I recently had a fascinating conversation with Dr. Steven Hsu, theoretical physicist and serial entrepreneur, about computational genomics and genetic engineering.

There were a number of really interesting takeaways which I think would be of interest to LWians.

  • Apparently, genetic engineering in the form of embryo selection is very far along. They can already select for things like eye color and propensity for diseases. The only reason they're not selecting for intelligence yet is because there is such a strong stigma and because there's not currently enough labeled training data for the algorithms.
  • The algorithms involved are simpler than you might think, it mostly appears to be basic regression models regularized with LI penalties (as those handle sparse data much better)
  • Dr. Hsu does think it's worth worrying about this technology's potential to worsen inequality, and we kick the ethics of that around for a little bit. 
  • Epigenetics plays less of a role in determining life outcomes than many of us have been led to believe. 
  • We also discuss whether AGI or genetic engineering will lead to the first superhuman intelligence.

Hope you enjoy it!

New to LessWrong?

New Comment
23 comments, sorted by Click to highlight new comments since: Today at 4:26 AM

As someone who's tried a bit to attain it, this kind of selection does not appear to be readily available on the market.

I am currently writing a post about how to have polygenically selected children. I'll ping you when it goes up.

Relevant username? I just wondered if your name is actually Gene Smith, or you are a genesmith (a smith of genes), or just something else.

The latter

Note the key point "lack of labelled data". The tech is there, but the next mandatory step is gathering the requisite data. I have also considered and rejected this as an approach due to that lack. My timelines are short enough that I don't think this has any bearing on AGI risk.

After UK Biobank the progress on that front seems to have stalled (i.e. in the last ten years). A comparable US Biobank would probably hit the right size to get real predictive power for intelligence. 

This made me curious how much the UK Biobank cost. Remarkably, apparently only about $150 million. (Their funding page says their "core funding" has been £133 Million.)

In terms of impact on genomics, I subjectively think it's on a scale comparable to a particle accelerator or space telescope, but it has received more than an order of magnitude less funding.

I wonder whether there's a EA case for funding biobanks, given that they seem underfunded compared to other big science projects.

There is such a biobank in the US being formed right now. It's called "All of Us", and it will probably have more genomes from Africans and American Natives than any other biobank in existence.

Unfortunately, it has two problems:

  1. They have made it very hard to actually access the data. Getting an application for access actually approved is a huge slog, and researchers themselves actually have to pay for some of the costs.
  2. They aren't investigating intelligence at all

It would be incredibly helpful to have a large biobank (or even just a consumer genetics company like 23&Me) that was interested in intelligence. One such organization could triple or quadruple the accuracy of our current intelligence predictors.

i would like to ask how far the somatic genetic editing technology has gone, compared to gamete genetic editing like perhaps CRISPR.

There are now therapies to cure monogenic conditions like sickle cell anemia or beta thalassemia.

The therapies work by withdrawing your bone marrow, edit the stem cells in solution to insert the working version of the defective protein, give the patient a dose of really terrible chemotherapy designed to eliminate the remaining (un-edited) stem cells, then re-injecting those edited cells back into the bones of patients with the disorder.

It looks like in this particular study they actually used a lentivirus rather than CRISPR, but I've read of another study where they used CRISPR.

Last I heard though, the therapy cost $500k. At that price, you'd only ever use it for extremely debilitating monogenic mutations.

This suggests a technology with a poor level of readiness, both in it's safety and versatility (the chemotherapy is really bad for any enhancement, and even restricted to medicine, I would be surprised if most diseases could be worse. Also, monogenic alterations are almost never that impactful in life, especially positively. So it has extremely limited versatility.)

So it has extremely limited versatility

I obviously agree that in its current form this therapy has no potential for use as a means of enhancement (at least if we define enhancement as deviation above normal human baseline).

But it's worth pointing out that the chemotherapy step of the treatment was only necessary because ANY remaining sickle-cell carrying stem cells would continue to produce misshapen red blood cells which cause all of the symptoms and problems associated with the disease.

Other conditions are not all like that: going from no cells having a copy of some particular variant to SOME cells having a copy would still produce a positive effect.

The real reason I think somatic cell editing is unlikely to produce any large enhancements any time in the next 50 years is because you have to edit an ungodly number of cells all over the body to have an effect, and some of the modified genes won't do anything because they're only active during the developmental phase of a person's lifespan.

