Interactive explainer
The MatMul playground, where the mining happens.
Matrix multiplication is the arithmetic inside every AI model, and it’s the work Pearl’s miners do. Play with it below, then turn it into lottery tickets and try to win a (toy) block.
- Runs in your browser
- A simplified illustration
01 Multiply two matrices
Inputs A edit any cell
Weights B
Outputs C pick one
Pick any output to see the row and column that make it.
One at a time, or all at once?
A processor core works through the cells one after another. A GPU has thousands of cores and fills them in parallel, which is why AI, and Pearl mining, run on GPUs.
A 3×3 product takes 27 multiply-adds.
02 Now make it AI-sized
The work grows with the cube of the size68.7 billion multiply-adds for one product of two such matrices.
For scale: a language model does roughly one multiply-add per parameter for every token (a word or piece of a word) it writes, so a 70-billion-parameter model runs on the order of 70 billion per token. Modern AI GPUs perform tens to hundreds of trillions of these operations every second.
03 Turn the work into lottery tickets
A toy, using SHA-256 in your browser- Add noise. Mix small random, low-rank noise into A and B, so the work can’t be skipped.
- Multiply. Compute the noisy product, exactly the kind of work above.
- Hash it. Every row of the result becomes a lottery ticket.
- Check the target. A ticket starting with enough zeros wins the block.
- Remove the noise. Because it’s low-rank, subtracting it is cheap, and the useful answer A×B comes back intact.
About 1 in 256 tickets wins.
- Tickets tried
- 0
- Noise seeds
- 0
- Tickets per second
- –
- Press “Mine a toy block” to start turning work into tickets.
Toy block found!
What the toy leaves out
The real thing is much harder to fake.
- Real noise isn’t picked by the miner: it’s derived from the miner’s commitment to its matrices and from the chain.
- Real tickets come from a committed record of the computation, not from a few hashed rows, and real matrices are enormous.
- The winning work is compressed into a zero-knowledge certificate that every node can check quickly.
- Real difficulty is astronomically higher, and it adjusts after every block.
Why it matters
Same math, two jobs.
In Bitcoin, the work is hashing numbers that nobody needs. In Pearl, the work is matrix multiplication, which can be the very computation an AI customer is paying for. The noise and the lottery tickets ride along, and the useful answer survives.