Welcome to the latest in a long series of related questions.
The topic: usefully-large intelligence enhancement of adult humans.
The means: anti-inductive, novel, obscure, unsafe, risky, and/or crazy. (This thread is NOT medical advice.)
I’d be sorely disappointed if nothing new has come up since July 2022, so let’s see what’s come up!
EDIT: The full post is now up
Oh boy do I have a response for you.
I think it may be possible to significantly enhance adult intelligence through gene editing.
The basic idea goes something like this:
There are about 20,000 genetic variants that influence fluid intelligence
Most of the variance among humans is determined by the number of IQ-decreasing minor alleles someone has.
If you can flip a significant portion of those IQ-decreasing alleles to their IQ increasing counterparts, you can likely significantly increase someone’s intelligence
The effect size is going to be smaller than it would be if you made those same edits in an embryo because some of the genes you’re targeting are only active during development. But my best guess at the moment is that we would still expect a gain of several standard deviations. However I am not very certain about this because I have not yet gotten access to SOTA genetic predictors of intelligence.
There are a million little details to get into, especially those related to the delivery of an editing vector, avoiding a negative immune response and avoiding off-target edits. But after researching this with a couple of collaborators for the last month and a half, I am starting to think this is going to be possible.
What’s more, there are already several clinical trails underway right now that plan to use the same gene editing delivery platform that I have in mind for this kind of adult intelligence enhancement.
IF one could get this protocol to work, the actual experience of the procedure would be kind of magical: you’d literally get an intravenous injection (and possibly some medication to temporarily suppress your immune system) and your fluid intelligence would improve by a couple of standard deviations within about a week. I suspect it would take further months to years for the full benefits of the change to become clear, since crystallized intelligence is what really determines outcomes.
It’s difficult to predict how long it will take to roll out something like this in an actual human trial, but I think it’s plausible we could have something working within 5 years, which might be soon enough to significantly impact the trajectory of AI.
I’m working on a longer post about this, so I’ll ping you when it goes up.
The “weights” of crystallized intelligence are adapted for old substrate, changing the substrate might damage ability of existing “weights” to perform the old computations. So the experience might also be like getting dementia and then hopefully recovering into a smarter person.
How’d the payload get through the blood brain barrier?
Are there any alleles which if flipped would boost IQ but cause some major harm e.g. Torsion dystonia?
If so, would you offer people the choice of whether they wanted to flip said alleles?
Presumably there are a bunch of alleles correlated with personality. If a large chunk of them also correlated with intelligence, plausibly you’d kill the person’s psyche and replace them with a stranger with similiar memories. Do you think this could be avoided/isn’t a problem in the first place? If so, how do you know that?
Would you do challenge trials?
Would you do your work in a country with laxer rules, both de dicto and dejure, around challenge trials/genetic engineering?
If you succeed, or heck even during development, would you open source your process so that ~anyone can replicate/improve on your results privately?
How would this treatment affect the elderly? Like, would it make them several SD above the mean fluid intelligence amongst their age group or would it have a bigger impact because they’ve got deficits in fluid intelligence or what?
Buddy, if this looks like it might work out, I want to help out. So of course, I’d like to know more about this whole endeavour. Please contact me.
All good questions.
Using a targeting ligand like angiopep attached to the outside of a lipid nanopaticle. There are papers where researchers have already done this.
Yes, probably, though I don’t think it will be particularly common. The way to avoid this is just to incorporate a bunch of predictors for other traits into your allele selection criteria.
Obviously yes, though hopefully we just wouldn’t need to flip those alleles in the first place. I’m still uncertain about the maximum effect size.
Maybe? Our predictors for the genetics of personality still aren’t very good, so it would be hard to evaluate the expected effect size.
I’m not sure what that would mean in this context. I think the first human trials would probably target polygenic brain diseases like Alzheimer’s. If the delivery platform works, you could use it to dramatically adjust someone’s polygenic risk score for many diseases. So you might as well do the first trials on people who are going to die from a degenerative brain disease with no cure or effective treatment, for which in-vivo editing might offer a solution.
