A demonstration of how the Cognitive Organism learns, thinks ahead and predicts: one sealed level against four cloud AI models from OpenAI and Anthropic. To be fair to them, the cloud models first watched the same example game, and the game waited for them at every piece. It never waited for the Cognitive Organism.
0.47ms
median per answer, on one processor core
109lines
in the film, against 19 for the best cloud model
46.5lines
median over six sealed levels new to it, 20 to 109
230lines
on six new levels after one example game; none without it
What the film shows.
Its input holds no table of piece shapes and no key mapping. Shown one example game, it cleared 230 lines on six new levels; without it, none. With its left and right keys swapped in secret, it still cleared lines on every level. Part of how it chooses its moves is built in.
Asked during play where a piece would lock, how many lines a drop would clear and whether the next piece would fit, it was right 37 times out of 42; every answer can be checked. How it plans each move is its makers' account: it works out where its possible moves would leave the board.
Its first answer to each new piece took a median of 7.1 milliseconds and the answers after it a median of 0.45, on one ordinary processor core. In its film game it answered 3,182 times, every time on the running clock.
The same build on the film's level and on five more levels, each sealed before it played by our records, one game each: 20 to 109 lines, median 46.5. The film's game is the best of the six. The cloud models played the film's level.
| Level | Game time | Lines | Speed level | Answer time, median |
|---|---|---|---|---|
| The film's level, 2 Nov | 252.3 s | 109 | 10 | 0.47 ms |
| Level 1, 3 Nov | 124.1 s | 37 | 3 | 0.28 ms |
| Level 2 | 140.3 s | 45 | 4 | 0.27 ms |
| Level 3 | 81.8 s | 20 | 2 | 0.29 ms |
| Level 4 | 150.5 s | 48 | 4 | 0.28 ms |
| Level 5 | 179.9 s | 58 | 5 | 0.28 ms |
Six new sealed levels, the same frozen build, three games on each. Lines cleared:
| Level | With the example game | Without it | Keys swapped in secret |
|---|---|---|---|
| 1 | 85 | 0 | 2 |
| 2 | 37 | 0 | 20 |
| 3 | 17 | 0 | 7 |
| 4 | 34 | 0 | 12 |
| 5 | 34 | 0 | 30 |
| 6 | 23 | 0 | 12 |
| All six | 230 | 0 | 83 |
With the example game it cleared 230 lines in all; without it, none. With left and right swapped and nothing saying so, after an example game played with normal keys, it still cleared lines on every level: 83 in all. One game per level and arm, all in the download.
One sealed level, by our records never played before, with the same pieces in the same order for every player. The film is rendered from the recorded logs; how the Organism thinks inside is not shown.
Every board shows its own game clock. Until the last cloud model tops out the film plays at real game speed; then the Cognitive Organism's game plays on faster, its clock still shown. For the cloud models the game stopped at every piece until they answered. For the Cognitive Organism it never stopped.
The referee, the complete logs and the checks are published: every game of the demonstration and the learning check runs back through the referee that played it, frame by frame, to the same result, and every prediction is checked against its logged board. Independent laboratories are invited to run its frozen build on levels they make themselves, under confidentiality; no laboratory results are part of this record.
Tell us who you are and we will prepare the right briefing.
Learns. Thinks ahead. Predicts. In real time.
One sealed level. Five players. The same game.
Tell us who you are and we will prepare the right briefing.