Why You Won't Be Uploaded
If smarter-than-human AI is developed, some feel this isn't a problem because we will all be able to be "uploaded" by then. Economics suggests that even if the formidable technical challenges to uploading are ever resolved it will still only be a tiny fraction of the population, if any, that gets uploaded.
Cost of uploading
Putting aside the technical challenges for one moment. Let's look at the costs today.
Cost to store a brain representation:
86x109 neurons x say 103 distinct connections per neuron (not synapses) x 7 bytes (5 byte neuron id; 1 byte connection strength; 1 byte propagation time delay) = 6x1014 bytes
- Amazon S3 storage price (2018): $0.0125/GB-month infrequent access storage; $0.004/GB-month Glacier tape storage
- $90k/year to store or $30k/year for tape (today)
Cost to run a brain emulation:
IBM's TrueNorth project had plans to build spiking neuron hardware with 4 times as many neurons as the human brain
They didn't reveal the costs, but a reasonable estimate is, if produced in volume, around $80/simulated brain-hour (2016) (see HardwareOverhang)
- $700k/simulated-year (today)
Cost to scan a brain for uploading:
Brain volume = 1.2x10-3m3
Scan resolution: 5nm x 5nm x 25nm (sufficient to resolve synapses) = 1.9x1021 voxels
Current neural scanning technology uses scanning electron microscopes: 1x106 voxels/s (Kasthuri, 2015, Saturated Reconstruction of a Volume of Neocortex)
6x107 years of microscopy
- Cost of electron microscope: $2k/year (Ebay quotes used electron microscopes at $5k-20k; amortize over 5 years)
>$120b to scan (today - ignores electron microscope operating costs, down time, labor, and image segmentation costs)
Separate estimate: <$34t to scan (today - includes other unrelated costs; Lichtman lab granted $28m in 2016 to scan 1mm3)
- take $1t (today) to scan as a reasonable mid-point estimate
Only one brain might ever be uploaded
Costs of scanning brains, running emulations, and storing emulations will fall as technology progresses, but the very large difference between the cost to scan and to store (currently over a factor of 10 million) is unlikely to be change. The cost of scanning might fall rapidly, but it is unlikely to ever obtain parity. The difference is driven by the difference in the resolution required to scan a brain (the number of voxels) and the resolution required to store a brain representation (the number of bytes). These two numbers differ by over a factor of a million.
Given the expected cost differential, if it was possible to upload, once the first brain was uploaded it would be much, much, cheaper to make copies of the first brain and modify them, than it would be to upload additional brains. By the time uploading was affordable for the masses the proliferation of emulations would have made the world an unrecognisable place for humans (in The Age of Em, Hanson estimates any era of emulations will only last for 2-5 years before being replaced by something even more bizarre; and during any such era it seems likely that manipulation of the physical world (scanning) will evolve at a slower rate than manipulation of the virtual world (running emulations), with the cost of storing emulations falling somewhere in between these two rates).
Or none at all
It seems that construction of an emulation based on scanning small pieces of different types of brain tissue to discover general wiring organizational principles is going to have a long head start over scanning a complete brain. Scanning, say, 100 different 1mm3 volumes would reduce the price by a factor of 12,000 to $80m today. This is quite affordable for a government sponsored research project today. By the time uploading was ever affordable to the general public, the effects of such emulations would have transformed the world irreversibly.
Other approaches to scanning
The reason for the large difference between storing and scanning is scanning today needs to construct a structural connectome, which is much larger than the functional connectome needing to be stored. One promising technique for deriving the functional connectome is scanning the brain using DNA or RNA barcoding. Such techniques are close to becoming practical (see Using high-throughput barcode sequencing to efficiently map connectomes), but unfortunately while they might one day deliver a connectome, they don't include the weight associated with each connection. Knowing the weights between neurons is vital to uploading. Other scanning techniques, such as nanobot probes, might be able to directly deliver a functional connectome that incorporates weights, but for the foreseeable future such techniques remain the work of science fiction.