Nathan Hwang

9s of Cats

Epistemic status: value judgement.

The internet has a lot of cat pictures.

Let's say I upload a cat picture to Amazon's Simple Storage Service (S3). As of writing, their marketing materials claim that a stored object is 99.999999999% likely to stay securely stored in a year, which translates into a 50% chance of losing a given cat picture once every 70 billion years years[1]. In storage/networking jargon, this is 11 9's of durability, a sort of fast n' dirty logarithmic shorthand for stating how reliable a service is found by counting the 9s in a percentage. For example, 99.9% would be 3 9's.

This doesn't mean that Amazon is super optimistic and thinks the chance of total nuclear war or perfect storm pandemic is some tiny percentage. It's just that if civilization does collapse then former customers would want Amazon warriors over Amazon refunds. Conditional on the continued existence of Amazon, the business, they'll probably keep doing crazy replication schemes[2] to maintain those crazy guarantees.

However, smaller apocalypses will leave Amazon broken while humanity lives on[3]. In these futures, I could easily imagine children gathered around a working fin de sicle computer wondering why in the world this cat looks so grumpy?

So certain cat photos might in fact have 11 9's of durability, enough to live 9 lives over and over. What about humans?

What about humans? Looking at the 2014 CDC death rates, there are 823.7 deaths/100,000 people, working out to a 99.18% annual durability for a randomly selected human (American), for a measly 2 9s of durability. If you show someone a cat picture when they are 12, at best you can expect them to hold onto that memory with 2 9s of durability, because after that they are likely dead[4].

Cat pictures hold together with 11 9s: humans hold together with 2 9s.

It seems a little incongruous, yes? One is a chuckle-worthy image, and the other is a person.

I mean, there is a good reason, one is much more complicated than the other. Grumpy Cat herself will die far before her image does (maybe that's why she's grumpy?). We can barely simulate nematode neural systems, and even simply finding a human's brain connectome (connection graph) is still prohibitively expensive, much less running the entire graph forwards in time[5].

Instead of doing the naive thing easily suggestible by the S3 analogy and trying to scan people to replicate them across availability zones[6], we could simply extend their lives. For example, we boosted the general US life expectancy from 40 years to 80 years since the early 1800s. But note:

y(t) = a \cdot e^{-b \cdot e^{-c \cdot t}}

It's not even "fuck the natural logarithm", it's "fuck the double logarithm". If we find some fantastic intervention in a pill that reduces our relative risk of death by half without any side effects, that halves the b value, which means this only moves the curve over a few years[10]:

A graph showing two curves, one with normal humans,
and the other with humans that have half the relative risk (RR)

We'll somehow need to invalidate this model with our mental fists.

(At this point, I should point out that there are some people working on the problem with an eye towards halting or reversing aging[11], like The SENS Foundation and The Methuselah Foundation. They are nonprofits, and could always use more money: if nothing else, they could make a bigger incentive prize of the XPrize sort.)

But I didn't write this post to complain about our problems, I wrote this post because:

  1. coining "9s of cats" was too tempting to pass up.
  2. consider this a weak post-pre-registration[12] of an informal study I did for well supported longevity actions we common folk can do today. Sure, the things we do are still subject to the steep demands of the Gompertz curve, but we want to maximize our chances of hitting Kurzweil's escape velocity if/when it happens.

Stay tuned.


[1]  Note that this is for a specific object, and not for a set of objects. If you have 10 trillion objects, you might see one of them go missing in a year, and that would be within the guarantee.

[2]  If you want an example of the sorts of replication large tech companies use, you can check out Facebook's blob store.

[3]  Note that while I work for a competitor of Amazon, I don't intend for this to be a pleasant daydream, but a nightmarish one. Also, it bears repeating that I do not presume to speak for my employer, etc etc.

[4]  This doesn't even include things like Alzheimers, which destroy the people without destroying their bodies.

[5]  Contrast this with genome sequencing costs, which have dived faster than exponentially. Today, you can get your genome sequenced for around $1k (the cost is sitting behind some cost request, but I've heard from biologists that Illumina whole-genome sequencing is around that much. Veritas Genetics also has a quote for around that much). It's possible that high resolution scanning technology will hit a similar trend, but it might not.

[6]  Availability zones: broad sectors with non-overlapping support, the theory being that bringing down one zone doesn't bring down the others. Concretely, it would be harder to kill you for good if you had copies living in both Europe and Asia.

[7]  Quip appropriately lifted from Ra, the Space Magic chapter.

[8]  To be fair, that's going from "lol leeches for everyone" to "well, let's scrape your bones out and put them in another person, and hey presto, they stopped dying!".

[9]  More by Gwern on his longevity page.

[10]  Graphic generated using an R+ggplot2 script, available as a Github Gist. I use the same curve that Gwern does circa 2017.

[11]  There are arguments against extending human lifespans, like overcrowding, but that's silly. Droning on about the sanctity of death because it's the Dark Ages is fine, but defaulting to death because oh no there are problems to overcome is a damn defeatist attitude. If you haven't read Bostrom's The Fable of the Dragon Tyrant, it's a gentle storytale introduction to non-deathism.

[12]  A pre-registration, so I can't just sweep things under the table, and weak, because I've already done the bulk of the research and analysis.

Filed under: Uncategorized

Leave a Reply

All content is licensed under CC-by-nc-sa
Creative Commons License