The article by Robert Sparrow:

A series of recent scientific results suggest that, in the not-too-distant future, it will be possible to create viable human gametes from human stem cells. This paper discusses the potential of this technology to make possible what I call ‘in vitro eugenics’: the deliberate breeding of human beings in vitro by fusing sperm and egg derived from different stem-cell lines to create an embryo and then deriving new gametes from stem cells derived from that embryo. Repeated iterations of this process would allow scientists to proceed through multiple human generations in the laboratory. In vitro eugenics might be used to study the heredity of genetic disorders and to produce cell lines of a desired character for medical applications. More controversially, it might also function as a powerful technology of ‘human enhancement’ by allowing researchers to use all the techniques of selective breeding to produce individuals with a desired genotype.

Quote:

However, to my knowledge, the current paper is the first to explicitly discuss the possibility of the iterative use of this technology for reproductive purposes

The possibility was discussed in MIRI's "Uncertain Future" toy forecasting model back in 2009, and the analysis formulated a few years before that.

ETA: And further discussed in James Miller's recent book, "Singularity Rising."

New to LessWrong?

New Comment
34 comments, sorted by Click to highlight new comments since: Today at 11:46 AM

This Slate article talks about the issues that children of sperm donors have dealing with the fact that they don't know their biological father. Maybe we should prioritize the discovery of genes related to mental health, happiness, etc. so that eugenically bred kids will be cool with the fact that they aren't very closely related to their parents. (I assume adopted kids deal with similar issues, and some of them probably deal with their issues better than others... could a large-scale genetic analysis be useful?)

(Also breed them for humility and compassion so they won't be stuck-up despite knowing for a fact that they are just plain better than everyone?)

[-][anonymous]11y10

Being cool with not being directly related to your parents is a question of values, not mental health or intelligence.

If both types of gametes are possible to create from stem cells, it could make it possible in many more cases for the parents to also be the biological parents of the child, which could largely mitigate this harm. (Not wanting money being involved in one's conception is a different issue, and one I have less intuition for - I'm not squicked out at all by prostitution, for example.)

And I cited your paper and discussed iterated embryo selection in my 2012 book Singularity Rising (p. 98).

I mentioned this to Sparrow.

How has Sparrow reacted to the point that this idea was previously written up in some detail by MIRI?

As far as I can tell, his only public reaction is a line in Sparrow 2014: "A recent treatment of this topic by Shulman and Bostrom^6 calls the same technology 'iterated embryo selection' - a name that Matthews and Fujita et al may prefer.". So, he never acknowledged being scooped.

Although after more looking into it, if we're going to argue about priority, it looks like IES was actually first proposed a decade before MIRI did, in Haley & Visscher 1998's "Strategies to Utilize Marker-Quantitative Trait Loci Associations" - their Figure 5c is unambiguously IES.

[-][anonymous]11y20

Also, I'm curious if someone has an (independent) answer for the questions I posed Randall Parker in this reply, especially the last one:

http://www.futurepundit.com/archives/009032.html#reply20130803061511

Alexander, probably you could select for intelligence, using the data from the Rietveld et al study. But if you made 10 embryos and picked the one with the highest predicted IQ the increase in average offspring IQ would be under (at most, if one assumed the Rietveld results were fully solid) 3 points. And since the process has not yet been routinized, you would have to undertake serious legwork, bringing together techniques from different labs. And note that the Rietveld results are only a few months old.

But bigger genetic studies should multiply the impact of embryo selection by at least several fold over the decade.

But bigger genetic studies should multiply the impact of embryo selection by at least several fold over the decade.

Now I'm wondering what a full cost-benefit analysis of embryo selection looks like. We know the benefit of each marginal IQ point is something like a few thousand dollars in lifetime income, we also know roughly how much variation can be explained by SNPs and how many relevant SNPs there are, which means we can calculate how many SNPs we'd expect for a given sample size of genotyping, and the genotyping sample size growth over time should be fairly predictable, so you can estimate an IQ boost vs time curve and then set off the gain from each marginal IQ point against the estimates of your sperm/egg deteriorating over time or storage or opportunity costs... Sounds like a thorny yet doable analysis.

