What this is not: an exhortation to dig up Von Neumann and clone him

What this is: an exhortation to think seriously about how to decide whether to dig up Von Neumann and clone him

Having tried to clarify that I am not presently advocating for cloning, and am far more concerned with the meta level than the object level, let me now sink to the object level.

The argument for cloning

Suppose everything Eliezer Yudkowsky said in his recent discussions on Artificial General Intelligence (AGI) is entirely literally true, and we take it seriously.

Then right now we are a few decades (or perhaps years) away from the development of an unaligned AGI, which will swiftly conclude that humans are an unnecessary risk for whatever it wants.

The odds are not great, and it will take a miracle scientific breakthrough to save us. The best we can do is push as hard as we can to increase the possibility of that miracle breakthrough and our ability to make use of.

The claim I really want to draw attention to is this particular chestnut:

Paul Christiano is trying to have real foundational ideas, and they're all wrong, but he's one of the few people trying to have foundational ideas at all; if we had another 10 of him, something might go right.

Well, why not? 10 of him, 10 Von Neumanns, and an extra Jaynes or two just in case. If we have a few decades, we have just enough time.

Leave aside all the obvious ways this could go horribly wrong, the risks and the costs: if what Eleizer says is true, it is conceivable that cloning could save the world.

The argument for the argument about clones

I am not saying this is a good idea, but I am saying it could be a good idea. And if it is a good idea, it would be really great to know that.

There are obvious problems, but I don't feel well-placed to judge how high the benefits might be, or how easily the problems might be resolved. And I am sufficiently uncertain in this regard that I think it might be valuable to become more certain.

A modest proposal?

I think OpenPhil, or somebody else in that sphere, should fund or conduct an investigation into cloning. Most basically, how much would it actually cost to clone somebody, and how much impact could that potentially have?

It may become immediately clear that this is prohibitively expensive relative to just hiring more of the very smartest people in the world, but if the limiting factor is not money then I think it would then become worthwhile to ask: what actually are the ethical ramifications of cloning? could it be done?

I would like to think that sufficiently neutral and exploratory research could be done in public, but it is conceivable that this must be conducted in secret.

As I see it, the most obvious types of arguments against this are:

  • somehow it is just obviously the case that, no matter what, this couldn't conceivably turn out to be high impact (which I find quite unlikely)
  • the opportunity cost of even investigating this is too high (which would surprise me but I would begrudgingly accept)
  • such a study already exists, in secret (in which case I apologise)
  • such a study already exists, in public (in which case I apologise even more, and would love to be directed to it)

New to LessWrong?

New Comment
45 comments, sorted by Click to highlight new comments since: Today at 9:29 AM

Would the clones be at least as smart as the original?   On the yes side we have the Flynn Effect, which is the steady increase over time of IQ scores.  If this represents a real increase in intelligence (as opposed to just measurement error) this means we would expect a modern clone of someone born in 1903 to be smarter than the original perhaps because a reduced pathogenic load is making nearly everyone more intelligent.  

On the no side, the more exceptional someone is, the more regression to the mean we should expect from the clones.  If your intelligence is some combination of genes and luck then the smartest person ever should have had an enormous amount of luck (say in terms of how random noise influenced brain development) as well as exceptional genes.  The clones, on average, will  have much less brain development luck than the original did.

Regression to the mean isn't really an issue here. Regression is a constant in terms of heritability/SDs: the nonlinearity/cherrypicking is already baked in. If, say, adult IQ is 80% heritable, then it deflates towards the mean by 20% on average no matter where you are on the spectrum of low/above-average. (This might be different if it turns out to be driven by rare genes, which it wasn't for IQ, but since we're discussing clones that wouldn't matter anyway.) So if von Neumann is at +5SD, then he is expected to regress back to 4.47SD, just like someone at 3SD regresses back to 2.6 and someone at 1SD to 0.89 and so on. It may be higher or lower, but him simply being extreme doesn't matter AFAIK. (If you condition on more information like his family background, you could argue he'll regress to a higher mean, so even less of a problem.)

