Pathway Announces New Brain-Inspired Artificial Intelligence Architecture
Pathway developed the BDH architecture on Amazon SageMaker HyperPod, moving beyond transformers to reason with sparsified local interactions in latent space.
Pathway, operating in the field of artificial intelligence, has developed a novel brain-inspired architecture that moves beyond traditional transformer paradigms. Named BDH, this system achieved notable results after being trained on Amazon SageMaker HyperPod.
New Brain-Inspired Architecture
As artificial intelligence systems take on more complex tasks, industry progress generally focuses on increasing model scale, data volume, and context length. However, Pathway followed a different approach from this trend by designing a brain-inspired structure.
Reasoning in Latent Space
Developed by Pathway, BDH and its extended version BDH-CQ can reason directly in latent space without generating intermediate text traces. The system works by learning from examples and refining solutions through iterative computations.
Scaling with Amazon Infrastructure
Operating seamlessly with popular frameworks like PyTorch, this new model leveraged the Amazon SageMaker HyperPod infrastructure to scale its training processes. Workloads were distributed across instances featuring NVIDIA H200 GPUs.
Performance and Cost Success
Pathway Co-founder and CEO Zuzanna Stamirowska stated that a 150-million-parameter model set a new standard on the ARC-AGI-1 benchmark by performing iterative reasoning in latent space.