Pacific Compute · safety compute for the open frontier
Open models lead the world, and the work to make them safe has no compute.
Every open model lab works under a compute budget, and that budget forces a choice between capability and safety. Capability wins almost every time. A neutral, safety-only cluster in Singapore pays for the safety side, for open models on both sides of the Pacific.
The problem
Open models are now the most-used in the world and are closing the capability gap. The people best placed to evaluate them have the least compute. Of the ten most-downloaded open models of 2025, only one shipped even one of three basic safety measures (Casper et al.).
The intervention
Checking a model is inference, not training, and costs a fraction of building one. A modest cluster could support hundreds of safety campaigns a year, where the field now runs a few dozen.
Containment
Not controlled compute for capability work. The cluster holds no proprietary weights, runs no capability training, and logs every job in public. It studies models that are already open, under a single licensed operator.
The ask
Support it early, in private: your name is never shown. Contact info@pacificcompute.org. (This text appears because JavaScript is off or blocked; the full site needs it.)