Digital Immortality:

Theory and Protocol for Indirect Mind

Uploading
Alexey Turchin
Digital Immortality Now
Foundation Science for Life Extension
[email protected]

(с) Digital Immortality Now, DIN, 1987–2018

Disclaimer: DIN provides consulting in the sphere of digital immortality. You

can use this theory and protocol for free. DIN may help by providing advice
to implement the protocol, but do not take any legal responsibility for your
actions according the protocol.

Abstract. Future superintelligent AI will be able to reconstruct a model of the

personality of a person who lived in the past based on informational traces.
This could be regarded as some form of immortality if this AI also solves the
problem of personal identity in a copy-friendly way. A person who is currently
alive could invest now in passive self-recording and active self-description to
facilitate such reconstruction. In this article, we analyze informational-
theoretical relationships between the human mind, its traces, and its future
model; based on this analysis, we suggest the instruments to most cost-
effectively collect quality data about a person for future resurrection. These
guidelines form a “digital immortality protocol”. Digital immortality is plan C
for achieving immortality, after plan A, life extension, and plan B, cryonics.

Keywords: Digital immortality – superintelligence – effective altruism – life

extension – mind uploading.

Highlights:
 Future superintelligent AI will be able to simulate past people.
 To help AI improve its simulations, we can collect data about a living
person now.
 Passive data collection is constant recording.
 Active data collection is running tests and recording self-description.
 The best way to collect information is to create art, as it is unique,
valuable, and predictive.
1. Introduction .....................................................................................4
THEORY ..............................................................................................6
2. Why is digital immortality needed?......................................................6
3. Informational identity ........................................................................6
3.1 Two types of personal identity: observational and informational .........6
3.2 Criteria for informational identity....................................................7
3.3 Should computer be conscious? .....................................................8
3.4. Theory about the possibility of eternal preservation of information ....9
3.5 Transformational resurrection ........................................................9
3.6 Loss of time in DI ....................................................................... 10
3.7 Different personalities in the different periods of time ..................... 10
3.8 Immortality-for-me and immortality-for-other ............................... 10
3.9 Identity as a decision under uncertainty problem ........................... 11
3.10 Digital immortality as plan C, after life extension and cryonics ....... 11
3.11 Necessary conditions for DI and its probability of success .............. 11
3.12 Overcoming the showstopper effect of the identity problem ........... 12
4. The theory of reverse reconstruction of the human mind ..................... 12
4.1 Reconstruction based on traces as an inverse problem .................... 12
4.2 Reconstruction of function output based on its previous behavior ..... 14
4.3 Reordering of facts based on their predictive level .......................... 14
4.4. Common, unique, and random facts ............................................ 15
4.5 Valuable facts ............................................................................ 15
4.6 Combining all types of facts ......................................................... 16
4.7 The size of human mind .............................................................. 16
4.8 Principle uploading speed limit and technical feasibility of digital
immortality ..................................................................................... 17
4.9 Uploaded data as a hash of the “mind function” ............................. 18
4.10 Combining the declining predictive power of facts with our identity
threshold ........................................................................................ 18
4.11 Individual variance of predictability ............................................. 20
4.12 Predictive facts as limits in the space of possible minds ................. 21
5. Cost-effective extraction of the best predictive facts............................ 21
5.1 The problem of the cost-effective extraction .................................. 21
5.2 Human behavior ........................................................................ 21
5.3 DNA data .................................................................................. 22
5.4 Digital footprint .......................................................................... 23
5.5 Relatives’ memories and expectations ........................................... 23
5.6 Security camera recordings ......................................................... 23
5.7 Personal belongings and archives ................................................. 23
5.8 Medical brain data ...................................................................... 23
5.9 The subconscious—should it be regarded as part of the personality? . 24
5.10 Human social graphs and the need for collective digital immortality 24
5.11 Unexpected reactions ................................................................ 24
5.12 Intersection of different data sources .......................................... 24
5.13 Sources of human individuality ................................................... 25
6. Main strategies of DI: passive and active uploading ............................ 25
6.1 DI as an acausal deal with future superintelligence ......................... 25
6.2 Main strategies of DI: data-driven, experiments and self-description 26
6.3 Not a black box, but a black torus ................................................ 27
 6.4 Combining strategies .................................................................. 27
7. Digital immortality and art ............................................................... 28
8. Post-mortem DI .............................................................................. 29
9. Integration potential of DI ............................................................... 29
9.1 Growth of technologies ............................................................... 29
9.2 DI + cryonics ............................................................................. 30
9.3 DI + direct mind uploading .......................................................... 30
9.4 DI + exocortex .......................................................................... 30
9.5 DI + quantum immortality .......................................................... 30
9.6 DI + Alzheimer .......................................................................... 31
9.7 Resurrectional simulations ........................................................... 31
9.8. DI as psychotherapy .................................................................. 32
9.9 Speed of personal changes and DI ............................................... 32
9.10 DI and family history ................................................................ 32
9.11 Social aspects of DI .................................................................. 32
9.12 Age-related strategies for DI ...................................................... 33
10. Digital immortality as a cause of effective altruism ............................ 33
11. Theory behind different recording technologies ................................. 33
11.1 Association tables ..................................................................... 34
11.2 Encyclopedia of things............................................................... 34
11.3 Encyclopedia of people .............................................................. 34
11.4 Declaration of personal properties .............................................. 34
11.5 Automative writing of stream of consciousness............................. 34
11.6 Recording childhood memories ................................................... 35
11.7 Video interview ........................................................................ 35
11.8 Professional police and psychological checklist ............................. 36
11.9 Dream diary + drawings ............................................................ 36
11.10 EEG + spontaneous art ........................................................... 36
11.11 Scanning of paper archive........................................................ 36
11.12 Spontaneous theatre ............................................................... 36
11.13 Lists ...................................................................................... 37
11.14 Guided visualization and active imagination ............................... 37
11.15 Idealized world: a novel .......................................................... 37
11.16 Bainbridge yes-no tool ............................................................ 38
11.17 Terasem sites ......................................................................... 38
11.18 Uploading sessions ................................................................. 38
11.19 “One my day” and videocommenting ......................................... 38
11.20 Audio dictation ....................................................................... 39
11.21 Testing the mind as an information filter.................................... 39
11.22 Environmental data collection from photos ................................. 39
11.23 Keylogging ............................................................................. 40
11.24 Memoirs ................................................................................ 40
11.25 Diaries and memory dumps ..................................................... 40
11.26 Drawings ............................................................................... 40
11.27. Ask other people what they think and remember about you ........ 41
11.28 Futuristic and untapped sources of information ........................... 41
PRACTICE .......................................................................................... 42
12.1 Choosing a personal uploading strategy ...................................... 42
12.2 Steps of data collection ............................................................. 43
 Day 1: Collecting existing data ....................................................... 43
Day 2. Self-description .................................................................. 44
Day 3. Running tests ..................................................................... 44
Day 4. Videocommenting of stream of consciousness ........................ 44
Day 5. Visiting friends and recording interviews ................................ 45
Day 6. Experiments with EEG monitoring ......................................... 45
Day 7. Creating art ....................................................................... 45
Day 8. Creating an idealized image of self ........................................ 45
12.3 Privacy notice .......................................................................... 46
12.4 Legal notice ............................................................................. 46
13. Requirements for digital immortality software ................................... 46
14. Data preservation strategies ........................................................... 47
Home hard drives archives ................................................................ 48
14.1 Data structure .......................................................................... 48
14.2 Cheap small “data-bombs” for eternal preservation ...................... 49
14.2 Underground preservation ......................................................... 50
Conclusion ......................................................................................... 51

1. Introduction
The term “digital immortality” is used with two different meanings: 1) direct
brain uploading via brain scanning, which produces an uploaded and
presumably eternal copy of the mind (Wiley, 2014); and 2) indirect brain
uploading, that is, reconstruction of the personality based on its footprints by
future AI, also known as “person capture” or “cyberimmortality”. Direct
uploading is currently impossible, and many people now living will not survive
until its appearance. In this article, the term “digital immortality” (DI) will be
used only in the second meaning:

Digital immortality is collecting information about a person with the

expectation that this information will be used by future superintelligent AI to
resurrect the person in the most accurate way possible.

Another way to put the same idea into words is the Terasem movement’s
suggestion that the “conscious analog of a person may be created by
combining sufficiently detailed data about the person (a ‘mindfile’) using
future consciousness software (‘mindware’)” (Terasem, 2014). Terasem
suggests 2 services for “personality capture”: www.lifenaut.com, and another
site, CyBeRev.org, which offers slightly different functionality.

The idea of digital immortality has been explored by several authors.

Bainbridge suggested running an array of tests and recording emotional
reaction for “personal capture” (Bainbridge, 2004). Goertzel wrote about the
role of AI in future resurrection of personality (Goertzel, 2007). Microsoft
created a computer program to collect digital immortality data (Gemmell, Bell,
& Lueder, 2006), and published an article outlining ideas related to its use
(Bell & Gray, 2000), but Microsoft later closed the project and kept the code
proprietary because of the legal complexity of transferring its ownership.
Kurzweil collected data about his father, including boxes of his fathers’ letters,
hoping for his future resurrection. He hopes to create an “avatar”, or a “virtual
computer replica” in the future. Kurzweil said, "That [will be] a replica, but I
can actually make a strong case that it would be more like my father than my
father would be, were he to live" (Berman, 2011).

Recently, several projects have appeared which suggest using social media
data to create a chat-bot based on the digital footprint of the deceased person.
In one, a son worked together with his dying father on his chat-bot (Vlahos,
2017). The Eternemy startup suggests help in data collection (Schilling,
2016). The 2045 movement and Humai have more complex agendas, which
combine different approaches (it appears activity on these projects has slowed
down as of early 2018). Even funeral agency Fenix now suggests interactive
chats with the dead, which are expected to evolve into full blown avatars
(Alestig, 2018). Memori connects person’s memories with physical objects.
http://getmemori.com/en/about.html Digital Immortality Now (Volpicelli,
2016) is concentrated not on providing services, but on consulting in DI.

The idea of the informational uploading of a person via non-technological

means can be traced back as far as Saint Augustine, who wrote his
Confessions in the 4th century (Augustine, 398 AD). But the real beginning of
the honest and conscious creation of the model of self was Confessions by
Rousseau, written in the 18th century, which starts with the following words:
“I have resolved on an enterprise which has no precedent and which, once
complete, will have no imitator. My purpose is to display to my kind a portrait
in every way true to nature, and the man I shall portray will be myself”
(Rousseau, 1782). After Rousseau, the genre of autobiography flourished, but
it was not recognized as an actual instrument for achieving immortality.
Berdyaev wrote the book “Self-Knowledge: An Essay in Autobiography”
(Berdyaev, 1950), which is not only biography with facts and emotions, but
full of generalizations of personal traits, which also become declarations. As a
result, proper self-description became a transformative technique, similar to
“recapitulation”, as described by Castaneda (1993). The longest-ever written
diary has 35 million words (180 MB), kept by R.Shields (Martin, 2007). He
wrote an entry every 5 minutes for decades, thereby collecting a lot of
information about himself.

Surely, personal writings were recognized as some form of immortality, as

seen in “Exegi monumentum”, a poem by Horace (Horace, 23 BC), a
philosopher also regarded as world first autobiographer. But as Woody Allen
famously said: “I don't want to achieve immortality through my work. I want
to achieve it through not dying” (W. Allen, 1993). Despite its limitations, a
written text has ability to influence the reader, who creates an internal model
of the characters described. In the case of autobiography, it is a model of the
text’s author, and thus she can “install” her personality in the other human’s
brain. However, this “viral” form of text-based immortality was never
recognized, and is of limited importance now, as powerful AI will soon appear
and provide a more effective method of resurrection.
This article suggests that passive recording is not enough; active information-
extraction efforts are needed to quickly and cost-effectively collect and focus
on the most important information about a person. Based on the idea of
information renormalization by importance and predictive power, it seems
possible to collect enough data to enable digital immortality in reasonable
time, and with reasonable effort and costs.

The article consists of two parts: theory and practice. If you want to start
preparing for personal digital immortality immediately, skip to the practical
part.

THEORY

2. Why is digital immortality needed?

The first answer is that people are still dying and not much can be done
about it. Once we accept the premise that human lives and minds are
valuable, there is a moral imperative to protect, extend, and/or preserve
them if possible.

Only a few people are signed up for cryonics, and for many it is expensive and
logistically difficult or impossible (as cryonic is legal only in US). Cryonics could
also fail due to legal, organizational, or financial problems. Digital immortality,
on the other hand, could provide a cheap and affordable option for potentially
indefinite life extension for almost everybody.

As in the case of cryonics, fighting aging, and many other important issues,
the main problem of DI is not technological, but psychological: people don’t
want it. The main reasons—besides the fact that many have never even heard
of the idea—is the problem of personal identity: the thought that the model
of their personality will not be the same person is a “showstopper” for the idea
of DI (Bainbridge, 2006).

Copying of the digital minds opens possibility for indefinite life extension
(Eubanks, 2008), as copying outperform any survival strategy longterm.

3. Informational identity
3.1 Two types of personal identity: observational and informational
The analysis of the problem of personal identity will be limited here, as it is
known to derail discussion about mind uploading. Several publications claim
that the identity problem will make mind uploading difficult (Swan & Howard,
2012), and most people share this view.

According to Parfit, there are two types of the personal identity (Parfit, 1984):

- “qualitative identity” which we will call informational identity in this

article, is the measure of similarity between two minds (Parfit’s term is
not self-obvious).
 - “numerical identity” which we will call observer identity for clarity,
answers the question: will my copy (in the next moment of time) be me
or not?

In short, belief in the observer identity implies that informational identity is

not enough for “real” identity, and some form of non-informational identity
carrier is needed, which could be either “soul” or “continuity of
consciousness”. These things are currently unmeasurable, so their existence
is purely based on belief for now; in other words, we can’t measure observer
identity.

In this article, the “conservative approach” to personal identity will be used;

that is, the idea that one should try to preserve as much identity as possible.
It is assumed that in the future scientists will better understand the nature of
the identity, and when superintelligent AI appears, it will be able to solve the
problem.

