently upwards rather than random, because we are consistently trying to improve such things? (Though the families one would probably still be random)

2satt11y

I'm not sure we do consistently try to improve these things. Nutrition, yes. But lead exposure got appreciably worse between WWI and 1970-1975, at least in the UK & US, and shrinking families is a manifestation of the demographic transition, which is only semi-intentional.

0TRIZ-Ingenieur11y

Reading Steve Wozniaks biography iWoz I support your view that parents nowadays focus more on education in youngest years. Steve learned about electronic components even before he was four years old. His father explained to him many things about electronics before he was old enough for school. He learned to read at the age of three. This needs parents who assist. Steve Wozniak praised his father for explaining always on a level he could understand. Only one step at a time. His exceptional high intelligence, he cited a test IQ > 200, is for sure not only inherited but consequence of loving care, teaching and challenging by his parents and peers.

3ChristianKl11y

There no good reason to think about this effect based on individual anecdotes. We do have controlled studies about the effects of parenting and it suggests that it doesn't matter much.

0TRIZ-Ingenieur11y

You are right. I needed some time for reading. The Flynn effect documents long term rise of fluid intelligence. Parenting and teaching are improving predominantly crystalline intelligence.

[-]lukeprog11y70

The linked Hsu paper is a convenient introduction to the genetics of intelligence and other quantitative traits like height, and therefore serves as handy background reading for understanding the biological cognition section.

[-]KatjaGrace11y60

'Let an ultraintelligent person be defined as a person who can far surpass all the intellectual activities of any other person however clever. Since the improvement of people is one of these intellectual activities, an ultraintelligent person could produce even better people; there would then unquestionably be an 'intelligence explosion,' and the intelligence of ordinary people would be left far behind. Thus the first ultraintelligent person is the last invention that people need ever make, provided that the person is docile enough to tell us how to keep them under control.'

Does this work?

2paulfchristiano11y

Looks good to me, with the same set of caveats as the original claim. Though note that both arguments are bolstered if "improvement of people" or "design of machines" in the second sentence is replaced by a more exhaustive inventory. Would be good to think more about the differences.

2KatjaGrace11y

What caveats are you thinking of?

2CarlShulman11y

This application highlights a problem in that definition, namely gains of specialization. Say you produced humans with superhuman general intelligence as measured by IQ tests, maybe the equivalent of 3 SD above von Neumann. Such a human still could not be an expert in each and every field of intellectual activity simultaneously due to time and storage constraints. The superhuman could perhaps master any given field better than any human given some time for study and practice, but could not so master all of them without really ridiculously superhuman prowess. This overkill requirement is somewhat like the way a rigorous Turing Test requires not only humanlike reasoning, but tremendous ability to tell a coherent fake story about biographical details, etc.

0Jeff_Alexander11y

For me, it "works" similarly to the original, but emphasizes (1) the underspecification of "far surpass", and (2) that the creation of a greater intelligence may require resources (intellectual or otherwise) beyond those of the proposed ultraintelligent person, the way an ultraintelligent wasp may qualify as far superior in all intellectual endeavors to a typical wasp yet still remain unable to invent and build a simple computing machine, nevermind constructing a greater intelligence.

[-]V_V11y50

Economic history suggests big changes are plausible.

Sure, but it is hard to predict what changes are going to happen and when.
In particular, major economic changes are typically precipitated by technological breakthroughs. It doesn't seem that we can predict these breakthroughs looking at the economy, since the causal relationship is mostly the other way.

AI progress is ongoing.

Ok.

AI progress is hard to predict, but AI experts tend to expect human-level AI in mid-century.

But AI experts have a notoriously poor track record at predicting human-leve... (read more)

2TRIZ-Ingenieur11y

Organizations can become much more superintelligent than they are. A team of humans plus better and better weak AI has no upper limit to intelligence. Such a hybrid superintelligent organization can be the way to keep AI development under control.

2V_V11y

In which case most of the "superintelligence" would come from the AI, not from the people.

-1TRIZ-Ingenieur11y

The synergistic union human+AI (master+servant) is more intelligent than AI alone which will have huge deficits in several intelligence domains. Human+AI has not a single sub-human level intelligence domain. I agree that the superintelligence part originates primarily from AI capabilities. Without human initiative, creativity and capability using mighty tools these superintelligent capabilities would not come into action.

1KatjaGrace11y

Do you think AI experts deserve their notoriety at predicting? The several public predictions that I know of prior to 1980 were indeed early (i.e. we have passed the time they predicted) but [Michie's survey] covers about ten times as many people and suggests that in the 70s, most CS researchers thought human-level AI would not arrive by 2014.

3V_V11y

I thought that the main result by Armstrong and Sotala was that most AI experts who made a public prediction, predicted human-level AI within 15 to 20 years in their future, regardless on when they made the prediction. Is this new data? Can you have some reference on how it was obtained?

4KatjaGrace11y

That was one main result, yes. It looks like Armstrong and Sotala counted the Michie survey as one 'prediction' (see their dataset here). They have only a small number of other early predictions, so it is easy for that to make a big difference. The image I linked is the dataset they used, with some modifications made by Paul Christiano and I (explained at more length here along with the new dataset for download). e.g. we took out duplications, and some things which seemed to have been sampled in a biased fashion (such that only early predictions would be recorded). We took out the Michie set altogether - our graph is now of public statements, not survey data.

[-]AlexMennen11y50

I'd like to propose another possible in-depth investigation: How efficiently can money and research be turned into faster development of biological cognitive enhancement techniques such as iterated embryo selection? My motivation for asking that question is that, since extreme biological cognitive enhancement could reduce existential risk and other problems by creating people smart enough to be able to solve them (assuming we last long enough for them to mature, of course), it might make sense to pursue it if it can be done efficiently. Given the scarcity ... (read more)

2ChristianKl11y

Smarter people can also come up with more dangerous ideas so it's not clear that existential risk get's lowered. I think there already plenty of money invested into that field. Agriculture wants to be able to effectively clone animals and insert new genes. Various scientists also want to be able to change genes of organisms effectively without having to wait years till your mouse get's children. Getting information information about what genes do largely depends on cheap sequencing and there a lot of money invested into getting more efficient gene sequencing.

