from purely storage prespective:
right now you can buy 1Tb microSD card (1^12 bytes), which is about 7800x more than first 128Mb microSD card introduced in 2005, so x^15 = 7800 => x = 1.8 i,e 80% improvement a year (I know it's super vogue approximation but helps to get the order of magnitude). Extrapolate that for 17k Petabytes, and we get 1.8^y = 17000 => y = 17, so less than a century for sure.
anothing thing, that likely for AI the topology of the brain wiring is more important than raw organic data
I don't know if storage is the right metric, perhaps something closer to RAM is, or particularly VRAM. If that's the case then we have a bit farther than 17 years to go.
Compared to modern digital electronics, the human brain is "very wide but very slow". The equivalent clockspeed is just tens of hertz, maybe a few hundred for some areas. High fidelity simulations of the signal waveforms might require only a few thousand samples per second! Contrast that to a typical modern computer, which has clocks a million times faster than that.
Similarly, simulating a brain may not need high precision at all, and is likely amenable to various forms of compression. For example: a simulator could have a library of the behaviour of a million synapses, and simply interpolate between them to closely match each synapse in a brain scan. Hilariously, for extra efficiency, this kind of interpolation in a high dimensional space is something artificial neural networks are quite good at!
I envision a path where ordinary "machine learning" is used to automatically model small sections of the brain, such a ganglia, axons, synapses, etc... Similarly, ML techniques can be used to match scans to the previously learned library of these models. The final thing might just be a hundred 8-bit "parameters" per synapse, and there's 125 trillion of those, so about 12.5 petabytes. That's a lot of data, but you can buy a server right now with 12 TB of memory! Assuming a 30% increase in capacity yearly (thanks Mr Moore), the goal is just 26 years away.
Make an entry in your calendar: 2047 is the year we'll have human-equivalent AI...
Yes but not all parts have to be active simultaneously. Like our long term memories are stored on a long term type of media in the brain so could an artificial life agent store those on a different type of media, maybe it could be heavily compressed or self erasing over time
no, you're misreading this. First, you can't extrapolate that out, the sample was of cortical tissue which while expansive is just a thin layer. Second, thats the size of the imaging data, it doesn't mean anything about human compute in the slightest.
Hence why I said "raw organic data." I didn't say memories or exactly the space of an actual brain. Plus, it's just a simple fun way to think of something in terms we can understand. I as well as everybody else is fully aware it's not that simple.
let me put it another way; you could apply your analysis to detailed imaging of an apple slice and walk away with the same conclusion. it is fun but oversimplified to the point that it doesn't make any sense even if you squint. the size of the data is proportional to the quality of the image here, not to anything in the brain.
and honestly I don't think you or anyone else is fully aware it's not that simple. brains are.. just so complicated.
While this is a lot of data if you extrapolate to a full brain, a lot of signs point towards the data you'd need to actually simulate a brain being much smaller than that — possibly between 20-100 PB.
I agree, a rudimentary one at least should be possible with less. A human brain is not active all at once either and large parts of the brain are dedicated with decoding or other limiting factors of human hardware
Human genomes are typically stored as a diff from a reference genome (100GB -> 15MB), so you could do the same here.
Although if you can disentangle memories from the rest of the brain recording, you'd probably want to try running your own memory on Einstein's brain and see what that's like rather than actually use yours. It wouldn't be you, but a scan of your brain is not you either.
14 comments
[ 3.2 ms ] story [ 44.9 ms ] threadSomething tells me were still a few centuries off from "true AI."
Similarly, simulating a brain may not need high precision at all, and is likely amenable to various forms of compression. For example: a simulator could have a library of the behaviour of a million synapses, and simply interpolate between them to closely match each synapse in a brain scan. Hilariously, for extra efficiency, this kind of interpolation in a high dimensional space is something artificial neural networks are quite good at!
I envision a path where ordinary "machine learning" is used to automatically model small sections of the brain, such a ganglia, axons, synapses, etc... Similarly, ML techniques can be used to match scans to the previously learned library of these models. The final thing might just be a hundred 8-bit "parameters" per synapse, and there's 125 trillion of those, so about 12.5 petabytes. That's a lot of data, but you can buy a server right now with 12 TB of memory! Assuming a 30% increase in capacity yearly (thanks Mr Moore), the goal is just 26 years away.
Make an entry in your calendar: 2047 is the year we'll have human-equivalent AI...
and honestly I don't think you or anyone else is fully aware it's not that simple. brains are.. just so complicated.
Although if you can disentangle memories from the rest of the brain recording, you'd probably want to try running your own memory on Einstein's brain and see what that's like rather than actually use yours. It wouldn't be you, but a scan of your brain is not you either.