AIFP's compute verification plan: 1000x training slowdown via 10MB/s limits
Removing backend networks could slow rogue AI training without hurting inference—here's how.
jacob_drori's non-expert analysis of AIFP's Plan A focuses on inference-only verification: ensuring compute is used for inference, not training. The core proposal is physically removing the backend network in datacenters—the link that carries gradients during training—and imposing strict per-rack bandwidth limits on the frontend network. Since inference mostly sends prompts and responses, a 10MB/s cap (versus typical ~4TB/s backend bandwidth) barely affects inference, while a trillion-parameter gradient in fp8 would take roughly a day to transmit, massively slowing standard training.
However, attackers could use low-communication training algorithms like DiLoCo or SparseLoCo, which sync updates rarely. Covenant-72B, pretrained with SparseLoCo, compressed pseudo-gradients 146x and synced every 30 steps, suggesting a 10x training slowdown at 10MB/s. But frontier-scale models and thousands of replicas remain untested—an open question (OQ1). The plan also raises OQ2 around video generation's high egress. Overall, the strategy wins on minimal overhead and no secret sharing, but its real-world efficacy hinges on low-communication training scaling.
- AIFP's Plan A targets sub-0.1% unmonitored compute, aiming for a 1000x training slowdown with minimal inference overhead.
- Removing backend networks plus 10MB/s frontend caps makes trillion-parameter fp8 gradient transfer take ~1 day.
- Covenant-72B's SparseLoCo training suggests only a 10x slowdown at 10MB/s, but frontier-scale viability is unproven.
Why It Matters
Compute verification could cap rogue AI training, but low-communication algorithms may undermine it—critical for AI safety policy.