2skeptical_lurker11y

It might be more a problem of public perception, because at the end of the day people have to be willing to use these technologies. Whichever group funds embryo selection will be denounced by many other groups, so it may be wise to find a source of funding that is difficult to criticise.

-1PhilGoetz11y

One problem is that for that approach, you would need, say, standardized IQ tests and genomes for a large number of people, and then to identify genome properties correlated with high IQ. First, all biologists everywhere are still obsessed with "one gene" answers. Even when they use big-data tools, they use them to come up with lists of genes, each of which they say has a measurable independent contribution to whatever it is they're studying. This is looking for your keys under the lamppost. The effect of one gene allele depends on what alleles of other genes are present. But try to find anything in the literature acknowledging that. (Admittedly we have probably evolved for high independence of genes, so that we can reproduce thru sex.) Second, as soon as you start identifying genome properties associated with IQ, you'll get accused of racism.

4CarlShulman11y

You can deal with epistasis using the techniques Hsu discusses and big datasets, and in any case additive variance terms account for most of the heritability even without doing that. There is much more about epistasis (and why it is of secondary importance for characterizing the variation) in the linked preprint.

2gwern11y

? I see mentions of stuff like dominance and interaction all the time; the reason people tend to ignore it in practice seems to be that the techniques which assume additive/independence work pretty well and explain a lot of the heritability. For example, height the other day: "Defining the role of common variation in the genomic and biological architecture of adult human height" Seems like an excellent start to me.

1PhilGoetz11y

It would be better than nothing. I am grinding one of my favorite axes more than I probably should. But those numbers make my case. My intuition says it would be hard to mine a few million SNPs, pick the most strongly associated 9500, and have them account for less than .29 of the variance, even if there were no relationship at all. And height is probably a very simple property, which may depend mainly on the intensity and duration of expression of a single growth program, minus interference from deficiencies or programs competing for resources.

6CarlShulman11y

"My intuition says it would be hard to mine a few million SNPs, pick the most strongly associated 9500, and have them account for less than .29 of the variance, even if there were no relationship at all." With sample sizes of thousands or low tens of thousands you'd get almost nothing. Going from 130k to 250k subjects took it from 0.13 to 0.29 (where the total contribution of all common additive effects is around 0.5). Most of the top 9500 are false positives (the top 697 are genome-wide significant and contribute most of the variance explained). Larger sample sizes let you overcome noise and correctly weight the alleles with actual effects. The approach looks set to explain everything you can get (and the bulk of heritability for height and IQ) without whole genome sequencing for rare variants just by scaling up another order of magnitude.

1ChristianKl11y

That's just a matter of time till genome sequencing get's cheap enough. There will be a day where it makes sense for China to sequence the DNA of every citizen for health purposes. China has also standardized test scores of it's population and no issues with racism that will prevent people from analysing the data.

[-]KatjaGrace11y40

Brain-computer interfaces for healthy people don't seem to help much, according to Bostrom. Can you think of BCIs that might plausibly exist before human-level machine intelligence, which you would expect to be substantially useful? (p46)

0Liso11y

This is also one of points where I dont agree with Bostrom's (fantastic!) book. We could use analogy from history: human-animal = soldier+hourse didnt need the physical iterface (like in Avatar movie) and still added awesome military advance. Something similar we could get from better weak AI tools. (probably with better GUI - but it is not only about GUI) "Tools" dont need to have big general intelligence. They could be at hourse level: * their incredible power of analyse big structure (big memory buffer) * speed of "rider" using quick "computation" with "tether" at your hands

1Liso11y

This probably needs more explanation. You could tell that my reaction is not in appropriate place. It is probably true. BCI we could define like physicaly interconnection between brain and computer. But I think in this moment we could (and have) analyse also trained "horses" with trained "raiders". And also trained "pairs" (or groups?) Better interface between computer and human could be done also in nonivasive path = better visual+sound+touch interface. (hourse-human analogy) So yes = I expect they could be substantially useful also in case that direct physical interace would too difficult in next decade(s).

[-]KatjaGrace11y40

How would you start to measure intelligence in non-human systems, such as groups of humans?

6gwern11y

One proposal goes that one measures predictive/reward-seeking ability on random small Turing machines: http://lesswrong.com/lw/42t/aixistyle_iq_tests/

2paulfchristiano11y

Below you ask whether the definition of intelligence per se is important at all; it seems it's not, and this may be some indication of how to measure what you actually care about.

2skeptical_lurker11y

Maybe a good starting point would be IQ tests?

1NxGenSentience11y

I am a little curious that the "seven kinds of intelligence" (give or take a few, in recent years) notion has not been mentioned much, if at all, even if just for completeness.... Has that been discredited by some body of argument or consensus, that I missed somewhere along the line, in the last few years? Particularly in many approaches to AI, which seem to view, almost a priori (I'll skip the italics and save them for emphasis) the approach of the day to be: work on (ostensibly) "component" features of intelligent agents as we conceive of them, or find them naturalistically. Thus, (i) machine "visual" object recognition (wavelength band... up for grabs, perhaps, for some items might be better identified by switching up or down the E.M. scale and visual intelligence was one of the proposed seven kinds; (ii) mathematical intelligence or mathematical (dare I say it) intuition; (iii) facility with linguistic tasks, comprehension, multiple language acquisition -- another of the proposed seven; (i.v) manual dexterity and mechanical ability and motor skill (as in athletics, surgery, maybe sculpture, carpentry or whatever) -- another proposed form of intelligence, and so on. (Aside, interesting that these alleged components span the spectrum of difficulty... are, that is, problems from both easy and harder domains, as has been gradually -- sometimes unexpectedly -- revealed by the school of hard knocks, during the decades of AI engineering attempts.) It seems that actors sympathetic to the top-down, "piecemeal" approach popular in much of the AI community would have jumped at this way of supplanting the ersatz "G" -- as it was called decades ago in early gropings in psychology and cogsci which sought a concept of IQ or living intelligence -- with, now, what many in cognitive science consider the more modern view and those in AI consider a more approachable engineering design strategy. Any reason we aren't debating this more than we are? Or did I miss it in one of the p

0SteveG11y

Bring these questions back up in later discussions!

