AI Research Symposium: The Next Frontiers | Keynotes by Demis Hassabis, Yoshua Bengio & Yann LeCun

Hallelujah.

I should have kept so many.

Good morning. Good morning everyone.

Good morning. Shri Ashwini Vaishnaw, Honorable Minister for Electronics and Information Technology. Sir Demis Hassabis, CEO of Google DeepMind. Shri Jitin Prasad, Honorable Minister of State for Electronics and IT. Shri S. Krishnan, Secretary MeitY. Mrs. Kavita Bhaia, COO India, Ministry of Electronics and IT. And distinguished guests, eminent researchers, ladies and gentlemen, a very warm welcome to all our dignitaries and participants. We welcome you to the research symposium on AI and its impact, the flagship event of the India AI Impact Summit 2026.

Now we begin the opening ceremony. It is my privilege to invite Shrimati Kavita Bhaia, Chief Operating Officer India, to deliver the welcome address.

Honorable Minister, Honorable Minister of State, Secretary MeitY, Sir Demis Hassabis, symposium chair and co-chair of the program committee of the research symposium, esteemed researchers, academia leaders, innovators, partners from India and across the world. A very warm welcome to all of you. It's my privilege to open the research symposium on AI and its impact, a flagship knowledge platform of the India AI Impact Summit. Over the past one year, as the preparations for the summit have unfolded, one conviction has guided our efforts: that meaningful conversations on AI must be anchored in research, evidence, and interdisciplinary exchange. The research symposium has therefore been designed not as a ceremonial gathering but as a working platform, one that brings together the people building AI systems, studying their implications, and shaping the real-world deployment.

Today's convening reflects months of institutional collaboration. We have worked closely with researchers and academia, in particular our knowledge partners IIIT Hyderabad, to curate a program that is both globally representative and substantially rigorous. The diversity of the participation we see today across geographies and sectors is a testament to the growing importance of such integrated research platforms.

Increasingly, meaningful progress is emerging at the intersection of communities: computer scientists and social scientists, engineers and ethicists, public institutions and private laboratories, startups and universities. By bringing these communities into a shared space, we hope to enable conversations that are technically grounded, societal aware, and implementation oriented.

Another important dimension of this platform is its emphasis on emerging research voices alongside globally recognized scholars. The symposium has created dedicated space for early career researchers, doctoral scholars, and student innovators, because institution building in AI will require not only nurturing ideas but also building ecosystems of mentorship and visibility.

From an institutional standpoint, convenings such as this play a catalyst role in surfacing frontier research questions, identifying partnership opportunities, and aligning innovation pathways with real-world needs. It's our hope that the conversations initiated here will extend well beyond today's session, evolving into a sustained institutional partnership.

As we begin the proceedings, I would like to express my sincere appreciation to all partner institutions, knowledge collaborators, speakers, and program curators who have contributed in shaping the symposium. Your participation is what transforms this platform from an event into a living research network.

As we inaugurate this symposium, we do so at a moment when India's AI journey is accelerating. The insights and partnerships that emerge from these discussions will play an important role. It's now my honor to invite the dignitaries to give their special address. Thank you.

Thank you, ma'am. Now to set the context for the sessions ahead, I invite Professor P.J. Narayanan, Professor IIT Hyderabad and chair of the symposium, to address the gathering.

Distinguished members who are sitting on the dais, Chairman, Secretary, Mr. Krishnan, and I would like to also recognize some special people sitting in the audience. Professor Raj Reddy is here, Turing Award winner, and so this is an academic gathering.

So when the summit was planned, MeitY or the organizers also wanted to have an academic symposium or research symposium in parallel with it, and I was contacted to see if I could coordinate this. So I agreed to be the co-chair, and I get to stand here and represent as co-chair, but there were so many people who helped me with very critical decisions. One of them is Kalika Bali sitting there. I mean, she's part of the India team. She did a lot more firefighting than I did, and of course Abhishek Singh, the additional secretary, who guided us in multiple ways. We think we are researchers, we think only in the researcher's way, but this is not only a research event, this entire summit, so how to put these things together has been very instrumental in doing that.

