The Neuralink "Lossless" Compression Wars

Oh, this gets even worse when you consider the actual neuroscience. I worked in non-human primate primary motor cortex in grad school, and one of the fun things we discovered is that the line between "signal" and "noise" might not actually exist. Depending on how aggressively we filtered "noise" out from "signal", we found that our resulting data encoded different things about the monkey's behavior. The "noise" we're worrying about removing might actually be encoding important information - the "noise" might be more of a "context" or "background" for the "signal".

It could be argued that analog-to-digital conversion is not inherently lossy because it's occurring with a band-limited signal. Nyquist and Shannon found out in the 1930s that such a signal will be reconstructed perfectly.

The data is not audio, it’s 1024 electromagnetic signal channels combined together in an order that looks like noisy audio if you incorrectly look at the data as audio. The data is transmitted in an audio format for bitpacking convenience. Just like the spoon, there is no noise

I feel like there is a lot of confusion around a very simple engineering requirement here. Neuralink need to get data out losslessly just means data in == data out. The data isn’t actually sound and thinking in terms of noise and signal is making a judgement call around brainwaves which we don’t understand. I imagine they mostly put these streams into AI systems to get a value out so the “noise” might actually matter

Vinyl and other analogue formats aren't necessarily inherently lossless. Analogue formats have a "resolution" too. For example, in photo-/videography, we have surpassed the resolution of film with digital sensors.

The challenge is data compression. If you want to argue about the SNR of the recordings, then surely you need to do so purely in the domain of the information content of the signals going into the ADC. Unless I'm mistaken, that information is unspecified. Rather than deep dive facts about that, this video proceeds with assumptions about that source signal by analogy from audio.

Vinyl isn't lossless. Any recording is imperfect. Moreover, there is always a sampling rate high enough to be higher fidelity than any given analog recording tech.

This made me question the validity of Theo's audio engineering degree.

I didn't know you were an audio engineer, one more reason to love this channel! That was my thing for 15 years before I started programming.

This was so helpful. Thank you for explaining how lossless depends on point of reference. :)