0NxGenSentience11y

Will definitely do so. I can see several upcoming weeks when these questions will fit nicely, including perhaps the very next one. Regards....

-7almostvoid11y

0TRIZ-Ingenieur11y

Survival was and is the challenge of evolution. Higher intelligence gives more options to cope with deadly dangers. To measure intelligence we should challenge AI entities using standardized tests. To develop these tests will become a new field of research. IQ tests are not suitable because of their anthropocentrism. Tests should analyze capabilities how good and fast real world problems are solved.

[-]KatjaGrace11y40

If ten percent of the population used a technology that made their children 10 IQ points smarter, how strong do you think the pressure would be for others to take it up? (p43)

1skeptical_lurker11y

With diet, modafinil, etc this might already be the case. Sugar alone makes it more difficult to concentrate for many people, as well as having many other deleterious effects. Yet all many people do is say "you can have your chocolate, but only after you take your ritalin"

3paulfchristiano11y

I'm extremely skeptical of extracting even 1-2 IQ points (in expectation, after weighing up other performance costs) from these mechanisms. Changing diet is the most plausible, but for people whose diets aren't actively bad by widely-recognized criteria, it's not clear we know enough to make things much better. For the benefits of long-term stimulant use (or even the net long-term impacts of short-term stimulant use) I remain far from convinced. It seems true that more research on these topics could have large, positive expected effects, but these would accrue to society at large rather than to the researchers, and so would be in a different situaiton.

2skeptical_lurker11y

I seem to remember that eating enough fruit/vegetables alone raises your IQ by several points. But rather than IQ, stimulants affects focus and conscientiousness, which is just as important. You can still fail with an IQ of 150 if you can't sit down on focus on work. I would say the same is true of sugar. If you can spend more time focused on work, it might raise your IQ as a secondary effect, but this isn't necessary for a boost in effective intelligence.

2Lumifer11y

That seems highly unlikely. Links? Certain nutrient deficiencies in childhood can stunt development and curtail IQ (iodine is a classic example, that's why there is such a thing as iodized salt), but I don't think you're talking about that.

2skeptical_lurker11y

I'm not sure exactly where I read this, but here are some links with similarly impressive claims (albeit with the standard disclaimers about correlation not implying causation): http://healthland.time.com/2011/02/08/toddlers-junk-food-diet-may-lead-to-lower-iq It would help if they said what a 'unit' is. http://www.nootropics.com/vegetables/index.html

2Lumifer11y

These standard disclaimers are pretty meaningful here. The obvious question to ask of the first study is whether they controlled for the parents' IQ (or at least things like socio-economic status).

0skeptical_lurker11y

Indeed. But I don't have the time to read their papers (not that the article linked to the original paper), and its not my field anyway. From a practical viewpoint, good diet might give significant advantages (if not in IQ, then in other areas of health) and is extremely unlikely to cause any harm, so the expected cost-benefit analysis is very positive.

2Lumifer11y

Oh, that is certainly true. The only problem is that everyone has their own idea of what "good diet" means and these ideas do not match X-)

3skeptical_lurker11y

I think most people agree on vegetables, in fact this is one of the few things diets do agree on.

1kgalias11y

What do you mean?

1skeptical_lurker11y

I mean, if you are oscillating between sugar highs and crashes, it is difficult to concentrate, plus it causes diabetes etc..

4kgalias11y

Is this what you have in mind? wikipedia

5skeptical_lurker11y

No, I have this in mind: http://www.ncbi.nlm.nih.gov/pubmed/17224202

2kgalias11y

I don't have time to evaluate which view is less wrong. Still, I was somewhat surprised when I saw your first comment.

1skeptical_lurker11y

Upvoted for not wasting time!

[-]KatjaGrace11y30

If parents had strong embryo selection available to them, how would the world be different, other than via increased intelligence?

6CarlShulman11y

A lot of negative-sum selection for height perhaps. The genetic architecture is already known well enough for major embryo selection, and the rest is coming quickly. Height's contribution to CEO status is perhaps half of IQ's, and in addition to substantial effects on income it is also very helpful in the marriage market for men. But many of the benefits are likely positional, reflecting the social status gains of being taller than others in one's social environment, and there are physiological costs (as well as use of selective power that could be used on health, cognition, and other less positional goods). Choices at actual sperm banks suggests parents would use a mix that placed serious non-exclusive weight on each of height, attractiveness, health, education/intelligence, and anything contributing to professional success. Selection on personality might be for traits that improve individual success or for compatibility with parents, but I'm not sure about the net. Selection for similarity on political and religious orientation might come into use, and could have disturbing and important consequences.

3skeptical_lurker11y

Presumably many other traits would be selected for as well. Increasing intelligence has knee-jerk comparisons to eugenics & racism, so perhaps physical attractiveness/fitness would be selected for more strongly. Since personality traits are partially genetic, these may be selected for too. Homosexuality is partially genetic, so many gay rights movements would move to ban embryo selection (although some people would want bi kids because they don't want to deprive their kids of any options in life). Sexuality is correlated with other personality traits, so whatever choices are made will have knock-on effects. Some would advocate selecting against negative traits such as schizophrenia, ADHD and violence. Unfortunately, these traits are thought to provide an advantage in certain situations, or in combination with other genes, so we might also lose the creatively that comes with subclinical psychosis (poets are 20x more likely to go insane than average), the beneficial novel behaviour that comes with ADHD, and the ability to stand up for yourself (if aggression correlates with assertiveness). I know one should not generalise from fictional evidence, but it reminds me of the film 'demolition man' where society has evolved to a point where there is no violence, so when a murderer awakes from cryonic suspension they cannot defend themselves. Far better film than Gattaca.

