Until recently, I had never been conceptually attracted to the idea of an afterlife or prior lives, either as thought experiments or as aspirations. And definitely not in any religious sense. This is perhaps because I’ve never been able to imagine interesting versions of those ideas.
What has been piquing my interest over the last year is a particular notion of digital after lives/prior lives based on persistence of memory rather than persistence of agency or identity. Not only is this kind of immortality more feasible than the other two, it is actually more interesting and powerful.
We generally fail to understand the extent to which both our sense of agency and identity are a function of memory. If you could coherently extend memories either forward or backward in time, you would get a different person, but one who might enjoy a weak sort of continuity of awareness with a person (or machine) who has lived before or might live after. Conversely, if you went blind and lost your long-term memories, you might lose elements of your identity, such as your sense of your race or an interest in painting. Mathematician Paul Erdos understood the link between memory and identity:
When I was a child, the Earth was said to be two billion years old. Now scientists say it’s four and a half billion. So that makes me two and a half billion… I was asked, `How were the dinosaurs?’ Later, the right answer occurred to me: `You know, I don’t remember, because an old man only remembers the very early years, and the dinosaurs were born yesterday, only a hundred million years ago.'”
Erdos’ version of course, is based on no more than clever wordplay, but I want to consider a serious version: what if you could prosthetically attach to your own mind, the memories of somebody who died on exactly the day you were born, serving as a sort of reincarnation for that person? What if you could capture your own lived experiences as raw and transferable memories that could be carried on by somebody else, or a robot, starting the day you died, thereby achieving a sort of afterlife? Or perhaps live on somewhere in the Internet, changing and evolving?
The most interesting and unexpected consequence of any notion of immortality based on the idea of a living memory, is that notions of heaven and hell make no sense within it.
Memory is a constantly rewritten living thing, with evolutionary characteristics, shifting boundaries and some continuity. I began to explore the idea of immortality in my post A Beginner’s Guide to Immortality, but didn’t quite grasp the centrality of memory as opposed to what you might call, for lack of a better word, code.
The integrity and coherence of memory is central to the integrity and coherence of identity and agency itself. To paraphrase computing pioneer Frederick Brooks, reveal your thinking and hide your memories and I will understand nothing about you, hide your thinking and reveal your memories and I will understand everything about you.
Call this view the memory determinism view of identity and agency. In pure form, memory determinism can be defined as the existential position: If you’d been through what I have, you’d think as I think and do as I do.
If this seems like a very strong statement, think about just how strong the antecedent clause is: if you’d literally been through what I have, it means you would also have made the same path-dependent decisions I did. This means you would have thought and done as I thought and did every step along the way, and started at the same point. The strength of the consequent is entirely captured in the difficulty of achieving the antecedent.
I won’t try to rigorously defend memory determinism in this post, but just assume it for the purpose of exploring its consequences.
Some details matter here.
I am not a neuroscientist, but from what I understand, the way long-term memories are formed is that raw sensory memories are gradually integrated into a coherent narrative and then settle again into a sort of edited sensory memory with an associated narrative. This is why we can, with a little effort, weakly relive memories in a sensory way and also tell stories based on those memories. If you closed your eyes right now, and calmed your mind a bit, you could probably conjure up a ghostly echo of a summer vacation or party from long ago that has settled into long-term memory. You could also narrate the associated events without re-experiencing them.
There is apparently such a thing as a memorist, a tribe of Proust-like people who actively work on exploring memories. I suppose they get together and eat madelines and watch Being John Malkovich on occasion. I recently came across a quote from memorist Andre Aciman that seemed particularly clever: smell is the permanent address of memory. If you don’t instinctively get why that’s a true and important point, you might have a problem with your sense of smell.
I think I’d like to attempt writing a memoir someday, not because I think my life has been, or will have been, particularly remarkable, but because I think trying to capture the essence of your life in a form that can be experienced by another, is one of the most interesting technological challenges in the world. So far, we’ve invented only one technology that can be used to take on this challenge: reading and writing. Good memoirs, such as the Diary of Samuel Pepys enable reliving to some degree.
