Comment by infogulch

So they get down from 1.58 to 1.48 bits per weight by exploiting the fact that actual weights in practice are 0 51% of the time. Neat.

If ternary llms work out and are baked into hardware as custom silicon I bet they'll be shockingly efficient.