2paulfchristiano11y

If ture, this would be somewhat surprising from a certain angle. As if saying "selecting for what's on the inside is too superficial and prejudiced, so we should be sure our selection is only skin-deep." I would bet against selection for things like sexual orientation or domesticity, and in favor of selection for general correlates of good health and successful life outcomes (which may in turn come along with other unintended characteristics).

3skeptical_lurker11y

This is, admittedly, a bizarre state of affairs. But if we were to admit that IQ is meaningful, and could be affected by genes, then this gives credence to the 'race realists'! But we can't concede a single argument to the hated enemy, therefore intelligence is independent of genes. QED. Attractiveness OTOH is obviously genetic, because people look like their parents. I concur. I would however bet in favour of a large argument over sexual orientation.

1diegocaleiro11y

Yvain has a biodeterministic guide to parenting. Some people would do the same things: http://squid314.livejournal.com/346391.html

1okay11y

Gattaca, except everyone is actually superhuman and nobody cares about whether you'll have a heart attack at thirty except your doctor.

[-]KatjaGrace11y30

Ambiguities around 'intelligence' often complicate discussions about superintelligence, so it seems good to think about them a little.

Some common concerns: is 'intelligence' really a thing? Can intelligence be measured meaningfully as a single dimension? Is intelligence the kind of thing that can characterize a wide variety of systems, or is it only well-defined for things that are much like humans? (Kruel's interviewees bring up these points several times)

What do we have to assume about intelligence to accept Bostrom's arguments? For instance, does the cl... (read more)

4RomeoStevens11y

Is intelligence really a single dimension? Related: Do we see a strong clustering of strategies that work across all the domains we have encountered so far? I see the answer to original question being yes if there is just one large cluster, and no if it turns out there are many fairly orthogonal clusters. Is robustness against corner cases (idiosyncratic domains) a very important parameter? We certainly treat it as such in our construction of least convenient worlds to break decision theories.

2PhilGoetz11y

There is a small set of operations (dimension reduction, 2-class categorization, n-class categorization, prediction) and algorithms for them (PCA, SVM, k-means, regression) that work well on a wide variety of domains. Does that help?

4Punoxysm11y

Not that wide a variety of domains, compared to all human tasks. Specifically, they can only handle data that comes in matrix form, and often only after it has been cleaned up and processed by a human being. Consider, just the iris dataset: if instead of the measurements of the flowers you were working with photographs of the flowers, you might have made your problem substantially harder, since now you have a vision task not amenable to the algorithms you list.

2PhilGoetz11y

Can you give an example of data that doesn't come in matrix form? If you have a set of neurons and a set of connections between them, that's a matrix. If you have asynchronous signals travelling between those neurons, that's a time series of matrices. If it ain't in a matrix, it ain't data. [ADDED: This was a silly thing for me to say, but most big data problems use matrices.]

5Punoxysm11y

The answer you just wrote could be characterized as a matrix of vocabulary words and index-of-occurrence. But that's a pretty poor way to characterize it for almost all natural language processing techniques. First of all, something like PCA and the other methods you listed won't work on a ton of things that could be shoehorned into matrix format. Taking an image or piece of audio and representing it using raw pixel or waveform data is horrible for most machine learning algorithms. Instead, you want to heavily transform it before you consider putting it into something like PCA. A different problem goes for the matrix of neuronal connections in the brain: it's too large-scale, too sparse, and too heterogenous to be usefully analyzed by anything but specialized methods with a lot of preprocessing and domain knowledge going into them. You might be able to cluster different functional units of the brain, but as you tried to get to more granular units, heterogeneity in number of connnections per neuron would cause dense clusters to "absorb" sparser but legitimate clusters in almost all clustering methods. Working with a time-series of activations is an even bigger problem, since you want to isolate specific cascades of activations that correspond to a stimulus, and then look at the architecture of the activated part of the brain, characterize it, and then be able to understand things like which neurons are functionally equivalent but correspond to different parallel units in the brain (left eye vs. right eye). If I give you a time series of neuronal activations and connections with no indication of the domain, you'd probably be able to come up with a somewhat predictive model using non-domain-specific methods, but you'd be handicapping yourself horribly. Inferring causality is another problem - none of these predictive machine learning methods do a good job of establishing whether two factors have a causal relation, merely whether they have a predictive one (within

0PhilGoetz11y

First, yes, I overgeneralized. Matrices don't represent natural language and logic well. But, the kinds of problems you're talking about--music analysis, picture analysis, and anything you eventually want to put into PCA--are perfect for matrix methods. It's popular to start music and picture analysis with a discrete Fourier transform, which is a matrix operation. Or you use MPEG, which is all matrices. Or you construct feature detectors, say edge detectors or contrast detectors, using simple neural networks such as those found in primary visual cortex, and you implement them with matrices. Then you pass those into higher-order feature detectors, which also use matrices. You may break information out of the matrices and process it logically further downstream, but that will be downstream of PCA. As a general rule, PCA is used only on data that has so far existed only in matrices. Things that need to be broken out are not homogenous enough, or too structured, to use PCA on. There's an excellent book called Neural Engineering by Chris Eliasmith in which he develops a matrix-based programming language that is supposed to perform calculations the way that the brain does. It has many examples of how to tackle "intelligent" problems with only matrices.

