[ExI] Mind Uploading: is it still me?

[Ben Zaiboc]

On 26/12/2025 12:59, John Clark wrote:
> In 2024 researchers at Harvard and Google sliced one cubic millimeter 
> of a human brain into 5019 slices and then used an electron microscope 
> to take photographs of each slice, they then used the resulting 1.4 
> petabytes (1,400 trillion bytes) of data to construct a 3-D model of 
> that cubic millimeter. They got a X-Y resolution of 4 nanometers (a 
> DNA molecule is 2.5 nanometers thick) but the Z resolution(thickness) 
> was only 30 nanometers. That probably would be good enough 
> resolutionfor an upload but unfortunately it was just for one cubic 
> millimeter, the average human brain contains about 1,400,000 cubic 
> millimeters.

On the surface, this sounds quite discouraging, but I'd just like to 
make a brain-dump on the topic, having been thinking about it for a while:


The x-y resolution mentioned was overkill, by at least 10 times, and the 
z resolution less so, but still probably higher than necessary. Let's 
say it was just about right, though. That means approx. 14 trillion 
bytes for 1 cubic mm.

Times that by 1.4M for the whole brain (almost certainly not actually 
needed (for our purposes), for several reasons, and as discussed 
previously, a generic human brain model could probably cut down on that 
considerably, with individual variations laid on top of it), so we get 
14 x 10^12 times 1.4 x 10^6 = 19.6x10^18 bytes (? please check this, I'm 
on dodgy ground here, given my mathematical challenges). Say around 20 
exabytes.

That's a lot, but I reckon it can be reduced a fair bit (a lot, 
actually), considering that most of the brain is white matter (axons), 
not grey matter (cortex, where the vast majority of the synaptic 
information needs to come from), and all that's needed for the white 
matter is to record each axon's origin and destination (including 
branchings), and axons tend to be about 300nm in diameter, so the 
resolution doesn't have to be so high as for the cortex. Basically, 
what's needed for the bulk of the brain is a wiring diagram (a whopping 
big, complex one, but still nothing more than that. Probably).

Plus the fact that the original resolution of the scan wouldn't be the 
same as the resolution of the data derived from it, and therefore the 
storage needed. Lossless compression algorithms as well as dynamic 
intelligent feature extraction would reduce the amount of scanned data 
on the fly, so you might be scanning 14 trillion bytes per cubic mm but 
condensing that into a few megabytes or less, without losing anything 
relevant to reconstructing the neural structure and function. There'd be 
no point recording every 300nm voxel of white matter to produce a 
connectome map because that would be like creating a raster diagram when 
a vector diagram is far more efficient. I think there are parallels here 
with the way our brains process vision, extracting and compactly 
representing specific features of our visual fields, and ignoring the rest.
Yes, the scanning would have to be very detailed, but as soon as a 
structure can be detected, and linked to prevous structure, that 
detailed data can be discarded.

When you think that we'd be effectively scanning trillions of organic 
molecules, and converting their positions into a datastream, you can see 
there's plenty of scope for compression, just for starters.

Unknowns include whether variations in the myelin sheaths are important, 
and whether the support cells (glial cells) need to be recorded as well 
or not. My money's on Not.

It would probably be very useful to devise a scheme where different 
aspects of the brain scan are condensed into separate maps, that can 
each be stored in a relatively compact way, to later be integrated and 
expanded into a functioning structure when it comes time to actually 
build the upload.

So overall, I expect that the 20 exabytes estimate to capture all the 
relevant data from an individual brain is vastly greater than what will 
eventually be really needed, once we figure out what needs to be 
captured, how to encode and store it, and what can be recorded as 
variations on a 'canonical human brain' model (which could potentially 
make /enormous/ savings on the amount of data needed to encode an 
individual brain (if that turns out to be a feasible idea!)).

Another factor will be the time needed to scan an entire brain. And 
there's also the problem of the scanning method dumping heat into the 
tissue surrounding the area being scanned, potentially messing up the 
structure and chemical environment.

We probably need entirely different scanning methods to what we 
currently have, or it wouldn't be remotely practical. I've no idea how 
to approach that problem.

All this makes me more convinced than ever that cryonic suspension alone 
won't be enough to prepare a brain for the kind of scanning that will be 
needed for uploading. Aldehyde stabilisation, or something equivalent, 
might just be the start.

Now I'm wondering if and how a brain could be dismantled - as in 
separated into distinct sections - without damaging anything essential, 
either before or after preservation. Then each section could be 
separately scanned, in parallel. Hmmm...

If axons were elastic, I wonder if a 'reverse convolution' could be done 
on the cortex. Cortices, rather. I expect cerebellum would need to be 
uploaded as well, which presents its own problems (smaller neurons) and 
opportunities (much more regular structure than elsewhere in the brain).

Or, if the axons could somehow have markers set in their connections to 
the cortical layers, then be removed for separate scanning, then the 
cortices spread out into what, four or five big thin sheets ??

Any other ideas?

-- 
Ben

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20251227/39693db6/attachment.htm>