Artificial intelligence has learned to see, speak, and even write poetry, but it still hasn’t learned how to evolve on its own. Despite their billion-parameter brilliance, today’s large language models remain static. Once trained, they stop learning. That’s the paradox at the heart of modern AI: the smarter our models become, the less they resemble the thing that inspired them—the human brain, which continuously learns and adapts.
But one research team believes it may have found an answer and built it directly into a model.
Pathway, a Palo Alto–based AI startup led by complexity scientist Zuzanna Stamirowska, claims its new AI architecture, Baby Dragon Hatchling (BDH), represents a breakthrough in adaptive reasoning. Designed to mirror how intelligence naturally emerges in the brain, BDH mathematically maps how neurons interact to form and refine understanding over time. The result, according to the company, is an artificial reasoning system that doesn’t just process information, but evolves with it.
“Current LLMs are re-living Groundhog day (if you know the movie). They are trained once then wake up every day with the same state of memory (and potentially with access to a large library of notes), without having any consistent learning that could happen over time,” Zuzanna Stamirowska, CEO of Pathway, told me. “BDH has as much ‘memory space’ for their context as they have for long-term knowledge. It opens the way to systems that get better “on the job”, by solving problems, gradually over time. Like us humans.”
In its paper, Pathway says it has bridged how attention works in both neuroscience and machine learning. The team found that BDH simulates how reasoning develops, showing neurons interacting to find the next relevant idea, much like how the brain directs focus across its networks.
Moreover, because BDH’s activations are sparse and easier to interpret, researchers can see which neurons represent which concepts, a property Pathway calls “monosemanticity.” This transparency, the company says, could make future AI systems simpler to audit and regulate.
“There is a mismatch between the capacity of engineering systems and databases to ingest fresh data, and the inability of Large Language Models (LLM) to process it in a way that makes them gain insight or experience. In fact, current enterprise deployments often combine the two components: a “static” LLM, which does not improve its skills over time, with lookup access (retrieval) from an external database,” Adrian Kosowski, chief science officer at Pathway, told me. “Due to the architecture change in how BDH handles context, for many use cases, we see BDH as the way to bypass this limitation – enabling contextualized reasoning at enterprise scale.”
The Future of AI Lies in Biology, Not Just Code
BDH is built on a core neuroscience principle called Hebbian learning, often summed up as “neurons that fire together wire together.” In the human brain, repeated co-activation strengthens the links between neurons, turning simple activity into complex thoughts, memories, and behaviors. Pathway has translated that biological process into code.
In BDH, each artificial neuron works independently but connects locally with others. When certain connections activate repeatedly, they strengthen, forming pathways that represent learned ideas. Over time, this creates what scientists call a scale-free network, a self-organizing structure that stays stable even as it grows or processes new data.
“BDH goes back to first principles and inspirations behind neural networks – how a distributed, complex system of simple agents (neurons) can learn by applying local rules which require no external synchronization,” Jan Chorowski, chief technical officer at Pathway, told me.
Pathway’s leadership team brings serious technical depth. CEO Zuzanna Stamirowska, a published complexity scientist, co-authored a forecasting model for global trade networks in the Proceedings of the National Academy of Sciences (PNAS). CTO Jan Chorowski, who previously worked with Nobel laureate Geoffrey Hinton, often called the “Godfather of AI”, helped pioneer attention mechanisms for speech at Google Brain. Meanwhile, Kosowski has published extensively across computer science, physics, and biology.
“The way interactions of particles in physics lead to a global structure, and the way distributed systems perform computations at large scale, are in fact strikingly similar. We were looking for a similar way to explain intelligence: how to go from “programming” the behavior of individual neurons, to a brain-like system which displays intelligent behavior?,” said Kosowski. “Such an approach leads both to more predictable behavior of intelligent systems at scale, and opens the door to new ways of training and evaluation of performance.”
