Pathway’s brain-inspired architecture development on Amazon SageMaker HyperPod
Amazon Web Services Paulo Aragão ● Covered by 2 sources
Pathway built a brain-inspired model on AWS SageMaker HyperPod that reasons without long text traces. It hit 29.5% on ARC-AGI-1 at $0.0007 a task.
Based on reporting by Amazon Web Services, Paulo Aragão — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Pathway is trying to move AI past the familiar transformer playbook. Instead of making a model think by spitting out more and more tokens, its BDH system does the work in latent space, refining answers without building a long verbal trail. The company says that matters because today’s models pay for reasoning in context, latency, and compute.
BDH, short for Dragon Hatchling, is built around a brain-inspired design: sparse, local interactions, persistent state, and connections that act a bit like synapses. Pathway says only about 5 percent of neurons are typically active at once, which is meant to keep inference lean. The model also updates its internal memory during inference, so it can work through new problems without fine-tuning or retraining.
That architecture is not just a theory exercise. Pathway says BDH-CQ, its reasoning system built on top of BDH, reached 29.5 percent pass@2 on ARC-AGI-1 at a cost of US $0.0007 per task. The benchmark uses a few before-and-after examples and asks the system to infer an unknown visual rule, then apply it to a new grid. In Pathway’s telling, the model’s recurrent latent computation changes the cost-accuracy tradeoff for that kind of task.
Amazon SageMaker HyperPod is the other half of the story. Pathway uses it to scale training, and says the setup helps its applied AI scientists share compute in a resilient, scalable, and cost-effective way. The company also points to observability tools like Amazon Managed Service for Prometheus and Amazon Managed Grafana, plus EC2 p5en.48xlarge instances, EFA, and an EC2 UltraCluster to keep the distributed training machinery visible and manageable. That’s not glamorous, but it’s exactly where these systems either become practical or become a lab demo.
The pitch here is bigger than one benchmark. Pathway is betting that intelligence does not have to be token-heavy, and that reasoning can live in state rather than in a stream of text. If that holds up outside curated demos, it would make a lot of current AI infrastructure look rather expensive for the amount of actual thinking it buys.
My take — AI-written commentary, not fact-checked reporting
This is the kind of AI work that deserves more attention than yet another bigger transformer with a shinier name. Token-heavy reasoning is becoming a very expensive habit, and Pathway is poking at the obvious flaw: maybe the model isn’t dumb, maybe the architecture is just wasteful. The industry loves to worship scale; biology, annoyingly, keeps getting a vote.
Read more about this at: Amazon Web Services