Essay · Information · Thermodynamics

Time Travel,
From a Zelda Speedrun

A deterministic minigame, a fixed seed, and the cheapest version of the past the universe will actually sell you.

The fixed seed(100, 100, 100)
RNG calls before the squids appear: 0
An essay in Jason's voice · ~9 min read
scroll

There's a minigame in The Legend of Zelda: The Wind Waker called Sploosh Kaboom. It's Battleship. You fire at a grid, three squids are hidden somewhere on it, and you have a limited number of shots to sink them all. For a 100% speedrun you have to win it twice, and it sits about thirty minutes into a run that takes hours. For years it was the thing that killed otherwise-perfect attempts, because the board looked random. There are 604,584 possible layouts and every one of them can come up. You guess, you pray, you reset.

In 2020 a group of runners, led by Linkus7 with the heavy lifting done by an ML researcher named Peter Schmidt-Nielsen, stopped praying and reverse-engineered the game. They pulled the board-generation code out of the binary, watched memory in an emulator, and found the thing underneath the randomness. Wind Waker doesn't roll dice. It runs a Wichmann-Hill pseudorandom generator, and it always seeds that generator with the same three numbers: 100, 100, 100. One generator, shared globally, called dozens of times per frame for everything the game does. By the time you reach Sploosh Kaboom it's been called something like ten million times.

The board was never random.

It was the ten-millionth-and-something output of a fully deterministic function that started from a fixed seed. The only reason it felt random is that nobody knew which step of the sequence they were on. Observe a single board, locate yourself in the precomputed sequence, and you can read off every board that follows. The squids stop hiding.

So play it yourself. Fire blind — or reveal the seed and watch the chance collapse into something knowable.

Live demo · Sploosh Kaboom
SHOTS FIRED0
SQUIDS HIT0 / 9
BOARD SOURCErandom
STATUSfiring…
watermisssquidseed-known

Random board: every layout is possible, so blind firing is a coin-flip.Reveal the seed: the same fixed start always produces the same board — click it twice and the squids land in the exact same place. Show probability runs the actual solver idea: take every layout still consistent with your hits and misses, and fire where the most of them agree a squid should be.

The question hiding in the seed

A speedrun is a universe with an arrow of time

The state evolves forward, deterministically, from initial conditions. "Going back" in that universe doesn't mean reversing the river — it means re-running the river from the seed. You don't rewind ten million RNG calls. You reset to (100, 100, 100) and replay.

That distinction is the whole game. There are two ways to imagine going back, and they have wildly different price tags.

Determinism, in one line
Seed
100,100,100
Generator
Wichmann-Hill
Tick
10,000,000
Board
#48,213

Seed plus rule plus step-count equals board. No dice are ever rolled. Knowing where you are in the stream is the entire trick — and it's the entire trick for time travel too.

The first way is to store and restore: record the complete state at the moment you want to return to, hang onto that recording, and later write it back over reality. This is the time travel of movies — you carry your memories and your scars back across the boundary, and the world is restored bit for bit.

The second way is to re-seed: throw away the recording, keep only the generating rule and the tiny seed it started from, and regenerate. This is the speedrun. It's also, it turns out, the only version the physics can afford.

Two ways back · very different bills
Option A

Store & restore

Record the whole universe. Carry the snapshot. Write it back.

cost: ~10⁹⁹ joules & nowhere to put it
Option B

Re-seed

Keep the rules. Keep the seed. Re-run from the start.

cost: a few constants & the laws of physics

Same goal, opposite economics. The interesting work is figuring out why the first one is so expensive — and the cost is measurable, in joules.

A closed universe can't afford its own memory

The information is the universe

Start with a closed universe — one box, fixed energy, nothing in or out. Roughly the one we live in. To do store-and-restore, you need somewhere to keep the recording, and the recording is not small. The holographic principle caps the information in a region by the area of its boundary: about one bit per four Planck areas. Run the numbers for the observable universe and you get on the order of 10¹²² bits — and that's not a reserve, it's the budget already in use. The universe is, informationally, close to full. It is its own maximum-resolution snapshot. There's no spare room inside it for a second copy of itself.

That's the storage problem. The energy problem is worse. Information isn't free to manipulate: Landauer's principle, confirmed in the lab since 2012, says erasing one bit dissipates at least kT ln 2joules. Tiny — right up until you multiply by 10¹²². Rewinding the universe meanserasing the information that distinguishes now from then. So price it as an erasure.