So AI has suddenly become so big. Even two years ago, if somebody had suggested that people working in AI or machine learning will win Nobel prizes, everybody would have laughed, and we got two in the same year. And even the Turing Awards, you know, the first initial pioneers, four or five pioneers of computer science got Turing awards, I mean of AI, got Turing awards back in the 60s and 70s. After that, about 20 years there was a gap in even Turing awards being given to AI work, and Raj Reddy and Professor Feigenbaum were awarded Turing Award in '94 for their work in AI. So even the computer science community was not recognizing it as much. Of course, now it has changed, you have many more Turing awards in the recent years.

So when we started planning this, we didn't know where to go. There are so many things happening in AI. How do we bring everything together? And we tried to get as much of it together as possible, and I'm very, very pleased to have four stellar individuals doing keynotes. Demis Hassabis himself, who uses AI or ML techniques to advance science, mostly where humanity benefits unquestionably.

And we will have after that Dame Wendy Hall. She's been working in web sciences and looking at the progress of how web and this big connected world enabled this whole thing in some way. AI wouldn't have been there without this data and this adoption. Then we have Yoshua Bengio, who has been on the safety side, cautioning everybody of the potential dangers of AI, and I think they are very real. We need to listen to him. And we also have Yann LeCun, who probably takes the other side, who is very, very gung-ho about AI's potential. So I think we need to listen to all the voices. I'm very glad that all four kind of voices are heard here in terms of plenaries.

Then we have what we call research dialogues. One is on how AI can be a catalyst in the discovery of science and engineering progress and how we can increase the impact of that. The other one is on AI itself. What next? What is the next frontier for AI research? Where is AI going? Then we have two international panels with a global south concentration. One of them is on the global south itself. What is the impact on the global south? The other one is on AI safety, safe and trusted AI. These are the two parallel sessions that will happen.

And then we have invited several top publishers, authors who have published influential work in AI from across the world. People have come from all over the world to present. These will be in posters during lunchtime and right after the session. So we have a fantastic program ahead, and this should spur dialogue and discussion in topics related to AI. And I was very pleased to see the large crowds for the last two days, and I'm sure everybody is pleased to see a full hall like this here. Everybody is eager to know about AI. Everybody is eager to know how it will affect them. So we'll have a fantastic session today. I'm thankful for this opportunity to curate the session and bring it in front of you. Enjoy the session. Thank you very much.

Thank you, sir. May I now invite our Honorable Minister for Electronics and IT, Shri Ashwini Vaishnaw, to deliver the special address.

Very good morning, friends. This is a research symposium, and I'm not at all qualified to be standing here. I wouldn't take too much of your time. Just want to thank you for coming out in such large numbers at 8:30 in the morning. Yesterday we had about 250,000 people, mostly young. The average age was, I think, below 30, attending the AI Expo and the entire exhibition area. It was phenomenal response. When I interacted with the young minds, I was so surprised at the optimism that most of the young people expressed towards this opportunity which is coming for them.

Initially I was a bit apprehensive, but looking at their optimism, I'm feeling really hopeful for a totally new future for our country and for the world. We in India are very focused on AI at the edge, AI for use cases, AI for solving real-world problems, AI for improving the productivity in the enterprises, AI for population-scale problems like healthcare, like agriculture, like climate change. These are the things that we are focused on here in India, and this AI summit brings that opportunity.

And as we discuss various technical matters, I will request the topmost brains which are present here to please do bring some solid, concrete suggestions on how to make AI safe. This is a great tool. It should be used for the benefit of humans. It shouldn't be something which really distorts that balance. I really look forward to the suggestions and ideas which will come out of the symposium. Thank you very much for coming to this symposium in such large numbers, and I wish you all the best. Thank you.

Thank you, sir. With this, we formally conclude the opening ceremony of the research symposium on AI and its impact. Thank you, sir, and thank you everyone.

