Cryonics As Untested Medical Procedure
If you’re trying to prevent information-theoretic death by preserving the brain it’s critical that the information that makes you be “you” actually be preserved. If you could freeze the brain in a way that did keep around the necessary information then some future civilization might be able to recover the person or the memories, but if the information is gone it’s gone for good. The problem is, this is an untested medical procedure, and it’s not something we should expect to get right flying blind.
In freezing a brain there are obvious things that can go wrong. For example, if you just cool it down to below freezing the water in the cells will turn to sharp little ice crystals, disrupting synapse structure and making a huge mess. We know about this now, though, so since the early 2000s cryonics organizations have used “cryoprotectants” which are able to vitrify the brain tissue and reduce [1] ice crystal formation. Beyond these known problems, however, there are many aspects of the brain structure that might or might not be relevant. Is information stored in the positions of proteins within the cells? Are phosphorylation states significant? What scale of preservation is sufficient?
Our normal approach is to try something, see if it works, fix apparent problems, and try again, each cycle getting us closer to something that does work. With cryonics the “see if it works” step isn’t there, and there’s only “check for known failures”. So what we should expect is that the current process will be “good to the best of our knowledge” and then repeatedly our knowledge will expand about what matters and the process will need to be updated.
(Situations where current preservation technology fails to preserve something we know is required are actually kind of nice, because they’re as close as we get to cryonics as an experimental science. Those are the cases when the process can actually improve because the feedback loop is temporarily closed.)
Imagine if in the development of In-Vitro Fertilization an inexplicable barrier stopped researchers from continuing any experiments past the “combine egg and sperm” stage. Instead they worked out something they thought was as good as they were going to get, documented it, and started freezing hopefully-fertilized eggs. How likely would it be that later we would be able to take these frozen eggs and complete the process? Much more likely would be that something unknown was wrong with the beginning of the process and these eggs would actually not be usable. Given that the brain is so much larger and more complex than these zygotes I expect the odds in the cryonics case are much worse.
Cryonics depends on a complex medical procedure developed under conditions of minimal feedback. Expectations for success like 80% or even more likely than not seem incredibly optimistic. When you can’t test the output of a process because you don’t know what counts as correct output it’s very unlikely you’ve got the process right.
(I also posted this on my blog.)
[1] I say “reduce” instead of “eliminate” because as far as I can tell no one has actually taken random samples from a human brain that’s been preserved with vitrification. There are ethical reasons why the cryonics organizations would not want to do this, but there being reasons why we don’t wish to run a test doesn’t mean we can act as if we already know the answer.
- The Brain Preservation Foundation’s Small Mammalian Brain Prize won by (9 Feb 2016 21:02 UTC; 72 points)
- 's comment on Open thread, Dec. 1 - Dec. 7, 2014 by (1 Dec 2014 12:05 UTC; 22 points)
- 's comment on Rationalists Are Less Credulous But Better At Taking Ideas Seriously by (23 Jan 2014 23:28 UTC; 5 points)
There are projects pointing the way towards testability:
OpenWorm and Nemaload are working on uploading C. Elegans and determining exactly what you need to get an accurate simulation. This is necessary (but probably not sufficient) to figuring out what you need to preserve in a human brain to get approximately-you back.
The Brain Preservation Foundation is running electron microscope analysis of preserved brains to see exactly how well cryonics and chemical fixation preserve things right now.
A while back I did a survey of the state of nematode emulation research to try and understand what this meant for whole brain emulation. At the time David Dalrymple was very optimistic about his Nemaload project; don’t know how it’s going though.
For the current state of the Brain Preservation Foundation’s prizes, two Reddit AMA comments by Ken Hayworth are interesting:
And on chemopreservation:
This is good to hear. It will be even better to hear that it’s been well-replicated.
Has anyone noticed that “success” vs. “failure” is not really binary in cryonics? The question is not whether information is preserved but how much information is preserved. In other words, the hypothetical future civilization restoring the cryopreserved person is going to get some person, the question is how similar she will be to the original. One limiting case is zero preserved information which means “restoration” is going to produce a random person. Another scenario is in which only genetic information is retained, so we get essentially an identical twin of the original. If the restoring entity has wide knowledge about the culture to which the person belonged it adds a lot of information. If it has access to e.g. blog posts / tweets / facebook posts / less-wrong comments made by the person, it has a whole lot more information. Who knows, maybe you can get quite close without any physically preserved brain at all.