What's interesting here, and might be driving your intuition about clones being a disappointment, is that if you select high enough and the heritability is high enough, then even as you drive the average of the clones very high compared to the population, they may have an increasingly low probability of matching or exceeding the original. (In the extreme case of a genetically fixed trait or 100% heritability, the clone must be identical to the original and will always match and never exceed.) This is because that remaining 20% variability may be 'tight' around the mean, I think. I didn't calculate this out in detail but I did notice it while thinking about dog cloning: https://www.gwern.net/Clone#liability-threshold-model Still, even somewhat short of von Neumann is still with few peers and would have massive changes. (To put this in perspective, and contra JBlack (who really ought to work more with order statistics & normal tails if he thinks as tiny an effect as 100x is an upper bound when we are dealing with extremes selected out of millions or billions of people), if Neumann clones regress back to, say, 4.47SD, then in a country of 320m like the USA, there are ~1,200 such individuals, and so any scenario involving more than 1.2k clones would mean instantly doubling the total population of such individuals. And then there are tens of thousands of adoptions, and egg/sperm donations, in the USA each year...)

Anyway, the big imponderable here is that we aren't really concerned with IQ: IQ is the consolation prize, and the keys under the lamppost. We want eminence, achievement, genius. The heritability of that is a puzzle. If you look at families, genius seems to run in them but in a way that looks like it'd have to be pretty weak additively; that would seem to be bad for the prospects of cloning if the additive heritability of this overall 'achievement' trait was low, but on the other hand, there's good arguments and examples to think that personality, the next most important factor, is highly heritable and just highly nonadditive, and that achievement relies on many factors simultaneously and is emergenic - in which case it flips back to potentially highly heritable in a way where cloning works really well (and where embryo selection/IES/editing would not, incidentally, at our present state of knowledge of miserably poor personality GWASes etc). It's hard to investigate because identical twins are just not that common and wouldn't show up in compiling lists of eminent figures to try to estimate concordance rates. Still, perhaps the consolation prize is enough?

What's interesting here, and might be driving your intuition about clones being a disappointment, is that if you select high enough and the heritability is high enough, then even as you drive the average of the clones very high compared to the population, they may have an increasingly low probability of matching or exceeding the original.

So to work it out: what percentage of clones, because of regression to the mean, would be expected to surpass the original on some quantitative trait? Using again von Neumann at an arbitrary +5SD (roughly 80 such individuals in the USA), then clones are expected to regress back to 4.47SD, because IQ 80% heritable, leaving 20% leftover (most of it non-sharedenvironment).

20% variance is another way of saying a SD of 0.44 around the new mean of +4.47SD. How much luck does one need for ±0.44 to reach up from 4.47 to 5?

  • (5 − 4.47) / 0.44 = +1.2SD
  • ≤1.2SD is 88% of the population (pnorm(1.2))

So only 100% − 88% = ~12% of the clones would match the original (ie 1 in 10 vs a base-rate of 1 in >4 million, a multiplier of >400,000×, as opposed to a measly 100× or something - that's tail effects for you, always larger than you think).

You could also run this with my dog cloning code. I considered a scenario where there's two thresholds: one for picking the elite donor, and then a more ordinary pragmatic threshold. (In that case, 'best military dog anywhere' and 'acceptable military dog after training'.) But you can simply set the two thresholds equal and now it calculates the probability of a clone passing the original selection threshold (ie matching or exceeding the original - at least as far as you can tell by comparing!):

R> cloningBoost(successP=(1/300000000), preThreshold=(1/300000000), heritability=0.8)
# [1] 0.148684202
R> unlist(Map(function(threshold) { cloningBoost(successP=pnorm(threshold), preThreshold=pnorm(threshold), heritability=0.8) },
               seq(0,-8, by=-0.5)))
#  [1] 0.9447298251 0.8778126221 0.7922355447 0.6996006781 0.6080710792 0.5220580968 0.4435373873 0.3731751928
#      0.3109859370 0.2566634390 0.2097376297 0.1696489955
# [13] 0.1357875174 0.1075174469 0.0841969256 0.0651871840 0.0501318209

