How can we fully harness
the compute of thermodynamics?

id Software's real E1M1, running as a 25,000‑cell probabilistic lattice. Every pixel, the textures, the lighting, the monster AI — and even which wall you see — is a sample from a Potts Hamiltonian relaxed by block‑Gibbs. Drop the temperatures and watch the whole world dissolve into noise and re‑condense. Nothing on screen is a fixed image. (Desktop recommended.)

A frozen LLaMA-3.2-3B run as an energy-based sampler. Clamp a prompt and an ending; watch the middle condense out of the noise. Information flows backward through the sentence — something no ordinary chatbot can do.

A persistent stochastic solver. Add constraints by clicking while it runs — it never restarts, and when you over-constrain it, it publishes its best honest compromise instead of giving up. Anytime, interactive constraint solving.

The same runtime on the classic constraint grid: type in your newspaper puzzle, clamp the clues, and watch the blanks relax into place — solving to zero conflicts, and degrading gracefully on a contradictory puzzle instead of failing silently.

A denoising thermodynamic model with every pixel exposed as a real measurement of uncertainty and energy — the honest, see-through version of the demo everyone else renders as a flat grid.

Extropic's own published codon-optimization energy, run inside this live runtime with real E. coli / human / yeast codon tables — optimize the COVID-19 spike for a human cell and watch every gauge. Reproduced against their baseline to the decimal (444.11 vs 444.0).

The hardware side. A bit-accurate golden reference model of a p‑bit sampler datapath — fixed-point field, sigmoid LUT, LFSR compare — scored live against the exact Boltzmann distribution. Inspect every number, download a reproducible receipt, and score a real device's own sample stream for pass/fail. The same core passes 12,864/12,864 vectors in synthesizable Verilog.

The full writeup on ThermoBridge / TASB: a frozen LLaMA‑3.2‑3B running with its attention replaced by provably-exact Boltzmann sampling, no retraining. χ² goodness‑of‑fit against the analytically enumerated target (mean p ≈ 0.497), a blend that is bit-exact at α=0, the measured fidelity envelope in K and α, four silent corruption bugs found and fixed — and a full accounting of the ~10× cost and the hardware the whole thing is contingent on.