0TRIZ-Ingenieur11y

lukeprog linked above the Hsu paper that documents good correlation between different narrow human intelligence measurements. The author concludes that a general g factor is sufficient. All humans have more or less the same cognitive hardware. The human brain is prestructured that specific areas normally have assigned specific functionality. In case of a lesion other parts of the brain can take over. If a brain is especially capable this covers all cranial regions. A single dimension measure for humans might suffice. If a CPU has a higher clock frequency rating than another CPU of the identical series: the clock factor is the speedup factor for any CPU-centric algorithm. An AI with NN pattern matching architecture will be similar slow and unreliable in mental arithmetics like us humans. Extend its architecture with a floating point coprocessor and its arithmetic capabilities will rise by magnitudes. If you challenge an AI that is superintelligent in engineering but has low performance regarding this challenging requirement it will design a coprocessor for this task. Such coprocessors exist already: FPGA. Programming is highly complex but speedups of magnitudes reward all efforts. Once a coprocessor hardware configuration is in the world it can be shared and further improved by other engineering AIs. To monitor AI intelligence development of extremly heterogeneous and dynamic architectures we need high dimensional intelligence metrics.

1diegocaleiro11y

We have to assume only that we will not significantly improve our understanding of what intelligence is without attempting to create it (through reverse engineering, coding, or EMs). If our understanding remains incipient the safe policy is to assume that indeed intelligence is a capacity, or set of capacities that can be used to bootstrap itself. Given the 10¨52 lives at stake, even if we were fairly confident intelligence cannot bootstrap, we should still MaxiPok and act as if it was.

1skeptical_lurker11y

I disagree. By analogy, understanding of agriculture has increased greatly without the creation of an artificial photosynthetic cell. And yes, I know that photovoltic panels exist, but only a long time later.

2diegocaleiro11y

Do you mind spelling out the analogy? (including where it breaks) I didn't get it. Reading my comment I feel compelled to clarify what I meant: Katja asked: in which worlds should we worry about what 'intelligence' designates not being what we think it does? I responded: in all the worlds where increasing our understanding of 'intelligence' has the side effect of increasing attempts to create it - due to feasibility, curiosity, or an urge for power. In these worlds, expanding our knowledge increases the expected risk, because of the side effects. Whether intelligence is or not what we thought will only be found after the expected risk increased, then we find out the fact, and the risk either skyrockets or plummets. In hindsight, if it plummets, having learned more would look great. In hindsight, if it skyrockets, we are likely dead.

-2SteveG11y

Single-metric versions of intelligence are going the way of the dinosaur. In practical contexts, it's much better to test for a bunch of specific skills and aptitudes and to create a predictive model of success at the desired task. In addition, our understanding of intelligence frequently gives a high score to someone capable of making terrible decisions or someone reasoning brilliantly from a set of desperately flawed first principles.

1KatjaGrace11y

Ok, does this matter for Bostrom's arguments?

2SteveG11y

Yeah, having high math or reading comprehension capability does not always make people more effective or productive. They can still, for instance, become suidical, sociopathic or rebel against well-meaning authorities. They still often do not go into their doctor when sick, they develop addictions, they may become too introverted or arrogant when it is counterproductive or fail to escape bad relationships. We should not strictly be looking to enhance intelligence. If we're going down the enhancement route at all, we should wish to create good decision-makers without, for example, tendencies to mis-read people, sociopathy and self-harm.

-1Lumifer11y

What's wrong with that? ...and, presumably, without tendencies to rebel against well-meaning authorities? I don't think I like the idea of genetic slavery.

0SteveG11y

For instance, rebelling against well-meaning authorities has been known to cause someone not to adhere to a correct medication regime or to start smoking. Problems regularly rear their head when it comes to listening to the doctor. I guess I'll add that the well-meaning authority is also knowledgeable.

1Lumifer11y

Let me point out the obvious: the knowledgeable well-meaning authority is not necessarily acting in your best interests. Not to mention that authority that's both knowledgeable and well-meaning is pretty rare.

1SteveG11y

Really, what I am getting at is that just like anyone else, smart people may rebel or conform as a knee-jerk reaction. Neither is using reason to come to an appropriate conclusion, but I have seen them do it all the time.

1KatjaGrace11y

One might think an agent who was sufficiently smart would at some point apply reason to the question of whether they should follow their knee-jerk responses with respect to e.g. these decisions.

0NxGenSentience11y

I thought that this had become a fairly dominant view, over 20 years ago. See this PDF: http://www.learner.org/courses/learningclassroom/support/04_mult_intel.pdf I first read the book in the early nineties, though Howard Gardner had published the first edition in 1982. I was at first a bit extra skeptical that it would be based too much on some form of "political correctness", but I found the concepts to be very compelling. Most of the discussion I heard in subsequent years, occasionally by psychology professor and grad student friends, continued to be positive. I might say that I had no ulterior motive in trying to find reasons to agree with the book, since I always score in the genius range myself on standardized, traditional-style IQ tests. So, it does seem to me that intelligence is a vector, not a scalar, if we have to call it by one noun. As to Katja's follow-up question, does it matter for Bostrom's arguments? Not really, as long as one is clear (which it is from the contexts of his remarks) which kind(s) of intelligence he is referring to. I think there is a more serious vacuum in our understanding, than whether intelligence is a single property, or comes in several irreducibly different (possibly context-dependent) forms, and that is this : with respect to the sorts of intelligence we usually default to conversing about (like the sort that helps a reader understand Bostrom's book, an explanation of special relativity, or RNA interference in molecular biology), do we even know what we think we know about what that is. I would have to explain the idea of this purported "vacuum" in understanding at significant length; it is a set of new ideas that stuck me, together, as a set of related insights. I am working on a paper explaining the new perspective I think I have found, and why it might open up some new important questions and strategies for AGI. When it is finished and clear enough to be useful, I will make it available by PDF or on a blog. (Too len

[-]PhilGoetz11y30

What are the trends in those things that make groups of humans smarter? e.g. How will world capacity for information communication change over the coming decades? (Hilbert and Lopez's work is probably relevant)

A social / economic / political system is not just analogous to, but is, an artificial intelligence. Its purpose is to sense the environment and use that information to choose actions that further its goals. The best way to make groups of humans smarter would be to consciously apply what we've learned from artificial intelligence to human organiza... (read more)

-8almostvoid11y

[-]KatjaGrace11y30

Have we missed any plausible routes to superintelligence? (p50)

7PhilGoetz11y

Unexpected advances in physics lead to super-exponential increases in computing power (such as were expected from quantum computing), allowing brute-force algorithms or a simulated ecosystem to achieve super-intelligence. Real-time implanted or worn sensors plus genomics, physiology simulation, and massive on-line collaboration enables people to identify the self-improvement techniques that are useful to them. Someone uses ancient DNA to make a Neanderthal, and it turns out they died out because their superhuman intelligence was too metabolically expensive. Reading all of LessWrong in a single sitting.

