A SMALL EXPERIMENT IN BIG IDEAS
A computer
in a model.
What if a neural network could run a machine? Explore a tiny monitor executing through constructed Beam decoder weights, one instruction at a time.
The machine room.
Boot the monitor and replay a real checkpoint run. Each step below came from a model forward pass.
BOOT MACHINE TO BEGIN.
Inside the circuit
Watch the machine state move as the recorded instructions execute.
Four-bit addition wraps modulo 16. Examples replay captured executions.
When precision
changes the answer.
A small CPU experiment with a separate binary-clock attention circuit shows why execution needs verification. These are 128 immediate reads using one seed; they are not results for the pretrained Beam model.
See the recorded test ↗Constructed BeamAttention layer · CPU · seed 417
Under the hood.
A small machine, a precise boundary.
CPU transitions and attention-based memory reads run inside a hand-constructed two-layer BeamForCausalLM checkpoint. A host supplies device input, injects memory-write tokens, and renders the output.
This page is a recorded replay, not live inference. The checkpoint uses Beam’s architecture; it does not contain the pretrained 501B Beam weights. The program is a single-task monitor, not a general-purpose operating system.
Explore the code & checkpoint ↗