[ExI] The quick and easy path to uploading

Rafal Smigrodzki rafal.smigrodzki at gmail.com
Fri Dec 17 00:16:34 UTC 2021

The identity thread has been a perennial recurrence on Exi for decades and
it doesn't seem to make much progress. Recently, there were some
developments in neuroscience and deep learning that might shed light on the
subject of personal identity.

Obviously, our identity is not determined at the level of individual atoms
and molecules, and any discussion at that level is just barking up the
wrong tree. But, which is the lowest level of organization of the brain
that encodes a person is still a valid question. For a long time I thought
that the sum total of individual synapses, their strengths and the pattern
of connections between neurons was the level relevant to encoding
personhood. About 15 years ago or so, I changed my mind, and felt that the
details of synaptic connections are not really important, as long as the
neural network as a whole processes information in an equivalent fashion,
as measured by inputs and outputs. I wrote here on the list that it would
be OK for my upload to use a generic visual cortex rather than an exact
copy of my own visual cortex scanned from my brain. I even ventured that I
might be able to squeeze my person into a terabyte or so, with generic
neural networks doing 99.99% of creating my subjective experience, and only
0.01% of the upload structure being determined by information from my brain

This attitude to self has been recently bolstered by two developments:

The first new finding is that the pattern of synaptic connections in a
human brain is more dynamic than previously believed. Every time you access
a personal memory, the network creates new synapses and eliminates old
ones, yet subjectively we have a feeling of remembering the same events. My
earliest memories are from the time I was about 3 years old, and I have
revisited them countless times. The patterns of synapses I use to encode
those memories have probably changed multiple times but the subjective
experience feels stable.

The second development comes from deep learning. It turns out it is
possible to construct functionally equivalent copies of neural networks by
training a new network on a sample of inputs and outputs of the original
network. The copy will have a completely different pattern of connections
but it will almost always produce the same outputs from the same inputs, if
trained appropriately.

>From these two developments I conclude that it should be much easier than
previously expected to upload a person using a brain machine interface
(BMI), without the need to scan the microstructure of the brain.

To upload you first need to create a high-level image of the long-range
cortical and subcortical connections, perhaps using a high-resolution DTI
scan. Then you need to use a Neuralink device to insert a bunch (million?
hundred million?) electrodes into the white matter of the brain, probably
also into the basal ganglia and record the neural traffic going in and out
of every cortical column and basal ganglia neural cluster. No need to
directly read the synapses, which are changing all the time anyway. Use the
neural traffic to train millions of deep learning modules arranged in a
larger structure based on the DTI image. Keep this running for some time
(weeks? months? years?) until the deep learning does its job and is capable
of running a functional copy of the mind in shadow mode. As the biological
network starts shutting down due to aging, the digital copy keeps taking
over more and more of the processing and gradually replaces the original,
potentially keeping a puppet body going, and, of course, branching out into
virtual realities.

I thought that uploading was a far prospect, requiring an insane number of
doublings in computing power and an insanely laborious process of
post-mortem brain scanning. Now I think that uploading may be a realistic
option after only a few orders of magnitude of growth in BMI bandwidth. I
might have as much as 20 - 30 years of function left in my current
substrate, and in this time, Saint Elon willing, uploading could become
commercially available. Of course I would be an early adopter.

Cryonics is still an important part of the reasonable person's preparations
for the future but for people in their fifties it hopefully won't be ever

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20211216/e16f3914/attachment.htm>

More information about the extropy-chat mailing list