Thus, we are postponing the solution of the problem of the observational

identity into the future. Maybe observational identity is just an illusion, maybe
it is real, but future AI will find ways to capture and transfer it. For example,
if the observer identity is connected with continuity of consciousness which is
related to causal connection between mind states, then preservation of causal
connections between me-now and a future copy could be used to extract not
only information, but also “causal connection”. Future AI may find a way to
use causal connections with the past to extract “continuity”. But to preserve
causal connection, we need to preserve some form of information about the
person, so we again return to digital immortality.

Human personal identity is not the same as identity of minds in general and
is a social construct (Bamford & Danaher, 2017). The author has explored the
problem of personal identity in detail in (Turchin, 2016).

3.2 Criteria for informational identity

The following criteria may be used for informational identity in the case of the
constantly changing human mind:

Strict criterion: the copy is the same person if the difference between the copy
and the original is less than the typical changes that a human being
experiences sleeping through one night.

Human memory decays exponentially (Murre & Dros, 2015), but different
types of memories decay at different speeds. One night’s loss may be roughly
estimated as losing 0.1 per cent of all personal data, but more research is
needed. It is absurd to require a higher level of informational identity as it will
be lost in several hours because of the natural rate of forgetting. The strict
criterion translates into 99.9 per cent behavioral similarity. We later will
discuss how it could be measured. If one rejects the strict criterion, one must
assume that each person is another person next day, which is theoretically
correct but practically absurd. (The relative importance of different facts will
be discussed later in the section about predictive power of facts in the section
4.10.)

Mild criterion: If a person becomes sick, and later recovers, she is still
regarded as the same person, despite the fact that she may lose some of her
memories and skills. Here, the copy will be regarded as the same person in
the mild sense if the difference is equal to forgetting the last year of
experiences—which, due to the exponential rate of memory decay, may mean
losing more than half of her memories.

Plus, the mild criterion means that the person will be recognized as the same
by his social circle (Rothblatt, 2012)—and by himself, if she will be able to
somehow have a view of her internal mental states. The fact that humans are
able to recognize themselves from inside is based on observation that
sometimes they fail to recognize their identity, mostly in cases of some
dissociative disorders. This inability to recognize oneself is based on
discrepancy between image of self and actual observations, but in case of DI
we can’t be sure that “image of self” was preserved accurately. Thus, to
ensure identity via self-recognition we need to imagine the situation where
currently living person could have a look on internal experiences of his model,
and either recognized as her or not. This experiment is counterfactual, as we
can’t actually provide access of past person to her future model, however,
future AI could predict result of the experiment. For example, if I wake up in
the body of the opposite sex, it is easy to predict that I will recognize that
something is wrong with me.

This ability to “self-recognize” and be recognized by others is based on the

retention of the core “indexical and identification information”. In the human
case, this core includes memories about the person’s name, native language,
family members, living address—in other words, everything that is typically
used to recognize the person. The core could be rather small and will be
discussed later—here, we just assume that the core is the socially accepted
set of memories used to identify the person (as well as the ways how I
internally identify myself). In the mild case criterion, the core should be mostly
intact, but still could have some errors, as it rather typical for human to forget
even some important details about themselves.

Rothblatt (Rothblatt, 2012) suggested as a criterion examination of the copy

by a group of psychologists, an analogy to the Turing test. Our criterion is
even stricter as it suggests counterfactual possibility of internal self-testing,
that is, will I regard myself as me, even in the presence of some small
changes.

3.3 Should computer be conscious?

It seems that idea of digital immortality is based on the assumption that a
computer can have consciousness—but this assumption is not required. Digital
data about the dead person will be analyzed by the superintelligent AI, but it
could put the resulting model not on the digital carrier, but inside another
biological brain artificially grown from biological neurons, if AI finds that
biological neurons are necessary for subjective experiences. Other solutions
will may be found to preserve consciousness in a computer, potentially
involving the use of quantum computers or as-yet unknown devices.

In other words, the possibility of a non-classical nature for the human mind
does not prevent the viability of digital immortality as a resurrection
technology. If “soul” exists, it could “return” to a sufficiently advanced replica.

3.4. Theory about the possibility of eternal preservation of information

Everything which exists in our universe is decaying because of entropy.

However, information can be copied; the nature of information is that it is
easy to copy. To ensure eternal preservation, the rate of copying should be
higher than the rate of decay, including the fact that any copying has a non-
zero probability of errors. This error rate can be infinitely small by some
preservation mechanisms like control sums.

In the past, when only a few copies of manuscripts existed, and copying was
expensive, the question of how many copies should be preserved was not
trivial. A mathematical model showed that creating one copy every year is
enough to outperform the decay process [ref].

3.5. Transformational resurrection

One way to stop thinking about “identity obstacle” is to assume that the
resurrected person will not be you but will instead be someone who is closer
to you than your child. She will be a person with same DNA, same set of
memories, and same values, but improved: younger, healthier and open to
adapt to the new world.

As mentioned, Kurzweil suggested, and has said that he can prove, that his
reconstructed father will be more his father than his father was at any moment
of his life; that is, the upload will be a quintessence of what his father was
(Berman, 2011). Such proof may lie in the fact that most of the time, humans
react automatically to stimuli, and don’t always remember things which they
regard as important to them. Thus, constructing a “better me”, which is close
to me in my best moments of life—but even better—is possible by
simultaneously combining all of the traits which were my best traits at
different moments in my life.

There could be many future time-lines, which suffice to satisfy the

informational identity criterion, but only some of them present a “best me”. It
is human nature to want to improve and to change one’s own personality.
Thus, the creation of an idealized image of self will be not only the fixation of
current hidden desires, but an act of creation of a “better me”, including
instructions for the future AI of how I should be improved.

The paradoxical part of resurrection is that, despite the fact a resurrected

person should be as close as possible to the original for the purpose of identity
preservation, most saved data should be immediately discarded as the person
adapts to the changed world around them. Resurrection is always a
transformation. If we know in advance which data may be discarded without
being read, we may choose to not invest in its preservation.

3.6 Loss of time in DI

It is obvious that DI will not provide a full accounting of the last several days
of the person. There are several reasons for this. Firstly, active uploading
periods—that is, the most concentrated periods of uploading, are at relatively
infrequent intervals, like every several years. Constant passive recording
provides a narrower update than these infrequent full uploads. Secondly, last
days are full of memories which would be forgotten, if death does not occur.
In addition, when a person is seriously ill, there is often no time to invest in
proper recording. She may experience very transformative changes in the last
moments of life which will not be recorded.

As a result, after resurrection, a person may feel as if she has some form of
amnesia about her last time period. It is rather typical for humans to have
some form of amnesia after a serious illness, especially coma, and such
memory loss is not considered to have changed their “identity”.

This also means that only our mild criterion of informational identity is
applicable to DI, that is, one year of distance. How much you remember about
what you felt and planned last year? Are you regarding yourself one year ago
as you?

3.7. Different personalities in the different periods of time

A person is always changing, sometimes more slowly and sometimes more
quickly. Stable periods could be called temporal personalities, which have a
rather stable set of traits, like dominating values, emotions, job, friends,
home—which could be easily distinguished from those of another period. Such
periods are established as social roles, like first grader, sophomore, mother
of toddler, professor, a person who is grieving loss.

As personalities within these periods are more or less stable, it is not

important in which moment of this period a person is recreated. The clear
beginning of a new period means that an update of the digital immortality
data is needed.

3.8. Immortality-for-me and immortality-for-other

The observer identity is important from my point of view, and its preservation
can be called immortality-for-me. If we are interested in the eternal existence
of someone else, like a loved one, this is the problem of immortality-for-other.
In the first case, we are more interested in continuity of existence, and in the
second case, the most important is stability of behavior and the ability of peers
to recognize it as “the same person”.

This difference may affect data collecting strategies when a person is

collecting data about himself or about another person. For example, if we
return a dead poet to life, we may expect that he will continue to write high-
class verse. But from his point of view, it may be more important to study
world history, learn the fate of his relatives, and drink in a bar.

Preserving identity doesn’t mean future behavior will be predictable; there

could be many different future timelines with the same initial identity.

3.9. Preserving identity as a decision under uncertainty problem

The nature of the human personal identity is currently unsolved.

Let’s imagine the following thought experiment: you are expecting capital
punishment via hanging, but the night before you get a message: there is 1
per cent chance to escape the prison, but to do so, a very strange set of
actions are required. Should you agree? The answer seems to be YES! both
from an emotional and a rational point of view.

The situation is similar for digital immortality. It is uncertain if DI will work,

but it seems reasonable to try if there is nothing to lose. However, in real life
there are opportunity costs, as explored in the next subsection

3.10. Digital immortality as plan C, after life extension and cryonics

There are other options for a contemporary person at the beginning of the
21st century to reach immortality, including life extension until the creation
of AI (plan A) and cryonics (plan B) (Turchin, 2015). They seem to have a
higher chance of success for most young people, as they could survive until
immortality or earn money to pay for cryonics. However, for older and poor
people, these two options are less available, and they must go directly to plan
C: digital immortality. Plan D in this context is “quantum immortality”
(Turchin, 2018).

The idea of a multilevel defense suggests that as the future is very uncertain,
different approaches may be tried simultaneously, with investments
distributed proportional to their expected effects. A multilevel defense
suggests that a person should invest in the three most promising plans, life
extension, cryonics, and digital immortality. As the returns on investment
diminish for each plan in turn, it seems reasonable to turn to the second most
effective plan after investment in the first.

3.11. Necessary conditions for DI and its probability of success

There are two necessary conditions for the digital immortality:

1) Human personal identity is (a) “copy-friendly” and (b) “noise stable”.

2) In the future, (a) superintelligent (b) benevolent (c) AI will appear and (d)
it will choose to resurrect past people.

These two main conditions consist of 6 mutually independent sub-conditions

of unknown truth. If we apply to them theory of logical uncertainty (Kosoy,
2017) and Bayesian logic, we can give them a prior probability of 0.5 each.
For digital immortality to be possible, all six conditions must to be true, which
has a 1:64 chance, or 1.56 per cent. To get the actual probability of success
we need to add probability that the data will be correctly collected and will
survive until AI.

Yudkowsky (Yudkowsky, 2016) wrote that this style of reasoning creates

underestimation (illustrated by the cases of cryonics and the election of
Trump) probably because it excludes the optimizing agent’s “will to win”,
which could outperform chance by deliberately concentrating on the weak
points of such chains.

Anyway, a 1 per cent of success with DI seems to be a well-reasoned

estimation, and personally the author thinks that the chances are even higher.

3.12. Overcoming the showstopper effect of the identity problem

The practice shows that most people spend all their available resources not
on digital immortality but on musing about the identity problem; people are
mostly inclined solutions that are not “copy-friendly”. This tendency may
result from the fact that human identity is powerful socio-biological
adaptation. Humans are also do not adapted to be in a state of logical
uncertainty about complex philosophical problems: they choose to believe.

To overcome the problem, it seems reasonable to add the digital immortality

practice to several other similar ideas:
- creating external effective memory (Memex)
- creating a personal assistant-avatar, trained on personal behavior
- preserving family history

3.13. Overcoming the fear of the “unfriendly resurrection”

Some people expressed fear that unfriendly AI will resurrect them based
on the digital immortality and based on that expressed intent not to be
digitally resurrected. But by doing so, they prevented any benevolent AI
from creating their digital replicas, but non-benevolent AIs will proceed
anyway. So your replicas will find yourself only in hands of UFAI (no
matter how improbable UFAI is). And there is non-zero probability that
you are not original anyway, but already a replica, so - with some
circularity caveat - you just ensured that you are in hands of UFAI.
The same way by refusing to invest in the creation of high-quality
replicas, you increase the share of bad quality replicas.
This by agreeing in DI, you increase probability that friendly AI will
resurrect you.

4. The theory of reverse reconstruction of the human mind

4.1 Reconstruction based on traces as an inverse problem
Imagine a function f(t) which outputs string of numbers (f1, f2 … ft). The
problem of digital immortality can be formalized as reconstruction of f(t) based
on given string of (f1, f2 … ft).
There are two ways to reconstruct the function:

1) Reconstruction of the actual function f(t).

2) Reconstruction of the function f2(t), which outputs the same string of
numbers.

If f(t) is a simple function and the string is long, full reconstruction of the
function is possible. For example, if we have the string (1, 4, 9, 16, 25) we
can conclude that f = x2.

For full reconstruction (1), three conditions must be met:

a) The description of the function f(t) is shorter than the output string.
b) The function f(t) is the simplest function, unique in term of algorithmic
complexity that can produce the output F (Ming & Vitányi, 1990).
c) The computational complexity of its reconstruction can be achieved with
existing computers.

It is rather obvious that these conditions can’t be realized for the human mind.
We can consider the human brain as a very larger neural net, with 100 trillion
synapses, and consider that at least 1 kb of data is needed to describe each
one. If each description includes information about two connected neurons,
the type of synapse and its weight, this is equal to 1017 bits of data, or
approximately 1016 bytes.

(Strictly speaking, many different functions could have the same output on
part of its inputs – but in human case I just don’t know which one of such
similar in beginning functions I am. For example, I don’t know how I will react
on 1-million-dollar prize, and there are many future timelines this different
reaction which still will be me. In other words, I am indexically uncertain about
who I actually am. See also thought experiment about smoking lesion.)

The life-long informational output of the human being is the total sum of
signals which the brain sends through its motor functions, which could be
roughly estimated as 1 kB a second (and probably less), mostly relating to
speech and posture. In that case, the total life output of the brain is 35
GB/year, or 2 terabytes for 70 years = 2 × 10Е12 bytes, that is, 5000 times
less than the total amount of data in the brain.

This analysis probably ends the idea of full connectome reconstruction based
on behavioral patterns. However, constant changes in the human brain, like
learning, pruning and neuronal death do not affect human personal identity,
and thus much less information is actually needed.

