So, hello everyone and welcome. I'm joined by our lovely panel over here today and I think we should just jump straight into it. So, I wanted to first start off with a question, a general question addressed at each and every one of you. So, each of you grew up before AI was part of everyday life, but for me and for many other young people, it almost feels as if AI has always been around. So, before we get into those big questions, I want to hear in those terms, what's one thing about AI today that would have been unimaginable to you guys at my age?
Well, in my time, I was actually thinking about AI a lot as an adolescent because I was reading a lot of science fiction. And then when I started my grad studies, I got really into the science of it. But the one thing I would not have imagined is that we would get where we are today so quickly. I thought it would take centuries.
I think for me it's really access to knowledge. When I grew up, we didn't have internet, so it was even, you know, a book or someone you would ask when you had a question. And then I went to Stanford for grad school and I remember thinking the answer to any question I might have is within a 5-mile radius from where I live, which sounded incredible. And now with our large language models, that answer is actually at your fingertips most of the time. So, it's actually just been as Yoshua was saying, incredibly fast.
Thank you, Yoshua. And you, Yoshua?
I think that AI provides huge opportunities for younger generations. In Japan, the society is aging. So, there is a huge need for younger students, younger generation to work in the society, but AI can help them. AI can empower them. So, I think it's a huge opportunity.
Yeah, I think on one hand, if you think about when you do programming or you're learning how to code, it's learning that you can actually manipulate the world in different ways and improve by building tools. I think one thing I never really imagined about today, even though I grew up here I was already there just in different ways, was that it was just going to decrease that distance between people looking at the world, looking at a problem, and then trying to think about how to solve it, and now you can use AI itself as a tool to help you do that. So, that's magic. That feels like magic.
Thank you, everyone. So, the next thing I wanted to ask you guys, I think we'll start with Joshua. So, it's kind of to do with speed and safety. So, you've helped build modern AI and now you're one of its most prominent critics, warning that we're maybe moving too fast. Young people, as I, we've grown up with AI moving at exactly that pace. So, from where they sit, is speed actually the danger or is it something else?
It's not just speed. I mean, if you are driving a very safe car on a safe road very fast, it's fine. The problem is we don't have headlights. We don't know exactly where we're going and we're not even sure the car is okay. So, it's about the other factor that matters is incentives. Like what is it that is driving this car? What are the reasons why we're developing AI in this particular way and I'm afraid that the incentive structure right now, profit isn't exactly aligned with the well-being of humanity.
Um, and Joel, so for young people building their lives around these tools right now, does building responsibly look different to them than it does from inside DeepMind?
Yeah, this is a great question and first, I want to congratulate you because, as part of those young people, you've already done so much. I think we might be, within DeepMind, maybe thinking like your parents. Many of us have kids, so when we think about the young generation, we really think about our kids. And I do think that the responsibility frameworks that we put in place really have the user in mind and we're really trying our best to make sure that we measure, we evaluate those systems for a purpose for which they've been built. They have general capabilities and very often, in health care we talk about off-label use when a physician is using a device not in the way that it's been designed for. And that has a lot of very specific regulatory frameworks around it. But of course, in other fields, you don't have necessarily as many of those regulatory frameworks. And I think it really depends on what the purpose is for you using a tool. Are you using it to learn? Are you using it because you have a health-related question? And I think that has to be very, very intentional. The youth doesn't necessarily realize that; they're just using it because it's convenient. So we really have to be thoughtful in how we design and deploy those systems.
Yeah, thank you. Um, and I'll go to Bukowski first and then to Yutaka. So, does this tension between speed and safety look different for the young people in your countries?
Sure. I think for all systems, AI or technical, context changes. So one place it might just be that you designed it there with the people there and it works and it looks like 90% of the time it does the thing that it's supposed to. But then you could be in another place. Let's say you are somewhere in the Sahara Desert, and now your sensors don't work in the same way, and that's now information that's coming into an AI system. So you need to think about that. A lot of not just on the African continent or in South Africa, but you might be in Latin America or Southeast Asia, having low bandwidth means that the data that comes into AI systems might be slower and not really respond in the way that people are expecting. I work on language especially, and that's a thing where you can also miss cues that could lead to negative responses. The biggest thing in those areas is equipping people to be able to evaluate and assess if AI systems are fit for their purposes, and on the other side building up the capabilities that people can actually build or change systems in the way that they want so that they actually much better suit the populations that they're from.
Um and you, Tak, it's the same question to you.
Yeah, in Japan, the government is currently putting emphasis on physical AI. So many in the manufacturing companies, robotics companies are starting to invest into that technology. But as you know, those industries are very slow, very matured engineers work in the field, and they have a large stock of knowledge and skills developed for so many years. And they have to be fast. They have to quickly adopt the AI technology. So that is our challenge. And that's where the younger generation can come in. And maybe the collaboration between the younger generation and the mature engineers will be very important. And also the safety, the AI governance is very, very important for physical AI because it can affect physically directly to us. So with that respect, we have to think more about AI governance.
Thank you. So, I also wanted to touch a bit on mental health and dependency. Jow, if we could go to you for this. So, you've raised concerns about how young people are forming real dependencies on AI systems, the same way as what we're seeing with social media, and about those tools being built to agree with young people and tell them what they want to hear maybe instead of what might actually be useful. So, do you think that's already affecting young people's mental health or is it too early to tell?
No, absolutely. And I think this is something we touched on in the UN report that we talked about earlier this week in the global dialogue. Now one in four conversations that the world is having with chatbot touches on health, mental health, or wellness. And all of us have been interacting with those systems. They use the same language that we use to talk to our friends and colleagues. And indeed, they are still authentic. They are trained to continue the conversation and to somehow please you. The goal is that you as a user come out with a good experience as all the products we build. People design them so that we're happy with them. And I think the hard part is that it's not always good for us. Having something that we like doesn't necessarily mean that this is the best interaction for us as human beings. And I think that's really the challenge that the industry is facing in designing the systems. The exact same systems that can actually help with mental health issues. I think there were some other sessions earlier today speaking about that. Are there systems that can also be incredibly harmful when people having discussions with them for hours at a time and end up in bad places. And so the industry has been trying to put safeguards in place and bring them to hotlines for example, but there's still a lot of work to do before this can be done in a way that is beneficial for everyone.
I think this is a really important question, but it's important both because of the impact it has on young people and vulnerable people in general, but also because it touches on something central in how those systems are trained. And in a good part of the training process, they're actually trained to please human labelers. Now, sounds okay like we want to be pleased, but the problem is pleasing someone sometimes could mean lying to that person. Or not saying something that they should know, for example, in the case of someone who has psychological depression. And we really need to have more researchers involved in figuring out why this is happening and how we can fix the way that the systems are trained so that we can avoid this kind of dishonesty that comes with pleasing people.
Yeah, I 100% agree with AI trying to please instead of telling what's actually might be useful to someone.
