Perceptron theory · the thermodynamic body · noise as engine
The perceptron in noise
The spin body gave one stochastic magnet a sigmoid. Wire a whole network of them together and something larger appears: a machine that doesn't compute an answer so much as relax into one. Thermal noise jostles every unit; the couplings — the weights — tilt the energy landscape; and the system wanders, then, as you cool it, settles where the energy is low. It samples the Boltzmann distribution directly. Here the heat isn't the enemy you fight to keep bits stable. The heat is the engine.
weights = couplings (the energy landscape) · compute = thermal relaxation to P(s) ∝ e−E/kT · knob = temperature (anneal) · output = a sample
✓ STRONG
Boltzmann, Ising, annealing. That a thermal system samples e−E/kT is bedrock statistical mechanics; simulated & quantum annealers exist (D-Wave). The math is a century old and exact.
◐ MIDDLING
Thermodynamic hardware. Dedicated noise-native chips (Extropic, Normal Computing) are early prototypes — promising for sampling and Bayesian inference, unproven at scale.
◔ FRONTIER
Beating the GPU. Whether thermodynamic silicon wins on real workloads — versus just doing on-chip what a GPU samples in software — is open.
I · The landscape, and the heat that explores it
Picture the state of the machine as a ball on an energy landscape the weights have shaped. Cold, it sits in the nearest dip. Warm, thermal kicks send it hopping over ridges, exploring. Over time the fraction of time spent at each spot follows P(state) ∝ e−E / kT — low-energy states are visited most. Drag the temperature down and watch the wandering collapse onto the minimum: that's annealing, and it's how the machine "decides."
the energy landscape · the thermal ball hops (hot) or settles (cold) · the histogram = the sampled Boltzmann distribution
exploring
regime
landscape= the weightsnoise= temperatureanswer= a low-energy sample
High T is creativity — it escapes bad local dips. Low T is commitment. Schedule the cooling well and the ball finds the global minimum; cool too fast and it freezes in a so-so one. Annealing is the whole art.
II · A network that samples
Now the real thing: many binary units, each flipping with a probability set by its neighbours' pull and the temperature — P(+1) = σ(2·field/T). No unit decides; the population does, settling toward states the couplings favour. This is a Boltzmann machine: set the weights and it doesn't return one number — it draws samples from a whole distribution. That's native generative modelling and inference, in physics.
stochastic units (ferromagnetic couplings) flipping by σ(2·field/T) · energy trace below · cooling settles it onto a low-energy consensus
—
system energy
0.60
temperature
III · The gift, and the bill
The gift is that noise does the work: sampling, optimisation, Bayesian inference fall out of letting a physical system equilibrate, no algorithm grinding it out step by step. The bill comes due as Landauer's principle — every bit this machine erases dumps at least kT·ln2 ≈ 2.87 zJ of heat. Which raises the obvious next question, and the final body: what if you never erased at all?
the gift= native sampling / annealing / inferencethe engine= thermal noise, not fought but usedthe bill= Landauer kT·ln2 per erased bitstatus= theory solid, hardware early