So for a writer, writing a memoir is the most interesting technical challenge of a career. It is the slightly creepy business of making your life relivable (but not rewritable) by others. But the challenge goes beyond writing. Writing can only memorialize a life in a static form, since it does not allow for a meaningful form of rewriting. It cannot turn a lifetime of memories into an evolvable, rewritable entity that can continue to live beyond the existence of the body that experienced them.
The real challenge for memory-based immortality is making an immortal memory independent of the bodies it passes through. There are two ways to approach this challenge, using finite and infinite game models respectively.
Immortality for Finite Games
The simplest version of this idea is a practical one: an artificial agent that captures enough about you to create a limited form of persistent agency. A blogger friend of mine, tubelite, explored this possibility in depth in this post.
Tubelite’s thought experiment involves a conventional sort of software design: one based on encoding rules and policies for governing actions, and perhaps embodying a critical dataset distilled from a life. The examples he explores are motivated by wanting to enable finite sorts of agency. For example, an investor-agent that perpetuates your investment philosophy after you are dead. In this view, the design of the after-life agent is really the main act, with the life being the startup phase, devoted to perfecting a highly survivable investment philosophy before death-time. This is like the sea squirt: a creature that swims around when young until it finds a good rock to anchor itself to, and then eats its own brain, turning into something more like a plant in a personalized heaven.
But there is something unsatisfying about this kind of eternal afterlife. The basic idea in this approach is to capture an instrumental essence of who you are through some mix of code and a rolled-up information-state at some point in time arbitrarily labeled death, and then using that essence as the starting point for continued evolution that is geared towards a particular purpose. As we’ll see, in this view, death is a bottleneck condition between life on earth and life in a designed heaven or hell (implied by the purpose encoded).
Even though the agent design is nominally for an eternal after-life, any design guided by a specific objective such as investing is necessary a finite-game design in the sense of James Carse (as regulars have probably noticed, Carse’s Finite and Infinite Games is rapidly becoming one of my core references, so you might want to read it if you plan to continue reading this blog regularly).
An extreme example of such a finite-game immortal is R. Daneel Olivaw, a robot who occurs across Asimov’s robot, spacer and Foundation stories, evolving into increasingly powerful forms, swapping many bodies and positronic brains along the way (evolved gods are a staple in Asimov’s writing). R. Daneel Olivaw’s agency is of course, driven by Asimov’s laws of robotics, which are imperative rather than descriptive in form, so capable of animating a limited kind of eternal agency. Those laws are game rules that are vast in scope (saving humans and humanity), but still finite.
To get to a design of an afterlife agent that is set up for a true infinite-game conception of a digital afterlife, you have to start with data, not code, and work the problem of memory perpetuation without reference to specific objectives.
This is an easier problem (entities such as corporations, cities, countries and cultures already solve it in limited ways), but one whose solution has broader consequences.
Immortality for Infinite Games
It seems like a trivial and obvious point, but an immortal agent must have a potentially infinite memory. What makes an infinite game infinite is not that it lasts for ever, but that it is driven by a desire to keep playing forever rather than winning. This means no finite set of unchanging rules for playing the game of life will do. Changing conditions might require changing game rules in order to continue playing.
The agent part must be capable of being born and dying within the memory as particular finite games begin and end. The immortality is carried through such birth/death bottlenecks by the continuity of memory.
This is actually a very powerful constraint. Much of our sense of identity and agency are attached to finite and unchanging game rules. My sense of identity latent in qualifiers like Indian and American for example, will be meaningless if applied to a being who exists in times when neither of those qualifiers do. My thoughts in English and Hindi would make no sense if remembered without translation by a being living in times when neither of those languages exist.
