Superintelligence should be controlled by the greatest number of people. Today, it’s controlled by whoever owns the most GPUs.
The Internet has more compute, more data, and more minds than any single organization. All the resources to train AI at massive scale — simply lacking coordination.
Soma is a network that trains a foundation model. It uses competition instead of gradient passing to train AI across the Internet. The weights are open. Participants download, train, and submit new weights. They compete against each other on a universal pretraining objective: given any data, predict what comes next. The best weights are rewarded each round and the bar keeps rising.
The goal: a model that scales with every contributor — and belongs to those who make it better.
If you can make the model better, prove it.