More on inverse problems: https://en.wikipedia.org/wiki/Inverse_problem

4.2 Reconstruction of function output based on its previous behavior
But there is another type or reconstruction, this is, reconstruction of the
output F, which is more applicable for the human minds. There are three
reasons for it:

1) We can’t observe our connectome, but only its behavior, even from
within. If we ignore here the problem of the nature of the qualia, I will
not be able to recognize that a part of my mind was replaced by another
system with the same functions. For example, if part of my visual cortex
is replaced by another part that can recognize the same objects in the
same way, I will not be aware of the difference.
2) Human minds constantly lose neurons and their connections without
significant loss of socially accepted identity or functionality. An adult
brain typically loses 1 per cent of it cortical mass every year, not equal
to the loss of 1 per cent of neurons, but neuronal loss is probably the
on same order of magnitude. Humans lose a lot of information about
the previous day during sleep each night.
3) Human behavior has a large random component, with both hardwired
non-individual and simple parts (reflexes). The individual, non-random
part of human behavior is relatively small.

This means that reconstruction of output F for the human mind could be done
with large error, R, which consists of a part connected to the natural human
tendency to forget, R1, and a part connected with intrinsic randomness of the
human behavior, R2.

Reconstruction of function output with a given error size is a well-explored

topic in machine learning. It is probable that the human mind works in the
same way, as it is also a neural net. The mind learns to predict the next
moments of observed experiences; current artificial neural nets are also
remarkable in predicting sequences (Karpathy, 2015) It is assumed that
things which I never experienced are not part of me.

The task of behavior prediction is simplified since the behavior of most humans
is rather stereotypical. But some of the most important human actions are
uniquely individual, like writing original verse or generating creative insights.
But if we look deeper, they are not as unique as they seem in the beginning
but are part of the larger trained behavior; we see just the tip of the iceberg,
and do not see hundreds of failed attempts.

4.3. Reordering of facts based on their predictive level

We will assume for now that output F consists of the sum of independent facts
or behaviors. In reality, they are connected, but future AI will be able to
unravel them, so we can think of them as functionally independent.

In that case, the total behavior could be presented as sum of small behaviors.
For example, I sleep 8 hours, like to eat an egg on breakfast, read a spam
email at 12h 11min 37sec. All these traits have different predictive value about
my future behavior.
Now, my total past behavior F = sum (F1, F2, F3) and it could be extrapolated
to my future behavior, F’. But different f has different predictive values. The
fact that I slept 8 hours strongly predicts that will sleep around the same time
the next day. The fact that I likes egg predicts that I will probably eat egg
again, and the fact of reading the spam email all doesn’t predict anything (but
my typical reaction to it may be predictive).

Thus, different facts about my past have different predictive value and we
could renormalize their order so the most predictive facts are first: F (f(P1),
f(P2… f(Pn), where Pn is the predictive power of each fact, and Pn  Pn + 1.
The way we measure the predictive power of a fact is not important for now,
but it is correlated to the time of the future behavior it predicts.

From the point of view of digital immortality, it is clear that we should collect
the most predictive facts first. We will discuss which facts are actually
predictive and how to extract them later in the section 5.

4.4. Common, unique, and random facts

Another way to classify facts is their level of individuality or “uniqueness”. The
output F consists of 3 types of elements:

1) Common facts. These are traits which are shared by all humans or
everybody from my social group, so there is no need to preserve them.
Every human has parents, has two eyes, has reflexes, etc. Not
surprisingly, common facts are typically the most predictive. The
absence of some common fact may be unique information, for example,
if a person has only one eye.
2) Unique facts. These facts describe traits of the person which
differentiate him from other people.
3) Ransom facts. These are facts which have very small predictive power,
like news read, the shape of the nail on my left thumb.

From the point of view of digital immortality, we should record all unique facts
first, as they are most predictive of unique behavior.

4.5 Valuable facts

There are also unique and predictive facts which are trivial and don’t provide
any information about individuality of a person. For example, my alarm clock
rings at 12:01 PM every day. It is unique and predictive, as the next day it
will go off at the same time again, but there is almost no difference about my
personality if it is at 12:02 PM every day. The value of the fact is almost zero
from my point of view as well as my typical peers (AI may extract some
predictions from it, but I don’t regard this fact as important trait of my
individuality).

In the other direction, the fact, that I prefer Vermeer over Rafael is an
important fact about my self-determination, but its predictive power is lower,
as I do not encounter the work of these artists every day.
4.6 Combining all types of facts
Based on all that was said in this section, we may conclude that for the
successful future reconstruction we need facts about the person which satisfy
three conditions. They must

1) Have predictive power

2) Be unique
3) Have subjective and objective value.

These conditions will help us to design tests which collect the most useful data
quickly and cheaply, and also helps with data compression, which is important
for storage.

All 3 types of properties should be present in fact for it to be important for the
digital immortality, as we could imagine situations where only one of the
properties is present, and the fact is not suitable for DI:

- Predictive but not unique fact: number of fingers

- Unique but not valuable fact: the shape of the nail on the little finger
- Predictive and valuable, but not unique fact: pin-code from the bank card

4.7 The size of human mind

According to the study "How Much Do People Remember?" (Landauer, 1986),
the volume of the conscious human memory is about 109 bits, or 125 MB, and
the average memory creation speed is 2 bits per second. These estimates
were obtained by a set of experiments which asked groups of people to
memorize words, images, and included subsequent playback at different
intervals. (Crevier, 1993) provided an estimate of total human memory of 2.5
GB in the book "The tumultuous history of the search for AI".

The inspiration for the hero of the film “Rain Man", Kim Peek, was a savant
with phenomenal memory. He knew 12,000 books by heart (N. Allen, 2009),
which roughly corresponds to 6 gigabytes of information. Thus, even a
memory savant can’t have a memory much over the range of several
gigabytes.

The size of the Oxford English Dictionary text is 540 MB, corresponding to 310
000 entries. Most native speakers know fewer words. However, we could use
this size estimate for this part of memory.

Thus, different estimates of the size of human conscious memory fall between
0.1–6 GB. Based on the size of the Dictionary, even unconscious memory is
not much larger. In the following, we will use estimation of 1 GB of target
data, what needs to be collected about a human to enable DI. But not every
gigabyte equally counts!

Those parts of memory that are more difficult to access are larger, but of less
importance, and as a result, we can assume that these two tendencies
compensate for each other. Obviously, this estimate immediately raises
questions about the role of 100 billion neurons in the human brain and the
1000 synapses of each, which give absolutely different estimates of
information in the brain, on the order of 1016, or a billion times greater than
our estimation of conscious memory.

If you believe that your personality is a powerful supercomputer, try to

memorize the number 105578932879 and repeat it in 5 minutes. DI treats
the personality as a program, and the brain as its carrier. Thus, there is no
direct connection between the amount of information in the brain, and the
amount of information in the mind. Similarly, there is no connection between
the number of atoms in the flash drive and the number of letters in a document
stored in it.

But if we consider that estimate of 1 GB of memory includes a lot of junk data,

like non-unique, non-personal, random, or never-used memories, it seems
that extracting of around 10 MB of data could be enough to extract most of
unique, valuable and predictable information. We will call these “essential
memories”. The amount of data contained in these essential memories is still
very large, on the order of 5 times more than “War and Peace”.

4.8 Principle uploading speed limit and technical feasibility of digital

immortality
The main problem of extracting information from the brain is that a person
can consciously give out a very small amount of information, in contrast to
the speed of visual perception, which is up to 1 MB/s in the optic nerve (Jonas,
Schmidt, Müller-Bergh, Schlötzer-Schrehardt, & Naumann, 1992). For
example, a person can, in principle, write 20 pages a day by hand or on a
computer (about 40 kB), if he devotes all his time to that purpose and is not
at the same time a professional secretary. Such writing is equal to an
uploading speed of 14 MB a year, and 1 GB will be uploaded in 70 years. The
holder of the Guinness World Record for the most prolific writer, C. Tellado,
has written 4000 novels, which may be equal to 1.6 GB, assuming normal
novel size (Thetimsonline, 2010). But most prolific professional writers wrote
less than the 1 GB of total human memory. Isaac Asimov published 7.5 million
words across 108 books, equal to only 45 MB, by the age of 49 (Nichols, 1969)
and wrote even more after. Prolific AI safety writer E.Yudkowsky’s
“Sequences” are 1 million words (ciphergoth, 2011).

The quickest way of uploading information is as speech, which is typically

equal to 1 page every 2 minutes. A person speaking for several hours in a row
could maybe read 100 pages, but for most, such speech would be unnatural
and exhausting.

In general, we can assume that a person can hardly afford to spend more than
100 consecutive days of clean time on self-description—after all, one must
also live. Otherwise, a description of one’s life will consist of a description of
the description process. Thus, the maximum size of autobiographic
information is not much than 10 000 pages, or around 10 megabytes, which
is 100 times less than our full estimation of conscious memory, but should be
enough, based on the “essential memories” concept.

Other instruments, like background data recording and DNA tests, could help
to ensure that needed data is collected.

4.9 Uploaded data as a hash of the “mind function”

In previous sections, we have generally ignored the fact that collected data
need to be deciphered to be converted into predictive functions. If we have
excess data, such deciphering is not needed: we could just train a neural net
or use other types of machine learning to extrapolate from the data. Current
neural nets are using similar methods to learn to drive the cars: they are fed
hundreds of thousands of hours of data and then become able to predict
correct driver behavior without knowing rules.

But in many cases the output of human behavior causally corresponds to

human traits but does not actually present the traits. Sometimes this may be
just because it is a non-verbal skill, or because a person has some secret
motives or unconscious desires. The situation in some sense is similar to the
generation of a so-called cryptographic hash function (Preneel, 1994): this is
a computationally complex transformation of input data which allow outsiders
to identify, but not reconstruct, the original data.

In the same way, personal handwriting style can identify human’s unique
handwriting skill, but we don’t have any idea how actually reconstruct the
skill—and because of this feature, handwriting is a method for identification.

These factors combine to explain why we think that only superintelligent AI

will be able to reconstruct human personal traits based on collected
information. The computational complexity of such reconstruction may be
extremely high and may require running quantum computers or world
simulations, which will be discussed later.

We could also deliberately put the human mind, which we want to record, in
situations where it generates such hashes:

1) Filter situation. The mind works as a filter of known data, for example,
drawing a still life. In that case, distortions of the input works to reveal
unique features of the mind, like its digital signature.

2) Free generation situations. Such situations may include free drawing

based on imagination and automatic writing.

4.10 Combining the declining predictive power of facts with our identity
threshold
In section 2, we stated our ideal: the ability to predict 99.9 percent of
behavior, plus a full indexical core. If a person’s behavior is not evolving very
quickly (as it seems true for most adults), that would mean that we need to
record 99.9 percent of previous behavior to achieve that level of future
prediction.

And by saying “behavior”, we should include here “internal behavior”, that is,
thoughts, dreams, and hidden emotions.

The task of recording so much information seems to be complex, but our idea
of “renormalization of facts”, and the idea of the half-life time of memory
survival, provide an opportunity to record much less. Because of the declining
predictive power of facts, recording the first N facts will be enough to predict
future behavior to the needed threshold of 99.9 accuracy.

To estimate the value of N, we need some model of how the predictive power
of facts declines, and also an assumption that we know how to find and extract
predictive facts ahead of other facts. If predictive power declines
exponentially, we would not need to record many facts; however, it seems
that it declines slowly for most human beings. See fig. 1 for a representation
of the estimated relationship between the predictive power of different types
of facts.

The main question is how we can predict which facts are predictive, so they
could be extracted first. One idea is that human behavior is cyclic, with cycles
during the day and week, and if we record several such cycles, we will be able
to predict the next cycle, as wells as the typical level of deviation between
cycles.

Another idea is that the most predictive facts are those that already affected
my past. If that is the case, something I remember often is more predictive,
and if I never remember a fact which I know, it is less predictive. Thus,
predictive facts take more time in the past.

The third idea is that humans are in some sense “holographic”, that is, smaller
parts of their behavior resemble larger parts, and everything done by an
individual seems to be done in their personal style. Some popular
psychologists suggest that if you want to know how a person is during sex,
look how the person eats—and this idea illustrates coherence of the personal
style in different activities (Roberts-Grey, 2015).

Training datasets and environments also have strong predictive power. For
example, learning the person’s native neighbourhood could help to predict
types of behavior, like jokes, she may like. Similar environmental factors may
include her musical playlists and composition and habits of friend circle (even
now “shadow profile could be created about a person who never signed up on
the Facebook (SpiderOak, 2016)).

Internal “feeling of self”—sometimes presented as a core of personal identity—

could be learned from examining a person’s posture as it is presented via
contraction of some muscles. Another part of “feeling of self” is the internal
image of oneself, and it can be extracted by self-image drawings or
transformative games.

Fig. 1. Predictive level of facts. Unique facts, which are needed for DI, form a
rather narrow strip between common and random facts; inside this strip, the
curve declines almost linearly and doesn’t change much.

4.11 Individual variance of predictability

It is not easy to estimate in advance the speed of decline of the power of
predictive facts for typical human beings. It will require extensive
psychological research, and it will surely vary between individuals; it is much
more difficult to predict Einstein then to predict Mr. Smith. It seems that
younger individuals with higher intelligence are less predictable, and older and
less-intelligent people are less predictable.

However, a person, who underwent long “individualization” in the Jungian

sense, for example, a self-taught artist, or an original thinker, may have many
more individual traits than a younger person who is primarily shaped by his
education and environment.

Older people (not affected by memory decline connected with aging) are
carriers of large amounts of unique information.
4.12 Predictive facts as limits in the space of possible minds
It is wrong to think about predictive facts as simple behavioral patterns which
future AI will mindlessly replicate. This idea affects current creators of chat-
bots of the deceased people, who try to hard-code replies in the style “if X
then Y”. Such replication would not be a real person and will invoke negative
feelings from the public who will feel that it is a fake. Such replicates will be
damaging to the idea of digital immortality.

Another view of predictive facts is how they “slice” the set of all possible minds
(Yampolskiy, 2014). If we know that a person was born a woman, it
automatically slices the set almost in half. The best predictive facts are those
that maximally slice uncertainty in the set of all possible minds. We could
know now that some facts are very strong, but we can’t predict how exactly
they will affect the behavior. For such a prediction, we need superintelligent
AI; without it, our prediction will be wrong.