The thermodynamic bill · log scale
Erase the matter bits (~10⁹⁰)~10⁶⁷ J
Total mass-energy of the universe~10⁷⁰ J
Erase one universe of information (10¹²²)~10⁹⁹ J
0universes' worth of energy to rewind one
= 2.6e-23 J

A closed universe cannot send its full state backward for the same reason you can't fit a stadium inside one of its own seats. (Honest caveat: the 10¹²² figure is the holographic count, dominated by horizon degrees of freedom. If you only had to reset ordinary matter — ~10⁹⁰ bits — the bill drops to ~0.2% of the universe's mass-energy. Astronomical, but no longer impossible. Which bits you think you need to reset is doing a lot of work.)

The arrow is a gradient, and gradients don't loop

Geometry permits the loop. Thermodynamics forbids the edit.

General relativity does not forbid going back. Real solutions to Einstein's equations — Gödel's rotating universe, certain wormholes — contain closed timelike curves, paths that loop to their own past. The geometry is allowed. The trouble is the entropy gradient: the arrow we actually feel, the one that tells a broken egg from a whole one, always climbs. To close the loop you'd need a stretch where disorder spontaneouslydecreases — the one thing the second law won't hand you for free.

Entropy can't climb back to where it started
watching entropy climb…

The Novikov self-consistency principle and the quantum models (Deutsch; Bennett–Schumacher post-selection) all share one move: they let you visit the past but neveroverwrite it. They permit the trip and forbid the edit — which is the same wall, approached from a different side. The edit is the recording. The recording is what costs 10⁹⁹ joules.

The open universe, and the price of a fresh box

Maybe you go open — many-worlds, eternal inflation, a multiverse spawning bubbles. It doesn't help the way you'd want. Many-worlds branches forward; it never backs up the tree. And the Farhi–Guth work on creating a universe in a lab found you can nucleate a baby universe from a tiny seed, but it needs an initial singularity behind a horizon, at energies we can't reach. The seed is small; the apparatus to store a whole snapshot somewhere external is bigger than the universe you're copying. You've just moved the storage problem up a level and made it worse.

Both branches dead-end in the same place. And that's when the speedrun answer stops being a cute analogy and starts looking like the shape of the only affordable option.

Re-seed, don't rewind

The cheapest past money can buy

The runners never stored a single board. They found the seed and therule, and from those two small things regenerated any board on demand. The entire ocean of randomness compresses to a constant and an algorithm. This has a name outside of Zelda: Kolmogorov complexity. The true information content of a system isn't the size of its state — it's the length of the shortest program that regenerates it.

0
bits of full game state
vs
~23 bits
seed + step count
≈ 0× compression

Hold the laws of physics (a textbook's worth) plus a seed (the low-entropy initial conditions of the Big Bang), and you don't return to the past — youre-run it. No storage tax, because you stored almost nothing. No erasure bill, because you're generating, not overwriting. The arrow stays a gradient; you just start a fresh climb from the bottom.

The catch, which is the whole question

Is the universe a pseudorandom generator?

Wind Waker is reproducible from a seed because its randomness is pseudo-random. Wichmann-Hill is deterministic: same seed, same sequence, forever. That's why (100, 100, 100) is enough. Whether our universe is like that is one of the genuinely open questions in physics.

Two re-runs from the same seed
Same seed → same history. The squids land in the same place. Re-seeding gets you thisuniverse back.

If quantum mechanics is fundamentally stochastic — if a measurement genuinely rolls a die that wasn't pre-loaded — then there is no seed. Re-running the initial conditions gives youa universe, not this one; the squids land somewhere else, and the compression trick fails because there was never a short program to begin with. But if the universe is deterministic at bottom — hidden variables, some superdeterministic substrate, "it from bit" all the way down — then it is a Sploosh Kaboom board. Enormous, beautiful, and in principle re-runnable from a seed and a rule, with everything that ever looked like chance turning out to be the ten-billionth output of a function we haven't reverse-engineered yet.

Which is a better punchline than I expected to find. The speedrunners' whole achievement was discovering that what looked like randomness was a deterministic sequence nobody had located themselves inside of. The question of whether time travel is even coherently possible — as re-seeding, the only affordable kind — reduces to the same question, scaled up: is the universe rolling dice, or running a generator with a seed we can't see?

I don't know. But I know which one I'd bet a speedrun on.

(100, 100, 100)