We know the benefit of each marginal IQ point is something like a few thousand dollars in lifetime income

~1% income boost per point. See this report for an almost identical analysis of the benefits of programs to prevent low birth weight. With very low discount rates/return on other investments they estimate a value of between ~$3k-14k for each 1% increase in earnings. By that metric (and metrics used for preschool interventions) it should be cost-effective in a few years but is not so now (particularly in the absence of any clinic offering an organized service).

If one has a higher discount rate (they use rates of 2-4%, and wage growth rates of 0-2%), like a typical parent, then it would not be appealing.

On the other hand, this isn't taking into account non-wage gains (health, longevity, institutionalization, marriage prospects, well-being).

Well, one problem with simply regressing IQ against income is that to an extent the gain is coming from what look like positional or zero-sum effects like displacing someone at an elite university; that's one reason I've been compiling relevant papers for a while at http://lesswrong.com/lw/7e1/rationality_quotes_september_2011/4r01 I am more interested in the marginal increase in net societal wealth from IQ; if all IQ does is help on play games better, it's not something I care about increasing.

By that metric (and metrics used for preschool interventions) it should be cost-effective in a few years but is not so now (particularly in the absence of any clinic offering an organized service).

Yes, probably, but part of the point of the analysis would be estimating what 'a few years' would be (5 years? 15? 30?); if sample sizes are growing exponentially to match the apparent exponential fall in price of genotyping, that looks different from a linear or log growth scenario (eg. perhaps from constant funding, or from datasets disappearing after a certain amount of time).

Well, one problem with simply regressing IQ against income is that to an extent the gain is coming from what look like positional or zero-sum effects like displacing someone at an elite university

The numbers are higher if you don't control for education. Remember that on all the major theories of returns to education educated workers are more productive: learning (IQ helps to learn), ability bias (which is just saying that controlling for education understates direct IQ effects), and signaling. On signaling accounts the signals have to be honest enough that employers don't find it profitable to ignore the educational credentials.

Externalities are interesting, but hard to estimate (except for obvious individual-level things like net fiscal contribution, crime, etc), and probably not a primary focus for parents.

if sample sizes are growing exponentially

Yes. There are large samples scheduled to be online within 2 years that could give a 10x in sample size.

I'm not going to continue this thread any longer though, as many of the pieces should be addressed in a forthcoming paper with Nick Bostrom.

I'm not going to continue this thread any longer though, as many of the pieces should be addressed in a forthcoming paper with Nick Bostrom.

Alright, I'll wait for that to come out then.

[-][anonymous]11y20

Thanks gwern and Carl. From what you've said here, my confidence that this is (at least in principle) possible now has risen. I do wish somebody with credentials would go on the record about it--I guess that's what your paper is for.

[-][anonymous]11y20

How would this be useful for reproductive purposes over and above embryo selection? You can see almost no actual phenotypes in an embryo other than "growing right" versus "going horribly wrong".

You can sequence for genotypes, and do selection based on the results of genomic studies.

[-][anonymous]11y00

I read this and all I can think of is the recent family 23andme results with half-opposite results of what they 'should' be for risk factors given known family history back three generations indicating that whatever is causing the risk is completely unknown to genetic science, and a recent study in which only 15% of the genetic basis for height (known to be >80% responsible) was discovered spread over more than 400 loci... don't get me wrong there's hardly any reason anybody should be conceived with simple common mendelian diseases these days, but if you wanted to actually change much other than that or gross differences in brain or liver chemistry you need to actually observe phenotype.

and a recent study in which only 15% of the genetic basis for height (known to be >80% responsible

Note that the proportion explained has been scaling very nicely with increasing sample size, (going from ~0 to current levels over about an order of magnitude scaling in sample size, and with several orders of magnitude left to go), and that the efficacy of a single generation of embryo selection goes with the standard deviation, not the variance, so one should take the square root of "heritability explained" numbers to estimate efficacy.

