Dragon Hatchling
Summary
Dragon Hatchling, also referred to as BDH and Baby Dragon Hatchling in the public narrative, is Pathway’s recurrent attention/state-space sequence-model architecture based on sparse positive activations, a large mutable fast state, and a graph/neuron-particle interpretation.
Role In The Wiki
Use this page as the object card for the model family. The source page carries the evidence details, limitations, narrative caveats, and agenda mapping.
Relation To Foundation TSFM Agenda
Dragon Hatchling is an architecture hypothesis for maintained latent state and sparse fast-state updates. It is not direct evidence for time-series foundation models, numeric time series, event streams, or action-conditioned world models until it is tested with those interfaces.
Evidence
Overlap Notes
Dragon Hatchling is closest to the wiki’s recurrent-state and test-time-memory branch because inference updates a persistent fast state. It differs from Mamba, Mamba-2, and Mamba-3 because its state update is framed through attention-like sparse graph dynamics rather than the structured linear SSM path. It differs from Titans, ATLAS, MIRAS, and MesaNet because its mutable memory is the core architecture state rather than an added test-time-memory module or local optimization solver.
For Alex’s work, the important question is whether this style of high-dimensional sparse state can track regimes, topology, exogenous context, and action effects in multivariate time-series systems.