I don’t know the answer to this yet. Most of the pioneering work in this field has happened in the US, so my suspicion is it would start here. Pretty much every step except perhaps the human trials could be done in the US.
I would have to think about this one more. If this tech actually works it would become incredibly valuable. Because you could use the delivery platform to basically target any disease of your choosing by adjusting someone’s polygenic risk score. There are still questions around whether you can repeatedly dose someone with the editing agent without provoking an immune response, but there are plausible solutions to all the problems I’ve looked at so far. My biggest concern though relates to AI; I really don’t want this tech to be used to speed up development of world-conquering AI. The main reason I’m interested in the tech is because I think it could potentially result in people capable of solving some of the more difficult problems in AI alignment and coordination around its appropriate use. If I were to open source it I would lose control of that.
And I think in reality there will be a lot of stakeholders in the technology so the decision won’t be mine alone. If I want to bring something like this to market, I’ll need to raise quite a bit of money and work with a lab to run animal trials. So there will be many other stakeholders with input into that decision. I suspect open-sourcing the tech would also diminish the financial gain. 8. Hard to answer, but my default assumption is they would be just like an elderly person with a very high IQ. 9. I’ll let you know when the main post goes up.
What do you mean by challenge trials in this context?
In the vaccine context, a challenge trial is a trial where you expose people who get the vaccine to the virus.
That makes little sense in the context of intelligence improvement.
E.g. pay people who are planning on Euthanasia to undergo the treatment, and if it fails then handle the euthanasia.
Do such people partake in dangerous medical trials today? I couldn’t find anything about it from some quick googling but if such practices don’t already exist, I doubt we would be able to start them for what would be, in most cases, an intervention not designed to cure a disease or treat a disability.
I’m not sure why you would call that a challenge trial.
I would also be surprised if that’s something you can effectively do.
Big if true, thank you for agreeing to ping me about this!
The “delivery vector” alone would be the big progress, I’d imagine, though I’m no expert. My experience so far is: knowing high school biology + watching that one YouTuber who “”“”cured”””″ his lactose intolerance with simple gene editing + the HackerNews comments about that YouTuber being a risk-blind idiot who was giving himself a cancer risk.
I wish this endeavor the best and will follow closely!
This is a bit of a tangent but what made you pick those ‘experiences’ instead of just watching some lectures from biology profs and reading some books?
I think saying I “picked” those experiences is giving me too much credit haha. In reality, it’s more “my time and mental energy are constrained, and what little I had left was put into AI alignment”. The rest was more free-floating attention; if I came across it (a youtube video, a Wentworth post, the Intelligence Enhancement tag on LW) I’d pay more attention, but seeking out lots of biology background knowledge is more on the “strategically become an intelligence-enhancement researcher myself” path. Not saying I’ll never do anything like that, but I’m too constrained to go after that in a remotely serious way right now.
Loosening such constraints and increasing my mental capacity are, in fact, key reasons why I keep making these threads.
This would be very exciting if true! Do we have a good (or any) sense of the mechanisms by which these genetic variants work—how many are actually causal, how many are primarily active in development vs in adults, how much interference there is between different variants etc?
I am also not an expert at all here—do we have any other examples of traits being enhanced or diseases cured by genetic editing in adults (even in other animals) like this? It seems also like this would be easy to test in the lab—i.e. for mice which we can presumably sequence and edit more straightforwardly and also can measure some analogues of IQ with reasonable accuracy and reliability. Looking forward to the longer post.
No, we don’t understand the mechanism by which most of them work (other than that they influence the level of and timing of protein expression). We have a pretty good idea of which are causal based on sibling validation, but there are some limitations to this knowledge because genetic variants physically close to one another on a chromosome are highly correlated, by which I mean if you can usually predict the value of one variant using the value of the other. Maybe the simplest way to describe this would be to say that we know the approximate location of causal variants, but aren’t always certain exactly which of a few neighbors is causing the effect.