6paulfchristiano11y

It seems like the characterization of outcomes into distinct "routes" is likely to be fraught; even if such a breakdown was in some sense exhaustive I would not be surprised if actual developments didn't really fit into the proposed framework. For example, there is a complicated and hard to divide space between AI, improved tools, brain-computer interfaces, better institutions, and better training. I expect that ex post the whole thing will look like a bit of a mess. One practical result is that even if there is a strong deductive argument for X given "We achieve superintelligence along route Y" for every particular Y in Bostrom's list, I would not take this as particularly strong evidence for X. I would instead view it as having considered a few random scenarios where X is true, and weighing this up alongside other kinds of evidence.

-6diegocaleiro11y

[-]KatjaGrace11y30

If a technology existed that could make your children 10 IQ points smarter, how willing do you think people would be to use it? (p42-3)

9lukeprog11y

Shulman & Bostrom (2014) make a nice point about this: As table 2 in the paper shows, the American public generally opposed IVF until the first IVF baby was born, and then they were in favor of it. As of 2004, only 28% of Americans approve of embryo selection for improving strength or intelligence, but that could change rapidly when the technology is available.

[-]gwern11y100

On the other hand, we could point to Down syndrome eugenics: while it's true that Down's has fallen a lot in America thanks to selective abortion, it's also true that Down's has not disappeared and the details make me pessimistic about any widespread use in America of embryo selection for relatively modest gains.

An interesting paper: "Decision Making Following a Prenatal Diagnosis of Down Syndrome: An Integrative Review", Choi et al 2012 (excerpts). To summarize:

many people are, on principle, unwilling to abort based on a Down's diagnosis, and so simply do not get the test
the people who do abort tend to be motivated to do so out of fear: fear that a Down's child will be too demanding and wreck their life.
Not out of concern for the child's reduced quality of life, because Down's syndrome is extremely expensive to society, because sufferers go senile in their 40s, because they're depriving a healthy child of the chance to live etc - but personal selfishness.

Add onto this:

testing for Down's is relatively simple and easy
most people see and endorse a strong asymmetry between 'healing the sick' and 'improving the healthy'
You can see this in the citation in Shulman &a

... (read more)

3KatjaGrace11y

An interesting datapoint, thanks. One big difference in favor of selection for intelligence relative to testing for Down syndrome is that at the point where people don't get a Down syndrome test, they have a fairly low probability of their child having the disease (something like 1/1000 while they are youngish), whereas selection for intelligence is likely to increase intelligence.

7paulfchristiano11y

28% is a pretty large number. I expect that in more abstract framings such as "improving general well-being" you would see larger rates of approval already, and marketing would push technologies towards framings that people liked.

1PhilGoetz11y

I'm used to Robin Hanson presenting near / far mode dichotomies as "near mode greedy and stupid, far mode rational". But perhaps far mode allows the slow machinery of reason to be brought to bear, and most people's reasoning about IVF and embryo selection is victim to irrational ideas about ethics. In such cases, near mode (IVF is now possible) could produce more "rational" decisions because it bypasses rationality, while the reasoning that would be done in far mode has faulty premises and performs worse than random.

3KatjaGrace11y

Interesting, I always interpreted Robin as casting near in a positive light (realistic, sensible) and far more negatively (self-aggrandizing and delusional).

1PhilGoetz11y

People in far mode say they will exercise more, eat better, get a new job, watch documentaries instead of Game of Thrones, read classic literature, etc., and we could call those far-sighted plans "rational". Near mode gives in to inertia and laziness.

9Lumifer11y

We could, but we really should call these plans lies for they intend to deceive -- either oneself to gain near-term contentment, or others to gain social status.

[-]KatjaGrace11y30

Did you change your mind about anything as a result of this week's reading?

6selylindi11y

Remember that effect where you read a newspaper and mostly trust what it says, at least until one of the stories is about a subject you have expertise in, and then you notice that it's completely full of errors? It makes it very difficult to trust the newspaper on any subject after that. I started Bostrom's book very skeptical of how well he would be handling the material, since it covers many different fields of expertise that he cannot hope to have mastered. My personal field of expertise is BCI. I did my doctoral work in that field, 2006-2011. I endorse every word that Bostrom wrote on BCI in the book. And consequently, in the opposite of the newspaper effect, I dramatically raised my confidence that Bostrom has accurately characterized the subjects I'm more ignorant of.

6Sebastian_Hagen11y

How close we are to making genetic enhancement work in a big way. I'm not fully convinced of the magnitudes of gains from iterated embryo selection as projected by Bostrom; but even being able to drag the average level of genetically-determined intelligence up close to the current maxima of the distribution would be immensely helpful, and it's informative that Bostrom suggests we'll have the means within a few decades.

-6almostvoid11y

[-]KatjaGrace11y20

If parents could choose their preferred embryos from a large number, with good knowledge of the characteristics the children would have, how much of this selection power do you think those who used this power would spend on intelligence? (p39)

2diegocaleiro11y

The first few generations would test distinct traits. However states would have a strong incentive to increase intelligence. Diligence or conscientiousness may be even more important, since it gets more things done.

0TRIZ-Ingenieur11y

Some parents will use a selection option as soon as any is available. To get a child genius like Einstein is attractive to these strange people. First 'selections' will be based on semi-knowledge. Descending childs will suffer from many unintended side effects. Unless we do not fully understand how intelligence and 'hard wired' cognitive features are coded in DNA and epigenetic activation patterns we should not start any selection. Many countries have laws prohibiting embryo selection but this barrier seems to weaken recently. We should not tear it down unless we are precisely knowing what we are going to do.

