AI is the tech revolution of our time, and with it comes trillions of dollars of economic value. Or so they say.
The thing is, according to the United Nations, more than a hundred countries around the world are lagging way behind. They're not even close to adoption, and many of them are in the Global South.
The other thing as well is that within the U.S., 30% of those using the internet are using it at speeds not even fast enough for Zoom.
So essentially, billions of people around the world are at risk of being left behind in this AI revolution. So what gives? How do you bridge that gap, and how do you bridge that gap fast enough? Because this is a technology that's moving really fast.
Lila, let's start with you. How do you get access to people, as many people as possible, to do it at scale? Starting with the easy questions. We have 20 minutes.
You know, this is actually something that is really important. We have to make sure as we're building AI that we're thinking about the global implications of this technology. Part of this means how do you develop the models as well with the world in mind? It's not something that we can think about as an afterthought. For example, one of the activities we've been doing here in Singapore is working with AI Singapore to ensure that Southeast Asian languages are well-represented.
And that means taking one of our open source models and fine-tuning it, you know, getting the data right first, then fine-tuning the model and making it available to a broader audience. To me, this is, you know, AI is not a technology where we have national borders. You launch a model and it's available worldwide.
So we have to think about everything from access to literacy to actually making the models easy to interact with, things like making it multimodal from the beginning, which we've done with Gemini, which is you can talk to it, you can type it, you can take a picture. And being able to do that in many different languages, being able to transform that output. So you're lowering the barrier for people in general to use it.
There's still a lot of work to do, but we're, you know, we're pleased to be making this progress. And it's only happening because we're doing it in collaboration with local ecosystems that are helping with that, like last-mile access.
But massive barriers, right? Digital infrastructure, AI infrastructure, key compute capacity. Those things will decide and shape who really gets full access to AI.
Well, I think if you characterize it as AI haves and have-nots, very few countries in the world will be able to have all the elements of an AI ecosystem that are desirable, that are good. It's a given. It's a reality. I don't think it's going to change.
But I also think we are not without agency within each of our own countries. We do have the ability to reshape things, and some of the areas that are worthwhile reshaping in order to be more ready for an AI future is investing in infrastructure, making it happen. You may not be able to do it on your own, but there are organizations like the World Bank, IMF, even the United Nations Development Program that are very alive to this problem and leaning forward to help.
At the same time, you also want to think about creating the right conditions for the technology to be used in a responsible way. So at the right time, putting in place guardrails is something that every country can think about doing.
But to my mind, the single most important area of intervention, of worthwhile effort, and where you must exercise the most agency is really in developing capabilities. And the capabilities have to be thought of in a holistic way. Capabilities in the workforce, capabilities in enterprises, capabilities within academia, actually even capabilities in AI governance within government itself, and different kinds of capabilities, entrepreneurial capabilities, engineering capabilities. There is a whole range of things that need to be built up in order for a country to enable its citizens to benefit from AI and to have the best chance of moving ahead.
Will you always get to what the next country is able to achieve? Maybe yes, maybe no. But that shouldn't hold us back from trying. We can try, and it is about access.
If you take a look at what we're hearing from the small and medium enterprises, 99% of the GDP in Singapore's economy, 70% of the workforce, they're saying that is just too expensive. The rising costs of GPUs, compute capacity. We know Google has taken that into consideration. How are you engaging with SMEs to counter the difficulties that they're going through?
Well, you know, within Google DeepMind, we're the AI innovation engine for Google products. So one of the things that we really are working on with our Google, our colleagues across Google, is how do we integrate our AI tools into existing programs? You know, in some ways I wish we could actually rebrand. I like it if it's a good feature and it's actually being useful, it's just part of an everyday workflow. So the first step I think is like actually getting it integrated and removing that barrier.