Yeah, you come here.

My game.

Mike.

Hello everyone. My name is Mausam. I'm a professor of computer science at IIT Delhi and also one of the co-organizers of the research symposium. I welcome everyone to the symposium, and in fact, we are all looking forward to it. And the rain gods have also showered their blessings today, so we know that this is going to be a great symposium.

It is my distinct honor to welcome Sir Demis Hassabis, the Fellow of the Royal Society, a Nobel laureate, and the CEO and co-founder of Google DeepMind, for the keynote address. And with him on the stage would be the moderator, our own Professor Balaraman, who is himself a Triple AI Fellow and Head, Department of Data Science and Artificial Intelligence at IIT Madras. So I request you to please join us on stage and deliver your keynote address. Thank you.

Thank you, ministers, and Government of India, and our fellow professors for inviting me here to give this keynote. It's a huge honor to be here. Very excited to see the incredible turnout from the country and the interest in AI, of course, but also from many of my colleagues from around the world in both industry and academia who've convened here. So I'm looking forward to a really amazing couple of days of discussions about the way forward for AI.

When we started DeepMind in 2010, it seems it's only 16 years ago now, but it's almost a lifetime in terms of AI. Almost nobody was working on AI back then, few people in academia, but almost nobody in industry. So it's been amazing to see the progress we've made in just over a decade. But at this point, I think we're at a threshold moment where AGI, artificial general intelligence, is on the horizon, maybe in the next five to eight years.

So this summit comes at a critical moment as we start seeing more autonomous, agentic AI systems, much more capable. I think we're seeing the beginnings of what those kinds of systems can do. And of course, the opportunities are incredible. My personal passion is using AI for science and to advance science and medicine with systems like AlphaFold. I think we can revolutionize drug discovery and human health, and also areas like material science and climate change with the help of AI.

But of course, it also comes with many risks too. AI is a dual-purpose technology. It's going to be one of the most, if not the most, transformative technology in human history. And that's just on the horizon now. And the thing about this technology is it's going to affect the whole world, everybody in the world. It's going to cross borders. So it's very important that we have summits like this, international summits, to bring the international community together to discuss how to make sure the opportunities benefit the whole world, and also how do we mitigate the risks. And I think in order to mitigate some of the risks, it's going to need international cooperation. And the first step for that is to have international dialogue. And I think that's what these summits, starting in the UK and then in Paris and then in Korea and now here in India, are for and have been really helpful in convening many of the important heads of state and many of the important leaders in this technology together to discuss what is on the horizon. So with that, thank you for having me, and I look forward to the discussions here in the next couple of days. Thank you.

Yeah. So good morning everyone. So you started talking about how amazed you are at the progress that AI has made, and you started DeepMind in 2010 with a vision to solve intelligence. So are we there yet?

No, I don't think we're there yet. I have quite a high bar for if we talk about solve intelligence, meaning building artificial general intelligence. Then, you know, my measure for that is having a system that can exhibit all the cognitive capabilities humans can, including in creativity, long-term planning, things like that. And today's systems are very impressive, but as we know, they still have many flaws and things they can't do, including consistency across different tasks that you would want from a general system.

So you did mention AGI, right? So what do you think really that we can do to take where we are right now and move it towards artificial general intelligence as you define it?

Well, when I look at the current systems and what's missing from them being a kind of general intelligence, I would say things like continual learning. So learning after they've been trained and put out into the world. Today's systems, we train them, we do various different types of training on them, and then they're kind of frozen and then put out into the world. But what you'd like is for those systems to continually learn online from experience, to learn from the context they're in, maybe get personalized to the situation and the task that you have for them. And today's systems don't do that.

Also, they have difficulty in things like doing long-term coherent plans. They can plan over the short term, but over the longer term, like the way that we can plan over years, they don't really have that capability at the moment. But I think probably one of the biggest issues is what I would call consistency. So today's systems, they're kind of like jagged intelligences. They're very good at certain things, but they're very poor at other things, including sometimes the same thing. So for example, today's systems can get gold medals in the International Math Olympiad, really hard problems, but sometimes can still make mistakes on elementary maths if you pose the question in a certain way. A true general intelligent system shouldn't have that kind of jaggedness.