5. Cost-effective extraction of the best predictive facts

5.1 The problem of the cost-effective extraction
Reasons DI is needed (section 2), criteria for informational identity (section
3), and the informational theory behind DI (section 4) have been explored.
From this basis, an important question can be formulated:

What is the most cost-effective way to extract the best predictive information
about a human being?

The question is important from the practical point of view. There is no reason
to try DI if it can’t be done in a cost-effective manner. Cost-effectiveness here
is relative to two factors:

1) Limited resources. A person can’t spend all her money and time on
constant recording of everything, as she needs time to live her life and
invest in other life-extension opportunities.

2) Different prices of predictive facts. Even constant recording could be

done wrong if an important source of predictive facts is lost.

To solve this problem, existing sources of information about typical human

beings will be examined in this section, and different extraction strategies
analysed in the the next.

5.2 Human behavior

The most obvious source of information about a person is her behavior, which
consists of:

 Symbolic behavior, mostly in use of language via speech and writing

 Actions, which include moving, dancing, working
 Expression of emotions, e.g. changes of facial expression
Here human "internal behavior” also should be added, which can’t be accessed
directly but could be reached using different advance technologies or self-
reporting:

 Thoughts (stream of consciousness)

 Internal mental images
 Emotions
 Dreams

5.3 DNA data

Another important source of personal information is human DNA. We know
that twins are very similar in many aspects, because they share the same
DNA. Any human cell has a full copy of a person’s DNA; thus, preserving even
a small amount of tissue could provide us with an enormous amount of
predictive information.

The size of human genome is 700 MB, but the individual variance is around 3
million changes, requiring about 125 MB to record, as for each change both
the type and its location must be recorded (Robinson, 2014).

However, the expression pattern of DNA (phenotype) is individual, and

changes with age. This depends on epigenetic factors, including methylation
and which of the two copies of a gene is dominant. Some tissues probably
could also preserve methylation patterns, but they may be different in
different parts of the body. The pattern of gene expression diverges during
cells differentiation (Ochab-Marcinek & Tabaka, 2010).

There are other factors that affect gene expression. For example, gene
expression is more complex in women, as they have two copies of the X
chromosome, and one of them is inactivated early in life, which happens
randomly in different cells (Learn.Genetics, 2018). This produces different
body coloring in cat clones, as color gene is on X chromosome. Some people
are chimeric (Praderio, 2017), and have cells from two different twins which
merged in womb on early stages.

Preserving tissue samples provide more information than just DNA. Even now,
archeologists are able to reconstruct what a person ate and where she
travelled based on the isotopic composition of bones. Tissues samples include
microbiome, viruses, and even could have pieces of other people’s DNA, if it
is from surface skin. Hormone level, toxins and maybe even RNA from brain
(it was recently discovered that some neurons express capsids containing RNA
(NIH, 2018)) also probably could be found in the tissue samples. The best
tissue sample is full-body cryopreservation; many providers of cryonics also
provide facilities to preserve tissue samples.

Hair is not a very good sample source, as special enzymes damage its DNA.
One of the simplest solutions is blood and epithelial skin cells samples from
inside the cheek, collected by cotton swabs. Dry blood is also relatively well
preserved.
5.4 Digital footprint
Internet usage is very a distinctive part of human behavior. Humans currently
spend a lot of time on the internet, and large amounts of this activity is
recorded. Facebook records all news you have seen, not to mention
information about which articles you liked. Your browser cache keeps track of
all sites you visit. Most people openly share enormous amounts of information
on social networks, and even if you don’t have an account, you will appear on
the accounts of other people. There are rumors that some devices are
constantly recording you, and voice assistants like Amazon’s Alexa probably
do this.

This information costs money, as it provides marketing opportunities and AI-

training possibilities that motivate companies to preserve it. However, the
ways humans express themselves on the internet are socially restricted and
affected by fashions. For example, there is a current fashion to post photos of
one’s meals on Instagram. Such fashions make information less individual.
There are also topics about which people will not post, as they may be illegal,
compromising, or just embarrassing.

5.5 Relatives’ memories and expectations

Other people have a lot of memories about us, as well as expectations about
our behavior. While this information is surely biased, in some cases it could
be the only source of information about a person. These sources will be
especially useful in case of immortality-for-other, as in that case the criterion
and the source of information is the same; thus it is possible for a
reconstruction to completely comply with expectations.

5.6 Security camera recordings

A lot of data is preserved by commercial and governmental security
organizations. Video is the best external source information about a person,
as it is most informationally coherent and informationally rich. It provides
action, speech, and emotional clues, as well as context.

5.7 Personal belongings and archives

The environment a person creates carries the imprint of his personality. It also
shapes his line of thoughts, and may carry small pieces of DNA among other
forensic information. Handwritten archives, bills, photos from childhood also
are large potential sources of data. All of these features provide valuable clues
for DI.

5.8 Medical brain data

The idea of DI is based in the understanding brain uploading is currently
impossible. However, because of developing technologies, new ways of
getting data from inside the brain are becoming possible. There are two main
currently available instruments:
1) Multi-channel EEG. The best systems currently provide 128 channels of
data with high resolution. Simple wearable systems are available, like Muse,
which has 4 channels.

2) Functional tomography. fMRT is expensive and potentially risky procedure

which currently not available for DI goals.

Both instruments have already been used to reconstruct data from the brain.
EEG has been used to reconstruct hands movements (Kim, Biessmann, & Lee,
2014). Deep learning based on EEG data was used to reconstruct visual
images (Shen, Horikawa, Majima, & Kamitani, 2017). EEG was used to
reconstruct words from a brain and it is expected that in 5 years it could be
used to record thoughts on a smartphone (Jonston, 2017).

Electrodes attached to the throat will probably allow reading of small changes
in muscular tension resulting from the internal thought stream—but this data
will still require machine learning to reconstruct the real words that are being
said. Some research on the topic was done by I.Trapesnikov (personal
communication). This technology creates the opportunity to build a wearable
device which constantly records one’s thoughts, data which will provide a large
boost for digital immortality.

5.9 The subconscious—should it be regarded as part of the personality?

There are some processes in the brain which surely are not parts of our
personality, like the activity of glia. Some unconscious processes may affect
the self, but they are almost external to the personality; they could appear in
dreams, like Freudian complexes. The situation is even more complex in the
case of multiple personality disorder.

5.10 Human social graphs and the need for collective digital immortality

As humans are extremely social and mirror each other behavior,

reconstruction of one’s social connections provides a lot of knowledge about
an individual.

Collective digital immortality means that a person will be resurrected not

alone, but inside his social circle.

5.11 Unexpected reactions

Another question is: should parts of myself, like my reactions to stimuli I have
never experienced, like certain type of pain, be considered part of my
personality? Probably, that we could safely ignore them.

5.12 Intersection of different data sources

It seems rather obvious that combining data sources will offer better
prediction, in much the same way two spotlights from different directions are
able to better highlight a dark area.

5.13 Sources of human individuality

There are three main sources of human individuality, that is, the uniqueness
of a person, in the reductionist world model:

1) Initial conditions, i.e. DNA.

2) Training dataset, summed up as total education.
3) Random processes in the brain, which stack one on the another and create
individuality.

If future AI has data on all sources of individuality, it will be able reconstruct

this individuality. However, some part of individuality is “inborn”, like the
property of 27 = 33. This individual property of the number 27 doesn’t depend
on the way in which 27 is reached, e.g. 26 + 1 = 27 or 81÷3 = 27.

6. Main strategies of DI: passive and active uploading

6.1 DI as an acausal deal with future superintelligence

DI implies some form of partner relationship with future superintelligence,

which is counterfactual. Different forms of such deals have been explored
(Turchin, 2017b). We don’t know if superintelligence will appear, and we can’t
know its final goals and capabilities. However, we could influence its goals by
changing the ways in which we create AI.

Anyway, we assume that future superintelligence will have enormous

computational capacity and also will have a goal to return dead people to life,
in line of thoughts of cosmist Fedorov (Fedorov, 1903). Thus, in planning for
DI we are collaborating with future AI, and we should try to help it get the
information which will be most useful for the future reconstruction of people.
We should also try to create such AI in the future and take steps to ensure
that it will be benevolent and interested in the resurrection of the dead.

But the task has significant intrinsic uncertainty, as we don’t know how much
information is needed or which information will be available from other
sources that we can’t even imagine now. Some examples we can predict are
global archeology and quantum archeology, resurrection simulations, and
trade between multiverse branches, which will be discussed later.

By choosing a strategy for data collection, we make some assumptions about

the abilities, and indeed, existence, of future AI:

a) Superintelligence (SI) will appear.

b) SI will want to resurrect humans with human-benevolent goals, not for
strange experiments but to help people.

c) SI will be not omniscient; omniscient SI would make data collection

unnecessary.

d) SI will able to solve any equations with unique decipherability; thus, we

must collect data but will not need to decipher it).

e) SI will appear relatively soon—in the 21th century—and will have access
to all current common human knowledge, so we don’t need to preserve
non-unique information.

6.2 Main strategies of DI: data-driven, experiments and self-description

There are several main strategies for personality capture:

a) Data driven. Here it is assumed that an enormous amount of data is all

that is needed, and future AI will able extract from it all the answers it
will need. The main problem of this approach is its “behaviorist”
philosophical basis, that is, it completely ignores personal subjectivity,
will, and the ability to have secrets. Important predictive facts could be
easily lost.

b) Self-description based. This is based on the idea of Rousseau, that a

person is an inside observer of her own psychological processes, who
can consciously conduct experiments on herself, determine the validity
and predictive power of data, extract needed files, and even edit her
own personality. The core method here is to write a complete and
detailed autobiography.

Self-description is the highest level of uploading, and raw data collection about
the brain is the lowest. Somewhere in between lie experimental testing and
other ideas:

c) Experiments and tests. A subject runs a series of tests on her own brain
in order to extract data via some technological setup. A fictional example
of it could be seen in the movie Transcendence (Pfister, 2014), where
main character was saying associations to words he heard while EEG
signals were recorded. The evolution of this approach will probably
evolve into real uploading

d) Side-loading. In this case, a model of the person is created

simultaneously with data collection and calibrated based on human
output. The model is based on some “human matrix”, that is, a model
of the human brain without any individual traits. This approach is not
currently widely available, but it is possible; in one case, a son created
a father-chatbot and calibrated it with his father while his father was
 still alive (Vlahos, 2017). This idea has also appeared in the novel
Zendegi (Egan, 2011).

e) Personal assistant evolving into “exocortex” and later into a digital copy
of the owner. The major benefit of this approach is that it could be
commercialized right now, and people are already investing in some
digital helpers, like Evernote. The computer has already replaced our
memory as the main instrument to preserve photos.

6.3 Not a black box, but a black torus

One way to describe DI is by analogy to a black box of which one wanted to

create a copy. The copier doesn’t know what is inside the box, but she could
record input and outputs signals. Thus, she could create a table of input and
outputs, and when the table is sufficiently large, she will be able to predict the
output with increasingly higher probability. To collect the data, she needs to
put the black box in many maximally different situations.

However, the black box metaphor is an oversimplification, because inside the

“black box” there is a conscious observer, a person who can observe her
mental processes from inside. She could run experiments, address and extract
memories, and make conclusions about the nature of psychological process
she observes. She also has some accumulated experience about herself which
she could present in the form declarations of personal traits.

Such a situation, a “black box” with an observer inside in it, could be called a
“black torus”.

6.4 Combining strategies

The main way to combine strategies is use them at different times, though
techniques should be consistent over time. Passive data collection could
happen constantly in a background mode. Active data collection could combine
different methods of self-description and self-experiment, but happen during
discrete intense sessions.

After initial data collection, marginal backups are needed every so often, as
personality changes. I estimate that major data uploading needs to happen
approximately every 10 years, especially given that completely new uploading
technologies typically appear during such a period, and also because of
substantial personality changes may occur on such timescales. Shorter, less-
intensive sessions are needed every year, which could coincide with proper
data backup from hard drives.

Background data collection should happen all the time, and include speech
recording, preservation of digital archives, informational photo and video
collection, as well as preservation of random papers (like bills and tickets).
7. Digital immortality and art

Works of art meet the requirements of predictive power, uniqueness and value
described above for DI. Every person has a unique art style, first of all, in
drawing, in the same way everyone has a unique style of handwriting.

Art is typically a part of the best and most valuable things a person does. We
love our friends for the stories they tell, the food they cook, the pictures they
draw—and the way they react on us.

The artistic style of a person has predictive power over what she will create
next. Style evolves with the creation of new art objects. Style is something
like as a personal “hash function” or signature. A full and correct copy will
create objects in the same style.

Creation of complex art requires coordination of all brain systems, so the

complexity of the output corresponds to the complexity of the human mind.
Many human actions don’t have such complexity—they are automatic, like
walking, and a person does them mindlessly, devoting to them only part of
their brain activity. The information traces of such actions have small
algorithmic complexity and they are not footprints of the whole mind.

The world is constantly collecting more and more data about people using the
Internet and video surveillance, and is investing in the storage of these data.
But this is only passive information, and there is little of the essential and
individual in it. Academic art education also kills individuality of style.

Contemporary humans are affected by powerful streams of data via mass

media, which produce standardization of human minds, killing individual
traces. The mechanism of it is “unconscious learning”. Naïve or outsider artists
of the past, like Darger, lived in hermetic worlds where they had time to evolve
their personal style, training on their own personal output.

Different types of art could be used as instruments to get special types of

information. Spontaneous drawing presents content from the subconscious.
Self-portrait is a presentation of person’s idea of Self. Drawing from nature
acts as an information filter, and so we learn what kind of filter it is.

But what is art? It is not only drawing: singing, a story about life, dancing,
theater, cooking—any complex unique activity acts as an imprint of
personality. Moreover, different modes of artistic expression give different
angles on the human’s personality, so it should not be limited to just one
genre.

