The next phase of AI will need to draw on deep science to enable new advances

Artificial intelligence has advanced with extraordinary speed. In recent years, powerful AI systems have transformed sectors ranging from software development and finance to logistics and scientific research.

Much of this progress has been driven by scale: larger models, larger datasets, larger computing infrastructure, and unprecedented computing power. These systems have demonstrated remarkable capabilities, and the rapid adoption of AI across industry and daily life is a staggering achievement.

But that same scale has also revealed the limitations of current AI architectures. Concerns are mounting around energy consumption, infrastructure, interpretability, reliability, security, safety, and cost. Frontier models now require enormous amounts of computational power to train and operate, raising questions about long-term sustainability. At the same time, many AI systems remain difficult to fully understand, predict, or validate internally, which is particularly concerning in high-stakes domains such as healthcare, finance, and scientific research.

The next phase of AI development must address these concerns, rather than simply scaling up existing systems. We must develop fundamentally new approaches to artificial intelligence.

Historically, major technological advances have emerged from deeper scientific understanding, rather than incremental scaling and optimization efforts. Physics, in particular, has repeatedly reshaped how humanity understands computing, energy efficiency, and complex systems. From Geoffrey Hinton’s foundational work in neural networks to Art McDonald’s Nobel-winning discovery that neutrinos have mass – a finding that rewrote the Standard Model of particle physics – Canadian scientists have a strong legacy in making such advances a reality. Statistical mechanics, information theory, and quantum mechanics have all contributed foundational ideas that later transformed technology and industry. Today, many of the central challenges emerging in advanced AI resemble the kinds of problems physicists have long studied.

Modern AI models are extraordinarily complex dynamical systems. Understanding them more deeply requires new theoretical frameworks that extend beyond current engineering approaches. Foundational scientific research is becoming central to the future of AI.

Across the global research ecosystem, physicists, computer scientists, mathematicians, and information theorists are beginning to explore new approaches to machine intelligence that prioritize efficiency, robustness, interpretability, and specialization rather than scale alone. Future AI systems need to become more targeted, more energy efficient, more scientifically grounded, and better suited to specific high-value applications.

Physics offers something uniquely valuable on this front: a way to understand what is happening under the hood of AI systems. As researchers such as Yoshua Bengio advance more interpretable and trustworthy AI, foundational questions remain about how complex models learn, generalize, and fail. Physics has a long history of uncovering hidden structure beneath apparent complexity. Applied to AI, that same lens will reveal the principles needed to build systems that are more efficient, robust, and understandable by design. Independent institutes such as Perimeter Institute for Theoretical Physics are uniquely suited to pursue this work, stepping back from commercial pressures to explore the fundamental science of machine intelligence. That is our purpose: asking the fundamental questions that drive future discovery.

Canada is well positioned to contribute to this next phase of development. The country has internationally recognized strengths in artificial intelligence, theoretical physics, mathematics, and quantum information science. Canada’s independent research institutes are globally connected scientific centres that attract, train, and retain the essential talent needed to work across disciplinary boundaries, and they can make connections that commercial settings are unable to prioritize.

That interdisciplinary capacity matters, because some of the most important advances in AI will likely emerge at the intersection of fields that were previously considered separate. 

Commercial AI development will continue to play an essential role in bringing powerful tools to market. But maintaining leadership in the field will also require sustained investment in upstream scientific research to expand the boundaries of what future AI systems can ultimately become. The next generation of AI will depend not only on building larger models, but on understanding intelligence, complexity, and information more deeply than before.