[-]selylindi11y10

This chapter seems like the right place to add this to the conversation. Near the end of the book, Bostrom suggests that a lot of work should be put into generating crucial considerations for superintelligent AI. I've made a draft list of some crucial considerations, but it's definitely the sort of thing that should grow and change as other people make their own versions of it. Biological superintelligences didn't really make my list at all yet.

[-]KatjaGrace11y10

Bostrom says it is probably infeasible to 'download' large chunks of data from one brain to another, because brains are idiosyncratically formatted and meaning is likely spread holistically through patterns in a large number of neurons (p46). Do you agree? Do you think this puts such technology out of reach until after human-level machine intelligence?

2PhilGoetz11y

It may be possible to decode the encoding used in different brains. Representations in particular sensory modalities are localized to a few square inches of cortex.

0nkh11y

I think it's probably enough of an obstacle that it's more likely an AGI will be developed first. In that sense I do agree with Bostrom. However, I wouldn't say it's completely infeasible, rather that it will require considerable advances in pattern recognition technology, our understanding of the brain, and our technological ability to interface with the brain first. The idiosyncratic morphology and distributed/non-localized information storage make for a very difficult engineering problem, but I'm optimistic that it can be overcome in some way or another. We've already had some (granted, very limited) success with decoding imagery from the visual cortex through "dumb" (non-AGI) machine learning algorithms, which makes deeper interaction seem at least possible. If we can make advances in the above-mentioned fields, I would guess the biggest limitation will be that we'll never have a standardized "plug'n'play" protocol for brains--interfaces will require specialized tuning for each individual and a learning period during which the algorithms can "figure out" how your brain is wired up.

[-]KatjaGrace11y10

Even if there were an easy way of pumping more information into our brains, the extra data inflow would do little to increase the rate at which we think and learn unless all the neural machinery necessary for making sense of the data were similarly upgraded. (p45-6)

This seems far from obvious to me. Firstly, why suppose that making sense of the data is such a bottleneck? And then even if making sense is a bottleneck, if the data is in a different form it might be easier to make sense of.

Intuitively, things that are already inside one's head are much ea... (read more)

5skeptical_lurker11y

The visual cortex can handle huge amounts of data input, but the amount of data one can output by typing/writing/drawing is orders of magnitude lower, suggesting that data output is the lower-hanging fruit for BCI.

5KatjaGrace11y

Two datapoints: I find reading and typing faster to both be very useful, suggesting my brain is not bottlenecked in its capacity to understand things at my natural rate of reading or typing.

3NxGenSentience11y

I'll have to weigh in wiith Botrom on this one, though I think it depends a lot on the individual brain-mind, i.e., how your particular personality crunches the data. Some people are "information consumers", others are "information producers". I think Einstein might have used the obvious terms supercritical vs subcritical minds at some point -- terms that in any case (einstein or not) naturally occurred to me (and probably lots of people) and I've used since teenager years, just in talking to my friends, to describe different people's mental processes. The issue of course is (a) to what extent you use incoming ideas as "data" to spark new trains of thought, plus (b) how many interconnections you notice between various ideas and theories -- and as a multiplier of (b), how abstract these resonances and interconnections are (hugely increasing the perceived potential interconnection space.) For me, if the world would stop in place, and I had an arbitrary lifespan, I could easily spend the next 50 years (at least) mining the material I have already acquired, generating new ideas, extensions, cross connections. (I sometimes almost wish it would, in some parallel world, so I could properly metabolize what I have, which I think at times I am only scratching the surface of.) Of course it depends on the kind of material, as well. If one is reading an undergrad physics textbook in college, it is pretty much finite: if you understand the presentation and the development as you read, you can think for an extra 10 or 15 minutes about all the way it applies to the world, and pretty much have it. Thinking of further "applications" pretty much add no value, additional insight, or interest. But with other material, esp in fields that are divergent and full of questions that are not settled yet, I find myself reading a few paragraphs, and it sparks so many new trains of thought, I feel flooded and have a hard time continuing the reading -- and feel like I have to get up and go wa

[-]KatjaGrace11y10

Bostrom is pessimistic about brain-computer interfaces. (p48) Do you agree with his arguments?

2diegocaleiro11y

These arguments have been made by quite a few authors over the years at the Edge question. The most robust prediction is that we won't defeat evolution in energy efficiency to absorb content. We will not create machine interfaces so good that they beat our sense organs.

0TRIZ-Ingenieur11y

Ramez Naam wrote two Sci-Fi novels on that issue: Nexus and Crux. Drinking a silvery liquid of communicative nanobots is enough. The bots autonomously find their way into the brain and connect to neural cells. Regarding non-invasiveness this vision might get acceptance. Unsolved technical issues are: building such nano bots, supplying energy and dynamic long range communication. I fully agree with Bostrums scepticism. Low social acceptance of Google glass shows that humans with brain interconnection might face similar repercussions. Via BCI you could 'videotape' what you see with your eyes. A BCI will probably not be so easy deactivated than switching off your Google glass.

[-]KatjaGrace11y10

Do you think the consequences listed in Table 6 (Possible impacts from genetic selection in different scenarios) are accurate? (p40) What does 'posthumanity' look like? What other consequences might you expect in these scenarios?

[-]KatjaGrace11y10

Do you have further interesting pointers to material relating to this week’s reading?

[-]KatjaGrace11y10

What did you find most interesting in this week's reading?

3okay11y

Iterated embryo selection was pretty interesting. I wonder if there is anything viable about inserting new / activating the growth of neurons / synapses into the human brain, particularly into specifically targeted areas, like the section(s) where people do math.

[-]KatjaGrace11y10

What did you find least persuasive in this week's reading?

