Research Engineer / Research Scientist, Compute Accounting

This is a fully remote position, open to applicants in Singapore, +1 more country.

📋 Description

• Create challenge protocols to evaluate the extent of external accountability for an AI compute cluster.

• Assess whether real accelerators can meet challenge protocols through the development of microbenchmarks and challenge kernels.

• Analyze high-bandwidth memory performance under adverse conditions.

• Verify if the measurements support claims that are suitable for international agreements.

• Collaborate with cryptography experts and colleagues in hardware security.

• Develop candidate challenge functions and evaluate honest-response latency across various real cluster topologies.

• Measure and minimize the spare capacity that a cluster can maintain while successfully passing challenges.

• Investigate adversarial compression of frontier-model weights and key-value caches.

• Design and assess post-training strategies that elevate compression costs without significantly impairing model performance.

• Measure hidden memory capacity and propose protocol modifications to make it ineffective to conceal.

• Expand verification processes from memory usage to computation through proof-of-useful-work methodologies.

• Conduct end-to-end testing of protocols against live inference clusters and document any failures encountered.

• Create red-team/blue-team protocol competitions and establish red-team timing assumptions.

• Disseminate research outcomes, including negative findings.

⛳️ Requirements

• Experience in GPU performance or systems measurement, including expertise in kernels, microbenchmarks, memory hierarchies, or distributed timing, or a background in cryptography with a readiness to engage hands-on with hardware.

• Strong measurement practices, particularly in evaluating how measurements can be manipulated.

• Capability to account for adversarial actions while optimizing system performance.

• Comfort working on protocols that may not succeed and publicly sharing negative results.

• Ability to excel in unclear, early-stage settings.

• Experience in CUDA or similar low-level accelerator programming may be beneficial.

• Familiarity with compression, information theory, or model quantization may be advantageous.

• Background in timing-based protocols, proofs of space or time, or distance bounding may be a plus.

• Experience in characterizing undocumented hardware may be a benefit.

• Knowledge of LLM inference internals, including batching, key-value cache management, and scheduling may be advantageous.

• Candidates must be eligible to work remotely from Singapore, the UK, or another approved remote location; UK visa sponsorship is available for relocation.

🏝️ Benefits

• Competitive benefits and leave policies.

• Visa sponsorship in the UK for candidates seeking relocation.

• Flexible remote work arrangements.

• A globally distributed team with adaptable working hours across different regions.