Pathway released its brain-inspired BDH architecture on Amazon SageMaker HyperPod, saying it enables more efficient reasoning without generating intermediate text traces. It is the company's first major update to its architecture since the introduction of its foundational graph-based model.
Pathway reported a state-of-the-art result on the ARC-AGI-1 benchmark, measured with a 150M-parameter model. That compares with earlier results from traditional transformer-based models.
BDH is built on a graph of neurons that communicate through sparse, local interactions and targets complex reasoning tasks. Availability begins with early access for developers and researchers.
"Today’s AI pays a steep token cost for reasoning, but that cost is imposed by architecture, not by any law of intelligence," said Zuzanna Stamirowska, CEO and co-founder, Pathway. The model’s recurrent computations in latent space allow it to process longer sequences more efficiently.
The announcement follows growing demand for more efficient and scalable AI systems. Pathway’s work highlights the potential of brain-inspired models for advanced reasoning in production environments.
Pathway did not say how the model performs in real-world applications, and it raised questions about the long-term scalability of the architecture. The company plans to continue refining BDH with further updates.
Source: awsml