Also see recent genetic results such as this indicating that common variants are adequate, given larger sample sizes.

This is discussed in the section 'Three Applications'.

[-][anonymous]11y10

Have people thought much about the social consequences of this?

I mean, people produced through mature iterated embryo selection (or actual genetic editing) would be geniuses and more athletic and conscientious. Now this in itself is not necessarily that new. We have geniuses today. But there would be more and they might be smarter than today's geniuses, so the variance in the population would rise. Steve Hsu has speculated that in theory you could make someone with a 900 IQ!

Edit: Actually, the correct figure was 100 + 30*15 = 550.

Edit 2: Still wrong. 100+30*10=400 seems true to Hsu's idea, though I agree with Carl below that it's likely to overestimate.

William Gardner thinks that credentials would probably reflect ability more since innate differences would be more obvious. So you would get less social mobility.

There is another side effect that John Maxwell IV points out and which I haven't seen discussed much: these people might not be genetically related to anyone else. What does society look like when Harvard is filled with genius "adopted" kids, and people who want genetically related children can no longer hope for them to lead? Does today's social elite just accept kids with a lower rank? Also, do religious or other traditional folks accept a social agenda maybe set by people who relate less, or not at all, to the kinship bonds that many religions emphasize? Or do we get separatism or war?

The big potential pluses of these technologies for people alive today look to me like network effects and externalities from being geographically near enhanced populations. But with separatism these effects go away and you are left with probably lesser impacts like gains from international trade.

In general I'd like to make a list of pros and cons. I have a feeling that many Less Wrong folks take the "pro" side. (Or am I wrong and most just think this is completely inevitable and so not worth debating or having a position on?)

For someone on the "pro" side: is your biggest reason wealth for (young) people alive today? Or is it altruism--creating minds that would be capable of better experiences or more meaningful pleasures from a utilitarian viewpoint? How do you answer my other questions here?

Steve Hsu has speculated that in theory you could make someone with a 900 IQ!

Hsu cited genetic distance measures suggesting that with respect to common additive genetic effects on IQ, there were 30-40 SNP differences per standard deviation of measured IQ difference. Since the standard deviation should go with the square root of the number of sites, he speculated that one might theoretically get 30-40 deviations of additive genetic effects, citing animal breeding successes that produced changes like that for milk production or the like.

But you should not be placing high credence on 800 points of IQ gains.

First, 40 standard deviations of IQ would be 600 points, and that's the high end estimate from a noisy measurement.

Second, the standard deviation of additive genetic influences is smaller than the phenotypic standard deviation (including environment, non-additive variation, etc), more like 10 points than 15.

Third, additivity often results from low frequencies of alleles (2 copies don't give twice the effect of 1, but if the frequency is low, then aggregate results are dominated by lone copies), including low frequencies of several alleles at different loci with similar effects on the same pathway.

Fourth, the animal examples of intensive breeding are for traits that were not valuable in the wild: producing extreme amounts of milk, eggs, or meat at the expense of the ability to survive in the wild. It looks like there was more selection for easy wins in intelligence than for those farm animal traits. Gains could be had from relaxed constraints on energy use and the like (see Bostrom's wisdom of nature paper), but generalizations from deformed farm animals bear only so much weight.

people might not be genetically related to anyone else

A large set of donors would be their great-great...-grand-parents genetically, and people really do love their adoptive parents, or parents who used donor sperm and eggs.

The big potential pluses of these technologies for people alive today look to me like network effects and externalities from being geographically near enhanced populations. But with separatism these effects go away and you are left with probably lesser impacts like gains from international trade.

