When can we train an
ASL-2
level AI model over a peer-to-peer network?
Current estimation:
Checklist:
Large-scale, Byzantine-tolerant decentralized stochastic gradient descent
Communication-efficient workload scheduler for a particular training workload of an ASL-2 model
Implementation of a decentralized data loader for requisite data
Implementation of a peer-to-peer protocol for high-performance numerical computing
Incentive structure of an ASL-2 training workload
Observed social coordination resulting in meaningful training progress
ASL-2 level capability observed as a direct result of training progress