Warning: Not very coherent! Just published after thinking about the title question for an hour.
How do you get a digital version of your identity? This probably means uploading your brain somehow. What is your brain? A bunch of cells. Where's the information that your brain stores? Right, that's mysterious.
Inputs and Outputs
Okay, suppose we upload the brain. What do we need to do after that? We need to connect it to some inputs and get its outputs.
- Do we know how input/output channels work? Are inputs different channels from outputs, or can they overlap somehow? What are all the points where brain touches non-brain stuff?
- For a good existence, this would require simulating a shitton of stuff. Tough.
- This isn't the crux yet though. It would be good to upload humans and store them so they don't die, even if we can't simulate well yet.
Figuring Out the Connectome Is Obviously Not Enough
Don't get distracted by research that maps how all neurons connect to each other. We don't care about just simulating an arbitrary human. We want to help a particular human not die.
What Does “Storing the Brain Digitally” Look Like?
We want something like "converting neurons/synapses to weights in an artificial neural network".
How Simple Is Each Neuron?
If we know that on some fundamental level, neurons respond to inputs in the same way always, then we can perturb the system and figure out what each neuron does.
It turns out they're not simple. Dale's principle was wrong. Co-transmission is actually true, neurons can release multiple neurotransmitters. Different input signals might change which neurotransmitters are released.
ChatGPT says: "However, it's important to note that the neurotransmitters a neuron releases are only one part of how neurons process and transmit information. Other important factors include the neuron's electrical properties, the receptors on the postsynaptic neuron, the presence of modulatory signals, and many others."
So just trying to somehow measure what each neuron does in a simple way is fucked.
Okay, but There’s Nothing Mysterious About a Neuron. It’s a Deterministic System, Governed by Physics of What’s in Its Environment and What’s Floating Around Its Cytoplasm.
Yeah, the problem is definitely solvable in theory. The above logic tells us that perturbations probably won't work well enough. We could alternatively just fucking figure it all out. But we probably can't be confident in coming up human-understandable principles; we want to just simulate it all after copying state exactly.
At that level, we don't have to worry about simulating things yet. Just think about: how do we duplicate a cell, exactly?
Actually, Maybe Perturbations Aren’t So Bad. We Probably Don’t Care About Really Unlikely Events in a Single Neuron, Right? Or Really Low-Magnitude Changes in Different Responses.
Yeah, something like Taylor series for a neuron... Check this Wikipedia page out. Don't these models all break down if there are different neurotransmitter released that interface differently with different connected neurons?
But we'd want to be pretty sure that we're accurate. There are probably important edge cases that really change behaviour. Some of these things might be really hard to truly incorporate because of how many input surface there are for a large, complex chemical system like a cell. But a lot of the desired behavior is encoded in the genome, right? Ah, maybe not. Just structures that have good properties will be encoded. The genome won't tell us what's so great about what these structures can do.
Here's an example for how we might actually measure perturbations. But yeah, the tools we use to measure should be ones we're confident are sufficient. It could be that other things the tools can't measure (eg I think this paper only measures calcium release) are really important.
Are We Sure That Thinking on the Level of Neurons Is Correct?
This might be the wrong way to look at it. What's another way to look at it?
- We could zoom out or zoom in. Zooming in probably won't be a big enough shift in perspective.
- Zooming out, is there some way we can capture what is most essential about a person? Eh, I don't care about what's most essential. I want to stay alive exactly as I am! What I want is to preserve all the information-processing parts of me, and we think this is all happening in the brain which is just a bunch of neurons in some connection with some weights.
Subproblem: How Do Things Learn? That’s a Core Part of Life, Updating Your Weights Somehow.
How do we read neuron "weights"? What do we know about how C. elegans neurons even work?
Reading isn’t enough. We want to know how it learns, so simulated worm can learn.
- Well, the cleaner solution would be if the state of brain that we read
itself captured how neurons/synapses should be updated. And it should know
that right?
- Indeed, if we don't solve it like this, we can't be sure we've even uploaded properly!
Uploading a Nematode Would Probably Teach Us a Lot
- should be much easier to accurately simulate its brain's inputs and capture outputs