Bengio has pivoted his research to explore the risks posed by AI.Credit: Joel Saget/AFP via Getty
Yoshua Bengio is a computer scientist at the University of Montreal in Canada. In 2019, he won an A. M. Turing Award — considered the most prestigious honour in computer science — for pioneering the ‘deep learning’ techniques that are now making artificial intelligence (AI) ubiquitous. Last month, he also became the first person to top 1 million citations on Google Scholar.
Bengio has since turned his focus to exploring the risks posed by AI. He chairs an international panel of advisors in this field, which includes representatives from 30 countries, the European Union, the OECD and the United Nations. The group issued the International scientific report on the safety of advanced AI earlier this year.
Nature met with Bengio in London to talk about the potential and pitfalls of the technology he helped to invent. The following is an edited version of the conversation.
Among the many papers that you’ve written, are there any that you’re particularly proud of?
The string of papers that I co-authored on language modelling and attention — that started in the late ‘90s — on how we could introduce attention mechanisms in neural nets to make them more ‘system 2’, meaning more deliberate, and not just intuition machines.
Is that the attention technique at the heart of a 2017 paper1 from Google researchers that introduced transformers — the technique that became the ‘T’ in ChatGPT?
Yes, but I would mention also another paper, which doesn’t get nearly as much attention — and that's the work on curriculum learning2, in which a machine is trained by feeding it data in a particular order rather than at random. It has become the standard way of doing things. The inspiration for me was learning in animals.
The existential risk posed by uncontrollable AI was not at the top of your worries until a few years ago. What has changed?
When ChatGPT came out, in November 2022. It took me two or three months to realize we were on a path that could be extremely dangerous. And although I was initially pleased to see that deep learning had finally reached that milestone, I realized that because of the nature of these systems, we didn't know how to make sure they would behave in the ways that we want.
... continue reading