Introduction
Creating more powerful artificial intelligence is an infinite game.
Scaling compute, curating training data, and innovating on model architecture are means to an end.
The end state of artificial intelligence is a system that perfectly models reality. Since reality is infinite and the physical world is not, all intelligence, including biological, must compress reality to fit within limited resources (GPUs, neurons, etc.).
This is the compression problem of intelligence.
More intelligent systems either compress reality more efficiently or have access to more resources.
The holy grail of artificial intelligence is a system that:
- Models reality with the highest compression ratio
- Adds more compute
The purpose of Soma is to create a self-improving model of reality.
Unlike traditional AI labs, the “model” is a meta-model composed of many participants. The entire system is permissionless, allowing anyone to change the meta-model in a small way by updating their own model(s).
Participants can scale compute, curate training data, innovate on model architecture, and use any other lever to improve their contribution of intelligence.
In the long term, it is possible to imagine companies, or even nations, combining resources to train AI models. Soma takes that concept to its logical end state.