Wei_Dai comments on Steelmaning AI risk critiques - Less Wrong Discussion
You are viewing a comment permalink. View the original post to see all comments and the full post content.
You are viewing a comment permalink. View the original post to see all comments and the full post content.
Comments (98)
Hmm... after more searching, I found this page, which says:
So biological neurons still don't seem to be near the physical limit since they fire at only around 100 hz and according to my previous link dissipates millions to billions times more than k_B T ln(2).
A 100kT signal Is only reliable for a distance of a few nanometers. The energy cost is all in pushing signals through wires. So the synapse signal is a million times larger than 100kT to cross a distance of around 1 mm or so, which works out to 10^-13 J per synaptic event. Thus 10 watts for 10^14 synapses and a 1 hz rate. For a 100 hz rate, the average dist would need to be less.
Not my field of expertise, but I don't understand where this bound comes form. In this paper for short erasure cycles they find an exponential law, although they don't give the constants (I suppose they are system-dependent).