3SteveG11y

I am a big Bostrom fan, but I am not sure why brain-computer interfaces were downgraded so much in this chapter. The text seems to overestimate the risks of brain implants today. Given another twenty or more years to work out the issues, these risks will fall even more. It also seems to underestimate the benefits of some rather boring upgrades. Having a direct link to several ordinary software tools like MS Excel would make us a lot more intelligent. We would also gain a lot of ability by being able to directly control machinery. Just having the equivalent of the Notepad application in there would permit me to ace many tests, remember people's names and stop forgetting where I parked my car and how many calories I've eaten today.

4PhilGoetz11y

I second this. Just being able to remember what I've read would amplify my intelligence by at least one order of magnitude. I appreciate the argument that a brain-computer interface wouldn't give you much beyond what you'd get by sitting down at a computer, but (A) being able to google with my cell phone made me significantly smarter than only being able to google at my computer, (B) being forced to state questions in formal language would GREATLY clarify peoples' thinking, (C) expanding my short-term memory store might greatly enhance my intelligence, and (D) if the BCI is able to use pattern-recognition on my memory's current contents versus the entire knowledge of humanity, pointing out analogies to systems described in books I haven't even read, that would be tremendously useful.

1Sebastian_Hagen11y

An easier way to deal with that is by improving VR to the point that people can spend virtually their entire waking lives at (probably stationary) computers with high-end I/O devices. Interface mobility is only an advantage if /physically moving around/ is worth doing, and we can probably remove a lot of the draw of that a lot easier than we can make BCIs work well. How are BCIs a major help with this? Re. (C) and (D): Agreed, but: 1. You can already get a non-crappy approximation to (C) at a computer, for instance by keeping open a window with some facts you're trying to keep in mind. 2. If I understand Bostrom correctly, his contention is that going much beyond this level of convenience with BCIs would be hard; you'd need to do some very tricky interfacing (since it isn't a usual I/O channel), and the tech to pull that off is likely to be AI-complete or close to it, itself.

2paulfchristiano11y

This seems right, but nevertheless the gains are relatively small compared to bread-and-butter improvements in the design of tools like spreadsheets. The overhead for doing any of these is rather small at the moment (perhaps 30s a day each?) and you only don't it because respectively (1) it's disallowed because that's the point of tests, (2) it's bad signaling, which largely defeats the point of remembering names, (3) the benefits are very small and/or you are unaware of how cheap it is for normal use cases (4) the benefits are very small and the main difficulties are measurement issues. I think it's not a coincidence that none of these are very important to your economic productivity (though I understand that this may in part just be because you wanted to choose generalizable examples).

1SteveG11y

Sorry, Paul, but Excel gives the ability to able to remember millions of arbitrary facts and make vast arbitrary calculations without putting pen to paper. It's clearly economically beneficial and, if used properly, is probably enough to ace any standardized test.

1Lumifer11y

I don't think so. Take someone stupid, give him a laptop with Excel full of whatever data she wants to put into it, and let her take a standardized test with more relaxed timing (to account for searching in that spreadsheet). I don't think she'll ace the test, in particular things like reading comprehension or logical puzzles would not be made easier by having large tables full of data available.

0SteveG11y

Always good to have skeptics to stretch your creativity! So, the counterfactual as it stands right now is that we're giving somebody additional mental powers through high-speed access to software hooked directly into the brain. We're not assuming this technology includes advanced AI that does not presently exist. We're sticking for the most part with software that we have now, but we would be safe to give x1000 of existing hardware capabilities. We could give them an internet connection, but let's say that's cheating. We will not allow them to utilize anybody else's genius or just any available database. For now let's say that they can download large, structured data sets which others have built into their brain-interfaced computers in advance, but they cannot access outside sources in real-time. OK, so someone like this is going to study for a test. They can study in an ordinary fashion, but we can also build dozens of spreadsheets for them to use during the process. First of all, any time any question relies on vocabulary, they are going to have that piece in place. They will have a definition of every word or unusual phrase at their immediate beckon call. That solves a lot of reading comprehension problems, but maybe not all of them. What about those questions where there is a passage to read and, for instance, she has to discern something about the author's intentions? Here, we get to give whatever kind of custom solver we might choose to provide. For example, we can give her an ability to accumulate a score all of the emotional words in the passage. It's a standardized test, not a general test of problem-solving ability. Therefore, she gets to include a lot of previous test questions and templates for answers in her data. How much of an advantage will this provide? She is never going to make an error in arithmetic or algebra, and she will be able to perform these functions very rapidly. She gets to immediately convert different kinds of units, one to

0Lumifer11y

Um, no. That's part of the issue -- we're not giving her access to additional mental powers. We're giving her easy, fast, and convenient access to some information tools. Her mental powers remain the same -- if her working memory is limited, it remains limited. Being able to look up things in a second does not imply a large working memory. If she gets confused with longish logical chains, direct access to Excel isn't going to help. Etc., etc. Oh, but she will. Go talk to, say, accountants -- people who professionally use Excel and have been doing it for a while. Ask them if they ever make an arithmetic error :-) Well, Excel includes VB which is Turing-complete. So you could treat Excel as a general-purpose computing environment and provide her with an narrow AI which, basically, solves the test for her. But I don't think that's what we are talking about :-/

0TRIZ-Ingenieur11y

Embryo selection for intelligence might easily lead to autism and other psychological defects. A single failed trial will cost millions if life long full care has to be provided for this poor human being.

[-]NxGenSentience11y00

Intra-individual neuroplasticity and IQ - Something we can do for ourselves (and those we care about) right now

Sorry to get this one in at the last minute, but better late than..., and some of you will see this.

Many will be familiar with the Harvard psychiatrist, neuroscience researcher, and professor of medicine, John Ratey, MD., from seeing his NYT bestselling books in recent years. He excels at writing for the intelligent lay audience, yet not dumbing down his books to the point where they are useless to those of us who read above the laymans' level in... (read more)

[+]almostvoid11y-130

LESSWRONG
LW

LESSWRONG
LW

14

SRG 4: Biological Cognition, BCIs, Organizations

14

14

Summary

Notes

In-depth investigations

How to proceed