FAQ
FAQ
Aren't we just funding capability work by proxy?
Compute is fungible: if a lab was going to spend money on evals and we do that work for free, that money may go back into capability. Two things hold that in check. The cluster itself cannot do capability work: gradient updates are blocked by default, and the one exception, safety fine-tuning before launch, is approved job by job and logged where every other participant can see it. And compute incentivizes research direction: a broader base of safety compute means more safety work gets done.
Is there urgency?
Every month the cluster goes unfunded increases the risk of an open model that is deeply unsafe being released, weights and all, onto the open internet.
Why Singapore?
Singapore is one of the few jurisdictions with the legal, diplomatic, and technical posture for this kind of facility, and the machinery is already running: a national AI Safety Institute since May 2024, the AI Verify assurance lineage, and a public red-teaming record.
Why not just fund METR, Apollo, and the AISIs directly?
For the Western evaluators, fund them too. They need it. The cluster does something direct grants cannot: a neutral venue where Chinese open models can be evaluated alongside American ones, on a shared setup, with one public record.
Why no government money?
Because trust builds faster without it. A lab can take the compute without a sermon and without a spotlight. Private philanthropy is the only funding.
Who runs it?
A small team operates the cluster, and the evaluation work is drawn from the wider safety community. An established US 501(c)(3) has offered fiscal sponsorship and the agreement is being drafted.