Snake at 120 moves a second: System 0 vs System One models, at game speed

Decision models (the "System One" family) are fast: tens of milliseconds where a chat LLM takes about a second. Is that fast enough? Give a decision a hard deadline, like a game frame, and you find out.

The setup#

The game. Snake on a 16 by 16 board. Every tick, the snake asks one question in plain words: "Food is 3 ahead and 2 to the left. Straight is free, left is a wall, right is free. Turn left, go straight, or turn right?" Each lane gets the same question, in the same words.

The clock. A lane asks for a move. The move lands on the first tick after its answer arrives; until then the snake keeps going straight. So a slow answer is not wrong, it is late, and late moves crash snakes.

The lanes.

• shad0w: a 2 KB table, trained on the hosted LLM's logged answers and certified at α = 5%.
• Laya, Kev-0.8B and Kev-4B. Open System One decision models, run in llama.cpp on the Apple M5 Pro's 16-core GPU, one at a time.
• The hosted LLM that shad0w learned from.

What happened#

Without a clock, the LLM is the best player of the four models (the decision models were given no examples, and play their own way). Its copy, shad0w, plays as well (over 8 games each). Then the clock starts:

• shad0w scored 28.0 at every speed, from one move a second to 120. It decided in about 1.6 µs per move, in-process on the CPU, so every move landed in its own tick.
• The hosted LLM dropped from 25.9 to 0.1 at 10 moves a second. At about 912 ms per answer, its first answer lands around the ninth tick, and most snakes hit the wall before its second.
• The small decision models are quick but play less well: Laya at about 16 ms and Kev-0.8B at about 23 ms per move make most frames up to 30 moves a second, so their low scores come from how they play, not from time (see the no-clock column). Kev-4B, at about 145 ms, falls behind from 5 moves a second.

The arena page replays every recorded game, move for move, with charts of score against speed, time per decision against frame budgets, and memory and model size for each lane.

Then we made it playable#

Snake shows the race; Pong lets you feel it. A shad0w learned Pong from 3,000 of an LLM's answers ("a ball reaches your side in 12 frames, 0.8 paddle-heights above your paddle's centre": up, stay or down). The whole paddle brain is a 6 KB table that agrees with the LLM on 99.8% of moves it never trained on. It runs in your browser and decides every frame in about 1.5 µs (measured in Node on the Apple M5 Pro); the panel on the page measures your own browser live.

Play it. Add six balls, shrink the paddles, and watch the panel: the decisions still take a sliver of each 16.7 ms frame.

What this does and does not show#

• shad0w plays like the LLM it learned from, at table speed. It does not invent a better strategy, and it can lose.
• The decision models were not fine-tuned for snake; with examples they would play better. The point of the race is time, and the no-clock column separates play from time.
• Every lane's numbers were recorded on one machine (Apple M5 Pro: 15-core CPU, 16-core GPU, 24 GB unified memory) on one day. The arena page lists exactly how, including what was not measured.

If your app makes decisions inside a loop (a game, a trading rule, a robot, a UI that reacts per keystroke), the frame budget is the question to ask of any model. Getting started shows how to put shad0w under your own decisions.