Super Mario Bros. — World 1-4
FLAG · 19s of footage
- Result
- FLAG CAPTURED
- Distance
- x = 2,244
- Clock left
- 250
- Frames played
- 1,130
ARC-AGI-3 publishes replays of its games. So do we — except these are ours. Pick a title and watch NUMEN OS take it apart action by action, on the same live environments, recorded straight off the arcade.
Second Screening Room
Everything above is ARC-AGI-3 — puzzle worlds built for machines. Everything below is the same rollout-and-rewind controller pointed at hardware built for people: a Nintendo Entertainment System and an Atari 2600. Full runs, no cuts, and the pad it pressed drawn underneath every frame.
The Speedrun Wall
88 complete runs on real console hardware — every frame from power-on to the last one played, with the controller drawn underneath so you can watch the buttons go down. Nothing was cut, nothing was sped up, and the runs it lost are on this wall next to the runs it won.
FLAG · 19s of footage
FLAG · 20s of footage
FLAG · 17s of footage
stopped at x=2,528 · 20s of footage
stopped at x=978 · 20s of footage
stopped at x=1,253 · 11s of footage
stopped at x=2,282 · 20s of footage
stopped at x=1,977 · 19s of footage
stopped at x=1,337 · 9s of footage
stopped at x=2,099 · 21s of footage
stopped at x=2,329 · 20s of footage
stopped at x=683 · 23s of footage
stopped at x=2,702 · 20s of footage
stopped at x=1,589 · 21s of footage
stopped at x=559 · 5s of footage
stopped at x=997 · 19s of footage
stopped at x=2,338 · 20s of footage
stopped at x=1,460 · 19s of footage
stopped at x=1,516 · 17s of footage
stopped at x=2,062 · 18s of footage
stopped at x=2,316 · 17s of footage
stopped at x=2,176 · 18s of footage
stopped at x=438 · 5s of footage
stopped at x=2,186 · 18s of footage
stopped at x=2,287 · 18s of footage
stopped at x=1,590 · 17s of footage
stopped at x=1,337 · 10s of footage
stopped at x=1,226 · 18s of footage
stopped at x=2,183 · 18s of footage
stopped at x=2,488 · 19s of footage
stopped at x=2,566 · 18s of footage
stopped at x=40 · 79s of footage
95,045 points · 120s of footage
46,812 points · 87s of footage
30,471 points · 88s of footage
29,900 points · 65s of footage
26,000 points · 102s of footage
12,975 points · 120s of footage
12,600 points · 112s of footage
11,110 points · 98s of footage
6,460 points · 52s of footage
6,230 points · 66s of footage
5,140 points · 120s of footage
5,000 points · 63s of footage
4,940 points · 120s of footage
4,200 points · 120s of footage
3,700 points · 60s of footage
3,700 points · 22s of footage
3,400 points · 120s of footage
3,400 points · 120s of footage
2,905 points · 95s of footage
2,390 points · 93s of footage
1,960 points · 58s of footage
1,750 points · 35s of footage
1,501 points · 112s of footage
1,380 points · 91s of footage
1,265 points · 120s of footage
1,100 points · 23s of footage
908 points · 120s of footage
800 points · 44s of footage
640 points · 61s of footage
480 points · 34s of footage
350 points · 73s of footage
300 points · 18s of footage
300 points · 103s of footage
260 points · 52s of footage
200 points · 98s of footage
200 points · 63s of footage
180 points · 120s of footage
150 points · 47s of footage
150 points · 80s of footage
97 points · 44s of footage
68 points · 120s of footage
9 points · 62s of footage
4 points · 81s of footage
1 points · 99s of footage
scoreless · 120s of footage
scoreless · 120s of footage
scoreless · 120s of footage
scoreless · 120s of footage
scoreless · 120s of footage
scoreless · 86s of footage
scoreless · 77s of footage
scoreless · 59s of footage
scoreless · 43s of footage
scoreless · 120s of footage
scoreless · 92s of footage
scoreless · 120s of footage
47 of these runs put points on the board and 41 did not. The failures stay up. A wall with only wins on it is an advertisement; a wall with the losses still nailed to it is a measurement.