How close does the restored person have to be to count as “the same” as the original? Of course most people would require at least some extent of memory restoration for it to count as “success”. However, I don’t think there is an unambiguous answer to this. What we have here is a continuous scale, not just two points “success” and “failure”.
“at least some extent of memory restoration …”
This seems like a useful criterion. I would say “all the memories of cryonauts are very likely permanently gone” which I think is stronger than “cryonics is unlikely to prevent information theoretic death”.
“Who knows, maybe you can get quite close without any physically preserved brain at all.”
By what ratio do you think cryogenically preserving the brain improves the chances that someone you identify with will exist in the far future?
I am reluctant to give a ratio but my guess is that the improvement is significant. Personally I am not thrilled by cryonics for a completely different reason, namely I’m not sure the value of restoring my life at a point in the future in which civilization has advanced much beyond its current state is more than the value of things I can do with my money in the present, in particular things that increase the probability this advanced civilization will actually come to pass. Also we don’t have cryonics in Israel so I don’t have to decide now anyway.
I assume the ethical reason is that one shouldn’t damage the brain of someone who gave it to you with the expectation that you wouldn’t damage it. This suggests a thought to me. I’m pessimistic about the success of cryonics procedures, which is one of the several reasons I haven’t signed up for it. But I approve of the attempt in principle. Perhaps it would be desirable/ethical/useful to donate my brain, upon death, for cryonic experimentation of exactly the sort you mention. If I believe the probability of success with current techniques is low enough, and if the information gained will help improve the techniques or provide a better estimate of their efficacy to those who come after, then it might be a net positive over getting frozen with intent to live.
I think the usual solution is to do those tests on animals. Vitrifying and testing a pig brain should give us enough info.
“one shouldn’t damage the brain of someone who gave it to you with the expectation that you wouldn’t damage it”
Definitely, but I think they would also refuse to experiment on a brain given to them without that expectation. The people at cryonics organizations don’t tend to take the common-on-lesswrong view that someone cryogenically preserved is probably dead but has a small chance of actually being recoverable. To them a “cryonaut” is stll alive. For example, Alcor’s website says:
So while you would view it as acceptable, knowing what you were getting into, and I would view it as acceptable, given the very high chance that your brain no longer contained you, people at Alcor or CI would likely find it unethical experimentation on a still-living human.
(I don’t know how old you are, but if Alcor or CI were willing to participate in a test like this it would probably make more sense to find someone already about to die so we could run it sooner.)
I hadn’t considered that. Are the techniques they use publicly documented in detail? Are there other organizations that do not hold such views, that would be willing and able to perform such studies, even if they don’t do cryonics themselves?
I’m 31. I don’t expect to die soon.
“Are the techniques they use publicly documented in detail?”
Yes. See this Reddit AMA comment by Ken Hayworth. 21st Century Medicine is applying similar techniques to preserving a mouse brain.
“Are there other organizations that do not hold such views, that would be willing and able to perform such studies, even if they don’t do cryonics themselves?”
Probably. And to be fair I haven’t asked Alcor or CI about this, so I could easily be wrong.
If IVF got to the ‘combine egg and sperm’ stage and they knew that they didn’t know what to do next so they froze it in liquid nitrogen, waiting for appropriate methods to come along, I wouldn’t think it terribly unlikely that it would, in the end, work.
The worse a job we do up front, the harder it is in the end. If we figure that they will eventually get really really atomically good at scanning, then the only question is, ‘is the information still there’. That’s where the 80% (or higher) figures come in. Also, some of the high end estimates aren’t about present tech but about best possible tech. Like, I’m around 98% sure that some cryo-like technique would work to preserve the information so that an arbitarily powerful but physically possible future entity could recover it. I’m noticeably less sure that our present techniques work (but still pretty confident).
Well, we can go back to the basics of what’s happening:
You have a 1.5kg sphere of water. You immerse it in liquid nitrogen.
How long does it take to freeze all the way through?
What does the pattern of ice crystal formation look like?
Repeat with water with cryoprotectants in.
Look at it from the other side: In order to achieve information-theoretic death, it is critical that the information that makes you be “you” actually be lost.
By “lost” we mean it has to be scrambled at least enough that superintelligent computronium dyson spheres aren’t going to be able to (reasonably) crack the code.
