In the latest episode of First Principles, Fundamental's Founding Advisor Marta Garnelo sits down with Razvan Pascanu, a researcher at Google DeepMind whose work spans optimization, learning dynamics, continual learning, and neural network architectures. Their conversation explores one of the central challenges in modern AI: how can models continuously learn and adapt to new information without forgetting what they've already learned? Razvan shares insights from his pioneering work in continual learning, examining the trade-offs between retaining knowledge, adapting to changing environments, and maintaining the ability to learn over time.
The discussion also explores out-of-distribution generalization, the connections between continual learning and meta-learning, and the challenges of applying these ideas to today's large language models. From the computational costs of different learning approaches to the role of model architecture and optimization, Razvan offers a perspective on what it takes to build AI systems that can generalize beyond their training data. Together, they unpack the fundamental questions shaping the next generation of more adaptive, efficient, and capable AI systems.
By Fundamental
@Fundamental
Research
Interviews