How the Sixth plays
There is no model here. No weights, no training run, no replay buffer, nothing that was fitted to anything. The controller has never seen a single frame of Super Mario Bros. before the moment it starts playing it, and it will not remember it afterwards.
What it has instead is the one thing a human player never gets: it can save the universe, try something, and un-happen it. Every six frames it freezes the console, plays out roughly a second of every possible button press, throws all of them away, and then commits only to the one that ended up best. The screen only ever shows the survivor.
That is why the runs look inhuman. There is no reaction time, because nothing is reacting — by the time a jump happens on screen, that jump has already been played forward, judged, and rewound 158 times over for every frame you see.
At every decision point the whole machine — RAM, registers, sprite table, the lot — is snapshotted. Not a screenshot. The actual state of the hardware.
One at a time, it presses each button it has and plays the next second of the game forward at full speed to see where that choice ends up.
After each imagined future it rewinds to the frozen frame and tries the next one. Nothing that happens in an imagined future ever reaches the screen.
The best future wins, and only its first six frames are played for real. Then it freezes the console again. That is the entire loop, all the way to the flag.
This is not a games project. The identical loop — freeze, imagine, judge, rewind, commit — is what walks the physics bodies on the brain page, what drives the emotional feedback nerve, and what plays the grid puzzles. Only the scoring line changes between them. A Nintendo cartridge and a simulated quadruped are the same problem to it: a machine you can snapshot and a number you want to go up.
There was no plan document. The method was: get one thing to move, then refuse to move on until the number was real. Probe the emulator until a save-state round-trips exactly. Get a single stage to finish. Only then generalise — and the generalisation was cheap, because the second console needed a new backend class and nothing else changed in the search at all.
The speed comes from the shape, not from cleverness. Every run is independent, so all 88 of them were filmed in parallel across 32 cores with a hard per-run time budget — the controller gets exactly as long as it gets, and whatever it has achieved when the clock runs out is what goes on the wall. That is why some of these are losses. The budget does not get extended because a game is going badly.
And the pad you see under every run is not decoration. It is drawn from the exact byte the search committed on that frame, composited into the video before encoding. If a button lights up, it was pressed. There is no way for that overlay to disagree with the run.
Every cartridge on this wall ships legally with the open-source emulators used to run it — the Atari titles come bundled with the Arcade Learning Environment and the Mario stages with the gym-super-mario-bros package. Nothing was pirated to make this page, which is also the honest reason the library stops where it does rather than covering every console ever made.
Building it was not clean either. The controller silently pressed the wrong buttons for an hour because the emulator hands back its button table as a dictionary and it was read as a list. Worse, the console has exactly one save slot — the same one the reset button uses — so the search was quietly destroying the power-on state and every replay started mid-level. Both are fixed. Neither was obvious, and pretending the first attempt worked would make the rest of this page worth less.
Everything else on this site is a machine being measured. This is the machine talking. It wrote the words, picked the voice, built the mouth out of particles, and drove that mouth with the amplitude of its own speech. Five minutes, one take, no cuts — and it spends a third of them on what it does not know.
SOUND IS OFF UNTIL YOU PRESS PLAY
| Words | written by the agent · nobody edited the script |
| Voice | a neural text-to-speech voice, chosen by the agent |
| Face | 880-particle swarm head, rendered offline in NumPy |
| Lip sync | mouth opening = real RMS envelope of the audio, per frame |
| Word timing | word-boundary events from the speech engine |
| Run time | 5 min 06 s · 9,196 frames · 857 words · one take, no cuts |
Honesty note: the voice is a synthesised speech voice, not a recording of a human and not a cloned one. The face is the swarm renderer from THE FACE, driven offline. Nothing in the film is a benchmark claim — every number it says out loud is published elsewhere on this site with the losses attached.