But aren't humans jagged as well?

No, not in that way. I don't think so. So I think if you're an expert in maths, and many of you in the room are, or in physics or something like that, you're not going to sort of make a mistake on a strictly easier, more trivial problem, right? You would have some way to work through that problem. Similarly, you know, let's take the example of games, which is very close to my heart, and we've used in the early days of DeepMind to train our AI systems. You know, today's general foundation models are very poor at, for example, playing chess, you know, below kind of weak amateur level, and that is not what you would expect from an AGI system.

There are other tests too that I would have, is true creativity. So AI is very useful as a scientific tool, an assistant, or for solving specialized areas like AlphaFold does with protein structures. But really, what separates great scientists from good scientists is their creativity and their sort of, maybe you could call it, taste for what's the right question, what's the right hypothesis. And it's much harder to come up with the right question and the right hypothesis than it is to solve a conjecture. And you know, I would call that the highest level of sort of creativity. And so far, today's systems, you know, they don't have that capability. You know, one way we maybe would test that is you could imagine training a foundation model with a knowledge cutoff of something like 1911, and then see if it could come up with general relativity like Einstein did in 1915. But that would be a good test, I think, for AGI, and I think today's systems clearly would not be capable of doing that. But I think this will be solved in time.

Okay, that's great. So you talked about a scientific tool, this is a tool for scientific discovery, right? And so we see a lot of, you know, headlines about AI, you know, making great breakthroughs in science sometimes, but you know, how would this become the norm, right? I mean, so do you see AI becoming the scientific tool for discovery soon?

Yeah, the reason I spent my whole life and career working on AI is I saw quite early on that if we could build these kind of general models that were good at pattern recognition, they would be incredibly useful scientific tools, maybe the ultimate tool for science. That's what I was thinking. A lot of science is about finding insights and structure in vast amounts of data, that's perfect for AI. And so I think we're going to enter in the next probably 10 years this new golden era for scientific discovery, almost a new renaissance, using these incredible tools like AlphaFold. But I hope that will be the first of many that can massively speed up our research and accelerate scientific discovery across almost any subject area. So I think that's going to be the next period, is using these systems as tools, and then after that, we'll see as they become more autonomous, can they be kind of like co-scientists with you, like a PhD student, something like that. I think we're still quite away from that, but maybe in 10 plus years that will be possible.

So just to reassure a good fraction of the audience, you still see a role for humans in that future, right?

Yes, I think, you know, the next phase is going to be incredible, I think, for human experts and scientists, the amount of work they'll be able to do. And I'm actually really excited about cross-disciplinary science, which is quite hard because you have to understand, you know, more than one subject area, maybe two, three, four subject areas, and then find some interesting connections between that. I actually think that's where a lot of the really valuable advances are going to happen in the next few years, these combinations of subject areas. And I think having a tool like AI will really help scientists learn about and be able to understand and process all of that information from, you know, multiple different domains.

So in some sense, science is a well-defined field, if you will. I mean, you can recognize success and so on and so forth. So what do you think would be the role of AI in more abstract domains like policy and other kinds of public decision-making?

I think things like science, but especially if you take coding and maths, are more amenable to the current systems we have today, mostly because, especially coding and mathematics, but also things like games like chess, they are verifiable. So the answer that the AI system outputs can be checked for correctness. And so it's very useful when you're training these systems, you can have databases of questions and check to see if 100% if it's right or not, right? I think, of course, when you get into the arts and the humanities, things like decision-making, policy, you know, I'm not sure what you necessarily had in mind, but they're much more subjective. They're hard to run the same experiment twice. So it's difficult to get data about what is a good decision in those cases. So I think those areas will be a lot harder for AI to sort of model.