So let’s say you dissolve the brain in acid. That is likely to be a good way to achieve information-theoretic death.
Leaving it to rot for a few days? Probably.
Freezing it in ice crystals? Maybe.
Vitrifying it? Probably not.
Any given bit of data is likely to be stored in multiple areas by multiple mechanisms, with lots of redundancy. Moreover, every time data is stored or accessed by some mechanism, there should be side effects, things you can infer the data from that aren’t part of the mechanism. The complexity of the brain works in our favor, not against—assuming we can develop good enough reductionistic models of the brain to account for all the details.
But we don’t have the details, do we? Suppose cryonics preserves the graph of neural connections; we can look back in 10000 years and know exactly how many neurons you had, and exactly which were connected to which others. Is that enough information to revive you, or someone who’s 90% identical to you? Who knows? We really have no idea.
Your point about redundancy is, I think, looking at it from the wrong angle. I would expect brain redundancy to handle random errors like “lost 2% of neurons”, but the idea that we would have multiple fundamentally different mechanisms for encoding memories seems evolutionarily implausible. If we simply haven’t preserved anything about, say, the thicknesses of the glial cells, or we know which synapses are present but it turns out they have different sized gaps and that’s important and we can’t recover that information, or any one of thousands of other pieces of brain biology that might turn out to be vital for encoding memory, then cryonics won’t work.
That’s not the mainstream position of actual scientists who study these things, at least not for the kind of procedures that cryocompanies use.
The phrase neural archaeology evokes the right kind of thinking here. Or Eliezer’s reference to secure deletion.
Eliezer’s secure deletion discussion:
This argument amounts to “it might be possible” or “you can’t prove it’s not preserved”. This is true, but it’s not a reason to think “it is probably preserved”.
Eliezer’s hard drive comparison is actually wrong. As I commented on Timeless Identity, Peter Gutmann, who wrote the original list of steps to securely erase a disk, is particularly annoyed that it has taken on the status of a voodoo ritual. “For any modern PRML/EPRML drive, a few passes of random scrubbing is the best you can do. As the paper says, “A good scrubbing with random data will do about as well as can be expected”. This was true in 1996, and is still true now.”
This isn’t directly relevant to the question of memory in a brain—but it wasn’t then either, because it just isn’t a very apposite analogy to use in thinking about this question.
I’ve also commented (not in the original thread, can’t remember where) that the hard drive is a very much cherry picked analogy. Substitute it with DRAM and you get the opposite result: information-theoretic “death” within minutes of power loss at room temperature, a few weeks or months at most at liquid nitrogen temperature.
Of course the human brain is neither a DRAM nor a hard drive. Rather than arguing from analogies I think it’s better to listen to actual domain experts: neurobiologists and cryobiologists.
Yep. I put up this hypothetical before: Drop an iPhone into liquid nitrogen, slice it up very thin. Now recover the icons for the first three entries in the address book.
At least in this case we would expect it to be possible for someone with enough money and time, with today’s technology. You should be able to recover the contents of the hard drive.
The domain-expert (Gutmann) says otherwise. At this stage, it’d really take an example of data recovery in practice, not just in “you can’t prove I’m wrong!” hypothetical.
(I’m assuming you don’t have an example to hand of having recovered data yourself in this manner.)
I read Gutmann as talking about what you should expect, security for the real world. I don’t see where they talk about someone willing to put in an unrealistically huge amount of effort. But maybe I missed that? Could you point me that way?
It is true that I can’t philosophically prove that arbitrary hypothetical technology that would achieve something currently nigh-equivalent to magic cannot possibly exist, nor can I philosophically prove the data isn’t there any more, yes. I can say there is no evidence for either, and expertise and evidence against both, and that “but you can’t prove it isn’t true!” isn’t a very good argument.
“For any modern PRML/EPRML drive, a few passes of random scrubbing is the best you can do … A good scrubbing with random data will do about as well as can be expected”
But what does that mean? Can someone with an STM and lots of patience still get the data back? Or is it just “gone for our purposes, with today’s technology”?
What you need to realize is that for 2 states to be distinguishable ever in principle, the states must be separated by an energy barrier taller than thermal fluctuations. Else the thermal noise is going to overwrite it randomly a zillion times a second.
