Analyzing the Distributed Training Efficiency of the Baby Dragon Hatchling (BDH) Architecture

In a lot of my job applications, I like to say that I do my own personal research on the latest machine learning techniques and architectures. It's about time that I back up that claim. I want to talk about a new post-transformer architecture called Baby Dragon Hatchling (BDH). From there, I want to look at why it may be a new contender specifically for distributed training and inference.

What is BDH [1]?

The paper for Baby Dragon Hatchling was released on arXiv in September of 2025. It was published by a few people from the tech startup Pathway [1]. Funny enough, I applied to this startup back in June of 2025 - I've definitely heard of them before, and they're composed of some very smart people.

The architecture is inspired by the brain

References

[1] Kosowski, A., Uznański, P., Chorowski, J., Stamirowska, Z., & Bartoszkiewicz, M. (2025). The dragon hatchling: The missing link between the transformer and models of the brain. arXiv. https://arxiv.org/abs/2509.26507