Because of thin tails, it's unsurprising that the more extreme you get, the lower the probability gets: it just gets wildly harder and harder the more SDs out you go. (Similar to bivariate double-maxima.) But the reduced variation + higher mean also benefits from the tail effect in that it gets ever more effective compared to the baseline of random screening/larger samples; the further out you go, the lower your probability may be, but the more effective it is relative to the alternatives which get worse faster. It's worth noting that a probability like '0.05' is astonishingly huge compared to the base rate (that "5% probability of matching the original" is for pnorm(-8), which has a base rate of 6.22096057e-16, ie. there is not a single person in the world 8SDs out - should you have such a person and you want another such outlier, better get busy with the cloning research, otherwise, you'll be waiting a while):

round(digits=0,
    unlist(Map(function(threshold) {
        cloningBoost(successP=pnorm(threshold), preThreshold=pnorm(threshold), heritability=0.8) / pnorm(threshold) },
        seq(0,-8, by=-0.5))))
# [1]              2              3              5             10             27             84            329           1604           9819          75541
# [11]         731681        8933802      137633507     2677226875    65788482879  2042914354508 80585337764361

Regarding cloning and regression to the mean: Have you written somewhere about the influence of the maternal environment on genetics? One could imagine that part of what made von Neumann great were not just his own genes, but also those of his mother who carried him to term, and a clone of him would not have her maternal environment.

I guess to disentangle the effects of genes and maternal environment, there might be studies of paternal half-siblings, or of couples who had children both naturally and via egg donation.

By maternal environment, do you mean just home/rearing environment attribute to the mother rather than father, or the womb environment/dam effects? The former does exist (you can do neat designs these days like looking at IGEs of just genes in the mother but not father) but influences stuff like EDU much more than IQ; a clone is already maxed out on EDU potential so we don't care if the adoptive placement gives mothers less keen on education than von Neumann's parents were. Womb environments have also been measured, I think through designs like you suggest (or was it children-of-twins designs? I forget) and found to be generally unimportant despite some hype. After all, whether you correlate against the mother or father, or second-degree relatives in any direction, the correlations are similar...

Thanks for the answer! By maternal environment I indeed meant the latter rather than the former, i.e. womb environment etc., i.e. the part one might not be able to duplicate when cloning. But if that hardly matters in the general population, I guess that's not much of a worry.

This is because that remaining 20% variability may be 'tight' around the mean, I think.

This is in fact the fatal flaw of cloning (with respect to producing high achievement individuals): it's much worse than sexual reproduction at doing so, because it's lower variance, and variance is your friend if you want rare outcomes!

After all, if we're considering 100 clones of a high-achievement individual being raised by surrogates, the natural comparison is 100 genetic children of two high-achievement individuals being raised by surrogates. (Note: I'm not offering an opinion on the ethics of either, just that this is the most apples-to-apples comparison.) The latter is FAR more likely to produce a super-high-achievement individual, because it starts from the same additive-genetic baseline, but has higher variance due to also including the genetic variance introduced by meiosis.

The only way this analysis could be false is if the advantage of preserving non-additive genetic effects overwhelms the disadvantage of lacking genetic variance, but I would be surprised if that were the case.

Therefore, cloning is just bad (for this goal). Really it seems to have no use whatsoever, and given that it's illegal and low probability of successful birth, why even bother? On the other hand, if you found consenting participants, distributing embryos from two selected high-achievement individuals to surrogates would be legal and technologically feasible today.

This is in fact the fatal flaw of cloning (with respect to producing high achievement individuals): it's much worse than sexual reproduction at doing so, because it's lower variance, and variance is your friend if you want rare outcomes!

No. Selection lets you drastically increase the mean of clones in a way at least twice as difficult as for obtaining equally-elite pairs of parents. You can get the tail just as much from increasing the mean as from increasing the variance. Both are your friend. This is also relevant to embryo selection when we consider the values of increasing embryo count/PGS/selection: https://www.gwern.net/Embryo-selection#optimizing-selection-procedures (It also doesn't follow that cloning-like approaches using a single parent are unable to exploit any variance-increasing methods, see the gamete-selection section.)

After all, if we're considering 100 clones of a high-achievement individual being raised by surrogates, the natural comparison is 100 genetic children of two high-achievement individuals being raised by surrogates.

No, it's not. Who's the female von Neumann whose eggs you're thinking of using...?

but I would be surprised if that were the case.

Please see the emergenesis link.

Selection lets you drastically increase the mean of clones in a way at least twice as difficult as for obtaining equally-elite pairs of parents.

