The AI revolution is here

One is a Université de Montréal professor and winner of what has been called the “Nobel Prize for computing,” the other a Université de Montréal alumnus and head of AI research at Apple. Brothers Yoshua and Samy Bengio, who immigrated to Montreal with their parents as children, sat down to talk with us about current developments in artificial intelligence (AI) and what the future holds.

AI is developing at a dizzying pace. Did experts predict this?

Yoshua Bengio: No, not at all! Looking back at AI’s trajectory over the past 20 years, we’re amazed at how rapidly it has reached the levels of performance we see today. The systems were so rudimentary back then. Even the developers are astonished. What we are witnessing is a revolution comparable to the advent of electricity.

What are some of the advances you never thought possible?

Samy Bengio: In the past year, two technologies that have been around for a while started to perform spectacularly. One is image-generation models, which can learn the characteristics of a corpus of data and then generate similar data. The other is large language models that can help us communicate: to write, summarize, translate or edit texts.

Now the image-generation models are being combined with the large language models to convert text into images and video. This has led to all sorts of applications for generating images and creating images and video. By combining these capabilities with other recent advances, it’s now possible to generate good quality code, spoken word and music. But while these new developments are great, they also come with some risks.

What sort of risks?

Yoshua Bengio: There could be unanticipated consequences. The market is being flooded with more and more AI tools that can imitate voice, text, even the appearance of humans. Yes, they can be creative tools that help us in our work, but they can also be used for unsavoury purposes. I’m worried about them being used to destabilize democracies. There was already a problem with fake news before these tools came along. What’s going to happen if it becomes even easier to produce enormous quantities of fake content? I think we need to be very cautious.

For example, AI tools that can mimic the voice and image of real people were used to create a fake video of Vladimir Putin that was disseminated in Russia and influenced the war on the ground. The video was somewhat lacking from a technical standpoint. Just think what will be possible in a few months as these tools become even more sophisticated.

Samy Bengio: Yes, there are bad actors. But there are also other, more subtle, dangers. These models aren’t perfect and they can make things up—hallucinate. It’s not out of malice or intent to lie; it’s because they aren’t competent enough yet. At the same time, they make their answers sound very convincing. If people aren’t aware of this problem, they’ll assume everything these models say is true. And often it’s not, as one defence attorney learned the hard way after preparing a court filing using a large language model. It gave him case citations that seemed trustworthy but were completely fictitious. One could imagine this sort of thing happening in other situations too.

Yoshua Bengio: These models can make decisions that are wrong in the absence of perspective. They can present something as absolutely true when it is, in fact, false.

But isn’t AI research now making advances in many fields?

Yoshua Bengio: Yes, researchers in all fields have an enormous appetite for AI tailored to their specific needs. Almost all disciplines are taking advances in AI and translating them into tools that can perform tasks the human brain can’t. If I were to give you a massive dataset of experimental results containing millions of data points, you would not be able to absorb it. But a machine could. I expect scientific breakthroughs in the next decade thanks to developments in AI.

Have the applied sciences seen remarkable advances in AI?

Yoshua Bengio: Biology has seen extraordinary advances in our understanding of how cells work and in drug development. In medicine, AI has become very accurate and efficient at analyzing medical images. These days, I’m working a lot with biologists and chemists, and also with researchers interested in modeling the Earth’s climate, weather and evolution using satellite data for all sorts of applications.

Samy Bengio: One example is the deep learning software AlphaFold, which debuted two years ago. It can model the structure of any protein and is facilitating the drug discovery process. Previously, it would have taken a Ph.D. student years of research to identify the 3D structure of just one protein. AlphaFold can predict the structures of almost all known proteins, and tests show a good match to the results of lab experiments. This is a major breakthrough in molecular biology.

Can AI help tackle the major challenges of the 21^st century, such as climate change?

Yoshua Bengio: Artificial intelligence offers no ready-made solutions to the climate crisis or the other major challenges of the 21^st century. That said, there is an enormous amount of work being done in a wide range of disciplines in which AI’s modelling, predictive and decision-making capacities are especially useful. Traditional modeling is very costly and difficult to implement at scale because of the enormous computational power required. It is often impossible to broaden the scope of the results as much as we would like, so how could it be done on a planetary scale? This is where AI can be very useful.

In Montreal, there are many projects underway to address climate change. For example, we are working with BrainBox AI, a company that designs AI systems to reduce energy consumption in buildings, a major source of energy use in Quebec.

We have also contributed to a number of research projects on ways to optimize the use of renewable energy resources. For example, researchers are working on forecasting when solar and wind power will be available in order to better manage their use and save energy.

With colleagues at McGill University, we are developing new materials for energy storage. The goal is to make the transfer of electrical energy in the form of hydrogen and in carbon capture more efficient.

Samy Bengio: We must not forget that one of the major challenges of the 21^st century will be AI itself, and its consequences for society. New jobs will be created and old ones transformed. We need to make sure everyone has the tools they need to learn how to use these technologies. Currently, a very small number of countries benefit from advances in AI, while the rest of the planet feeds these data-hungry models but receives nothing in return. This is a problem on many levels. Yes, we may be able to use AI to help solve this enormous problem that we’ve created, but in the meantime we have to focus on education.

What do you mean by focusing on education?

Yoshua Bengio: First, we can help democratize knowledge about AI. Université de Montréal has a number of initiatives to bring more African students here for internships and graduate degrees in AI. Many of these students later return to their home countries and set up companies. We have an important role to play in improving access to knowledge about AI.

Samy Bengio: When it comes to getting a job, people who have learned how to use AI will have an advantage over those who haven’t, and this will be true around the world.

Yoshua Bengio: AI tools are being applied in more and more fields. People with dual expertise, who have an understanding of these tools and are experts in a field in which they are applied, will be in especially high demand. I believe that in a few years, we will need such people even more than AI researchers who develop basic algorithms. It’s never too late to go back to school and train in an AI specialty!