H64 neural policy · browser CPU
I trained a tiny neural network to play Snake by watching a teacher, making its own mistakes, and learning from the situations it actually got into. This page runs the finished model entirely in your browser.
Both systems turn context into scores for what should happen next. The scale and task are radically different.
Tokens from a conversation
28 measurements of the board
A score for possible next tokens
A score for left, straight, or right
Patterns in language and data
Patterns that tend to stay alive and reach food
Often billions of parameters
6,211 parameters · 24.3 KB
First, an expert teacher demonstrated good moves. Then the neural model drove by itself. Whenever it wandered into unfamiliar or awkward situations, the teacher labeled what it should have done. Those hard examples were added to the next lesson. That loop is called DAgger.
The model gets a compact dashboard: danger near each move, direction to food and tail, current heading, open space, obstacle distances, body length, and distance to each wall.
That makes it tiny and fast, but also limits what it can learn compared with a full vision model.
There is no Hamiltonian cycle, search tree, or hand-authored route. It makes one local decision at a time. It can still trap itself because it does not plan an entire game into the future.
It also is not told whether the board wraps. It has to infer behavior from the measurements it receives.
The deployed model is the exact FP32 checkpoint, exported into a raw binary and evaluated by a hand-written JavaScript runtime.
100K teacher states plus six DAgger rounds of 20K each.
Action accuracy on a fixed, held-out 10K-state validation set.
6,211 FP32 weights. No ONNX or ML framework in the browser.
Sizes 10, 12, 16, and 20; 24×24 was held out.
The model never trained on 24×24 boards, yet over 300 seeded games it still averaged 81 food with wrapping and 56.4 with solid walls. That is not perfect play—but it is learned behavior transferring to a larger world.
95.07% does not mean a 95% win rate. It means the model picked the teacher’s move 95 times out of 100 on individual examples. Snake is a long-horizon game: one bad move can end a run hundreds of steps later.