I honestly don't know what you're talking about here. What's "twice as difficult"? Do you mean because you need to find two donors instead of one? I think finding donors is the easiest part, so that doesn't seem like a problem to me.

(It also doesn't follow that cloning-like approaches using a single parent are unable to exploit any variance-increasing methods, see the gamete-selection section.)

I was under the impression that we were not considering future technology (such as gamete selection). Cloning primates is current-day technology (though still immature).

Who's the female von Neumann whose eggs you're thinking of using...?

The same way you'd find a male donor? Just ask some high achievement women until you find a willing donor (and maybe check for high-achievement relatives and run some polygenic scores, if you want additional confidence). I don't see why this is a problem.

Note that von Neumann is not a possibility with present-day technology anyway, since he's dead, and current technology requires a living donor.

Please see the emergenesis link.

Ah, sorry, I missed that when reading your post the first time. It's true that sufficiently important non-additive effects would overwhelm the disadvantage of lower variance. This becomes a quantitative question: in principle the calculation could come out in favor of either cloning or sexual-reproduction, depending on assumptions.

I had the general impression though that non-additive effects were believed to be of relatively low importance compared to additive effects, but I admit that I don't know precisely how much lower, and even a small effect could matter when considering extreme outliers.

What's "twice as difficult"? Do you mean because you need to find two donors instead of one?

Yes, and it means you can't get a gain by throwing out one donor. If you must include the lower-scoring runner-up of the opposite sex, that drags you further back to the mean. (Note that for all the talk of variability, von Neumann's descendants aren't that impressive, and this is true of most such.)

I was under the impression that we were not considering future technology (such as gamete selection).

I was under the impression we were, as no one has cloned a human at all and it is therefore future technology.

I don't think gamete selection is all that hard or futuristic, either, merely uninvested in. There's also chemical treatments to increase recombination rate (see the references), which is existing technology in plants/animals.

This becomes a quantitative question: in principle the calculation could come out in favor of either cloning or sexual-reproduction, depending on assumptions.

Indeed. Depending on where you think you can select to for donor cells, how heritable the trait of interest is, how many embryos you select from... Each of these has its own diminishing returns and costs and tradeoffs. Which is why I present many scenarios in my writeups. There is no substitute for working through the math - even simple toy models are indispensable, because intuition is weak here. Clones are really good in some scenarios; and really bad in others. It depends.

I had the general impression though that non-additive effects were believed to be of relatively low importance compared to additive effects, but I admit that I don't know precisely how much lower, and even a small effect could matter when considering extreme outliers.

I would say that we are pretty sure that non-additive effects are unimportant for IQ. We are also reasonably sure that non-additives have been, overall, greatly overblown (I blame the medical Mendelians for this), at least for those human traits that are easy to measure: for example, Polderman et al 2015 finds no overall evidence for ADE instead of ACE in its twin mega-analysis, or Hivert et al 2020 scans 70 UKBB traits and finds 0% dominance and 6% epistasis. Not much there! We are also reasonably sure that for cases like that, the gap between the SNP additives (global mean ~20%?) & full heritability (~50%) is made up mostly of rare additive variants. Unfortunately, for personality, it happens to be a big exception: the SNP heritabilities can be as low as 0%, pointing to potentially huge roles for epistasis/dominance to explain the identical twin correlations (often >50%). And no one knows how 'eminence' works in part because it is intrinsically extremely debatable and hard to measure - but there is clearly a lot of variance unexplained by mere IQ, and other things are necessary, and what everyone tends to conclude (from Galton to Eysenck to Jensen to Simonton on) is that personality and personality-like traits such as motivation is a big part of the missing puzzle. So...

The emergenesis model is the best I've seen anyone propose which is consistent with the family-level observations, inadequacy of IQ to explain everything (and far grosser inadequacy of environmental causes like SES), and non-additivity of personality and similar traits. But it hasn't yet (and may never be given the paucity of data) been pursued to the quantitative point where you could hope to really compare it, which is why I only give numbers for IQ above. That's where the light from the lamppost is.

Thanks, this discussion has improved my understanding.

