[smolder]

I would prefer that we figure out how to make the massive vector blobs called AI work more efficiently, versus throwing tons of hardware at what we can barely understand in perpetuity. It doesn't sit right with me that putting lots of floating point ability in edge devices for running these approximate-at-best models is considered the way forward for computing.

[ericd]

Yeah, bitnets already shift those matmuls to integer math.

[smolder]

Thanks for turning me on to this subject. It's interesting. I took a look at this paper: https://arxiv.org/abs/2310.11453

[ericd]

No problem, glad you found it interesting, I think it's pretty exciting stuff. This is a more recent paper by the same people, where they tweak slightly to ternary (-1, 0, 1), and they seem to be able to maintain network performance: https://arxiv.org/pdf/2402.17764