But there is something about reducing the barrier of like the literacy aspect. And part of this is it's been quite amazing to see in the past few years how many different resources are available in order to be able to use the technology. So one is like just getting it integrated. The second is removing the barriers, as we talked about in terms of just, you know, making the context window, which is the point, larger so that you can input a lot of different information in a natural way. And then the other one is making sure that the users actually have the skill to do it.
Well, I think SMEs are a very heterogeneous bunch and there is no one single way of characterizing them in terms of their adoption. Likewise, there will be different pathways. I'll tell you a story that was quite encouraging for me. This is a hole-in-the-wall along Killiney Road. They are a small eatery and I happened to chance upon them one day and I asked them how business was and we ended up talking about AI and they told me something quite unexpected.
They said AI has been helping them do better business. I said, how? He said, well, you know, I have to run promotions and I've been running promotions in the past year. Some are more successful, some are less successful. And as it turns out, I now go to the platform, which has an assistant I can use, and I can ask it based on my past sales record to tell me which of my past promotions have been more successful. And that has helped me tremendously to plan the next promotion.
The point of raising this example is to say that not every single person is going to need its own access to GPUs. In this case, it's not even GPUs, it's just an app that was made available to this little eatery and an AI tool that was made available through it. There will be other services where being able to use a fine-tuned model could be meaningful.
If you take the fact that Singapore's GDP still has about 20% contributed by manufacturing and many of these SMEs perhaps engaged in precision engineering work, we now have an AI centre of excellence for manufacturing that has recently developed a common model that does predictive maintenance for CNC machines, which many of these SMEs will use. So the point I'm saying is that really there will be different methods of us being able to support the SMEs.
Some will come through tools that are made available by bigger platforms. Some will be through efforts that are targeted at the sector, targeted at the kinds of companies in them. I think it's more important that we look at how we can make meaningful differences to the way the businesses are operated. It is still about levelling the playing field and there are different options out there.
If you take a look at China, it's open source. You know, they say innovation is far cheaper, and then you have closed. How do you weigh the pros and cons of the two? We understand the concerns over data security, cyber breaches, but there are pros to an open system.
Yeah, it's something that we have a long history in actually making the technology available from the early days of our research into open source models, whether it's something we call Gemma or even AlphaFold for the scientific community to make it broadly available. With over 190 countries, over 3 million researchers using this to advance a number of fields, whether it's in food security, pollution, dealing with industrial waste, or even understanding diseases to advance therapeutics. And in many ways that in and of itself is lowering barriers because now it's democratizing access to this information.
But we also have to think about every case is different. So this balance of how can we seek the benefits while also managing the risks. And as the large language models and the multimodal models get more capable, this introduces a new level of risk. So when we released AlphaFold, our protein prediction model, we actually sought the outside expertise of scientists, academics, Nobel laureates in this space to say, what are the risks? We've assessed it this way. What do you think? What's the right way to deploy the technology?
So it's kind of on a case-by-case basis. As AI gets more general and capable, this is something that we're doing extra testing. So we've just released today Gemini 3, which is not only our state-of-the-art, most advanced model, but also the most tested. And I think that's part of the pace of the technology right now, because we still have issues with factuality, reasoning, understanding the context. These are things that are not going to be solved right away. So it's this balance of how do you make sure the technology is available so people can build off of it, they can advance their businesses, their economic opportunity, learning, etc., while still managing the risks.
I think if you are talking about widespread adoption, having access to open models are very important. And you think about the thousands and perhaps in time to come millions of applications that a very wide variety of organizations are going to have to use and individuals are going to have to use. Without access to open models, I think that would be quite problematic.
On the other hand, if you're talking about a financial institution, if you're talking about an organization that has a lot of intellectual property or security concerns, expecting them to only rely on open models is clearly not going to be feasible. So I think the future will see a mix of these different kinds of access be made available, and we must do our part to try and ensure that access remains open as widely as possible if we want to see widespread adoption.
I'd like to touch on education, but before that there is a poll. I would love for everybody to participate. The question is: how will AI most exacerbate global inequality? A) mass job displacement, B) concentration of power in a handful of companies and countries, C) insatiable demand for energy, and D) widening capability gap between advanced and developing economies.
