WHAT IS TRAINING@HOME?

Folding@home, but for training AI.

Open weights, free for everyone

Every model trained on Training@Home is released as open weights. Your computer’s spare time helps build better open-weight AI models that anyone can download, use and build on — for free.

Most recent problem and answer

No problem has been solved yet.

Updated every time this page loads. Only answers that were accepted appear here.

The idea

Training a neural network — a CNN, a transformer, anything — is mostly one operation repeated billions of times: multiply, add, and now and then divide to rescale. Training@Home splits that arithmetic into small independent problems, and lets anyone lend their CPU or GPU to solve them.

Each problem is one cell of a big matrix product: take a row of numbers and a column of numbers, multiply them pairwise, add everything up, divide by a fixed scale. The numbers are whole numbers (fixed-point), so every computer must get exactly the same answer — a laptop CPU and a datacentre GPU agree to the last digit.

Problems and points

Every solved problem earns points for the computer that solved it, from 1 to 10 points per problem according to how much work it takes. We count the basic steps in one problem — an add, subtract, multiply, divide or exponent step is 1, a comparison or square root counts 2 — and each extra point means roughly twice the work.

Problem	Steps	Points
Matrix product cell, rows of 16 numbers	33	6
Matrix product cell, rows of 256	513	10
Matrix product cell, rows of 2,048	4097	10
Square: x² ÷ d	2	1
Fourth power, then divide	4	2
ReLU: max(x, 0) ÷ d	3	2
Integer square root	40	6
Argmax of 16 numbers	32	5
Argmax of 256 numbers	512	9

The leaderboard ranks computers by points. Each computer appears under its own computer name (its network name, or whatever it is started with using --name) — so pick a name you are happy to show publicly.

How it works

The worker program asks this server for work. A request is always a bulk batch: at least 100,000 problems at a time.
It solves them on your CPU (or GPU) and sends the answers back.
The server checks the answers and keeps the good ones. Results from every CPU and GPU are public: the dashboard lists every device and what it solved, and finished jobs are downloadable at /training/api/job/<id>/results.bin.
The training run (on the owner’s laptop) downloads the finished results and takes the next training step.

There are always about 100 million problems open, unless training has reached its end.

Keeping it honest

Anyone can join, so answers are not taken on trust. About 1 in every 10,000 problems is quietly re-solved on ke9gdp.com’s own servers. If 25 of a device’s answers differ from the server’s, that device is warned; if more than 50 differ, it is banned and its results are discarded. A correct CPU or GPU never differs, because the arithmetic is exact.

Your computer, your rules

The worker asks permission on first run, can be limited (--threads, --throttle), and stops whenever you close it or press Ctrl+C. It only does arithmetic on numbers it downloads from this server; it reads none of your files. This site stores a hashed hardware ID, your computer’s name, your CPU/GPU model name and thread count (shown publicly on the leaderboard), and a hashed copy of your network address (never shown).

Download the Windows worker (.exe) unzip, run, answer the one question, leave it running

Open the dashboard Finished models

← KE9GDP