Model machine
A small computer that runs binary models. Each weight is just on or off, and this page can run a toy model.
Run it here
File toy_bmlp.tbw1 · 102,400 B · 24-byte context · 2 binary layers of 512 · 524,288 one-bit weights
File sha256 11c1c4b177a2308ee64732335335e044e4cf32ce0ddef75210b0d5fa4f5ed568(registered value, checking…)
Output sha256 0aef22dfb94c5a76f0aa1eba5ac3f3a1852ba1982a1965a5235b696579a7acca (200 B)(reference answer; checked after 200 characters)—
Everything is computed in your browser: no server connection, nothing on chain, no wallet.
Check it yourself
- Look at "File sha256" above. It should be
11c1c4b177a2308ee64732335335e044e4cf32ce0ddef75210b0d5fa4f5ed568. - Click "Write 200 characters" once. "Output sha256" should be
0aef22dfb94c5a76f0aa1eba5ac3f3a1852ba1982a1965a5235b696579a7acca. - Run it again on another computer or in another browser. The result is the same. The page uses integer arithmetic only.
What it is
- A binary model: each weight is only +1 or −1, so multiplying becomes "compare two bits, then count the matches".
- We designed the circuits ourselves: processor core, binary multiply-add unit, memory controller, display and serial port. So far they run in simulation. No physical board exists yet.
- At boot it fetches the runtime and the model from a tape container, and checks each fingerprint and signature. If a check fails, it goes back to the last version that worked.
This toy model
| Item | Value |
|---|---|
| Size | 102,400 bytes |
| Looks back | 24 bytes |
| Hidden layers | Two, with 512 binary units each |
| One-bit weights | 524,288 |
| Training text | Stories written by the Bard's model, originally from tiny-tiny-stories (MIT) |
It exists only to run the whole path end to end. The English it writes doesn't read smoothly yet.
Progress
- Done: fetch, check, start the runtime and write text, in the emulator and in circuit simulation, byte for byte identical to the reference implementation.
- Done: on an ordinary computer, the same boot loader runs a ternary model with 1.7 billion parameters (Bonsai from Prism ML, Apache-2.0), and the output is the same as running it directly.
- Not done yet: running on a real board, measuring speed on real hardware.