Why imagine separatism? Singapore, which has one of the smartest governing elites, is unusually open to low-skill guest workers, and generally smarter people are less xenophobic. Countries with higher income, education, and test scores contribute more to global public goods, give more foreign aid, and take more immigrants.

Re impacts, you're leaving out the impact on global technological and institutional development, production of information goods (pharmaceuticals, entertainment, designs, software), as well as resolution of global catastrophic risks, which affect people everywhere.

Possibly relevant paper suggesting there may be at least one genetic modification with the potential to make large improvements in IQ (~25 IQ points)

http://www.ncbi.nlm.nih.gov/pubmed/17237123?dopt=Abstract

Most reported associations with intelligence are from underpowered studies and have not replicated in larger samples. That site did not show up in a recent study of 100,000+ people, and generally hasn't been replicated over the last 6 years as sample sizes have exploded.

ETA: Gwern's comment below suggests a relatively plausible explanation for non-replication.

If you look at the fulltext, this is for a mutation analyzed via the old standard Mendelian techniques, not a brute-force SNP analysis, as your link refers to:

Fortunately, the process of Mendelian inheritance allows a further robust test of the causality of the Arg844His mutation to be performed, avoiding the usual problems of confounding from other variables that might aggregate in a familial fashion but might be unlinked to the mutation. To test the relevance of the mutation, affected and unaffected first‐degree sibs in each generation (II:2, II:6, II:8 vs II:4; III:5 vs III:7; IV:3 vs IV:4, IV:5) underwent identical cognitive testing (table 1​1).). The mean age of affected members (53.8 years) was higher than that of unaffected members (41.1 years), acting conservatively against any superior cognitive performance in the affected group, given the known effect of age.^21 We tested the significance of the association between affected status and every measured cognitive phenotype following a randomisation approach that accounts for the kindred structure.^22...The single score for the overall difference in all mean phenotype values between the affected and non‐affected subjects was significant (p = 0.006). For individual phenotypes, all except those for immediate story recall, delayed story recall and semantic fluency had p values <5%: vocabulary (p = 0.014), digit span (p = 0.020), similarities (p = 0.039), phonemic fluency (p = 0.050), cognitive estimates (p = 0.012), immediate story recall (p = 0.112), delayed story recall (p = 0.087), immediate verbal learning (p = 0.048), delayed verbal learning (p = 0.042), semantic fluency (p = 0.220) and Hayling test (p = 0.010). As expected, given that the three subphenotypes of VIQ were all significant, when we separately tested VIQ itself this was also significant (p = 0.014)....Could the RIMS1 mutation be a chance finding, unrelated to the eye or cognitive phenotype? The intrafamilial distribution of cognitive measures argues that the detected mutation is most probably causative, especially as it segregates with both the eye phenotype (which becomes clinically symptomatic) and the cognitive enhancement. Co‐mingling in each outbred generation of mutation‐carrying and wild‐type sibs each with respective enhanced or normal cognitive phenotype and the respective co‐segregating impaired or normal visual phenotype renders extremely unlikely the possibility of an intrafamilial founder effect unrelated to the RIMS1 mutation, as supported quantitatively by our modelling, beyond linkage analysis alone.

And the paper provides an excellent reason that there have been no replications so far (italics added):

We now report on the functional and structural effects of mutation in the eye‐ and brain‐expressed gene RIMS1, through the study of individuals from a family already reported to have retinal dystrophy caused by RIMS1 mutation.^12,13 To our knowledge, this is the only family so far reported with such a mutation: the eye phenotype is homogeneous in the family, and has been documented in detail.^13...We show that a mutation in RIMS1 is associated, in the only reported kindred with any RIMS1 mutation, with significantly enhanced cognitive function in at least the verbal (likely to be related to general ability, g) and executive domains. RIMS1 is an excellent candidate gene to influence cognitive function.