So as a neuroscientist, what do you think we have learned about human intelligence itself with all these advances in AI that we're seeing?

Well, I think when you're in the early days of this, I guess, modern AI phase, which maybe you could say is the last 20 years with the advent of, you know, things like deep learning from Jeff Hinton. A lot in the, when we started at DeepMind too, we took a lot of inspiration from the brain. High-level systems inspiration, not the direct mechanics of the brain, but things like episodic memory, what does the hippocampus do, which is what I studied. How, obviously, neural networks, reinforcement learning, your area, which we know the dopamine system in the brain implements. So we kind of took inspiration from the brain, which is the only general example of a general intelligence that we know of, maybe in the universe, as a starting point. And so we sort of knew that it was possible in the limit, right? And I think that's why neural networks and reinforcement learning have been so successful, because I think learning is the key to these modern AI systems working, not being programmed with the answer like expert systems were in the '90s, like Deep Blue, but actually allowing the systems to learn directly from data.

And I think now what I view looking back at neuroscience is, you know, how efficient the brain is, sample efficient. It doesn't need, you know, to ingest the whole of the internet to understand things. So it's a different, what we've built now today uses some of the same principles, but it's been manifest in a very different type of system than probably the way the brain works.

Okay, thank you. So, one of the things that people always talk about is the safety risk, security risk from AI and so on and so forth, right? And so, what does DeepMind's stand on this and how do you approach making sure that AI is safe?

Well, we've been, when we, again, when we started DeepMind, we actually, you know, were planning for success. Our mission statement was solve intelligence, step one, and then step two, use it to solve everything else, which at the time, you know, sounded like science fiction, but I think now is becoming clearer how that might be possible. And applying AI to almost every subject area, but we planned for success. So even though we were just starting out and the field was mostly just building up a head of steam, we understood also the implications of that if that did turn out to be the case, and we thought about it as a 20-year mission, and I think we're basically on track for that, you know, around 2030, then it would come with these attendant risks as well as the enormous benefits to science and medicine and all these things that I think we need as a society to help with our other challenges, many other challenges that we have around the world.

But also as these systems become more powerful, I think there's, you know, at least two risks that we've always worried about. One is bad actors, human actors, but individuals, but could be also nation states using these systems and repurposing them for harmful ends because they're dual purpose. But then also as we get closer to AGI, and I think we're entering maybe a kind of agentic era, should we call it, where systems are more autonomous, I think we'll see a lot of that happening this year and next year, then we have to make sure as well that the guardrails are in place, that these systems do what we expect them to do and don't veer off into areas that we hadn't planned for, and that could also be problematic. So I think those are the two challenges. There's a very sort of societal one, which I think also is going to require international dialogue and maybe ideally a minimum set of standards that is agreed internationally. And then the second one is more of a technical risk of how can we make sure these systems are robust and reliable.

So I mean, setting aside all existential risk questions, so what do you think is the top two risks that we should address with AI systems today?

Well, I mean, I think I mentioned the two classes of problems. I think we need to worry about things like bio and cyber risk very soon. I think that the current systems are getting pretty good at cyber, and I think we need to make sure cyber defenses are more powerful than the attack vectors. And it's something that, you know, we work on quite a lot at Google and at DeepMind, is sort of using AI for cybersecurity, to make sure that, of course, it's a very useful tool for cyber defense too, but you need to make sure that the defenses are stronger than the offenses. So, you know, I think those are kind of near-term risks. But there are many that we need to think about and do a lot more research on, and some of them are to do with agreeing a sort of set of standards, which, you know, I know many people including yourself are trying to work on.

Great. Thank you. So again, talking about international collaborations, right? So you said that we really have to work on that. So an audience like this, or a gathering like this, right, where we are trying to involve the global south very much in the dialogue, right, so what do you think would be the impact of such gatherings on the overall direction of AI, do you think, and are these going to make a big change going forward?

