Researchers at Pathway (Adrian Kosowski, Przemysław Uznański, Jan Chorowski, Zuzanna Stamirowska, Michał Bartoszkiewicz) posted 'The Dragon Hatchling' (BDH) to arXiv on 30 September 2025, describing a language model architecture built from a scale-free network of locally interacting neuron particles rather than standard Transformer blocks. The authors report BDH follows Transformer-like scaling laws and matches GPT-2 performance on language and translation tasks at equal parameter counts from 10M to 1B on the same training data, while admitting a GPU-friendly formulation. Working memory at inference is said to rely on synaptic plasticity with Hebbian learning over spiking neurons, with individual synapses strengthening for specific concepts. Activations are sparse and positive, and the authors demonstrate monosemanticity on language tasks. Code is released at github.com/pathwaycom/bdh.

• arxiv.org2026-08-11