If you believe…

There are few things more important right now than creating a shared understanding between US and Chinese labs of why AI safety is an urgent problem to solve.

Labs are torn between capability and safety work, and more often than not, they choose capability. Post-training is yielding such incredible returns, it's very difficult to commit to robust safety work as each RL runRL runA reinforcement-learning training run: the post-training step that turns a base model into a stronger assistant, and the step labs compete on hardest. contributes directly to the success of these labs.

It works because it gives compute-constrained researchers compute. Compute incentivizes research direction, and everything is downstream from compute allocation. Researcher incentives to do safety work rise, growing an indigenous AI safety expertise inside the frontier labs.

The compute can be spent only on safety. Inside that, no one tells a researcher which risks matter or which methods to use. We restrict the cluster, not the research.

The ledger measures the result. Today 6 of 19 reviewed releases have an independent evaluation, and none has one published by its own developer. Those are the two columns this compute exists to fill.