We recently had the opportunity to hear from Jan Chorowski, a Mila alumni now CTO and co-founder of Pathway, an AI-driven data processing platform.
Tell us about your academic and professional path.
I’ve spent my career exploring how neural networks can get closer to how human intelligence actually works. Early in my research, I was the first to apply attention mechanisms in speech recognition, an approach that helped shape the field of sequence-to-sequence learning. I had the opportunity to collaborate with Geoffrey Hinton and Yoshua Bengio, whose work deeply influenced my own understanding of how learning and reasoning emerge in distributed systems.
After roles as a tenured University professor, at Google Brain and Mila, I co-founded Pathway, where I now serve as Chief Technology Officer. Co-founding Pathway allowed me to bring research ideas out of the lab and have a tangible impact on the world. My focus is on turning research breakthroughs into scalable AI systems that can reason, adapt, and learn continuously in real-world environments.
What are you working on now?
At Pathway, we’re developing AI that thinks in real time and systems that evolve with experience instead of remaining static after training. Our latest research introduced Baby Dragon Hatchling (BDH), a new “post-Transformer” architecture that bridges artificial and biological intelligence.
BDH is designed to generalize over time, learning continuously, sustaining long chains of reasoning, and adapting safely in dynamic environments. We see it as a foundational step toward AI that can understand, reason, and self-improve like humans - a model for living intelligence. Our technology is already being used in real-world applications, including NATO’s adaptive planning systems and La Poste’s live logistics optimization.
BDH’s brain-inspired design allows it to scale efficiently, with inductive biases aligned to how biological systems learn. We’ve confirmed its properties on models comparable in scale to early LLMs, such as GPT-2. We’re now developing production-grade reasoning models based on BDH. Together with partners, such as AWS and Nvidia, we need to put them in the hands of our enterprise design partners as soon as possible.
Tell us a bit about your time at Mila.
I had a very productive time at Mila. The atmosphere was very inspiring: everyone was working on something important and it was easy to discuss ideas - at tea-talks, informally in smaller groups or at lunches. Mila fostered a sense of camaraderie that helped everyone achieve great results through collaboration and open exchange of ideas.
My Mila project I am mostly proud of is the application to the attention mechanism, developed for translation, to speech recognition - I started working with Dzmitry Bahdanau and Kyunghyn Cho on replicating their result on speech and it soon became the start of a new collaboration in which we have scaled the networks from working on sentences of a few dozen words to speech samples with thousands of frames. It was possible because Mila fostered a culture of trust and mutual help. I was trying to replicate this culture in all my future teams.
What advice would you give to new Mila students, especially those interested in launching their own startups?
Startups and research have one thing in common - it is very important to choose what you work on. The main difference is that in research, we, members of the scientific community, are the main recipients of our work. It takes a bit more empathy to understand what truly matters and how to create value for others. In startups, we must impress our users, the users of our products that will be different from us. Thus it takes a bit more empathy to understand what matters and what is important to create value.
What is the biggest challenge that you think AI faces today?
AI is the great interpolator - during training it indexed and internalized all information we have, collectively as humanity, produced and it is using it in constructive ways. However, new challenges will arise when AI will start to take part in our world and interact with it.
One interesting question is will AI learn on its own, by interacting with the world or with its own thoughts. Is it possible to get a “Ramamujan AI” able to see new patterns without first having to consume all the data that there is? I believe learning and knowledge discovery from little data is the next frontier: in fact new training data is very scarce.
Brain-inspired architectures like BDH are a step in this direction, with similar inductive biases as the brain, we will get closer to the brain's data efficiency. Learning from amounts of data comparable to humans, of course, requires from us a number of other steps linked to training and reasoning, that we are working on.
What are you looking forward to as AI continues to develop?
I’m most excited about the shift from systems that imitate intelligence to those that can actually develop it - AI that grows, reasons, and learns continuously, much like we do. As these systems become more context-aware and capable of lifelong learning, I see them unlocking new forms of creativity and discovery rather than just automation. Imagine AI tools that help scientists uncover new ways to produce clean energy, or accelerate breakthroughs in sustainability. That’s the kind of progress I’m eager to see. AI not as a replacement for human intelligence, but as a catalyst that amplifies it.