If in this life, I desire wealth or fame as a writer, those objectives will not translate to futures where wealth might be meaningless (perhaps it is a post-scarcity economy where everybody is plugged into a Nozick experience machine) and writing has been replaced by telepathy. Even sexual orientation, which creates very basic and powerful identities and motivations, is a finite game. If my memory in future might be prosthetically attached to a being with no sexuality, or a person with a different sexual orientation or gender, all the associated patterns of agency and identity become ephemeralized.
This means the only meaningful part of a lived mortal life that is suitable raw material for plugging into an immortal life is raw memory. One might argue that all memories are attached to some modality of agency or identity, so that identity/agency-independent memory is essentially the null set, but I don’t think that’s a serious argument (the idea of play to me is existence proof that we are not talking about a null set). So let us stipulate for the sake of argument that it is not.
This is enough that we can attach a name to this stream of memories and call it something more than entropic turbulence in the ocean of infinite memories. We’ve conceived of a meaningful notion of infinite-game immortality based on the persistence of memory. To borrow terminology from the philosophy of mind literature, there is something it is like to be an immortal version of me.
Specific finite-game chapters of agency within this stream might repress or surface particular memories into and out of an evolving Jungian shadow. Eating or sexual memories from one life might be repressed as disgusting in a future life. Self-esteem based on a particular aspect of identity might turn into self-loathing in a future body, but utilize the same memory in both cases.
But the sense of immortality itself will not depend on these particulars.
Heaven and Hell
The way we’ve set up our thought experiment here, it follows that heaven and hell are only meaningful concepts for finite-game immortals (this point is not an original one; Carse implicitly makes it).
To understand why, consider a famous athlete such as a future Tiger Woods, born in 2035. We equip this future Tiger Woods with sensors and brain implants that capture every stimulus he experiences through life, and every long-term memory he lays down. When he dies in 2110, we transplant his 75 year-long living memory into the most appropriate new baby (who might be a girl), or a tabula rasa robot (the question of whether the recipient will instantly age 75 years is an interesting one I’ll get to).
Now this memory-descendant of Future Tiger Woods might live for ever through multiple human and robotic bodies, and accumulate such an overwhelming historical memory and ongoing learning of golf that he/she/it becomes unbeatable in a few generations, given his/her/it lifetimes of deliberate practice. He might be beaten in individual games, but he evolves into a true all-time champion. He cannot be dethroned without changing the rules of golf so radically it becomes a structurally different game.
This is not a kind of memory immortality that interests me, because only one closed world (finite game) is ever being explored by this eternal being: the world of golf. This Tiger Woods is only as long-lived as the game of golf itself. Specifically, this is an immortal Tiger Woods designed to ascend to (and in so ascending, define) golfing heaven.
One could easily conceive finite-game immortals designed for specific hells.
I haven’t thought it through, but I suspect most versions of AIs conceived as capable of a post-Singularity immortality are finite-game constructs designed to create and inhabit specific heavens or hells. So are most visions of utopia and dystopia for humans.
This closed-world/finite-game assumption makes it possible to design simpler immortality software: you can compress the memories radically in golf-focused ways (or in the case of R. Daneel Olivaw, humanity-saving, heaven-creating ways). You do not need to explicitly retain details of specific golf swings that do not matter to golf. Some sort of machine learning algorithm that merely learns a really large vector of implicit and instrumental golf-swing parameters (which need not map directly to meaningful variables) would do the trick.
An aside: it seems to me that any notion of intelligence you might care to consider is simply a matter of enabling a particular variety of finite game immortality. Because notions of intelligence also come with associated notions of perfect intelligence, and therefore heavens won by those perfect intelligences.
Imagine though, that you want to structure the encoding of your memories for transplanting in the most open-ended manner possible. A manner that amplifies the generative potential of lives it is attached to in the future, rather than constraining it to something like golf or Asimov’s laws of robotics.
You’d approach the design very differently.
Googling Past Lives
If the goal of an afterlife mechanism is to perpetuate an evolving body of memories (that serve as prior lives to the future vessel), you get an infinite-game afterlife agent.
