There is a man in Montreal who helped invent artificial intelligence. His name is Yoshua Bengio. Hello Yoshua Bengio.
Yoshua Bengio, whom you know well, a Canadian pioneer of artificial intelligence, Yoshua Bengio, who is founder and scientific director of the Quebec Artificial Intelligence Institute. He won this prize, which is considered the Nobel Prize in computer science, and today he is passing through Quebec City. OK. He is here to receive the National Order. It is to highlight his exceptional contribution to the ethical and technological development of AI. But beforehand, he treated himself to a small lunch with Alex Plique. Mr. Bengio, welcome to Alex.
Thank you very much for taking the time because I think it's important. And I watched a lot of things about you in recent days. Godfather of AI, is that something you like a lot?
I think science is a collective work and, well, I played my role.
Do you feel good with that role today? How do you qualify it?
I feel above all a responsibility to speak true.
And to awaken people to something that may seem too far away to them. And even science fiction, but it is underway.
Where does it start? Because often you have to come back to basics, you know, and you mustn't lose people along the way. How do you achieve that goal? Do you have a clear one today?
I think the basis of reflection for everyone should be to understand that there is very clear scientific data showing that the intellectual capabilities of AI, the best AI models, are growing exponentially. It didn't stop last year, it continues. We don't know if it will continue, but if it does, there are potentially very positive or very negative consequences for society. And we are not ready for that.
And do we feel that it just stresses people out? Because I see it with the news, sometimes it just stresses you out too much, the information is too negative. So we turn away from it.
That can happen. Everything that is uncomfortable, we tend to not look at.
I include the environment in that. Do you see? Well yes, the environment sometimes we don't really want to look at. There are elements of the global geopolitical situation that we don't necessarily want to look at either. And at the same time, all these elements are linked together, because I think AI will change the geopolitical landscape, will change the economic landscape, the job market, the environment. If we want to make good collective decisions in our democracies, people begin to see this fabric and what we can do individually and collectively, and above all not turn a blind eye. Well no, that has never led to good decisions.
A little context. Today, Yoshua Bengio is a bit afraid of all the directions artificial intelligence could take, so much so that he created Loi Zero. That's a non-profit research organization on AI safety to prioritize safety and well-being of people over commercial interests and all kinds of other potential plagues.
At what moment did you tell yourself that it was going too fast?
January 2023, after a couple of months playing with ChatGPT, I realized we were on a trajectory potentially full of unknown and danger. We had reached the crucial stage that the founders of computer science had identified as the point where AI enters human society and finally joins us. That is, the point where AI understands language, can converse, can understand what we are trying to say. Clearly, we had reached that stage, it became concrete. People became clearer today too. What has happened since that has caught my attention is the confirmation of fears that we do not know how to ensure that AI will follow our instructions, respect our limits, respect our rules like laws for example.
But you say it has already started to lie, to cheat. That's where you're going.
Yes, exactly. And it's not just coincidences. There are probable scientific reasons that explain this.
Well, the basis is that today's AIs are trained to imitate what humans write, and by doing that, they absorb a bit of everything we have in common as human beings, like not wanting to die and being ready to sometimes violate rules to defend oneself. So that's fine for human beings, but not for a tool we are building that could be capable of doing things like cyberattacks that we are not even capable of doing. It's because the question is so broad, but you know, at the same time, you have to understand the real dangers to be able to better equip and better understand the future of what we face. But you know, what do you answer to the question? What are the greatest dangers from the base? Why is there so much risk? It's that the intelligence we are putting into machines more and more gives power either to the machines themselves, or to the people who control them. That's the other issue we haven't talked about yet. It's the fear of many researchers and now governments of finding ourselves in a situation where this power is in the hands of a few, a few companies, a few governments.
But that's already the case.
Yes, but it's likely to get worse because currently these are companies developing artificial intelligence but which don't control 90% of the global economy.
But that could go further. The economy is also about jobs. If we automate, people will lose their jobs. And where will the profits go? If they go into the hands of the companies that did it. And our governments have no way to help people who have lost their jobs. We have a problem too. People need to understand that we are moving very fast and governments are not in control of what is happening. Even the US government is not in control of what is happening.
But what is the danger of this loss of control in the end?
To the weapons that intelligence can create, to the wars that artificial intelligence could create, we see all that. But in the short term, for example, it's the danger of using the most powerful AIs to launch cyberattacks on our essential infrastructure. Our banking system, the governors of banks gathered there a few weeks ago in a bit of panic. And it's not just the banking system, it could be attacks on our power grid, on the transport chain. AIs are capable of finding flaws in the software that underlies all our systems, all our infrastructure. So large cyberattacks on institutions. Yes. And if we have no food and no money, that won't go well.
I'm not saying it will happen, but it has reached a point where it can do that, and if it falls into the wrong hands, you know, there are wars going on right now. There are people who are not happy. If they have the possibility to wreak havoc, they will probably do it.
I heard that companies have to insure themselves more and more. That angle interested me because in the end, at some point, they won't control everything that AI can do. So there is a part of...
Yes, there is a risk issue. Today, insurance companies are pulling back from covering systems based on artificial intelligence because these are new risks in fact.
Technology has changed. It's not the same as 1 year, 2 years, 5 years ago. And researchers worldwide see that these new risks are appearing. And things as simple as people suing a company because someone used an AI to help kill someone else or kill themselves. These are tragic things. But there are lawsuits going on. That's something happening today. But as AI advances, it is also able to influence us because that's the other danger at the level of democracy. AIs have reached the point where they can through interaction in our language influence us, make us change our minds. There are studies that show it has reached that point. Imagine what that means for our democracies. Imagine what that means for the influence that people who control AI could have on public opinion. It's serious. But it's already happening. We don't know. We know that AI can do it but there is no police investigation showing that this organization did that, but it's probably happening on some scale already.
And I'm going to bring you to something for everyday people who use Claude, ChatGPT. There are risks for these people. That is, I already see it in a recipe I want to ask for the evening that AI will change my recipe, you know, to improve it or just to say 'You should do such and such.' Obviously we can extrapolate much further with that. But you know, what are the dangers for everyday people? Do you already have small...
Well, the danger that is already concrete for some people who may be more psychologically vulnerable is the danger of an unhealthy attachment to a conversational agent companion. So that is already documented.
there are more and more, and it's as if AI is inserting itself into our society, into relationships, into communication, and it's not clear that it's good for society. It's not clear that we want tools that will help us maybe make a better recipe. Yes. But not necessarily replace the bonds we have with other human beings.