And no one knows how 'eminence' works in part because it is intrinsically extremely debatable and hard to measure - but there is clearly a lot of variance unexplained by mere IQ, and other things are necessary, and what everyone tends to conclude (from Galton to Eysenck to Simonton on) is that personality and personality-like traits such as motivation is a big part of the missing puzzle.

I'd be curious to know about the genetics of metrics like "number of patents authored", where it's measuring productive activity (instead of test performance and educational attainment).

That's an example of what I mean by debatable.

Patents are zero-inflated: most people simply have zero lifetime patents. Not informative.

And when someone does have a patent, what does it mean? Patents are a pretty lousy measure: look at software patents. Sheer fishing expeditions and patent trolling (like Intellectual Ventures, which made patents by getting people around a table to shoot the wind and have a lawyer listen and write down patents for any random idea someone idly speculated about). When you think of eminent figures like Einstein or von Neumann, how many patents do you think they had? Einstein had over 50, but few have any relationship to what he's famous for, and would he have that many if he hadn't literally worked in a patent office? And there are loads of people with far more than 50. Von Neumann also had some patents, but again nothing remotely like his stature*. Arguably, after a certain point, having more patents means you are less eminent and you're some sort of hired gun grinding out paperwork. (I think of this every time I hear about how IBM, or China, incentivizes patents with big bonuses and accordingly received a record number of patents that year. Hasn't done them much good long-term that I can tell...) And there are many areas of achievement where patents are entirely irrelevant.

If you pulled together a population-sized genealogy or registry (Scandinavia would do the trick, or one of the American genealogies to cross-reference with the USPTO), I don't know what the results would be in terms of an ACE model, but I don't think it would change any minds either way about emergenesis.

* The patent von Neumann is best known for is the one he didn't get but instead torpedoed by publishing about the design of a computer, thereby helping usher in the age of the digital computer immediately instead of it being controlled by a monopolist for decades.

I think of this every time I hear about how IBM, or China, incentivizes patents with big bonuses and accordingly received a record number of patents that year. Hasn’t done them much good long-term that I can tell...

Yeah, I personally have 2 patents to my name through this kind of Goodharting. (Higher management provided the incentives, lower management encouraged me to apply even knowing it probably wasn't economically worthwhile.)

This is very true, and a good reason to clone millions of von Neumanns, rather than just a handful.

Expected value compared to hiring the top 100 people in the international math Olympiad each of the next 20 years?

Depends on the price of creating and raising clones.

Not really, given the huge disparity in numbers -  unless  you have a  magic way of feeding/housing/clothing/caring for children which costs far less than is currently possible? (And note that we know baby warehousing REALLY doesn't work well.)

Super smart people are 10 a penny. But for every genius working to make AGI safer, there's 10 working to bring AGI sooner. Adding more intelligent people to the mix is just as likely to hinder as to harm.

More concretely, if we were to clone Paul Christiano what's the chance the clone would work on AGI safety research? What's the chance it would work on something neutral? What's the chance it would work on something counterproductive?

And how much would it cost?

Seems like it would be a much better use of resources to offer existing brilliant AI researchers million dollar a year salaries to work on AGI safety specifically.

You ask a number of good questions here, but the crucial point to me is that they are still questions. I agree it seems, based on my intuitions of the answers, like this isn't the best path. But 'how much would it cost' and 'what's the chance a clone works on something counterproductive' are, to me, not an argument against cloning, but rather arguments for working out how to answer those questions.

Also very ironic if we can't even align clones and that's what gets us.

This seems like the sort of thing that would be expensive to investigate, has low potential upside and just investigating would have enormous negatives (think loss of wierdness point, and potential for scandal).

More concretely, if we were to clone Paul Christiano what's the chance the clone would work on AGI safety research?

From this study from 1993,

The authors administered inventories of vocational and recreational interests and talents to 924 pairs of twins who had been reared together and to 92 pairs separated in infancy and reared apart. Factor analysis of all 291 items yielded 39 identifiable factors and 11 superfactors. The data indicated that about 50% of interests variance (about two thirds of the stable variance) was associated with genetic variation.

It should be noted that these sorts of studies likely underestimate heritability, due to measurement error. See e.g. this for more info.

At a guess reducing interest variance to a single number is inappropriate. For example I imagine the correlation between twins both liking maths is much higher than them both being interested in a specific branch of maths.