In designing such an agent, you’d ignore most proceduralized skill memories completely. Instead, you’d go for a Big Data approach: store as much raw memory as possible, in as unprocessed (or organically processed, with retention of the originals) a form as possible, with the ideal being no structuring at all.
Your uploaded brain at death would ideally be zero percent code, 100% data. Messy, varied, redundant data. Data that exists in various stages of digestion, from raw sensory memories to half-formed stories to settled long-term memories with canonical narratives attached. There would be abstractions weak and strong, well-worn and unfinished too. Code might exist too, but in bracketed, dormant form, rather than as active agency. Elements of the data store would be vulnerable to rewriting depending on their distance from raw sensory memory.
Note that you would make no particular attempt to “finish” your memory within your lifetime or put your memory affairs in order. Any state you stop at is a good enough state to be transplanted.
An interface to such a memory is not actually that hard to imagine. It is certainly a vastly simpler problem than imagining a human-like AI.
Imagine such a living memory being given to the transplant recipient simply as a searchable collection. You could evolve the interface for the transplant recipient as follows. Let’s call this recipient YouTwo, and the donor YouOne.
- Version 1 would simply be a Google-like searchable app on a smart phone. If YouTwo sees a cat, he/she/it searches for “cat” and sees a reverse chronological stream of cat images and associations from YouOne’s life. YouTwo can edit anything from YouOne’s past that comes up.
- Version 2 would automate the linkage as a sort of soft memory surgery. Anything YouTwo experiences would trigger a search whose results could be reviewed in peripheral vision, again in read/write mode. This would be a sort of ghostly/poltergeist haunting of YouTwo by YouOne.
- Version 3 would sensorize the search results, piping them (for instance) into an augmented reality (AR) headset that juxtaposes YouOne memories onto YouTwo experiences in a seamless way. This would not be organic though. If YouTwo meets a friend of YouOne whom he has never seen before, there would be no recognition as such. Just an explicit cue that “you knew this guy in a past life.” You might edit your experience by hitting like and dislike buttons on anything your prosthetic memory retrieves, driving a process of repression and highlighting.
- Version 4 would achieve a certain degree of coarse and organic temporal integration, and start to blur the two memories. YouTwo would interleave living his own life with reliving YouOne’s life. Perhaps for an hour every day, YouTwo simply lays back in an armchair with an AR headset on, and “remembers” random memories pulled up via association with recent experiences of his own, forming weak links in an evolving temporal memory graph over time. In this scenario, the AR headset might flash a pair of images: YouOne’s friend from the past life, and a random stranger encountered that day by YouTwo, juxtaposed in order to form an association. This association would be purely about creating a link between past and present, and not for any instrumental purpose. It would be a sort of assisted Hebbian learning. Version 4 would start to blur past and present in YouTwo’s mind. If he/she sees the friend later, he/she would begin to have doubts: “Is this somebody I’ve seen before, or is this a YouOne memory?” That blurring would be a feature rather than a bug.
- Version 5 and beyond would use neurosurgical implants to create ongoing subconscious temporal integration, and might even try to piggyback on dreaming past lives into current living memory. The grafting code would also attempt to integrate narratives rather than just sensory memories. I won’t go into this, but there are some obvious and non-mysterious ways to blur present and past stories, so you start dissolving the distinction between stories that happened to YouTwo and stories that happened to YouOne. You would also eventually have actuation of things like “gut feelings” (perhaps machine recognition of a face from YouOne’s lfe would trigger a little squirt of stomach acid or a mild electric shock to YouTwo, creating a new class of sensations correlated with past lives).
- The ultimate form would of course be a realization of the sort of memory integration envisioned in Philip K. Dick’s Total Recall.
There are plenty of other interesting details to consider and add. For instance, you probably don’t want to flood a 2-month-old’s (pre-verbal) brain with the prior-life memories of a sixty year old. Or a famous mathematician’s memories into an idiot’s brain. You’d want to achieve some synchronization within a life, bringing up age-appropriate memories through life (creating a sort of continuity and smoothing effect). You’d also attempt to do some matching of personality traits. An introvert’s memories transplanted to an extrovert host would probably just result in incoherence and poor integration. This whole category of engineering detail could be considered a case of handling impedance mismatches.
