Dust
Dust is Q Labs’ activation-space perturbation method for forward-only Transformer pretraining. It estimates token-local output errors and uses layer inputs to form optimizer updates instead of propagating a backward chain.
Contract And Limits
Token-level virtual populations, cached clean activations and attention-specific credit rules distinguish it from whole-model black-box weight search. Looped/recurrent training is proposed rather than demonstrated.
See the source evidence ledger for dates, status, pinned artifacts and numerical comparisons.
Official Artifacts
Relation To Foundation TSFM Agenda
This is adjacent training/evaluation context, not direct numeric time-series or action-conditioned world-model evidence. Forward-Only Sequence Training compares the mechanisms and defines the long-memory tests still needed.