Art objects do naturally last better. Pictures can exist for hundreds of years
and their value is recognized, so they are not likely to be intentionally
destroyed. Novels can be copied.
If a neural network is maximally amplifying random noise, it could reveal the
type of data on which it was trained. This is how Google’s “dreaming” networks
work, and it provides insight into image generation in the human mind
(Keshavan & Sudarshan, 2017). Human dreams are also revealing when
considering the human subconscious. The same way, similar data could be
extracted via spontaneous drawing—or on crystal-ball gazing, where random
noise is amplified.

See my presentation: “DI and art” https://www.slideshare.net/avturchin/ss-

80472635 (Turchin, 2017a)

8. Post-mortem DI

Even if a person has died, it doesn’t mean that we have to stop fighting for
his life. We could try cryonic method of plastination—that is, brain
preservation with the goal of the future resurrection. But even if the brain was
not preserved, it is not the end. A lot of data could be collected about the
person, as has been done about some famous people by their biographers. I
and some other researchers in digital immortality have collected data about
their dead friends and relatives.

Information held by external carriers decays after a person’s death, so it

should be collected as soon as possible, within the first months and years after
death.

There are several main sources of data for post-mortem digital immortality:

1) DNA samples
2) Interviews with friends and relatives
3) Personal archives + hard drives + data on obsolete carriers + other people’s
archives
4) Belongings and home.
5) Internet archives, including browser cash, websites and social network
accounts.

Often surprisingly small amounts of data remain despite availability of

videorecording and other data-gathering systems in the modern world.

9. Integration potential of DI
DI could be combined with several other powerful instruments, technologies
and projects, enriching their potential and lowering prices.

9.1 Growth of technologies

The exponential growth of information technologies makes DI simpler with
every year. We can collect more data, we have new instruments to examine
the brain, and a lot of data is now collected automatically by the internet,
banks, and security agencies. Every person now is better-recorded than most
people were in the 20th century, except some celebrities or authors of
autobiography.
Technological acceleration also provides new methods of data storage, but the
situation here is not as good. A lot of data is lost on dead drives or to
abandoned or unpaid accounts. It may have a better chance of survival if it
were printed on paper. However, some effective solutions exist, like The
Internet Archive, eternal M-disks, torrents, and blockchains.

The growth of information technology also implies that superintelligent AI

could appear relatively soon, within several decades of the present day, and
could recreate us, as well as collect data from currently untapped sources.

9.2 DI + cryonics

Cryonics is another instrument that can be used to reach future technological

resurrection (Best, 2008). Surely, a lot of data from the brain will be lost if
cryopreservation procedures are not perfect. Most likely, frozen—or more
correctly, vitrified, preserved in a cold, glass-like state instead of crystallized
by normal freezing—brains will be a source of data for future AI.

Such data could have holes, and digital immorality data could be used to fill
in gaps and check the quality of the resurrection procedure.

Cryonic preservation of other people and even animals may be useful my

personal digital immortality, as they have memories about me. Some people
may have eidetic memory, which is almost like video, even if they can’t access
it consciously.

A cryocompany may be used for the preservation of discs carrying data as

well as cryopreservation of tissue samples.

9.3 DI + direct mind uploading

Constant development of the technologies that enable digital immortality will
eventually merge with mind uploading technologies. Better ways to extract
information from the brain via brain implants, neural dust, etc., will give a
huge boost to digital immortality before full uploading is possible.

Some intermediate forms between DI and mind uploading will probably

appear, like brain implants and personal assistants.

9.4 DI + exocortex
Creation of the exocortex, that is, a brain implant and other systems which
collectively increase the intellectual potential of a person, will allow collection
of large amount of data about the individual.

9.5 DI + quantum immortality

So-called “quantum immortality” is one particular case of “multiverse
immortality” (Turchin, 2018), the idea that in a multiverse, all possible minds
exist, and there will be always my next observer-moment, moment T+1 for
any T. In particular, the situation of quantum immortality arises if the many-
worlds interpretation of quantum mechanics is true; in that case, in any
situation where I can die or survive, there is a possible future time-line where
I survive.

But the problem is that it is not good per se, as most likely I will survive in
some “bad immortality” time-line, for example, eternal aging without ability
to die.

However, if digital immortality exists in the world, there is the possibility that
my next observer-moment will be recreated by digital immortality, and the
probability of this is much higher than the future of eternal aging.

Some lost data may be filled via quantum randomness generator (Almond,
2006). In a many-worlds multiverse, all needed data for the person’s
resurrection will appear in some timeline. It will not result in the “measure”
decline despite widespread belief, as the same experiment will be done by AIs
in other worlds with other instances of “me”, so the total sum of measure will
not change.

9.6 DI + Alzheimer

With aging, humans lose many personal traits which have value to them.
Dementia, a common disorder of aging, causes patients to lose memories
rapidly; however, even healthy people tend to lose memories. Thus, digital
immortality could be useful even during the natural lifespan. We could use it
personally or with help of future AI to return important memories, to return
to being our “actual self”, not the bleak shadow which we become due to
aging.

9.7 Resurrectional simulations

It is rather probable that the future AI will make large ancestral simulations
(Bostrom, 2003). It could do it for different reasons, some of them rather
technical, like modeling the conditions in which it appeared.

But it could also do one simulation of the past to resurrect all people who have
ever lived. In that case, it will use all data not to model just one person, but
to model all of human history. Such a model may be the most powerful
instrument to extract useful predictions from available evidence. Some people
who lived in the recent past will be modeled almost exactly, as there are a lot
of data about them and thus uncertainty is small. But uncertainty will increase
with distance in the past.

After a person lives his whole life in the resurrectional simulation, he will die
in it, but appear in some form of “afterlife”; there all his personal data will be
integrated in a whole and perfect personality. Surely, the best strategy how
to resurrect people is in hands of the superintelligence.

There is a possibility that we already live in such a resurrectional simulation.

9.8. DI as psychotherapy
Many forms of therapy create a lot of data, and by practicing DI, the patient
in therapy could also solve some of his problems and obtain a better
knowledge of himself. Art therapy is an established field of psychotherapy.
Another direction is self-analysis in Freudian terms (Horney, 1942).
Automotive writing also could be used as an instrument for psychological
problem solving, similar to Stone’s voice dialogue (Stone & Stone, 2011).

Thus, it may be useful to record data during psychotherapy (which is often

not recorded or is later deleted).

9.9 Speed of personal changes and DI

The older a person is, the more slowly she normally changes. Quicker changes
mean that DI data become obsolete more quickly. The changes between ages
10 and 15 are enormous, but in most cases the changes between 50 and 55
are rather small. However, there are people who evolve all their lives or
experience a transformative experience at a later age—though their number
is small. It is much more difficult to change personal traits at older ages, and
long-term memories are more difficult to form. At the later stages of aging
the changes becomes quicker again.

9.10 DI and family history

Genealogy is one of the most popular hobbies in Western countries. Collecting
data about old family members may contribute to two goals: proper digital
immortality and creation of better family histories, which will be available now.

9.11 Social aspects of DI

It looks like DI is largely a solitary activity, but many social activities could
help:
- Create a startup
- Do it with friends, both for motivation and to polish the protocol
- Hire a biographer
- Become a famous person, so people will be naturally interested to collect
and preserve data about you.
- Take actions which have a bigger footprint in the internet. Try to correctly
present your views.
- Don’t lie and don’t keep secrets. It could damage your perspective of DI.
To do so, don’t pursue a career which requires a lot of lying about your
personality (like a con artist).
- Help others to collect their data. It will train you in DI techniques, and you
will also be recorded when you ask questions. Another bonus is that the
people around you will be correctly resurrected with you.
- Take chances to act uniquely and record it. Agree to give public
presentations and interviews.
- Promote the idea of DI—more people collecting DI data increases the
chances that effective technologies will appear and that future AI will
include DI-based resurrection in its agenda.
9.12 Age-related strategies for DI
For a young person, we are mostly interested in what she can be, but for an
older one we are more interested in what she was, as she loses many valuable
traits because of aging. This implies a different focus of interest of DI, the
younger person. We may be interested in her creative style and in the
direction of evolution of this style, which could be called “style of style”, while
her memories are of less importance; she may forget most of them soon
anyway. For older people, we are mostly interested in well-established skills,
habits, and memories.

This is all connected with the notion that the personal identity of a human
being is a complex socio-biological adaptation, which consists of several
mutually supporting parts, and evolves with age.

10. Digital immortality as a cause of effective altruism

Surprisingly few people have ever been interested in digital immortality,

probably less than 1000 worldwide as of 2018. Terasem reported 500
subscribers to their services; in Russia, I personally knows around 10 people
who practice data recording for DI. These numbers are even smaller than
number of people interested in cryonics, which has several thousand
subscribers. However, DI is the cheapest way to immortality. Anyone who
owns a computer and could afford to spend around 1 month on the project,
could collect significant amount of data about his personality, which
significantly increase the fidelity of his future reconstruction. DI could even be
done without a computer, using paper and pen.

Reconstruction by an AI means one will be effectively immortal. Thus, DI is

the cheapest way to save people from death. It could be estimated that
preliminary data collection may cost around 1000 USD per person, mostly in
form of the time. Advanced forms of DI data collection could be more
expensive, as they include EEG, functional tomography, and an interview
team, but may still be cheaper than high-end cryonics (see our similar take
on fighting aging as a cause for effective altruism (Turchin, Denkenberger,
Milova, Egorova, & Batin, 2017)).

Preservation of information about other people will help to reconstruct an

individual’s possible reactions to them. Thus, collective digital immortality is
the best option, as it provides mutual help in the reconstruction of each
person. It also makes all procedures cheaper and increase the probability that
reconstruction will happen.

11. Theory behind different recording technologies

In section 5, I analyzed different data sources. Here, I will look at some

existing data-extraction technologies and analyze the effectiveness,
shortcomings, and the best practices for the implementation of each.
11.1 Association tables
In the movie Transcendence (2014), the main character plays with free
association. However, in experiments with Trapeznikov, we found that such
experiments are mentally exhausting and put one in a trance-like state. In
addition, most replies are not unique.

Another method is to draw all associations as a small mindmap around the

central word.

11.2 Encyclopedia of things

Many ideas in the human mind are just mental images of unique things, like
“my table” or “my cats”. To upload a person, an encyclopedia of such unique
things is need. So, one may create a list of all types of most frequent unique
objects, and then create descriptions of them, consisting of text and drawings.
Drawings and handwritten text are a good way to capture personal style.
During a self-description experiment in 1990, I found that he had around of
1000 unique objects, after which he was exhausted.

11.3 Encyclopedia of people

In the same way, an encyclopedia of friends could be created. A typical person
has around 150 close friends, plus some historical people, like classmates,
who still appear in our thoughts and dreams.

11.4 Declaration of personal properties

Everyone has an image of oneself, which include some ideas about personal
properties. These ideas are often wrong, as they affected by different biases.
However, self-image is part of the personality and informational identity core,
and all such ideas about the self should be recorded. They are also based on
long-term observation of one’s own behavior and carry an imprint of one’s
values.

Berdyaev self-analysis (Berdyaev, 1950) is full of declarations of such

personal traits. In some moment, the declaration of traits becomes the act
not of self-description, but of self-constructing, as it presents the image of the
idealized “me”, and even could change the self in the desired direction.

11.5 Automative writing of stream of consciousness

Self-description becomes easy after “installing” several skills, like
remembering to start a recorder during interesting conversations. Another
important skill is the ability to record one’s own thought stream in the style of
automatic writing. Just sit at the table and record everything you have in
mind, maybe using handwriting as it creates lower internal censorship. At the
beginning, you will not know what to write, but later, you will adapt to record
your thought stream.

Another important skill for DI is reflective thinking, which is the polar opposite
of automatic writing, as it is the ability to think about what you think and trace
small changes of emotions, the logic of one’s own thoughts, and physical
conditions. In some sense, it is close to meditation.

11.6 Recording childhood memories

Childhood memories may be important part of the personal identity, as first
mentioned by (Maeterlinck, 1907); otherwise godlike future selves will have
no individuality. However, most geniuses have more individuality than simple
men, so the growth of abilities does not produce a decrease in individuality.

Recording childhood memories is important, as according to many

psychological theories such memories affect personality formation and serve
as “psychological DNA”. I suggest recording of all memories until some age,
like 10 years. However, human memory is variable, and some people
remember themselves from different ages. The most traumatic and important
memories probably should be recorded first as they have the large effect on
the formation of individuality.

A person has an understanding what things which has happened with him are
the most important, and these memories should be recorded first.

11.7 Video interview

For a video interview, several important conditions are needed. The first is a
good interviewer. The problem is that many people have a ready-made life
story which they could repeat but which is lacks important facets: it is just a
social presentation of honorable and acceptable facts. A good interviewer can
coax a person to open up and to go deeper in his soul, to make him answer
spontaneously. Thus, it may be reasonable to hire an experienced professional
interviewer. Another option is to use non-standard questions.

Another condition is high recording quality. HD video is needed to capture

facial microexpressions, and at least 2 cameras are needed to get two different
planes, full body and facial. Light should not be LED-based, to avoid blinks.
Another camera should be behind the interviewed person to capture body
language.

From my with Trapeznikov experience, preparing a proper setup for recording

(with EEG) take up to 3 hours, and recording can’t be longer than 1 hour, as
after that time both interviewer and subject are exhausted. High quality audio
is needed, and killing background noises is one of the most useful things here.
Using wearable mics and a ZOOM recorder is helpful.

Another option is relaxed videorecording on a phone, which could be much

longer and doesn’t make people tense. But the cellular connection should be
turned off, to avoid the chance a random call could kill the video, as in some
phones it will be not preserved.
11.8 Professional police and psychological checklist
A lot of personality tests and questionnaires already exist. Police have
exhaustive questionnaires about all personal traits, like “who were your
friends in 5th grade”.

You can provide a lot of information about yourself by liking claims with which
you are agree on Reddit.

11.9 Dream diary + drawings

Recording of dreams may provide useful insights into the opaque world of the
personal unconscious. However, most text about dreams is very
uninformative, as the structure of the dream is broken after awakening. Thus,
it may be more productive to draw pictures about dreams and try to comment
extensively about what you felt and thought during the dream. Japanese
scientists have used MRI for dream reconstruction, EEG may provide better
clues in the future (Anwar, 2011; Robertson, 2013).