86% of all genes are expressed somewhere in the adult brain, so I think it’s quite likely we can have a significant effect on brain function via editing. Ideally we would be able to establish some kind of prior on the expected effect size of editing a genetic variant in an adult brain (perhaps based on the timing of its peak level of expression throughout the lifespan?), but the tech could still work without this. It would just take more edits to achieve the desired benefit.
I also don’t think this effect could be replicated by medications because the effect genes have on cellular functioning can take into account the current cell state in a way that medications cannot.
As I wrote in my post on “How to have Polygenically Screened Children”, there are very few measurable non-linear interactions among genes that vary within the human population. The reason for this is actually pretty simple: genes with linear effects have an easier time spreading throughout a population.
It’s possible there are more interactions than we currently believe and that our sample sizes simply aren’t large enough to detect them, but the fact that we’ve found so few indicates that their influence must be small relative to simple additive effects.
I’ll talk about this in more detail within the post, but yes we have examples of monogenic diseases and cancers being cured via gene therapy.
Thanks for the response! Very helpful and enlightening.
This is interesting—I have never come across this. Can you expand the intuition of this model a little more? Is the intuition something like in the fitness landscape genes with linear effects are like gentle slopes that are easy to traverse vs extremely wiggly ‘directions’?
Also how I am thinking about linearity is maybe slightly different to the normal ANOVA/factor analysis way, I think. I.e. let’s suppose that we have some protein which is good so that more of it is better and we have 100 different genes which can either upregulate or down regulate it. However, at some large number, say 80x the usual amount, the benefit saturates. So a normal person is very unlikely to have 80⁄100 positive variants but if we go in and edit all 100 to be positive, we only get the maximum benefit far below what we would have predicted since it maxes out at 80. I guess to detect this nonlinearity in a normal population you basically need to get an 80+th order interaction of all of them interacting in just the right way which is exceedingly unlikely. Is this your point about sample size?
This is very cool. Are the cancer cures also monogenic? Has anybody done any large scale polygenic editing in mice or any other animal before humans? This seems the obvious place to explicitly test the causality and linearity directly. Are we bottlenecked on GWAS equivalents for other animals?
The way I think about the sample sizes needed to identify non-linear effects is more like this: if you’re testing the hypothesis that A_i has an effect on trait T but only in the presence of another gene B_k you need a large sample of patients with both A_i and B_k. If both variants are rare, that can multiply the sample size needed to reach genome-wide significance by a factor of 10 or even 100.
The ones I know of are.
If this tech works, it’s hard to understate just how big the impact would be, especially if you could target edits to just a specific cell type (which was been done in a limited capacity already). If you had high enough editing efficiency, you could probably bring a cancer suvivor’s recurrence risk back to a pre-cancerous state by identifying the specific mutations that made the cells of that particular organ pre-cancerous and reverting them back to their original state. You could even make their cancer risk lower than it was before by adjusting their polygenic risk score for cancer.
I really don’t think I can oversell how transformative this tech would be if it actually worked well. You could probably dramatically extend the human healthspan, make people smarter, and do all kinds of other things.
There are of course ways it could be used that would be concerning. For example, a really determined government might be able to make a genetic predictor for obedience or something and modify people’s polygenic scores for obedience. On the other hand, you could probably use that same technology to reduce the risk of violent criminals reoffending, which could be good.
I tend not to think too much about these kinds of concerns because the situation with AI seems so dire. But if by some miracle we pass a global moratorium on hardware improvement to buy ourselves more time to figure out solutions to alignment and misuse concerns, this tech could play a hugely pivotal role in that. Not to mention all the more down-to-earth stuff it could do for diseases, mental disorders, suffering, and general quality of life.
There’s so many cool ideas we could test if we were braver. Stuff like the variation of p53 in elephants which is a nonlethal variant of which your cells can safely have extra copies alongside the normal lethal variant to reduce cancer likelihood without negative side-effects of increasing undesired cell death.
