The edge stack · thermodynamic Physical AI

Sampling is the compute: let physics draw the samples

Generative AI is sampling from a distribution. A digital chip does matrix-multiply, then samples; thermodynamic hardware skips the multiply and lets thermal noise fall into the distribution directly. Here, walkers doing Langevin dynamics on an energy landscape sample a belief, the stochastic half of the edge stack, where the deterministic half is ternary.

Temperature

—

Samples / s

—

Modes found

—

Energy / sample

—

The sampler, walkers on an energy landscape · density ∝ e^−U/T

Energy, sample digitally vs let physics do it

Noise is the resource, p-bits fluctuating

Temperature —

Too cold → stuck in one mode (the mixing–expressivity tradeoff). Too hot → structureless. The right T samples the true belief. Energy: Extropic claims ~10⁴× vs GPU; Landauer floor kT·ln2 ≈ 2.8 zJ/bit.