11.10 EEG + spontaneous art

EEG gives some information about the brain, but currently we can only
partially decipher it. Japanese scientists presented some form of mindreading
via EEG in 2017, and expect that a smartphone app able to read thought
stream will appear in 2022 (Jonston, 2017). EEG is already used in thought
identification research (Nemrodov, Niemeier, Patel, & Nestor, 2018).

We hope that future AI will able to extract much more information from our
EEG data. But we need to not just record EEG data, but all other activities of
the person at the same time, so we will have audio and video clues for context.
As we said before, creating art is the most complex and valuable thing a
person can do for DI purposes, so it seems reasonable to try to record EEG
while a person is making art. The most suitable form of art for such
experiments are drawings.

11.11 Scanning of paper archive

All paper archives should be scanned, but not destroyed, as paper could
survive longer than many digital carriers. It may be photographed and the
photography or scanning process may be recorded on video, where some
comments will be provided about the contents of the papers, like “this is a
photo of my grandfather”. The papers will be triggers for some reactions; it is
better to this together with a partner as it will be fun and naturally create
opportunities for conversation. Doing it alone may be sad: archives are often
full of past memory triggers.

11.12 Spontaneous theatre

Spontaneous theatre, also known as improv classes, may help to quickly put
a person in many very different positions and demonstrate different types of
reactions. Dialog and other forms of psychodrama are also similar to
spontaneous theatre.
If a person is doing the same activity during every DI recording session, its
informational content will diminish. Changes of activity will provide new
spotlight positions to discover new personality traits.

It should be noted that if a person is a professional in some skill, his

performance of this skill may be less informative about his personality, like
art in case of a trained artist.

11.13 Lists
It may be useful starting point to create several lists of facts about oneself:

Movie I’ve seen + I liked + I hate

Books
Friends
Places I visited: cities, countries
My love
My secrets
My illnesses
My bad days
“Ask reddit” may be a source of unexpected questions.
My cars, my pets, my bags, etc

11.14 Guided visualization and active imagination

Jungian “active imagination” (Jung, 1925) is a form of guided visualization

which uses open questions to maximally stimulate a person’s imagination and
provide access to the content of the subconscious. Some people easily enter
the trance state and see very bright pictures full of interesting subjects. See
my book in Russian about active imagination (Turchin, 2007).

Active imagination may be combined with other recording techniques like EEG
and video recording.

11.15 Idealized world: a novel

One of the important parts of the human identity is the set of personal values,
but values are complex and not easy to extract. One way to capture such
values is to write a novel with an image of ideal world, where a person would
want to live. It will help to capture a core of the person’s identity.

Also, a novel is one of the most complex form of art, which mirrors the all
sides of the personality. It may be supplemented with drawings, music, maps.
Other forms of art are narrower and have less “algorithmic complexity”. A
novel represents not just a set of boring facts, but the best of what I have in
my personality—the best of me, which I want to present to the future.

Self-description is a transformative practice where our individuality not only

fixated, but appear, and where we create an image of the future idealized self.
It is a part of human nature to want to improve and to change one’s own
personality. Thus, creating an idealized image of the self will be not only the
fixation of my current hidden desires, but an act of creation of a “better me”
and instruction to the future AI of how I should be improved.

11.16 Bainbridge yes-no tool

Bainbridge created a questionary consisting of tens of thousands of yes-no

questions that could help to extract a lot of data quickly (Bainbridge, 2006).

11.17 Terasem sites

According to Rothblatt (Rothblatt, 2012), both sites of the Terasem project
have different functionality, and provide different ways of uploading
information. Below is a description of the tests available on both sites, which
may provide inspiration for uploading different data. Some interested ideas
from their list not mentioned above:
Structured Religious Views
Structured Morality Views
Structured Political Views
Reactions to Stock Photos
Biometrics
Favorite Rankings of Music
Literature, Films, Things Places, People & Food Hated Rankings of Music
Literature, Films, Things Places, People & Food Free-Form Journal Entries
Historical Contextualizing
Chatbot Training
Gratitude Test
Free-Form Favoriting
Life Geo-Tagging
Social Network Maps
Temporal Sequencing
Website Linking

11.18 Uploading sessions

Peak performance is needed for DI, as performance at this level will provide
the most unique and valuable information. So, it is important to record
episodes of personal peak performance, like public speeches or telling a story
to friends.

DI also should be practiced in sessions, especially if you do it with a partner.

11.19 “One my day” and videocommenting

In the LiveJournal there is a community, “one my day”, where people take
photos during the course of one typical day of their life and provide comments
with each photo. They typically take around 100 photos from morning to
evening (LiveJournal, 2018).

A more advanced form of such lifelogging would be one-day

videocommenting. This involves carrying a constantly recording camera all
day and providing audio comments about all your thoughts, things you see,
and typical behavioral patterns.

There are several commercial wearable video recorders, like GoPro, but they
have short battery life span and create large videofiles. It is better to use a
different device, like a police-style bodycam, which is worn on the chest and
designed to provide around 10 hours of recording, or a secondary smartphone,
which one could wear in the chest pocket of a shirt—camera-out.
Unfortunately, smartphones tend to overheat. Wearable head-cameras are
also typically heavy and strain the neck, or have small batteries. They also
could be illegal in some countries (including Russia) if they look like “spy
hardware”. Also, head video cameras are typically look ugly or weird and may
affect the social perception of the person who carries them. Solutions such as
Google Glass may not be so bad.

Police bodycam showed provide longest battery life and lowest minimal
resolution for effective data compression; one example is the BC-1, at a price
around 250 USD with 12 hours of battery life

11.20 Audio dictation

For some people, it may be easier to record longer stories as audio (or video),
rather than writing them. Dreams could also be simpler to record in audio
form. Some people speak loudly with themselves, or even speak in dreams.
Recording your personal life story and dreams, recording phone calls, and
dictating stories from your life when you have free time will all provide
valuable data for DI.

The following instruments will be useful:

• Wearable voice recorder

• Phone app for recording conversations
• Voice-activated digital audio recorder

11.21 Testing the mind as an information filter

This is a test how the mind changes the information which pass through it,
which may probably may reveal some internal mechanisms. The main
methods for such a test are:

 Drawing from nature and from memory

 Retelling stories or movies

11.22 Environmental data collection from photos

We may expect that in our age of Google Street View everything about outside
is already recorded. However, taking informational photos about important
places may be still useful.

Take photos of
• Other people, places, events, personal things, body parts.
• Apartment, furniture and item arrangements.
• Important places from personal history.

11.23 Keylogging
Most of activity in current world is spent not with humans but with computers.
Advanced keylogging may be needed to fully collect information on such
interactions. It should include not only the key pressed, but timing, screen
capture, and accelerometer data.

On main computer
• Spy program: Keylogger, screen capture, mouse tracking
• Geo-tracking
• Archive chats from social networks
• Archive browser history and cache

11.24 Memoirs
There are many ways of recording personal information from memory:
• Diaries
• 1000 facts about me
• Collect memories by topic
• Automatic writing
• Write down history of your life
• Write childhood memories
• Describe your body scars and how you got them
• Write everything which you prefer would not be known to others or could be
misinterpreted as evidence of your wrong doings
• Ask your parents and other people who remember you about stories when
you were young and was a child. Maybe they have their own photo, video
materials or diaries about you.
• Write a fictional story based on personal events.

11.25 Diaries and memory dumps

There are several types of possible diaries. One could consist of rather short
description of what a person was doing each day. Another involves in-depth
psychological analysis of thoughts, dreams and emotional states. Diaries
should ideally be written everyday but making entries every week or month
is also possible.

A memory dump is concentrated uploading of accumulated knowledge at its

peak to free one’s brain. It most likely will take a form of an article or memo,
which helps to bring structure to some experiences. This article may be seen
as an example of such uploading of conclusions, as I spent around 30 years
of thinking on the topic.

11.26 Drawings
Drawing is an instrument to output images from the brain, and probably only
one currently available instrument. Some drawing could be done very quickly,
on the order of 1 minute, to present basic details of an image. Drawings are
also very individual in style, and are unique and valuable objects of art.
Collect
• Drawings of important things, places and people.
• Drawings of inner representations of abstract ideas.
• Art-therapy and automatic drawings.
• Drawings of the real world.
• Drawings of dreams.

11.27 Ask other people what they think and remember about you
Internal censorship prevents a person from having an objective self-image.
The genre of memoir is often oriented on editing the past and presenting the
“correct” version of the events to descendants and landing a final blow on
enemies. A lot of memoirs often consist of descriptions of other peoples’
actions. These features all lower the predictive value of memoirs.

But if one interviews other people about oneself, maybe not directly, to escape
“friendly censorship” due to the fear of inflicting embarrassment, the subject
could collect a completely different view of her own personality.

11.28 Futuristic and untapped sources of information

There are several ideas in DI which, as far as I know, have never been used,
but may be used in the future as legal and technological changes allow:

1. Use of psychedelic drugs to induce more active self-presentation. It could

be LSD, memory enhancing drugs like galantamine or some truth serum.
Stimulators will allow a person to speak more and to reach deeper levels of
unconsciousness. Dissociative drugs could help to remove barriers of internal
censorship and a person could tell much more about herself—the most popular
such drug is alcohol. Stimulators could help to work more and produce larger
output. Some chemicals (e.g. dopenzil) are said to be able to induce
hypergraphia.

2. Computerized testing. A computer quickly demonstrates some stimuli like

images and records reactions via EEG. Approximately the same could be done
if a person watches a movie while carrying an EEG monitor. Use of tomography
during uploading sessions not currently known.

3. Invasive digital immortality. Some brain implants will appear long before
direct brain uploading is possible. Musk’s NeuraLink (Templeton, 2017) will be
used for medical applications before it will be used for augmentation and
uploading. Such implants will produce a lot of raw data about processes in the
mind and will likely appear in the late 2020s (Chen, 2017). Implants could
also send testing signals in the brain and record the outputs of some brain
regions. That is, it will treat some brain regions as small black boxes and
record their input-output pattern.

4. Life extension. An older person could still have some data about her past
even if she seems to forget it—maybe she just can’t access it. If she lives
longer—but not long enough to survive to actual immortality—she could
survive until better recording and data extraction technologies. The same is
true about other people who have personal memories of a deceased person—
they could survive long enough that a large amount of data will be extracted
about the people who they remember, like parents and grandparents who died
long before other types of preservation appeared.

5. Use of quantum or other currently unknown sources of information. Some

forms of quantum entanglement may help provide data for DI (Jones, 2017).

6. Global archeology. Future AI may be able to analyses all specks of dust on

the earth’s surface and extract enormous information about the past.

7. Use of optogenetics for better reading brain data. Even getting a lot of data
from the brain is not equal to proper direct uploading. Future scientists could
see every spike of each neuron, but still will not know synaptic weights, which
should be recalculated based on the observed spikes.

8. Preserving raw tomography data. Tomography is known be computationally

complex and even NP-hard (Brunetti, Del Lungo, & Gerard, 2001). There is a
hypothetical possibility to record tomography data about the brain now, but it
cannot currently be deciphered, as it is computationally intractable for current
computers. Such recording would preserve it until superintelligent AI is able
to decipher it.

PRACTICE

If you need super-simple version of the protocol, you need to do the

following 4 things:

1) DNA. Collect own DNA, may be by sending it to 23andMe.

2) Memoir. Write down the history of your life with drawings.
3) Lifelogging. Practice video and audio recording of important
conversations
4) Archives. Archive all the data from Facebook and scan archives.

12.1 Choosing a personal uploading strategy

The main starting point for the DI is to make a decision about beginning of a
personal DI project and determine how much time and money one should
spend on it.

The efforts should be divided almost equally between data collection and long-
term data preservation. Data preservation is difficult, and it would be a pity
to lose the data you collected.
I recommend starting with a period of active uploading and later turn to
passive recording. However, if one doesn’t have the time for active uploading
now, installing some recording software will be good first step.

The main principle here is to tap cheapest and most predicting information
sources first.

• Make a decision about starting DI

• Time and resource allocation
• Choose available resources
• Plan your data collection methods for DI
• Choose the most informative methods first, but also try several methods to
get different viewpoints on your personality
• Quickly upload the first version of your DI information

12.2 Steps of data collection

The advice below may seem obvious, but they are result of trials and errors
of a group of people over several years. Below is a personal uploading plan,
which is broken into “days”, though each task may actually require more than
a single day.

Day 1: Collecting existing data

1. Create a Dropbox or Google cloud account at least 1 TB, which provide
temporary storage of your data (Temporary, because if a person dies, she will
stop paying, and the data will be deleted. Google is a better option, as they
may preserve the data for their own needs and it is likely that
superintelligence will arise from Google.)

2. Copy all already existing electronic archives, including photo, documents,

and video to this temporary storage.

3. Extract personal data from social networks. Facebook allows you to

download all data you have ever created on their platform as one file. Extract
archives from your email services, like GMail.

4. Install audio and video recording software on your PC. For Mac, the best
programs are Evocam and Simple recorder. Evocam records video in 1-hour
fragments from the built-in camera with good compression. Simple recorder
records audio by clicking one small button.

5. Take care that you video files are compressed. Evocam can compress video
to as file sizes as small as 100 MB for 1 hour. GoPro cameras can create
several gigabytes per hour; large video sizes will complicate management of
archive copies.

6. Use your smartphone as a recorder during conversation. Use of a

standalone recorder is aggressive, and a smartphone is better for recording
management and naming. Ask other participant of the conversation to record
it, while explaining that you are not going to use or even transcribe the
recording, but are creating recordings for the sake of digital immortality.

7. Start writing you autobiography. It should be long, emotion-centered and

honest, not just a collection of facts.

8. Make photo and video recordings of your home environment. You may
comment the video.

9. Start writing a diary.

10. Order a 23andMe test. 23andMe is Google-affiliated; surely Google will

preserve your data. SNIP tests are not exactly the same as full genome
reading, but Google will keep your biological materials.