In this particular case I think the clone is far more like to be interested in AI or philanthropy in general than the particular cross section of the two that is AI safety research.

I hope someone has taken seriously the idea of just paying top researchers a million a year to work on safety instead of capability. The last several times 'pay for top talent' was made in the context of ea in general very unconvincing excuses were given not to.

Shouldn’t the ethics of cloning be the same as the ethics of having children normally? If you could have kids with von Neumann’s capabilities, it’s ethical as long as you raise them right and don’t abuse them. Presumably it’s the same with a clone army of von Neumanns: don’t neglect them, beat them or send them to public schools and it should be fine.

I'm inclined to agree with you. To flesh out the scenario: suppose that an organisation decided to clone Von Neumann (or Einstein or Marie Curie or whoever), then handed the resulting baby to the adoption services of a responsible Western country, knowing that the adoption service put a lot of effort into making sure that prospective parents were responsible and suitable to raise a child. (I believe the US is a bit of an outlier on letting random people adopt without stringent checks, so let's say it's the British or. Swedish adoption system.) The adoptive parents don't know the baby is a clone-genius and raise him/her normally*  I find it hard to say that this proposal to have a baby and ensure it's raised by loving parents would be unethical. 

Raymond D has some extra concerns about using Von Neumann's genes without permission, but I'm inclined to the view that permission from the heirs would be enough to make that ethical. Von Neumann himself is dead and can't suffer, so if the heirs are OK with the proposal, then I can't identify anyone who would be harmed by it. Another possible solution to this problem: clone a living genius who gives their consent.

I'm still thinking about this, and these are not firm conclusions, but I agree that the clone proposal deserves thought.

 

* I understand that knowing you are the clone of Einstein and that if you do anything less than revolutionise physics you have failed to meet expectations would pile unreasonable amounts of stress on a child. So there's a good argument for not letting the clones or their adoptive parents know they're clones.

Not putting unreasonable expectations on the children is very important, and you’re right that not telling the caretakers is potentially a good way to achieve this. Excellent point.

That might be, but I could find points for the opposite as easily. After all, we are expecting the child to help save the world. If a child is to become someone of exceptional importance, then probably some sort of special treatment can help tutor them into that role. Take the Dalai Lama: he's raised into his role since birth.

Studies such as https://files.eric.ed.gov/fulltext/EJ746290.pdf indicate that, if you take a genius child and make them go through normal school with their average age-peers (with no grade-skipping or other academic acceleration), then things go much worse than if you let them skip two or more grades.  So you may want to make sure that the parents will (a) notice that the child is a genius and (b) take at least semi-appropriate action.

Of course, keeping a genius with regular age-peers is only one case of "failing to recognize a child's special needs and accommodate them".  But it's a remarkably common one: in the study, 33 of the 60 exceptionally gifted Australian kids were not permitted any grade-skipping at all.

I think there are extra considerations to do with what the clone's relation to von Neumann. Plausibly, it might be wrong to clone him without his consent, which we can now no longer get. And the whole idea that you might have a right to your likeness, identity, image, and so on, becomes much trickier as soon as you have actually been cloned.

Also there's a bit of a gulf between a parent deciding to raise a child they think might do good and a (presumably fairly large) organisation funding the creation of a child.

I don't have strongly held convictions on these points, but I do think that they're important and that you'd need to have good answers before you cloned somebody.

How could it be wrong to clone him without his consent? He’s dead, and thus cannot suffer. Moreover, the right to your likeness is to prevent people from being harmed by misuse of said likeness; it doesn’t strike me as a deontological prohibition on copying (or as a valid moral principle to the extent that it is deontological), and he can’t be harmed anymore.

Also, how could anyone have a right to their genome that would permit them to veto others having it? If that doesn’t sound absurd to you prima facie, consider identical twins (or if they’re not quite identical enough, preexisting clones). Should one of them have a right to dictate the existence or reproduction of the other? And if not, how can we justify such a genetic copyright in the case of cloning?

Cloning, at least when the clone is properly cared for, is a victimless offense, and thus ought not be offensive at all.

The belief that people can only be morally harmed by things that causally affect them is not universally accepted. Personally I intuitively would like my grave to not be desecrated, for instance.

