Pathway, the data company building live AI that thinks in real time like humans do, has introduced a bold new step forward: Baby Dragon Hatchling (BDH), a “post-Transformer” architecture. This innovation directly addresses one of the greatest unsolved barriers to autonomous artificial intelligence—the ability to generalize over time.
Why Generalization Over Time Matters
Generalization over time is one of the defining hallmarks of human intelligence. It is the ability to sustain reasoning across long stretches of time, learn from past experiences, and integrate new information in order to make accurate predictions about the future. In humans, this property enables us to adapt to unfamiliar environments, anticipate outcomes, and transfer knowledge across domains.
In contrast, Transformer models are powerful but limited. Their reasoning remains tightly bound to their training data, and they are prone to “forgetting” or misapplying information when contexts shift. While they can produce astonishing results when working within static prompts or datasets, they lack the dynamic adaptability that true autonomy requires.
Pathway’s new architecture, BDH, was designed specifically to overcome this challenge.
A Scientific Breakthrough Inspired by the Brain
In a newly published paper, “The Missing Link Between the Transformer and Models of the Brain,” Pathway outlines how its researchers have formally mapped the principles of intelligence as they emerge in the human brain. Using these insights, they engineered BDH as an artificial reasoning system that mirrors brain-like execution.
The BDH model develops a modular structure that closely resembles a network of neurons. Remarkably, this structure emerges spontaneously during training, without the need for predefined architectural constraints. The resulting system demonstrates behaviors akin to the neocortex, the part of the mammalian brain responsible for higher-order cognitive functions such as memory, learning, perception, and decision-making.
This alignment with neuroscience provides BDH with a fundamentally different foundation compared to traditional AI systems. Where Transformers are optimized for sequence processing and attention mechanisms, BDH is designed to build knowledge dynamically, to sustain reasoning chains, and to adapt as new data flows in.
How BDH Works
At the core of BDH is a population of artificial neurons that respond collectively to incoming data. As information streams into the system, these neurons interact, form connections, and build layered structures of knowledge. Unlike Transformers, where reasoning is limited by fixed input windows and context lengths, BDH’s architecture is scale-free—meaning it can sustain reasoning chains for much longer periods of time.
This property ensures that BDH remains predictable and stable even as new or unforeseen information is introduced. The system continues to reason in line with its established structure, avoiding the brittleness that often plagues black-box AI models.
According to Zuzanna Stamirowska, CEO and co-founder of Pathway, this development arose from asking a fundamental question:
“Humans learn to reason through experience. Current artificial intelligence doesn’t. We asked ourselves: what is AI missing to mimic human brain function?”
The answer was not an incremental improvement on existing Transformers but a paradigm shift.
Adrian Kosowski, Co-founder and Chief Scientific Officer at Pathway, emphasized:
“We discovered that to achieve generalization over time, we needed a completely new architecture. That’s why BDH is not just a tweak to Transformers but a fundamentally different approach.”
Advantages of BDH Over Transformers
Pathway’s research highlights several critical advantages of the BDH architecture:
- Generalization Over Time
- BDH supports longer reasoning chains, enabling AI to sustain context and make inferences over extended time horizons. This overcomes one of the most significant barriers to developing autonomous intelligence.
- Predictability
- Unlike today’s opaque “black box” systems, BDH is designed with a provable level of risk. Its scale-free nature ensures stable reasoning patterns even as conditions evolve.
- Safety
- BDH addresses philosophical challenges such as Nick Bostrom’s famous “Paperclip Factory” thought experiment. By reasoning safely over long periods, BDH avoids runaway misalignment issues that could emerge in autonomous systems.
- Composability
- Multiple BDH-based systems can be combined, producing emergent capabilities. For example, combining two BDH systems could mirror the way a bilingual child naturally develops fluency across languages.
- Efficiency with Scarce Data
- Because BDH can extend its reasoning chain further than Transformers, it can draw meaningful conclusions from smaller datasets—a major advantage in domains where large-scale data is unavailable.
- Competitive Performance
- BDH is designed to run efficiently on general-purpose hardware but shows particular promise on specialized AI processors. Its potential for faster inference could reduce enterprise costs, especially in large-scale deployments. Importantly, BDH could generate outputs in a single long-running model with significantly lower latency.
Implications for Safe and Scalable AI
The introduction of BDH marks an important milestone in the pursuit of safe, scalable, and autonomous reasoning systems.
“Generalization over time is the foundation for safe and autonomous reasoning,” said Stamirowska. “With BDH, we now have a scalable model that can sustain long-horizon reasoning, expanding the AI market in enterprise.”
By embedding safety, predictability, and adaptability into its design, BDH directly addresses some of the most pressing concerns about advanced AI, including alignment, long-term autonomy, and the risks of unintended behavior.
Strategic Partnerships
To accelerate BDH’s development and deployment, Pathway has partnered with NVIDIA and Amazon Web Services (AWS). AWS will serve as Pathway’s preferred cloud vendor, providing the computational infrastructure necessary to scale BDH-based solutions. Collaborations with hardware leaders like NVIDIA could further optimize BDH performance on specialized processors, potentially unlocking dramatic gains in efficiency.
These partnerships highlight the company’s focus on not just theoretical research but also real-world application. By ensuring BDH can run effectively on enterprise-scale infrastructure, Pathway positions itself to bring this paradigm shift directly into the hands of industry users.
The Road Ahead
The launch of BDH underscores a growing recognition that the next leap in artificial intelligence will require more than larger datasets or bigger models. Instead, it calls for architectural breakthroughs that move beyond the constraints of the Transformer era.
Pathway’s approach suggests that the future of AI lies in systems that can think more like humans do—not by imitating surface-level behavior, but by replicating the deeper principles of how reasoning and intelligence emerge in the brain.
If BDH delivers on its promise, it could pave the way for AI systems capable of sustained, safe, and autonomous reasoning. This shift could transform industries ranging from healthcare and finance to logistics and energy, where long-horizon decision-making and adaptability are essential.
For now, the Baby Dragon Hatchling is still in its early stages, but it represents something rare in today’s fast-moving AI landscape: a true rethinking of the foundations of intelligence.
For more information on Pathway, please visit https://pathway.com/. For interest in joining our team, please visit: https://pathway.com/careers/.
About Pathway
Pathway is shaking the foundations of artificial intelligence by introducing the world’s first post-transformer model that adapts and thinks just like humans.
Pathway’s breakthrough architecture outperforms Transformer and provides the enterprise with full visibility into how the model works. Combining the foundational model with the fastest data processing engine on the market, Pathway enables enterprises to move beyond incremental optimization and toward truly contextualized, experience-driven intelligence. The company is trusted by organizations such as NATO, La Poste, and Formula 1 racing teams.
Pathway is led by co-founder & CEO Zuzanna Stamirowska, a complexity scientist who created a team consisting of AI pioneers, including CTO Jan Chorowski who was the first person to apply Attention to speech and worked with Nobel laureate Goeff Hinton at Google Brain, as well as CSO Adrian Kosowski, a leading computer scientist and quantum physicist who obtained his PhD at the age of 20.
The company is backed by leading investors and advisors, including Lukasz Kaiser, co-author of the Transformer (“the T” in ChatGPT) and a key researcher behind OpenAI’s reasoning models. Pathway is headquartered in Palo Alto, California.
