Grid-world
A box of cells. Agents step. Walls block. Tiny worlds, infinite lessons. Where every reinforcement-learning student has cried.
01
What is a simulation, exactly?
A simulation is a small world that obeys rules we wrote, populated by agents who do not know they are inside one. The simplest simulations are spreadsheets. The most ambitious ones are entire civilizations. Every simulation is a thesis about what matters.
Fig. 1.1 — A simulation, in three boxes and an arrow.
02
Agents, environments, and the loop between them
An agent reads from the world, decides, and acts. The world updates. The agent reads again. This loop — perception, decision, action — is the heartbeat of every simulated mind, from the simplest grid-walker to the most overdesigned strategy game.
- i.Perception — what the agent can see, smell, hear, or read. Most failures live here.
- ii.Policy — the function from perception to action. A table, a tree, a network — pick your weapon.
- iii.Action — the world only knows what was actually done, not what was thought.
- iv.Feedback — a reward, a state change, a signal. Without it the agent is just a clock.
caveat
If you confuse the agent's view of the world with the world itself, you are not running a simulation — you are reading a story. There is a difference, and it will eventually cost you.
03
Why simulate, when you could just train?
Real data is expensive, slow, and dangerous. Simulated data is none of those things, but it is also not real. The whole craft is keeping that gap small enough to matter without pretending it is zero.
reality
- true distributions
- collisions hurt
- edge cases come for free
- infinite hours per second
simulation
- chosen distributions
- collisions are data
- edge cases must be designed
- seconds per second, parallelizable
A good simulation is not a replacement for the world. It is a place where mistakes are cheap.
04
Five common shapes of simulation
Rigid-body physics
Newton at sixty hertz. Bodies fall, collide, scrape. Used for robots that will eventually bruise themselves on real chairs.
Population dynamics
Counts, rates, transitions. Disease spreads, predators eat, currencies inflate. Statistics in motion.
Multi-agent
Each mind has its own policy. Cooperation, defection, communication. The hard one. The interesting one.
Text-world
A novel that takes input. Players read; the world replies. The cheapest simulation per kilobyte ever invented.
05
Exercise — design a small world
Pick a question you cannot answer with a closed-form equation. Then design the smallest possible world that can answer it. Resist the urge to add features. Most simulation bugs are features in disguise.
step 1
State the question. One sentence. If you cannot, the simulation is not ready.
step 2
Choose variables. Two if you can, never more than five at the start.
step 3
Write the update rule. The simpler it looks, the longer it will live.
step 4
Run it too long. The interesting bugs only show up after the third coffee.
end
A note from the workshop
Every simulation is a small confession about what its author finds important enough to encode and what they are willing to ignore. There is no neutral world. There is only your world, drawn at scale, with a thick black border around it.
— SIM-AI / pedagogy desk
∎ end of pamphlet, please pass to a friend.