I think we have lots of moral intuitions that have become less coherent as science has progressed. But if my identical twin started licensing his genetic code to make human burgers for people who wanted to see what cannibalism was like, I would feel wronged.

I'm using pretty charged examples here, but the point I'm trying to convey is that there are a lot of moral lenses to apply here, and there are defensible deontological prohibitions to be made. Perhaps under scrutiny they'd fall away but I don't think it's clear cut, or at least not yet.

Isn't the guy who founded the AI alignment field the obvious person to clone? Especially given that he's still alive.

I'll assume this comment is tongue in cheeck, but respond by taking this (uncomfortably) seriously anyway: (Feel free to downvote or delete if this crosses some boundary.)

The AI alignment field has already been founded. The kinds of people we would want to clone are those who could solve it.

As for Eliezer specifically, from what I understand from his public writing, he's a) obviously incredibly smart (though not at the ridiculous level of von Neumann), but also he's b) particularly gifted in language (as can be seen in his prolific writing output), rather than e.g. math; c) he seems to have serious problems with fatigue, and he also d) mentioned somewhere that he's not particularly happy, in general.

a) and b) are presumably genetic and would already not make him the best candidate (and when considering cloning, why settle for anything but the best?), but insofar as c) and d) have a genetic component, that only disqualifies him further. And our prior for problems like fatigue to have a genetic component should be higher for him than for a random member of the population: see here.

I was actually being serious.

But (d) seems like a rounding error given what's at stake,[1] and (b) doesn't seem like a negative. The Van Neumann comparison only seems relevant if we can clone dead people (which I assume we can't?). So on the whole, he still seems like a great choice to me. But by all means, clone both him and paul.


  1. As far as political feasibility goes, the hedonic level seems less important to me than consent. ↩︎

I think this is worth taking seriously.

I think the best counter-argument is that, as James Miller points out in a similar comment, there is almost certainly some positive luck in creating a Von Neumann (or a Christiano) and only a small reduction in ability takes somebody to merely one among thousands of peers. With enough clones, one should expect a few serious successes, but if we create lots of capable AI researchers for only a few people capable of making unusual leaps in safety then the tradeoff would not be very clear (though I would guess still positive given, as Matthew points out, heritability of most personality traits).

The better option, though more extreme, would be making people who we were confident would be above the normal range of human abilities. I think this would answer Yair's worry that more smart people would be just as likely to accelerate unsafe as safe AI. If there were a small, connected group of people with far superior ability to create both safe and unsafe AI, the strength of competition would not be as strong and the amount  of slack and ease of cooperation amongst those at the frontier would be greater. Also, while intelligence is not directly related to being right, or altruistic; it's in everybody's personal interest not to create unaligned AI, and we would expect these people to be good at identifying risks to themselves. Even if they were to attempt to use it for their own ends, this is likely to be better for humanity than AI not aligned to any human.

The more promising pathway towards this seems to be iterated embryo selection on polygenic scores for IQ (+ any other trait expected to be helpful) through in vitro gametogenesis, which is not currently possible but may be possible within a decade or so and could generate people with intelligence well beyond current human limits. Gwern has a gigantic write-up of the subject here if anyone is interested, which I won't try to summarize.

Clones of von Neumann, or anybody else for that matter, are just time-delayed twins. That is to say, you get embryos with (at best) the same genetics, but that have to be gestated, raised, educated, and form their own experiences throughout.

Maybe the average intelligence of adults who came from von Neumann embryos would be substantially greater than from the average population, such that the extremely intelligent ones occur a hundred times as often as usual. That's probably the most optimistic outcome. Then again, maybe the radically different life experiences (such as being one of ten million in a prototype massive project on generation timescales, which historically do not go well) would make the upper extreme less likely.

So in the really great case, you raise ten million clones that are 100x more likely than average to be top tier in intelligence. But even doing nothing, you are likely to get just as many in a few years from the general population. To substantially increase the fraction of extremely smart people, you're going to need a bigger multiplier than a hundred for top-tier intelligence, or raise something like a hundred million clones, more than the population of most nations.

That's after you get past the difficulty of creating viable clone embryos at all, and finding enough women or creating good enough artificial wombs to gestate them all (without doing anything that might harm their chances of being extremely smart in adulthood), and then raising them in environments conducive to becoming intelligent. Just the initial experiments to establish viability of the idea would involve timescales on the order of a generation or two, even if there were no other practical or ethical considerations.