Or the function of seal glia cells in taking on some of the metabolic load of the neurons by having mitochondria to regenerate ATP and shipping the resulting ATP over to the neurons.
One thought: One could probably do mice studies where instead of maximizing a polygenic score, non-consensus variants are edited to reduce mutational load. If that had positive effects it would be a huge result.
That’s a decent idea. I had thought to do it with cows or some livestock animal for which we do have decent polygenic predictors, but mice are of course the obvious first step for animal trials.
Somatic gene editing was in cards for a while now, but I assumed that so far off-target effects would make that pretty risky, especially for a large number of variants.
What is the current situation regarding off-target effects for large numbers of edits?
Significantly better due to base editors and prime editors, neither of which induce double-stranded breaks
The next question would be the success-rate. Even successful somatic gene editing I’ve read about so far have modified only a small fraction of cells. Is it realistic to modify a double digit percentage of neurons in the brain?
I believe it is at least plausible that we should be able to do this. We have examples of mouse studies where they’ve modified autism-associated genes and showed significant reductions of behavioral issues. It looks like they were able to modify a particular gene in about 39% of the cells (86% for the treatment samples vs 49% for the control sample).
I suspect we could increase the efficiency to edit a greater proportion of the cells.
I do this this technological hurdle can be overcome with sufficient research effort. Do note that part of the challenge of somatic editing of human brain cells is the absolute number of cells to be edited, and the volume of space those cells occupy, which are exponentially different in large human brains vs small mouse brains. These factors trade off against things like immune response time, which doesn’t vary based on brain size. So yeah, the challenge in a human-size brain will be bigger. How much bigger? I dunno. Still, I absolutely think it makes sense to start with small cheap animal models like mice and juvenile zebrafish, and then move into larger models like pigs.
Name checks out—would like to read that post.
I certainly think that such a thing is possible. This is the general sort of thing I was studying back when I was trying to figure out human intelligence enhancement for the sake of helping solve the AI alignment problem.
I do think that successfully modifying a large percentage of existing neurons is a big challenge. Not insurmountable, but a big challenge in and of itself beyond just knowing what changes you’d ideally like to make.
I also think that getting adequate funding and permissions for doing this research, even just on animal subjects, within a 5 year time frame, is a substantial hurdle. To get approval for human testing? Whoa. There’d need to be some dramatic shifts in public policy. Not theoretically impossible but.… That seems to me to be the biggest hurdle. Of course, if you got as far as successful animal testing, and then got blocked on moving to human testing because of politics, and yet the treatment process was quite technically easy to accomplish.… well, there’s always the potential for some brave volunteers taking a boat trip into international waters and doing some private things they choose not to disclose.
Agreed. Hopefully the post I’m working on may catalyze interest among funders.
You can do almost all the human testing necessary to prove this technology in the context of other diseases and mental conditions.
Tier 1 would be pretty simple: find someone with a monogenic brain disease and try to get an editing agent into their brain to fix it.
Tier 2 would be to attempt to modify polygenic risk score. For example, imagine someone has treatment-resistant depression that doesn’t respond to therapy or medication, and they are suicidal as a result. We could potentially modify their genes to reduce the propensity for depression (particularly if their polygenic risk score is already unusually high).
Once you’ve done something like that, pretty much all of the pieces are in place. You will have tested literally everything except the genetic predictor for intelligence.
Good luck! Not a horse I’d bet on, but I’m glad you’re in the race. Diversity of approaches buys us micro-alignments! :-)
So I made this comment awhile back, though I admit being ignorant on how good modern somatic gene therapy is:
Am I just wrong here? Is somatic gene therapy really robust and error-free enough to safely edit billions of cells?
This is probably not true unless you start out with an embryo that already has a very high polygenic score for intelligence. To a first approximation, 15 IQ points = 75 edits of causal variants. And since we aren’t precisely sure which variant is causal, the needed number of edits will be higher (perhaps 100-300).