While you decide on your choice, let's talk about education. Minister, we talk about how you must prepare the next generation. How are you thinking about the possibility of the elite institutions of wealthy countries being the next, I guess, source of talent? Is there risk of that happening?
Well, there's definitely going to be competition to try and get as much of an advantage as possible in terms of raising the quality of human capital in each of our countries. And I don't think you can stop anyone from doing that. This sort of effort will take place with or without anyone, you know, putting a stop to it or trying to boost it. I think what people can attempt to do is to ensure that they themselves have the ability to learn.
And I think that in our education system, we do want to make sure that children have exposure at a young age, but in an appropriate way. The way AI is being used in our education system, I think, is being experimented upon. At one level, it was about enabling the teachers to be more productive, to be able to outsource, as it were, the more mundane aspects of their work, perhaps administrative work, to an AI assistant. That is not in and of itself harmful.
But it also kind of misses the point of, you know, using AI for education. The great value of AI is in personalizing learning in a way that attends to the unique strengths and shortcomings of every learner and trying to find ways to enable them to achieve better learning outcomes. So I'm very glad to see that my colleague, the Minister for Education, and I have had deep conversations about this, and our Ministry of Education is looking at it not just as a way to support teachers, but more as a way to support learning. That is the direction I think we all should try and move towards.
Now, is it something that every education system is able to do? I think we were just talking about this morning the fact that Google DeepMind has set up a research lab in Singapore and one of the potential areas for collaboration is identifying ways in which learning and education can be made more effective. And since we said in Singapore, AI for the public good for Singapore and the world, that being our vision, we hope that the benefits don't just accrue to students in Singapore, but potentially to students elsewhere.
And to that point, how do you cater to the underserved? I mean, you're an advocate for education and reskilling and skilling among such communities. What's your take on that?
Well, education is something that's really important to me. And just to comment on a few things there. I think that you have in many ways, starting with teachers, this productivity we're seeing around the world that teachers are really struggling to keep up with the workload as class sizes may get larger, as different skills may emerge within students. We just did a pilot, for example, in Northern Ireland, where with the government there and 100 teachers for six months, and they found on average they reported 10 hours of savings per week per teacher. That's pretty significant.
And what does that mean? It's not just about saving time, but it's actually letting the teacher do what they wanted to do in the first place, which is transfer their knowledge and passion for a topic to the students, which ultimately drives learning outcomes. So I think if we can come at this AI and learning from a different perspective, which is, yes, you have to start with the teachers, but the teachers are the magic in the classroom. So how do you actually free them up and then give them the tools to help with the learning outcomes and also meet the learners where they're at in their learning.
We all learn differently, whether you're a visual learner or you prefer to listen to podcasts or learn. So I think now that we have AI where it's again multimodal from the ground up, where you can now interact with content in a different way, you can learn about physics by learning about your favorite sport in that context, or instead of just giving an answer, being able to go multi-turn with the model. This is the type of technology that we think is really important.
It all starts with treating learning like a first-class science problem, right? It's not just about physics or biology. It's like, which at Google DeepMind we tend to focus on, it's actually investing in pedagogy and knowing that we come at it from an AI expertise. So who are the learning experts that we can go and talk with and make sure that our models and our AI development is grounded in this? We need to make the AI more factual. We need to make sure that it has the best of learning science built into it.
But we also need to work with the community because at the end of the day, this isn't about just layering AI on top of the existing education infrastructure. It's also rethinking about how are we going to train the next generation and how are we going to ensure that they have this as a tool that they use responsibly. They're not just offloading the learning but really engaging with it. So I think this is an area where one is we're still very early, two is it's extraordinarily high priority.
And then we think, we really believe that here in Singapore, especially the extraordinary history of workforce training, of lifelong learning, of education as an essential value of the culture, is really critical. That's an area we're really looking forward to collaborating more in terms of exposure.
