simulai.tech

A workshop for people who simulate things that don't yet exist. Physics that isn't. Agents that aren't. Worlds that might be.

This is not a product page. There is nothing here to purchase, no onboarding funnel, no gradient CTA. What follows is a record of experiments in differentiable simulation, probabilistic inference, and autonomous agents that run on attention the way old hardware ran on solder and spite.

margin note / 2026.04 — revision 7

GRADIENT METHODS

LIFTED INFERENCE

NUMERICAL RECIPES

# integrator.py
def leapfrog(q, p, dU, eps=0.01):
    p -= 0.5 * eps * dU(q)
    q += eps * p
    p -= 0.5 * eps * dU(q)
    return q, p

CH. 02

§ 4.2.1

// observer.ts
const field = tensor([n, n], 0);
for (let t = 0; t < steps; t++) {
  field = step(field, dt);
}

$ cat /doctrine.md

differentiable everything

We write simulators that know their own Jacobians. Every parameter is a handle; every handle can be pulled by gradient descent, by a Kalman filter, by a reinforcement learner that treats the world as a very slow training loop.

[01] Fluid dynamics under adjoint methods
[02] Rigid-body contact with smooth relaxations
[03] Agent populations as stochastic processes
[04] Protocol: suffer gladly, log everything

addendum / ledger no. 11

CH. 03

> ls /projects

current experiments

// 001

ember-solver

adjoint CFD · Rust + CUDA · 2,100 LOC

A differentiable Navier-Stokes kernel compiled to GPU shaders. Back-propagates through 10k timesteps in under a minute. Named for what it leaves on the heatsink.

// 002

oracle-minor

probabilistic programming · Python · 4,800 LOC

HMC sampler that treats itself as a subject of its own inference — meta-Bayesian, unreasonably slow, and occasionally correct in surprising ways.

// 003

tallow

agent framework · TypeScript · 1,300 LOC

Low-fat agents: no memory longer than a turn, no tools beyond reading a file, no personality. Surprisingly capable; disturbingly honest about what they don't know.

filed / stack 3, shelf 2

// ember-solver / kernel.rs
pub fn step(&mut self, dt: f32) {
    self.advect(dt);
    self.diffuse(dt);
    self.project();
    self.grad = self.tape.rewind();
}

STOCHASTIC VOL 2

CH. 04

> whoami

the workbench

Run by one operator, occasionally two. No VC. No roadmap. No Slack. Correspondence by mail or by patch. We read papers in full, not in abstract; we re-derive proofs before we cite them; we keep a physical notebook because sometimes the gradient is pencil-shaped.

If the thing you're working on doesn't exist yet and also shouldn't be a slide deck, this is the right workbench.

“ simulation is the epistemology of the 21st century. everything else is accounting.

cf. notebook VII, p. 41

CH. 05

$ echo "last transmission"

correspondence

No form fields. No mailing list. If the work speaks to the work, you already know how to reach us. For everyone else, patience is itself a filter.

pgp: 0x7A 4C 3E 91 F0 2D B8 12 A6 49 77 E3 0B 1C 5D 8F
mail: smoke <at> simulai <dot> tech
geo: 47.6N / 122.3W, 3rd floor, window facing alley
hours: 23:00 — 05:00, most nights

> session paused. not terminated._

end of file / no footer follows

# log.txt  // 03:47 AM
#
# converged. barely.
# residual 4.2e-7
# coffee: negative
# orange: gone
# will continue tomorrow

MANUAL / VOL I