It depends heavily on your editing vector and your target site. Base editors and prime editors have significantly higher ratios for on-target / off target edit rates. They also don’t introduce double strand breaks like CRISPR does, which means you won’t randomly get entire cells being eliminated if the DNA repair machinery doesn’t work.
I am writing a more detailed post about this that will answer most of these questions. In short, the answer to your question is maybe. It is much more plausible we could make edits in billions of somatic cells than I would have though even two months ago.
I think I may be almost crazy enough to volunteer for such a procedure, ha, should you convince me.
If three years has passed, and substantial capabilities progress has occurred and we still don’t seem close to a solution to alignment, and good progress was made on this project in animal studies.… I’d definitely volunteer. The EV calculation (for the safety of society and for my loved ones, not just for myself) seems pretty clear.
This comment was assuming causal variants are known, which I admit is a big gimme. More of a first-principles type eye-balling.
Recently I learned that the negative effect of sleep deprivation on cognitive performance seems to accumulate over several days. Five days of insufficient sleep can lower cognitive performance by up to 15 IQ points according to this source.
This definitely tracks with my experience, especially since (apparently) the normal work/school week is exactly 5 days long, meaning 5 nights in a row of restricted sleep.
Aerobic exercise. Maybe too obvious and not crazy enough?
Totally anecdotal and subjective, but the year I started ultrarunning I felt that I had gotten a pretty sudden cognitive boost. It felt like more oxygen was flowing to the brain; not groundbreaking but noticeable.
I could attest to this! Had the same experience doing multi year marathons..
I think that people should at least try to switch from social media/entertainment media to reading high-impact fanfiction including projectlawful and the second reread of HPMOR (idk if a third reread is valuable, I only reread HPMOR once and with a 6 yer gap). Social media might push people closer to burnout while only providing the illusion of being relaxing, by using gradient descent to optimize for whatever combinations of posts keep people coming back. It’s plausibly better to think of your daily slack time as “I will satisfy the valuable parts of my brain” via HPMOR and projectlawful etc. than to think of daily slack time as “I will satisfy the less valuable parts of my brain” like social media, base entertainment, etc.
I find it fairly likely that social media’s algorithms, by optimizing for use-maximization (which is highly measurable, in hours per day) will also induce a state of mind like “what now?” when you stop, or other ways that your routine feels hopelessly doomed or ruined if you don’t fill it with social media. This kind of thing isn’t the devs fault, it’s just that gradient descent finds all kinds of combinations of posts and content that causes the result of keeping people coming back and spending more time on the social media platform, even if it triggers urges in large numbers of weird ways. The best solution is to try several days doing something other than social media, that builds the habit of doing something else and moves your mind OOD relative to what the social media systems are built around.
This seems like one of the biggest bottlenecks to getting more intelligence out of the AI safety community ASAP.
Has anyone else looked at Sarah Constantin’s writings on Ultrasound Neuromodulation?
Remember that the most low-hanging-fruit intelligence enhancement is reducing “IQ decline” due to dumb reasons (eg microplastics, pollution, shitty diet, “default mode network noise”/trauma/excess central coherence/unaligned brainwaves)
[you can easily cut microplastic consumption by 50% with semaglutide]
Transcranial magnetic stimulation is worth trying (+not uncomfortable—you can do things while being TMS’d), as well as low-intensity focused ultrasound (openwater.cc), photobiomodulation, and high-frequency terahertz (THz) waves… Pollan’s “How to Change Your Mind” should have included these modalities too.
[low-intensity focused ultrasound is known to break ultra-crystallized structures in the depressed, making the brain more plastic]
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3085788/
Also Neuromyst for tDCS/tACS
https://cassondraeng.github.io/current.html
Plasmalogens as brain nutrition (they are SUPER-underinvestigated)
The effect sizes probably are not huge (like everyting else) but worth trying
also I have a friend who uses “100mg NSI-189” to be smarter which is like 10x the rec’d dose
short timelines only advance the argument for trying bromantane, cortexin, cerebrolysin… [some people have disproportionate returns, and some in the community have kits...]