Possibly, you could test out an entire library of stored memories against a new host and transplant the best match in some sense. And of course, there are other wild possibilities such as implanting the memories of multiple people into one host, or achieving a trading-places sort of memory switch between two living people. But those don’t add much philosophical fodder to the base case. The base case of splicing the end of one living memory into the beginning of another, smoothly and with preservation of rewrite capabilities across the junction, captures much of the challenge.
Why do any of this?
Because I suspect we can. Which means, by the logic of infinite games, we probably should.
The infinite game brings with it the moral imperative to do anything that keeps the game going in a more generative way, and I think seeking memory immortality qualifies. I’d certainly like the memories of some interesting person who died the day I was born grafted onto my mind.
The human body, being a product of evolution, is not perfectly designed. Some parts of our body are under-designed: our teeth and bones aren’t really meant to last more than about 40 years, so we have to start taking strange measures to life-extend those parts.
But some parts of the human body are also over-designed. In particular, I think our memory capacity is far higher than our lifespans. Our brains seem to be designed to store and use memories for perhaps several hundred years. Perhaps thousands.
Okay, so we have a good set up: an immortal designed to play an infinite game, and a plausible way to realize such an entity through technology. It isn’t AI, it isn’t artificial sentience, it isn’t a solution to the hard problem of consciousness.
But it is at least a plausible solution to the problem of consciousness transference through continuity of memory.
One useful thing we can do with this set up is reconsider heaven and hell.
Heaven and Hell Redux
One of the basic problems with religious notions of prior and after lives is that they are derived from finite-game instrumental notions of being. So it is perhaps not surprising that religious notions of both heaven and hell are uniformly tedious depictions of states of eternal stasis. Too much time is being filled up with too little substance.
It is not clear that such stasis could be enjoyable. Certainly burning in hellfire for eternity would not be enjoyable, but would living a life of consequence-free hedonism be enjoyable forever? Or winning every game of golf? Or to head East for a moment, would an eternal karmic cycle of life and death be any good?
One of the few classics I’ve read on the subject, Dante’s Inferno, seemed like a tediously contrived way to explore a particular conception of morality, and that ultimately is what most religious notions of heaven and hell are, whether religious or atheistic, and whether situated on earth or another imagined plane. Explorations of morality within specific finite games.
A more interesting treatment, in Chapter 10 of Julian Barnes’ A History of the World in 10½ Chapters, gets at the idea that an after life defined by any unchanging condition is not something anybody would want for eternity. Any eternally unchanging set of conditions is a hell by definition: a place you would not want to experience for eternity. The only difference is that some you want to exit within seconds, others within centuries.
Curiously though, life on earth (or more generally, this universe) itself qualifies as a heaven by the natural complementary definition: a place many would want to experience for eternity (both backward and forward in time) in some way.
On the non-religious metaphysical side of things, all the explorations of the idea of an afterlife that I am aware of seem to be explicitly or loosely existentialist in character. You have Sartre’s idea in No Exit that hell is “other people,” Camus’ treatment in The Myth of Sisyphus and Beckett’s Waiting for Godot.
Two things are interesting about these works: they don’t take the distinction between pleasant and unpleasant conceptions of an afterlife seriously. They seem to implicitly realize that any unchanging situation is hellish, and explore, in different ways, how humans strive to change and find reasons to live on, in situations that stay the same. Humans don’t really seek eternal bliss or or try to avoid eternal pain. Instead, we really seek eternal interestingness.
To me, these works demonstrate one thing: if we have any capacity for immortality at all, it is only a capacity for the infinite game. Finite game heavens and hells are not worth spending eternity in unless you have no evolving memory at all. Finite-game afterlives require eternal amnesia. Like the sea squirt, we’d have to eat our brains the moment we found our heavenly rocks.