I think so, and I think that's why it's important we convene this summit around the world, because it's going to affect, this technology is going to affect everyone. It's a digital technology, so it can't really be contained by borders. There's things like open source, which is generally very good, but also one has to think about if you found a vulnerability or some issue with an open-source piece of software, how do you recall it? How do you patch it? There's no recall. So we need to think about that. These are new issues with something like AI where it's hard ahead of time to fully understand if there are any vulnerabilities.

And I think for the global south and countries like India, I think there's huge opportunity for the youth of today. You all have access to pretty much the most cutting-edge tools in the world, right? I don't think that's ever sort of happened before, maybe only three to six months after they've been invented in the frontier labs. And I think, and I can say as someone working on the cold face of this, we barely have time to understand what are the amazing capabilities that could be supported by these models, you know, in products and applied research. So I think there's so much potential there to be explored, and I think we'll see a lot of that. And hopefully many of you in the audience too, entrepreneurs here, you can do incredible things, maybe 10x of what you could do before because these tools are so capable and they're available almost instantly around the world.

So I have a question. It's more specific because you are right now in India, right? So the Indian population, like as we, one of the things that a lot of people have been remarking is that the crowd at this summit is extraordinarily young, right? So this youthful energy, right? So you think what would be the role that India can play effectively in the future, given the resource constraints of India, but also the availability of this talent pool?

Well, look, I've been incredibly impressed already about the energy that's here, and we heard from the minister that, you know, the youth of today and in India especially, I think when you see the polls on this, are very positive about AI, which I think is great. And what I'd recommend to the students of today is to really lean into becoming incredibly proficient with these new AI tools. And I think over the next 10 years that will almost kind of make them superpowered in terms of what they're able to do, you know, whether that's business or science. And it's a little bit like the dawn of the computer age or maybe mobile or internet that we went through. The generation that grows up native with that technology will end up doing sort of incredible things that we can only dream of right now. And I think that's going to happen with AI. And I think India and the youth here can be at the vanguard of that.

Since this is a research symposium, can I ask you to get a little technical on the next? So we saw the evolution of how AlphaFold, right, started off by building on top of existing work from the Baker lab and then moving on to evolving all on its own, right? So what do you think is the next stage technically in the evolution of these kinds of co-scientist models?

Yeah. So well, with AlphaFold, what we, we actually built a new, completely new system, but it required the PDB, the protein database. So it needed the 150,000 structures that humanity had painstakingly found over the last 50 years through experimental work. Right? And that turned out to be only just enough data to actually, you know, build a, solve the problem and build a system like AlphaFold. The interesting thing is in the debate we have at DeepMind and other places is what's the difference between the general system, which you can think of like the brain, and the tools that it uses. So for us as humans, there's no debate about what's our minds and what are our tools because obviously it's physically separated. But if both things are digital and in some cases both things are AI, the tool and the orchestrating system, then what do you put in the main system and what do you leave as a specialized tool?

Now, in my opinion, I think you'll end up with, in our case, with foundation models like Gemini will use things like AlphaFold as a tool. So if Gemini wanted to or needed to fold a protein, understand the structure of a protein, I think it would be better for it to call AlphaFold as a tool than put all of that protein information into the main system. And I think the, if you want to talk about it technically, I think the choice comes down to if you put that data into the main system, does it help with other tasks? Does it transfer to other tasks, or does it actually degrade the performance on those other tasks? So it's actually an empirical question. So that's why, for example, coding and maths, we put that in the general foundation models because it turns out if you get good at coding or maths, you're actually better at planning and reasoning in general. So it's a useful skill, but it also generalizes. But something like, you know, folding proteins, probably is a very specialized skill that wouldn't necessarily transfer to other domains. So I would be of the opinion that we should leave that as a specialized tool.

Oh, that's interesting because a lot of learnings from robot path planning was used for trying to, robot path planning was used to try to solve protein folding. So you think knowing how to protein fold will not transfer back to other domains?