The other issue is that the closer are the states the less metabolic energy you’ll need to switch between them. Which makes something like neurons (evolved over a very long time) settle on an optimum where there’s no room for some weak residuals recoverable with some future technology that got more sensitive probes.
I.e. if the cryo-protectants happen to reset some bits, that information is gone. You have to hope that cryo-protectants do not actually reset anything, i.e. that nothing is forced from multiple states to one state.
edit: another issue. Individual ion channels, gap junctions, etc etc. combine more-or-less additively into final electrical properties of the neuron. When you need to know the value of a sum a+b+c+d+e , losing even a single variable of the sum introduces massive uncertainty in the result. It would’ve been a lot easier if those properties mirrored each other, like a=b=c=d=e , then we’d only need to preserve at least one, but as they combine additively, we need to not lose a single one.
As I note below, if you really want to hold on to this particular example for analogical purposes, it’s at a stage where “you can’t prove it’s false!” isn’t really adequate and you’d need to produce an example of recovering data in practice, not just hypothetically.
The data is potentially recoverable from hard drive wiped with zeroes solely because the underlying medium is capable of storing more information—the 0 overwritten by 0″, 0 overwritten by 1, 1 overwritten by 0, and 1 overwritten by 1 are, potentially, distinct states of the platter.
Synapses, on the other hand, store synaptic strengths on molecular scale, where if the state information is lost, it is completely gone, as the molecule which can be either in the state A or state B does not have state of “B but it used to be A” or “B and it used to be B”.
edit: bottom line is, the energy barrier between different values has to be considerably larger than kT for the data to persist at all . The synapses already store the data at an energy level low enough that if you go much lower—for the analogue of residual magnetization on a drive which is much less energetic than original magnetization—you’re in the region where the data simply can’t be stored.
No, the argument is “it might be difficult to recover[1], but it is incredibly unlikely that it’s destroying enough information to make it actually impossible to recover”. Which is true, and which logically implies “it’s probably preserved”.
[1] relative to an unclear standard, but presumably based in terms of computation need to recover brainstate as a multiple of whatever the fastest process which could recover a brainstate based on a perfect instantaneous view. (ie. a copy of every particle’s precise state as the brain fell asleep)
Why do you think “it is incredibly unlikely that it’s destroying enough information to make it actually impossible to recover”? Where did you learn it? Did the “secure deletion” argument convince you of it or is it something you believed before?
The secure deletion argument convinced me, yes. It’s a compelling analogy, in that it points out how difficult it is to actually destroy information, even in a minimally-redundant medium when you’re specifically trying to destroy that information. A process in a brain, which is a highly-redundant medium, when specifically trying not to destroy data, is incredibly unlikely to make information unrecoverable.
Hard drives aren’t minimally redundant. The size of the magnetic regions on the platter is bounded from below by the requirement that the heads must be able to read and write them while passing over them at a very high speed.
Furthermore, hard drives are a very stable medium: they are designed to reliably retain data for years or decades without power (possibly they may retain data even for centuries if the storage conditions are right).
I think it’s a bad analogy, and a cherry picked one. Contrast with how easy it is to delete data from a DRAM chip, and you’ll get why analogies with modern computer hardware don’t really make any sense for biological brains.
Except that the analogy is wrong: it’s quite easy to destroy the information. In practice, writing random data to a modern hard disk leaves it unrecoverable.
So:
Why did the idea that hard drives (a completely different form of data storage, something specifically designed to retain data across a wide range of conditions) are hard to erase make you think that it was hard to erase data from brains?
Now that you know that it isn’t hard to irrecoverably erase hard drives (even though they’re designed to retain data), how does that affect the analogy with brains? Why?
The analogy is not wrong. As you quote, it takes multiple passes deliberately trying to destroy the information to remove it.
Or an external degaussing magnetic field. Or heat. These methods make the hard drive unusable, but they reliably destroy information.
The weight of subject-expert opinion appears to be that it’s not recoverable unless and until you can show it is, in fact, recoverable. If you’re asserting otherwise, the first thing you’d need would be a counterexample.
I note also you’re supporting an expert opinion that agrees with you and denying an expert opinion that disagrees with you … when they’re linked opinions from the same expert.
No, I’m pointing out that the ‘expert opinion that disagrees with me’ doesn’t, in fact, disagree with me. The quote you yourself provided does not support your position.
“Friendly AI as Untested Medical Procedure”. :-)