Indeed, precisely as one familiar with criticism of SNPs (and with this being a mutation) would expect, the genotyping turned up nothing in their sample:

We sequenced the mutation‐containing RIMS1 exon 13 in a panel of 50 unrelated individuals with autosomal dominant cone–rod dystrophy, but did not detect any mutations. Common variation in RIMS1 (uncorrelated with the rare Arg844His mutation) did not influence cognitive function in LBC1921, for either genotype or haplotype (see supplementary ​tables 2 and 3​3 available online at http://jmg.bmj.com/supplemental). To determine if mutation in RIMS1 might account for the upper extreme of performance on cognitive measures, the entire RIMS1 gene was sequenced in the top‐scoring 5% of the LBC (24 individuals). Only one, previously unreported, SNP was found, in residue 592, exon 9: it was synonymous, conserving a glutamic acid residue in an unremarkable region of the gene.

This suggests the only way to further (ethically) investigate this would be to either sequence many millions of people at staggering expense in the hopes of finding a second family with this mutation who had not been reported in the literature as regularly going blind for no reason, or to mutate animals and hope the relevant systems are similar enough that the result is not too misleading. (The paper cites research showing animals with no RIMS1 as being stupider, so at least we do know the gene affects the brain somehow.)

The fact that the mutation causes you to start to go blind in your 20s also helps its plausibility as it handily eliminates the evolutionary argument - going blind is a pretty damn big deal ("Any evolutionary advantage of this particular mutation could be counterbalanced by the concomitant severe visual phenotype, albeit late onset.").

So, unless I've misunderstood some of the technical details, this actually strikes me as a plausible claim - albeit one that is useless for practical purposes like embryo selection since that exact mutation wouldn't occur for anyone & if it did you probably would not want to pick that embryo because going blind is a terrible thing to inflict on someone.

(Of course, you might argue that the benefit is worth it, especially since given another 30 years or so, when the degeneration really sets in, you might have the option of either replacing your eyes with improved computer vision assistance methods/implants or perhaps using CRISPR retinal therapy to try to remove the mutation from just your eyes and preserve your vision that way. It would be a very risky move, though, in assuming the IQ boost is real and that progress will be fast enough to save your eyes.)

Thanks Gwern.

I haven't read the paper myself, only been referred to it. It also violates my expectations about low hanging evolutionary fruit but the link with blindness provides at least some explanation.

Also cannot read your link.

Do you know is that same variant actually in present in the population, or is it just that variants in the same gene haven't shown association? It may be that only that specific variant has the gain of function phenotype, and that it hasn't been present at high enough frequencies in other populations to show up.

[-][anonymous]11y00

Sparrow also participated in a debate where he said this is unfeasible. The paper was received a month afterward.

My, opinions are changing fast these days.

http://artsonline.monash.edu.au/bioethics/public-debate/

ETA: And further discussed in James Miller's recent book, "Singularity Rising."

And of course Hsu has been talking about it for... well, a long time, anyway.

And of course Hsu has been talking about it for... well, a long time, anyway

Never, actually. Hsu has written interesting things about conventional embryo selection.

Ah. You were referring narrowly to just the in vitro part, not the iterated selection part. My bad.

What would non-in vitro iterated selection be, just normal breeding of adults?

I'm not sure what you mean by 'normal breeding of adults', but Hsu's proposal would be basically an augmented IVF: you do IVF as usual, sequence each of the embryos to get an estimate of IQ/whatever, and implant the highest-scoring one (usually referred to as just 'PGD'); the embryo develops into an adult, and you repeat the process.

Each generation could get a boost of maybe 5 points (it depends on how many embryos you can produce and how accurate your estimate is and how much of IQ turns out to be genuinely genetic), so in a century that could add up to quite a bit.

Ah I see, so the selection step is still in vitro, but not the entire lifecycle of an iteration. The wording had been confusing me. Thanks for clearing that up!