The startup recently raised $10 million in seed funding, led by TQ Ventures, with participation from Kadmos, Innovo, Market One Capital, Id4, and several angel investors, including Lukasz Kaiser, co-author of the original Transformer paper and a key contributor to OpenAI’s early reasoning models. Its technology is already in use. NATO employs Pathway’s systems to analyze live social and operational data, while La Poste, France’s postal service, uses them to improve logistics and delivery routes.
“If I tell you that AI should be more dynamic and in tune with the environment, wouldn’t you agree? It’s a no-brainer, as some of our investors said. This is fundamentally not how current AI works,” said Stamirowska. “In the case of Lukasz, it was also about our ability to translate rigorous science into practical, lasting impact. We have had great discussions with design partners in the enterprise who require deep personalization, models that learn on the job from scarce data, and the security of deployment.”
Could Self-Learning AI End the Era of Transformers?
In Transformer models, when new information appears, developers have to retrain or fine-tune the entire system. That’s why companies like OpenAI, Anthropic, and Google release numbered updates, GPT-4, Claude 3.5, Gemini 2, each one essentially a reboot of a static mind. Pathway says its new architecture changes that. Built for continuous generalization, the model evolves on its own, learning from ongoing experience rather than periodic retraining.
“BDH processes information in a localized way – its operations can be described exactly as local inference rules. Furthermore, it has been observed to spontaneously develop sparse representations – only a fraction of its units are active at any given time,” explained Chorowski. “Both strategies, information locality and activation sparsity, are employed by the human brai,n which has an unrivalled energy efficiency, using about 20W to reason about our daily lives by employing about a hundred billion neurons communicating over an ever-evolving network of hundreds of trillions of synapses.”
The implications could be both technical and economic. Retraining large models costs companies billions each year in computing power and energy. A system that learns continuously could make AI development cheaper, faster, and more sustainable. Because the architecture keeps critical data close to its processing cores, it reduces latency and slashes compute costs.
“BDH will benefit most from processing units in which the memory is meshed together with computational units, and we are closely looking at the trends in on-chip memory on Accelerators, such as Shared Mem on GPUs, or the Vector Memory on TPUs,” Chorowski added.
However, industry experts remain skeptical. They note that while BDH performs competitively with GPT-2–scale models ranging from 10 million to 1 billion parameters, it doesn’t yet demonstrate a clear scaling advantage over today’s leading architectures.
“Brain-inspired models are useful, but aircraft don’t fly like birds and submarines don’t swim like fish. Extracting design principles from nature is valuable, but literal interpretation may not work—especially in brain science where we don’t fully understand the link between structure and function,” said R. Ravi, professor of operations research and computer science at Carnegie Mellon University’s Tepper School of Business. “Explainability should be a prerequisite for public deployment, like safety factors in engineering. But no current model, including this architecture, comes close to meeting that standard.”
Likewise, Sid Ghatak, CEO of Increase Alpha and former White House AI policy advisor, called BDH a significant scientific milestone that tackles several core shortcomings of transformer-based models. However, he added that there’s still more work ahead before it can prove its long-term potential.
“While the approach does seem to address the specific safety concern of a model running for an infinite amount of time, e.g. the Paperclip Factory, I don’t believe that this approach necessarily delivers a ‘safe, autonomous reasoning system’, as it seems to have modeled the way a brain learns and reasons,” he told me. “Given its potential to adapt, change, and potentially evolve over time, I think this does challenge those current frameworks as it will become more difficult to contain and control using technology alone.”
For now, BDH remains an early-stage technology, but its potential extends far beyond Pathway’s lab. The real test will come with scale: whether BDH’s elegant equations can hold up under the complexity of trillion-parameter models and real-world uncertainty. If they can, Pathway may have sparked the beginning of a new era in AI, one where machines don’t just imitate the brain, but begin to think like it.
“By describing the emergence of reasoning from neuron-to-neuron interactions, we are creating a stepping stone towards advancing the entire field,” said Kosowski. “We believe that systems based on BDH will simply be more practical in many cases, offering more functionalities with better efficiency, especially in enterprise settings.”