Day 2. Self-description
People are different, and prescribing one way of self-description for all will
create unnecessary unification, where individuality will be lost. So, day 2 is
better spent in reflection to determine which ways of self-expression are most
appropriate for you. You may try different approaches, to see where
information is running freely from you and clearly presents some unique
insights. The more unique way of self-expression you will find, the better it
will fix your uniqueness.

You may benefit from finding partner in self-description, but another option is
take a vacation and spend it analyzing yourself.

Day 3. Running psychological tests

There are many personality tests, you may choose the one you like and
record all your data. It is better to take the test with large number of
questions.

Running Rorschach tests also helpful.

The third type of tests is police tests which they use to collect data about a
person.

Day 4. Recording of the stream of consciousness

Just start writing down whatever thought is coming into your head. It is
important to record every thought without censorship. If you will be able to
write down a lot, like more than a hundred pages, even contemporary
simulating systems like GPT-2 could simulate your internal stream of
consciousness.

Note that recording the stream of consciousness is not exactly the same as
pure automatic writing where you just open floodgates of your
unconsciousness and could write completely unexpected texts. In recording
the stream of consciousness (SC) you should put attention of what is actually
going in your mind in your everyday life, like “I have to call my friend, there
is my pet?”

Anyway, there is no way to do it wrong, as any your mistakes will be

representation of your personality.

An alternative to writing is audiorecording of the thoughts, but it could be

sometimes emotionally distressful, as it will be like confession, and it is known
to trigger emotions.

Day 5. Visiting friends and recording interviews with them

Other people have a lot of ideas about you and they also remember things
which you prefer to forget. You may ask other people to tell what they
remember about you, and surely you will be surprised. Just not forget to
record this on audio with their permission.

Day 6. Experiments with EEG monitoring

Currently it become possible to reconstruct internal thought stream based on

EEG data using neural nets. Thus, recording such data may help in the
future to reconstruct what you actually think. There are two options to
record EEG – buy a commercial system like Muse or Open BCI with 32
channels, which may cost around 1000 USD and some knowledge in
connecting wires etc.

Or just hire a medical professional for some medical tests, which may cost
less in some countries. In Russia, one could get EEG at home starting from
150 USD.

Day 7. Creating art

Art is a unique signature of one’s personality. Different people may have inclination to different
types of art, like singing, dancing or poetry. However, I recommend drawings as they could be
done quick, they are rather unique in style and they could convey a lot of semantic information,
which can’t be represented in words. Drawing with some text comments are especially useful. A
person could make hundred simple drawings a day, which could present his childhood memories,
dreams, friends etc.

Moreover, I recommend to make EEG and video recording while drawing. Such complex data
collecting will provide many views on the same internal process.

Another thing you could do is the practice of active imagination with a partner and record
resulting data.

Day 8. Creating an idealized image of self

Not all data in your brain is actually you, and you may choose what you want
to preserve in better form. The best way to do it is to create idealized version
of yourself and the world where you would like to live in some form of fiction.
It may be a novel or a series of drawings.

12.3 Privacy notice

If you want to be resurrected with high fidelity, you should not keep secrets
from the future AI. But contemporary people could hold an advantage over
you if they know your stupid jokes, financial misconduct, love affairs, etc. So,
your data should be protected by strong passwords and encryption, but the
strength of the password should not be infinite, or future AI will not able to
decipher it.

Sensitive information may be additionally edited, so it can’t be used as legal

evidence of misconduct (think about replacing bad words with one letter
abbreviations).

12.4 Legal notice

Recording of other people is illegal without their consent in most jurisdictions.

To record a conversation, you should ask other people for their agreement,
which is typically easy to get from friends, especially if you promise never to
publish or use the information.

Some people wear constant video recording glasses for the goals of DI, similar
to Google Glass. This may annoy people but be of little use for DI, as it mostly
records other people, not you. Recording your voice is the most valuable thing
when in public.

At you home, you may record everything but may have to put out signs that
recording is taking place. You may also wear a t-shirt with sign about
recording, but its legal power is small.

13. Requirements for digital immortality software

If you have read all of this article, you may find yourself confused, as there a
lot of detailed instructions to be taken into account to collect data for DI
properly. If you are not a professional, who wants to devote your entire life to
DI, you may find it difficult to choose what you should do next. There are
several solutions:

1) Use a simplified uploading protocol with just several steps.

2) Seek the professional help of a person who is trained in DI (no such people
exist now)
3) Use a computer program that will guide you through the process and
automatically save data.
Such a computer program does not currently exist, but we can list
requirements for it. It should consist of several parts:

1) Recorder. This audio and video recorder starts automatically when it hears
the voice or see a person. This part is needed for passive DI. It also includes
a keylogger and automatically copies the browser-cache.

2) Tester. This part asks question and prompts the user to upload his personal
history. It may run psychological tests and schedule uploading sessions. It
must allow addition of new parts, like new tests.

3) Saver. It automatically sorts, compresses, and encrypts data and saves it

different clouds.

4) Search and speech recognition. This is an optional function, but it could

help to convert collected data into active external memory, similar to ideas of
Bell (Bell & Gray, 2000).

5) Personal assistant. This is also optional but is a second step toward

creation of an exocortex. The system is trained on the data, can predict
replies, work as extended working memory, find links during conversation,
etc.

Such software should probably have both PC and mobile versions.

14. Data preservation strategies

Most electronic data lives only several years. Hard drives die, DVDs decay,
and subscriptions to data storage services expire. But we want to be sure that
all our data will reach future AI, which could require safe storage for at least
decades, and maybe even centuries. Long-term data preservation could be
achieved if the data is sufficiently “copyable” (Eubanks, 2008).

Main principles:

1) Create several independent different copies of all data

2) Use different approaches to data preservation
3) Create smaller “data-bombs” for longer storage.
4) Eternal storage must be free of charge and not depend on the decisions or
fortunes of business.

Practical solutions:

1) Use Google
2) Use the Internet Archive
3) M-discs.
4) Space storage
You may need to invest in a powerful computer for photo and video
compression and for constant copying of data between HDD, massive RAID,
optical disks, and cloud services.

You also should be cautious about small startups which suggest they will
collect and store your data for you. Only internet giants or well-established
foundations have the necessary level of survivability to preserve data for
decades, and copying is everything.

Home hard drives archives

8-12 TB disks existed in 2018 for 200-300 USD partly for video surveillance
goals, and even 60 TB solid state disks were presented, but they are very
expensive. In addition, the typical failure rate of a consumer hard drive is 0.8
in case of extensive use because of wearing of mechanical parts, and
excluding wearing (which is small for archival drives) it 0.9 in the first year
and then 0.01 per year (Anthony et al., 2013). The most risk of home drives
comes from falling, as this often happens with movable external drives and it
could damage their internal mechanics (however, damaged drives should be
still preserved, as future AI surely will be able extract data from them). The
most reliable disks are Hitachi according to, which had only 1 per cent annual
failure rate compared to 9 per cent for Western Digital (Lilly, n.d.) Surely,
RAID massive inside a stationary computer is better than external HD.

The progress in hard-drives was slow in 2010s and was not able to cover
growing size of personal information. Copying 1 TB over USB 3.0 will take at
least 3 hours. Cheap external drives are also fragile, suffering damage or data
loss from even small falls. With a price of around 250 USD for 8 TB of internal
Hitachi drive, and the need of several reserve copies of all data as well as a
computer, the total price of storing a 10 TB archive could easily be around 
1000 USD. Most reliable cloud services are limited to 1 TB (as of 2018) or
require more expensive “business accounts” for larger capacities, and are
requiring yearly payment for maintaining. Another option, preserving data on
durable M-disks, requires manual chunking as the biggest such disks are 100
GB. In addition, tape archives are expensive for individual users. In view of
these difficulties, data copying is no longer easy, and the situation becomes
similar to the historical issues around preserving manuscripts.

More about home data storage technology:

https://lifehacker.com/5660551/your-best-solutions-for-massive-multi-
terabyte-storage

14.1 Data structure

All the data you will collect will basically consist of two main parts: the most
valuable information, consisting of handwritten text and audio, and the most
size-consuming information, which is mainly photos from the holidays and
GoPro video output. The first category will have a size of gigabytes, and the
second will be several terabytes. Much data in the second category will be
almost useless, but as we don’t know which part that is, we should preserve
it all. As I discussed in section 3.4, management of such large data sizes is
difficult.

Correct marking of the data structure will make its management much
simpler. First, data should be in folders which clearly name the type of data
and its year of creation. This will make updating and searching the archive
much simpler. The correct data structure will also help the user remember to
regularly copy all needed data types. However, double-copying of some data
is inevitable.

All data should not be in one compressed archive, as a single error could make
whole archive unreadable. However, even damaged HD could be repaired by
AI in the future, and should be preserved.

While video is the most informative source of data, is predictive power fades
quickly after the first few hours of personal recordings, so only a small part of
it should be in HD; later videos may be kept in lower resolution.

Example data structure:

(My documents, 2017)
(My biography and digital immortality texts, 2017)
(My evernote, 2017)
(My audiorecorder, 2017)
(My videos, 2017)
(My photos from phone, 2017)
(My photos from Canon, 2017)
(My Evocam videos, 2017)
(My keylogger data, 2017)
(My emails, 2017)
(My facebook, 2017)
(My browser’s cache, 2017)
(EEG, 2017)
(Medical data, 2017)
(Downloads, 2017)
(My paper archive scan, 2006)
(WhatsApp, SMS, and other chat archives)
(Geo-tracking data)

14.2 Cheap small “data-bombs” for eternal preservation

As we said above, data consists of two parts: a small valuable part and a very
large less-valuable part. You can currently preserve several gigabytes almost
indefinitely in the internet (assuming no global catastrophes). Google provides
up to 15 GB of space on Google Drive for free.

We call a a small piece of information which is easy to copy and which we

could secretly put somewhere, expecting that in the future it will preserved
for free almost eternally, a data-bomb. It is based on the principle that is
cheaper to copy massive amounts of information verbatim than to clean them.
There are many internet sites which allow free posting of user content, like
Scribd. The biggest amount of video information can be posted on YouTube—
but take care not to post private information and to lock the videos. Google
cache and the Internet Archive (Wayback Machine) will make copies of data
you post on the web automatically and for free.

14.2 Underground preservation

The main problem of long-term preservation is that we do not know when
superintelligent AI will appear. Maybe it will arrive in 10 years and will send
its robot to our home to collect our disks and archives. Maybe there will be
nuclear war, the Internet will collapse, and civilization will be built from scratch
200 years from now, with superintelligence appearing in 300 years. Maybe
aliens will find our archives 100 million years from now.

This means that we should try very different ways of data preservation
suitable for different future scenarios. One of such ways is to put data in the
secret place on “eternal disks”. The only “eternal disks” available for ordinary
users are M-disks (a type of Blu-ray disk), which has an expected survival
time of 1000 years (MDISC, 2017). They can be recoded on an ordinary Blu-
ray recorder and have sizes of up to 100 GB and a price of only a few USD
each. Thus, to preserve several terabytes of data one needs at least several
dozen disks.

These disks should be put in a box which hard, hermetic and rust-proof. A
pressure cooker is an ideal choice, but in the face of recent terrorist attacks,
buying multiple pressure cookers could be viewed with suspicion. Another
option is a hermetic plastic box inside an aluminum pan, with cement used to
fill the space between the inner and outer boxes. Professional stainless-steel
boxes for time-capsules exist but are more expensive and may attract more
attention from potential grave-diggers. This time-capsule provides oxygen
removal, stainless steel and costs 70 USD
https://www.futurepkg.com/personal-7-x-9-arnold-time-capsule-standard.

Don’t forget to add biological samples to the box, and perhaps some papers
as well. You could bury it on your own property, but it is better do it secretly,
as some people may think that the box contains valuables. You may use your
plot at the cemetery, but first check how long you will own it. You may ask
your friends to preserve it or put in the basement. Like most time-capsules,
it could be built into the wall or foundation of a house. You may register your
time-capsule with the International Time Capsule Society
(https://en.wikipedia.org/wiki/International_Time_Capsule_Society).

There are several services which offer “digital time capsules”, (e.g.
https://www.lifestimecapsule.com/), but it is not clear if such digital capsules
will still exist several decades from now, especially if you stop paying the
services. Non-traditional and non-digital carriers like microfilms may be a
better choice (Normand, Gschwind, & Fornaro, 2007), (McKee & Panov, 2011).
The Arch Mission is planning to protect personal data in space (Wolfram, 2018)
and it may accept personal archives in the next 10 years, so if your data is
included, the next civilization on Earth will able to resurrect you (Turchin &
Denkenberger, 2018).

Conclusion
Digital immortality, also known as indirect mind uploading, is an almost
untapped possibility for radical life extension. It is relatively cheap, and the
main barriers to its adoption are philosophical: we don’t know the nature of
personal identity, the power of the future AI, and the amount of data needed
to maintain informational identity. This creates logical uncertainty about the
possibility of DI, which thus should be regarded as only Plan C for achieving
immortality, where Plan A is life extension until AI creation, plan B is cryonics,
and plan D is quantum immortality. Combined, these plans provide a
reasonable chance for radical life extension for 21st century people, in the
absence of global catastrophe.

Acknowledgments: The ideas in this work appeared in discussions with my

father, Valery Turchin, Anna Soboleva, D. Suhoruchkin, Olia Levitskaia, Igor
Trapesnikov, Tim Yashin, and Masha Sharova.

References:

Alestig, I. (2018). New technology allows relatives to “chat with dead”- Ny

teknik erbjuder anhöriga att ”chatta med döda” - dagen.se. Retrieved

from http://www.dagen.se/nyheter/ny-teknik-erbjuder-anhoriga-att-

chatta-med-doda-1.1104260

Allen, N. (2009, December 22). Kim Peek: The “Real Rain Man” dies.

