Research Engineer​/Research Scientist, Remote Compute Accounting

Some of the most consequential decisions of the AI era will depend on answering one question: can we verify what is happening inside AI datacenters? Answering it could unlock international cooperation, help countries protect their sovereignty, and enable trustworthy adoption of AI in high-stakes industries. Building tools that can earn trust across borders is an urgent technical and political challenge.

SASH’s Verification team builds and tests tools to verify agreements about AI. We prototype new verification mechanisms, lead international research collaborations, and work with policymakers around the world to show what these tools can do and inform how they’re used.

You’d join a small team combining technical expertise with international AI policy experience. Our technical team is led by Pascal Berrang, an Associate Professor at University of Birmingham, and brings experience from the Singapore Government. Our policy team brings experience from Oxford and the Centre for the Governance of AI, while our partners include experts from the Future of Life Institute and the University of Oxford.

The hardest question in verification is not whether a datacenter is running the model it says it is. It is whether it is also running something else. A cluster can serve declared inference all day and train an undeclared model on the capacity left over.

We are interested in software-only approaches to this problem. If a verifier can keep a cluster's accelerators demonstrably occupied (i.e., their high-bandwidth memory holding the state they claim to hold, their compute answering challenges that cannot be faked or precomputed in advance), then there is no spare capacity for an undeclared training run. Physics supplies the lever: information cannot travel faster than light, and some computations cannot be performed faster than the hardware allows, so the time a cluster takes to answer is itself evidence about what hardware it has and what that hardware is doing.

The nearest concrete version is a memory-state challenge protocol. The prover commits frequently to workload checkpoints, the verifier fires N unique nonces simultaneously, and each device must return a function of its claimed memory state within a deadline (a function designed to demand serious memory bandwidth and matrix multiplication). Round-trip times are recorded and a random sample is proven genuine. A related family, proof of useful work, pushes further: the declared workload's own computation becomes the evidence, so the accounting costs almost nothing to produce.

Compression is one of the open questions: weights or a KV cache engineered for fast lossless decompression would let a prover reconstruct state it was not actually holding, freeing the hardware for whatever else it wanted to do. Showing that workload state cannot be decompressed inside the challenge window, or modifying workloads so that it cannot, would unlock the approach. Beyond that, a prover may hold secret memory capacity in error-correcting-code cells or over provisioned regions;

the challenge has to be designed so honest response time stays feasible; and the deadline has to be short enough to block cheating while surviving network latency and kernel launch times.

You would own the question of how much of a cluster can be accounted for from outside it, and how little room can be left over.

Concretely, that means designing challenge protocols and then measuring whether real accelerators can satisfy them: microbenchmarking, writing challenge kernels, characterising what high-bandwidth memory can and cannot do under adversarial conditions, and deciding whether the numbers support a claim someone could stake an agreement on. The work sits between…