Epigenetics is somewhat comparable to Fetal Alcohol Syndrome, in that reasonable epigenetics (lack of fetal alcohol poisoning) is the norm and considered neutral. Unusually bad epigenetics (like Dolly the first cloned sheep had) can, like Fetal Alcohol Syndrome, reduce your potential quality of life and potential lifetime achievements. So... be careful not to mess up the epigenetics of your clones or you'll have a significantly less useful clone army?

Epigenetics is basically E-p-hack-netics. 

Totally unclear to me whether there are any results in that space that are robust. 

Dolly was probably totally fine. From wikipedia:

In 2016, scientists reported no defects in thirteen cloned sheep, including four from the same cell line as Dolly. The first study to review the long-term health outcomes of cloning, the authors found no evidence of late-onset, non-communicable diseases other than some minor examples of osteoarthritis and concluded "We could find no evidence, therefore, of a detrimental long-term effect of cloning by SCNT on the health of aged offspring among our cohort.

This matches my impression, though my impression is based more on "research smell" than any concrete analysis. Do you have good references?

I don't have a good reference and I left bioinformatics a couple of years ago, so I am open to being corrected by someone who is more in the loop. To me the general picture seems to be:

  1. There was a huge hype around epigenetics maybe 10,15? years ago, when epigenetics was supposed to explain everything. This hype has basically vanished. 
  2. The epigenetics results in human all seemed to look like: The effect skips a generation and only goes from grandfather to granddaughter, p < 0.038. Hence E-p-hack-netics. I think you just can't publish this kind of result any more these days. 
  3. I think our ability to read out methylation patterns has improved. For example the third generation sequencing technologies like PacBio sequencing were able to read methylation patterns directly. Third generation sequencing hit the market ten years ago and became generally accepted maybe 5 years ago, which would support a story where we gained the ability to measure accurately and the effects vanished. 
  4. There were results like with dolly, where later investigations overturned the initial theorising. 
  5. The general mechanism of epigenetics also came under fire if I remember correctly. I don't remember this very well, but something about methylation patterns resetting during oogenesis or embryogenesis?
  6. Many of the stories just didn't make sense from an evolutionary perspective. Epigenetically transmitted trauma? How does that evolve? 
  7. Generally smart guys became dismissive about epigenetics I think mostly because, like you, they caught a whiff of research smell. 

I am not very sure in the single claims, because they are based on things I read or heard over the years and generally quite a few years ago. But overall things just seemed to overwhelmingly point in the same direction. Though I do have to say that some of the mice results looked more robust, but of course those come from a much bigger sample size. 

I, too, would definitely be interested in a concrete analysis "What the hell happened to epigenetics" by an actual geneticist - ideally without a horse in the race. But right now I have a strong prior that epigenetics is mostly bunk. 

Whoa! That's a very different story from the one I was given back when studying genetics. Thanks for pointing that out.

Please add some tags into your video. Its a trivial inconvenience to people who're looking for specific content, reducing your total views.

Embryo selection is arguably not genetic engineering. Though iterative embryo selection would be close. 

My impression is that embryo selection is bound to be leapfrogged by dna synthesis or editing when it comes to super human intelligence. 

Embryo selection is arguably not genetic engineering. Though iterative embryo selection would be close.

It's a distinction without much of a difference in my opinion. It's true that embryo selection is somewhat lower risk because you don't have to worry about off-target edits and you're not introducing any genes that weren't already present in the organism. But (if you've done your editing correctly), the effect is the same: increase the frequency of certain alleles positively associated with a trait of interest.

My impression is that embryo selection is bound to be leapfrogged by dna synthesis or editing when it comes to super human intelligence.

It may be the case in the long run that editing or DNA synthesis overtakes embryo selection, but we're still trying to figure out how to CRISPR embryos without lopping off random chunks of chromosomes.

And when I last checked a couple of years ago, synthesizing a whole human genome would cost approximately $200 million just for the sequence alone without even taking into consideration the technical challenges of sticking all the oligonucleotides together and getting the whole thing into the nucleus of a zygote.

I would place heavy odds on iterated embryo selection or iterated meiotic selection working before we're able to fully synthesize genomes.

Interesting claim. We specifically asked him that and he didn't think that was the case, but you could be right!

There is very steady technological progress in both. And generally more potential. But that's only the technological side, where I think leap-frogging is likely or already happened. 

I think there are very significant political hurdles to actually applying genome synthesis or gene editing for intelligence in humans. He probably rightly expects that those won't be overcome, while embryo selection has an easier "in" via IVF where you have to select an embryo anyway.

Yeah. It's pretty crazy to imagine that DNA synthesis is happening anytime soon. Meanwhile, embryo selection is safe and effective now, and getting better all the time.