Lend your browser to train an AI, without it ever knowing what it is helping with.

Training an AI takes an enormous amount of arithmetic. This project spreads that arithmetic across ordinary browser tabs. Each tab does a small piece of maths on numbers that look like random noise, so it can’t tell what the work is for, and the server that asked for it checks every answer.

Watch it train live See what a worker sees

The idea in three steps

Scramble. The server takes a chunk of the training maths and disguises it. It shuffles the numbers and splits them into pieces that each look like random noise. One piece on its own says nothing.
Multiply. Each browser tab gets one piece and does plain multiplication on it, then sends back the result. The tab doesn’t know whether it’s helping to recognise handwriting, write text or anything else.
Check and combine. The server checks a short proof that comes with every answer. Wrong or faked answers are rejected and the sender is banned. Right ones are combined to recover the real result, and the model learns a little.

An everyday way to picture it

Imagine you need a huge sum done, but you don’t want any helper to know what it’s for. You cut the question into pieces, blur each one, and hand them to different people. Each does a bit of arithmetic and hands it back. Only you know how to put the answers together, and you spot-check that nobody scribbled in a wrong number. The helpers did real work and learned nothing about the project.

That’s the whole trick here, done with maths that makes the blurring and the checking mathematically solid rather than a matter of trust.

Why anyone would want this

Spare computers add up. Thousands of idle laptops and phones are a lot of processing power, and joining is just opening a web page.
Private by design. The people lending their computers don’t get to see the data, and the person who owns the data doesn’t have to hand it to strangers.
You don’t have to trust the volunteers. Because answers are checked, a dishonest or broken machine can’t quietly spoil the result.
There’s a paper trail. The whole run is written into a log that can’t be quietly edited afterwards. A contributor can be given a receipt for their work, and anyone can audit a stretch of the run.

Fine-tuning a language model, as a demo

The live demo includes a small language model that is being taught to add numbers by trial and reward, the same kind of training used to improve today’s open chat models. Volunteers have to opt in first, choose how hard their tab works, and can stop whenever they like.

It’s a tiny stand-in. A real model like Qwen or Mistral would need far more computing power, mainly from browsers with a graphics chip. Here’s what that would take.

What it does not do (yet)

It protects the numbers workers see from being understood, as long as only a limited number of them secretly team up. It does not hide the fact that some training is happening.
The proofs show an answer is correct. They don’t stop a worker from being slow or from leaving.
It is a research demo, not a finished product. The details and honest limits are here.

Explore

Live demoStart training, add nodes, and watch the model improve. What a node seesCompare a worker’s view with the real data, and test what happens if workers team up. ReceiptsProve a piece of work really went into the model. Real modelsWhat it would take to fine-tune something like Qwen or Mistral. How it worksThe mechanisms in more detail, with the limits. Speed testSee how fast your device does the maths.