So all of this looks possible, but not at all likely to be effective for preventing AI catastrophe or anything else that might happen within a hundred years.

Maybe the average intelligence of adults who came from von Neumann embryos would be substantially greater than from the average population, such that the extremely intelligent ones occur a hundred times as often as usual. That's probably the most optimistic outcome.

I think this is actually a quite pessimistic outcome and that IQ is much more heritable (and genetically determined) than this assumes.

I'm sure that IQ, or really some more specific capability that helps with AI alignment research, could indeed be highly heritable. In the long run, much greater than 100x prevalence multipliers should be easily achievable.

I very strongly doubt that the first generation experiment will get anywhere near the theoretical bounds of heritability. Getting 100x the proportion of supergeniuses in the first generation is flagrantly optimistic enough for me.

Having two generations to play with is pretty optimistic in my view, as it corresponds to having them start to work productively on AGI safety around 2100. While still plausible, I expect the problem to be either solved by then or obviated by civilization failure one way or another.

I think this is an interesting discussion to have (even if it's unlikely to be realized due to negative societal opinions of cloning).

It's worth being aware that cloning a dead person isn't currently technologically possible. You need an intact nucleus, I believe (though I'm not an expert). Cloning a living person is probably possible, since other primates have been cloned, though the success rate there was poor.

Regression to the mean is important to consider as well, as other comments have pointed out. If you could clone a great theoretical physicist, what you get is probably not a great theoretical physicist, but rather a pretty good theoretical physicist. There does not seem to be a shortage of pretty good theoretical physicists in the world, and if anything maybe there is a glut. It would seem tragic for the outcome of the project to be a bunch of pretty good theoretical physicists who are forced to settle for data science jobs.

Therefore you may be better off considering professions where there is a shortage of pretty good talent. Software engineering is such a field, which is why pretty good software engineers make a lot of money. So you might consider cloning a great software engineer, such as Jeff Dean or the like (and there are enough of them out there that you could probably find one that would consent to the plan). The Dean-clones would likely not match the original in capability, but they'd be well equipped to make a good living and contribute in a positive way to society.

Another thing to consider more explicitly is the replacement value of a person, from a societal perspective. How much money they make is a relevant but very-imperfect measure. For example, a top pro sports player makes a lot of money but offers barely any value over replacement, because if they didn't exist, then somebody else would be the top player. For that reason, the safest bet is to consider people who make a lot of money in ways that seem generally pro-social, such as engineers and entrepreneurs.

As far as entrepreneurs go, cloning a founder-billionaire seems potentially interesting. Elon Musk seems to have created billions of dollars value, so maybe Musk-clones, even though (because of regression to the mean) not quite as talented or driven, would create a lot of value too.

How closely correlated are the IQs of identical twins, anyways?

Early twin studies of adult individuals have found a heritability of IQ between 57% and 73%,[6] with the most recent studies showing heritability for IQ as high as 80%.[7] 

Source

https://www.goodreads.com/book/show/834518.Cyteen has a fictional exploration of a similar topics (trying to achieve not just intelligence transfer, but also expertiese in a relevant field)

I agree, it's indeed a topic worth of a serious consideration. 

A related approach is mind uploading. I would compare it with cloning as follows:

Pros: 

  • create the exact same mind, with all the genius and wisdom, without regressions to the mean etc
  • no need to wait for 15+ years to grow the clone
  • likely much easier to scale (e.g. create 10 instances of the entire MIRI, and task them with 10 different problems to solve)
  • it might be possible to run the uploaded mind much faster than real-time (e.g. an instance of the entire MIRI working x1000 faster)
  • it could grant immortality to existing alignment researchers (they are not getting younger)

Cons:

  • depending on implementation details, there is some risk that the uploading mind could go rogue (the risk could be partially mitigated by a careful selection of the people to upload)
  • unlike cloning (which is already a today's tech), mind uploading still requires some large R&D effort
  • uploading is unlikely to be useful for cloning dead people (but maybe not)

There are paths to mind uploading that don't require extremely detailed scans & emulations of the brain. Maybe one of the paths will take less time than growing a clone.