[djohnston]

one feature i dont understand about neuralink is the promised transmission of thoughts.

lets say i have a device that can encode activity of every neuron in my brain simultaneously. i think some variation of the thought "elephant," and the neuralink creates some digital representation of these signals.

then i send that digital encoding to your brain's neuralink.

how on earth are you going to think of ANY elephant, let alone my elephant? the structural nature of our brains is certainly different, the network of activated neurons aren't isomorphic, certainly not identical, right?

i am by no means an expert, but i dont see how you can ever take a digital signal and turn it into a conscious meat thought.

motor signals seem somewhat more tractable

[sudhirj]

There will likely need to be translation layer. Motor signals aren't necessarily tractable either, people can learn to walk again after even if the parts of the brain that govern walking are damaged.

The promise seems more that ideas can be transmitted with higher bandwidth. My signals -> word1, send word1 to you, word1-> your signals for word1 would be likely. Both of us would have to the system for word1, but if we could exchange it much faster than speaking, maybe even reading.

My issue with the system is that the line between reading and writing signals is very thin, and this these are hackable devices. Once the device learns a person's brain signals to make the legs run, it's easy to replay them on a bridge when facing sideways. There's no technology right now that's foolproof, and I'm not sure how any company can boast that it's secure enough — even if encryption is strong laws can still be enacted to force companies to hand over keys or install backdoors.

[visarga]

> how on earth are you going to think of ANY elephant, let alone my elephant?

Both brains sensors are calibrated on the same images, then a common representation can be found as intermediary. It has been done with neural nets translating into a common intermediary language, thus needing N encoders/decoders instead of (N-1)^2 separate networks. In the easiest case it can be a simple linear transform to map one vector space into another.

[djohnston]

wouldnt you need a distinct transformation for every thought?

[visarga]

the transformation of brain signals to the common space can be a neural net - and since neural nets are non-linear they can adapt to the inputs

you'd need to train a bidirectional adapter for every person instead of (N-1)^2 adapters for each pair of people

[djohnston]

got it, thanks for the clarification

[davidivadavid]

There's so much handwaving about this part, even at the functional level (how is it supposed to work at the most basic level for the user? what's it supposed to look like?), that I kinda tune out every time that comes up.

I'll start listening when they have a working definition of what a "thought" is or what "translating" it into a "universal language" even means.

[tudorw]

https://www.sciencealert.com/if-you-were-a-pokemon-whizz-as-...