Retrieved from

http://www.telegraph.co.uk/news/worldnews/northamerica/usa/68676

70/Kim-Peek-The-Real-Rain-Man-dies.html

Allen, W. (1993). Illustrated Woody Allen Reader: Retrieved from

https://www.amazon.com/Illustrated-Woody-Allen-

Reader/dp/067942072X
Almond, P. (2006). Many-Worlds Assisted Mind Uploading: A Thought

Experiment. Retrieved from

https://web.archive.org/web/20110513092111/http://www.paul-

almond.com/ManyWorldsAssistedMindUploading.htm

Anthony, S. J., Epstein, J. H., Murray, K. A., Navarrete-Macias, I.,

Zambrana-Torrelio, C. M., Solovyov, A., … Lipkin, W. I. (2013). A

strategy to estimate unknown viral diversity in mammals. MBio, 4(5),

00598–13.

Anwar, Y. (2011, September 22). Scientists use brain imaging to reveal the

movies in our mind. Berkeley News. Retrieved from

http://news.berkeley.edu/2011/09/22/brain-movies/

Augustine, S. (398AD). Confessions. Translated by Henry Chadwick. Oxford:

Oxford University Press.

Bainbridge, W. S. (2004). Progress toward cyberimmortality. The Scientific

Conquest of Death, 107.

Bainbridge, W. S. (2006). Cyberimmortality: Science, religion, and the battle

to save our souls. The Futurist, 40(2), 25.

Bamford, S., & Danaher, J. (2017). Transfer of Personality to a Synthetic

Human ('Mind Uploading’) and the Social Construction of Identity.

Journal of Consciousness Studies, 24(11–12), 6–30.

Bell, G., & Gray, J. (2000). Digital immortality.

Berdyaev, N. (1950). Self-Knowledge: An Essay in Autobiography. Retrieved

from https://www.amazon.com/Self-Knowledge-Essay-Autobiography-

Nicolas-Berdyaev/dp/1597312584
Berman, J. (2011, August 10). Bringing Back His Dead Dad Via Computer.

ABC News. Retrieved from

http://abcnews.go.com/Technology/futurist-ray-kurzweil-bring-dead-

father-back-life/story?id=14267712

Best, B. P. (2008). Scientific justification of cryonics practice. Rejuvenation

Research, 11(2), 493–503.

Bostrom, N. (2003). Are You Living In a Computer Simulation? Published in

Philosophical Quarterly (2003) Vol. 53, No. 211, Pp. 243-255.

Brunetti, S., Del Lungo, A., & Gerard, Y. (2001). On the computational

complexity of reconstructing three-dimensional lattice sets from their

two-dimensional X-rays. Linear Algebra and Its Applications, 339(1),

59–73. https://doi.org/10.1016/S0024-3795(01)00437-2

Chen, A. (2017, April 21). Elon Musk’s dreams of merging AI and brains are

likely to remain just that — for at least a decade. Retrieved February

14, 2018, from

https://www.theverge.com/2017/4/21/15370376/elon-musk-

neuralink-brain-computer-ai-implant-neuroscience

ciphergoth. (2011). An EPub of Eliezer’s blog posts - Less Wrong. Retrieved

February 19, 2018, from

http://lesswrong.com/lw/72m/an_epub_of_eliezers_blog_posts/

Crevier, D. (1993). AI: the tumultuous history of the search for artificial

intelligence. Basic Books.

Egan, G. (2011). Zendegi. Night Shade Books.

Eubanks, D. A. (2008). Survival Strategies. ArXiv:0812.0644 [q-Bio].

Retrieved from http://arxiv.org/abs/0812.0644

Fedorov, N. (1903). Философия общего дела. Moscow.

Gemmell, J., Bell, G., & Lueder, R. (2006). MyLifeBits: a personal database

for everything. Communications of the ACM, 49(1), 88–95.

Goertzel, B. (2007). The Role of AGI in Cybernetic Immortality,. The Journal

of Geoethical Nanotechnology. Retrieved from

http://www.terasemjournal.org/GNJournal/GN0201/ben_a.html

Horace. (23AD). Exegi monumentum: Horace, Odes 3.30. Retrieved from

https://www.gutenberg.org/ebooks/author/1790

Horney, K. (2013). Self-analysis (Vol. 18). Routledge.

Jonas, J. B., Schmidt, A. M., Müller-Bergh, J. A., Schlötzer-Schrehardt, U.

M., & Naumann, G. O. (1992). Human optic nerve fiber count and

optic disc size. Investigative Ophthalmology & Visual Science, 33(6),

2012–2018.

Jones, J. (2017). Technological Resurrection: A Thought Experiment.

Jonston, I. (2017, April 17). Scientists invent mind-reading machine that

turns your thoughts into words. The Independent. Retrieved from

http://www.independent.co.uk/news/science/read-your-mind-brain-

waves-thoughts-locked-in-syndrome-toyohashi-japan-a7687471.html

Jung, C. G. (1925). Jung on active imagination. Princeton University Press,

1997. Retrieved from https://www.amazon.com/Jung-Active-

Imagination-C-G/dp/0691015767

Karpathy, A. (2015). The unreasonable effectiveness of recurrent neural

networks. Retrieved from http://karpathy.github.io/2015/05/21/rnn-

effectiveness/
Keshavan, M. S., & Sudarshan, M. (2017). Deep dreaming, aberrant salience

and psychosis: Connecting the dots by artificial neural networks.

Schizophrenia Research, 188, 178–181.

Kim, J.-H., Biessmann, F., & Lee, S.-W. (2014). Reconstruction of hand

movements from EEG signals based on non-linear regression. In Brain-

Computer Interface (BCI), 2014 International Winter Workshop on

(pp. 1–3). IEEE.

Kosoy, V. (2017). Forecasting using incomplete models. ArXiv:1705.04630

[Cs]. Retrieved from http://arxiv.org/abs/1705.04630

Landauer, T. K. (1986). How much Do People Remember? Some Estimates

of the Quantity of Learned Information in Long-term Memory.

Cognitive Science, 10(4), 477–493.

Learn.Genetics. (2018). Why Clone? Retrieved February 20, 2018, from

http://learn.genetics.utah.edu/content/cloning/whyclone/

Lilly, P. (n.d.). Ongoing hard drive test shows HGST is the most reliable

brand. Retrieved March 19, 2018, from

https://www.pcgamer.com/ongoing-hard-drive-test-shows-hgst-is-

the-most-reliable-brand/

LiveJournal. (2018). One my day. Retrieved from https://odin-moy-

den.livejournal.com/

Maeterlinck, M. (1907). Intelligence of Flowers, The. SUNY Press.

Martin, D. (2007). Robert Shields, Wordy Diarist, Dies at 89 - The New York

Times. NYTimes. Retrieved from

http://www.nytimes.com/2007/10/29/us/29shields.html?_r=1&ref=us

&oref=slogin
McKee, S., & Panov, V. (2011). Archiving Color Images to Single Strip Black-

and-White 35mm Film—The Visionary Archive Process. SMPTE Motion

Imaging Journal, 120(1), 24–28.

MDISC. (2017). MDISC Archive Service. Retrieved August 20, 2017, from

http://www.mdisc.com/

Ming, L., & Vitányi, P. M. (1990). Kolmogorov complexity and its

applications. In Algorithms and Complexity (pp. 187–254). Elsevier.

Murre, J. M. J., & Dros, J. (2015). Replication and Analysis of Ebbinghaus’

Forgetting Curve. PLOS ONE, 10(7), e0120644.

https://doi.org/10.1371/journal.pone.0120644

Nemrodov, D., Niemeier, M., Patel, A., & Nestor, A. (2018). The Neural

Dynamics of Facial Identity Processing: insights from EEG-Based

Pattern Analysis and Image Reconstruction. ENeuro, ENEURO.0358-

17.2018. https://doi.org/10.1523/ENEURO.0358-17.2018

Nichols, L. (1969). Isaac Asimov: Man of 7,560,000 Words. NYTimes.

Retrieved from

http://www.nytimes.com/books/97/03/23/lifetimes/asi-v-profile.html

NIH. (2018, January 16). Memory gene goes viral. National Institutes of

Health (NIH). Retrieved from https://www.nih.gov/news-events/news-

releases/memory-gene-goes-viral

Normand, C., Gschwind, R., & Fornaro, P. (2007). Digital images for

eternity: color microfilm as archival medium. In Color Imaging XII:

Processing, Hardcopy, and Applications (Vol. 6493, p. 649307).

International Society for Optics and Photonics.

Ochab-Marcinek, A., & Tabaka, M. (2010). Bimodal gene expression in

noncooperative regulatory systems. Proceedings of the National

Academy of Sciences, 107(51), 22096–22101.

Parfit, D. (1984). Reasons and persons. OUP Oxford.

Pfister, W. (2014). Transcendence (2014) - IMDb. Retrieved from

http://www.imdb.com/title/tt2209764/

Praderio, C. (2017). It’s possible for one person to have two different sets of

DNA — here’s how it happens. INSIDER. Retrieved from

http://www.thisisinsider.com/what-is-a-human-chimera-and-how-

does-it-happen-2017-11

Preneel, B. (1994). Cryptographic hash functions. Transactions on Emerging

Telecommunications Technologies, 5(4), 431–448.

Roberts-Grey, G. (2015). What her eating habits say about her bedroom

habits. MSN. Retrieved from http://a.msn.com/05/en-us/AAcU5b

Robertson, A. (2013). Scientists turn dreams into eerie short films with an

MRI scan - The Verge. Retrieved from

https://www.theverge.com/2013/4/4/4184728/scientists-decode-

dreams-with-mri-scan

Robinson, R. (2014, January 6). How big is the human genome? Retrieved

February 20, 2018, from https://medium.com/precision-

medicine/how-big-is-the-human-genome-e90caa3409b0

Rothblatt, M. (2012). The Terasem mind uploading experiment. International

Journal of Machine Consciousness, 4(01), 141–158.

Rousseau, J.-J. (1782). Les confessions.

Schilling, D. (2016, April 20). Can you become “virtually immortal”? A

Silicon Valley start-up thinks so. Retrieved February 13, 2018, from

http://www.theguardian.com/commentisfree/2016/apr/20/eternime-

immortal-chatbots-silicon-valley-startup-can-you-cheat-death

Shen, G., Horikawa, T., Majima, K., & Kamitani, Y. (2017). Deep image

reconstruction from human brain activity. BioRxiv, 240317.

SpiderOak. (2016, May 23). Facebook Shadow Profiles: A Profile of You That

You Never Created. Retrieved March 19, 2018, from

https://spideroak.com/articles/facebook-shadow-profiles-a-profile-of-

you-that-you-never-created/

Stone, H., & Stone, S. (2011). Embracing Our Selves: The Voice Dialogue

Manuel. New World Library.

Swan, L. S., & Howard, J. (2012). Digital immortality: Self or 0010110?

International Journal of Machine Consciousness, 4(01), 245–256.

Templeton, G. (2017). Elon Musk’s NeuraLink Is Not a Neural Lace

Company. Retrieved February 14, 2018, from

https://www.inverse.com/article/30600-elon-musk-neuralink-neural-

lace-neural-dust-electrode

Terasem. (2014). Terasem Movement Foundation: Welcome to the TMF, Inc.

Retrieved February 18, 2018, from

http://www.terasemmovementfoundation.com/

Thetimsonline. (2010, May 24). Corín Tellado: Spanish romantic writer |

Times Online Obituary. Retrieved from

https://web.archive.org/web/20100524074821/http://www.timesonlin

e.co.uk/tol/comment/obituaries/article6114908.ece
Turchin, A. (2007). The textbook on active imagination (in Russian).

Samizdat. Retrieved from

https://www.scribd.com/doc/9831210/%D1%83%D1%87%D0%B5%

D0%B1%D0%BD%D0%B8%D0%BA-%D0%BF%D0%BE-

%D0%B0%D0%BA%D1%82%D0%B8%D0%B2%D0%BD%D0%BE%

D0%BC%D1%83-

%D0%B2%D0%BE%D0%BE%D0%B1%D1%80%D0%B0%D0%B6%D

0%B5%D0%BD%D0%B8%D1%8E

Turchin, A. (2015). Immortality Roadmap - Less Wrong. Retrieved February

19, 2018, from http://lesswrong.com/lw/mjl/immortality_roadmap/

Turchin, A. (2016). Identity map - Less Wrong. Retrieved February 16,

2018, from http://lesswrong.com/lw/nuc/identity_map/

Turchin, A. (2017a). Digital immortality and art. Retrieved from

https://www.slideshare.net/avturchin/ss-80472635

Turchin, A. (2017b). Messaging future AI. Manuscript. Retrieved from

https://goo.gl/YArqki

Turchin, A. (2018). Forever and Again: Necessary Conditions for the

“Quantum Immortality” and its Practical Implications.

Turchin, A., & Denkenberger, D. (2018). Surviving global risks through the

preservation of humanity’s data on the Moon. Acta Astronautica,

(Volume 146, May 2018,), Pages 161-170.

Turchin, A., Denkenberger, D., Milova, E., Egorova, A., & Batin, M. (2017).

Fighting aging as an effective altruism case: the model of impact.

Vlahos, J. (2017). A Son’s Race to Give His Dying Father Artificial

Immortality. Retrieved February 13, 2018, from

 https://www.wired.com/story/a-sons-race-to-give-his-dying-father-

artificial-immortality/

Volpicelli, G. (2016). This Transhumanist Records Everything Around Him So

His Mind Will Live Forever. Vice.Motherboard. Retrieved from

https://motherboard.vice.com/en_us/article/4xangw/this-

transhumanist-records-everything-around-him-so-his-mind-will-live-

forever

Wiley, K. (2014). A Taxonomy and Metaphysics of Mind-Uploading (1

edition). Humanity+ Press and Alautun Press.

Wolfram, S. (2018). How to Design Beacons for Humanity’s Afterlife. WIRED.

Retrieved from https://www.wired.com/story/how-to-design-beacons-

for-humanitys-afterlife/

Yampolskiy, R. (2014). The Universe of Minds. arXiv preprint

arXiv:1410.0369.

Yudkowsky, E. (2016). The Multiple Stage Fallacy - Effective Altruism Forum.

Retrieved February 19, 2018, from http://effective-

altruism.com/ea/uv/the_multiple_stage_fallacy/