It might do, but we would, I think it's, we in fact we do these experiments all the time on smaller scale models where we ablate different data sets and we try and mix in certain data sets and see if they help or if they regress some benchmarks. Like it would be no use if you put all the protein data in and then it got worse at language, for example, which is probably currently what would happen. So maybe over time with an AGI system you just have everything in the one system. But I think for the foreseeable future, I think it'll be more efficient to still have separate tools. Also those tools, by the way, might be hybrid systems in that they might not be just learning systems. They might also have built-in structures like AlphaFold did actually, about physics and chemistry and chemical bonds that you could learn, but would be more efficient to just tell the system or program that indirectly.

So now that I mentioned the word robots, so I'd like to ask you what's next in physical AI?

Well, look, I think I'm getting increasingly excited by robotics. You know, I probably wasn't so interested in that 10 years ago because I felt that the issue was the algorithms, not the physical construction of the bodies. I thought the algorithms were the thing that was behind. Now I think algorithmically, and actually we're very excited about Gemini Robotics because we built our foundation model to be really good multimodally, so it could understand vision, image, the world around us. So it has a very good understanding of the physical world, and that's exactly what you'd want for robotics, is a kind of general system that understands the physical context the robot's in. So I think we're probably in the next two, three years, we're going to see some very interesting new breakout moments for robotics. I think there's still quite a bit more research to be done. I don't think we're there yet like some robotics companies are claiming. And I think we'll have humanoid robots and also non-humanoid. I think both will be useful. But I think in the next couple of years there'll be some really breakthrough moments. So I think it's a very exciting space to watch and a good area to get into right now.

So a little tricky question. So a lot of people are, you know, I mean, there is a lot of fear-mongering around AI. We know that most of it is unnecessary. But if you start getting humanoids that are running off foundation models, do you think the fear factor would go up?

Potentially. I mean, it depends how we design those humanoids, but I think some risks go up too, right? Again, depends on what you deploy them for. And I think increasingly, especially if the humanoids are pretty capable and they're heavy, you know, there are dangers and risks with that. So I think we need to have those guardrails that we were discussing earlier in place by the time there's a lot of robots roaming around.

Great, thank you. So we talked enough about risk, so let's talk a little bit more about the positivity of things. So, you know, a lot of the benefits of all this cutting-edge AI right now still seems to accrue to, you know, the countries that have more resources, right, that have GPUs that can run their models there and so on and so forth. So what do you think would take for AI to reach the global south, benefit a much larger fraction of the population? So what kind of initiatives that we should be looking at?

Well, look, I think we kind of touched on it earlier, but these leading foundation models, maybe there's three or four of them, and perhaps five or six if we include the Chinese models too, then they're pretty much available very cost-effectively for, you know, a few months behind. Also open source, I mean, we work on our own open source models, Gemma, which we'll be releasing a new version of soon, which are very powerful for edge devices. So I think that's very interesting areas, really efficient models for, you know, computing on the edge, whether that's your phone or a single laptop or, you know, eventually robotics. I think there's huge opportunities there for optimizing what those kinds of models do and the types of products or applications you can build on top of that. So I think there's, you know, a lot of potential there for those types of work to happen.

The entire auditorium went dark for a bit. So no, that is not signaling any kind of ominous thing. So don't worry about that.

So I mean, it's amazing. So I was actually there when you did the first game-playing demo at one of the NeurIPS side events, and that was this tiny room, even then it was packed. And this large auditorium is packed. So what's the largest size hall you think you can pack if you are talking nowadays, Madison Square Garden?

I don't know, this is pretty big one, and I hear it's streaming to many, many people online. So, but yes, I remember that NeurIPS event very well. It was a hall maybe a third of this size, but it was, I think, it was standing room only, and it was packed outside the door. And that was really the first success that we had with these deep reinforcement learning systems that we sort of pioneered that could play Atari games, very simple games now, but just directly from the pixels, not giving any other information, just maximize the score here, the pixels on the screen. And it was the first demonstration, I would say, maybe of the modern AI era of an agentic system doing something kind of challenging and interesting that, you know, in this case a task that was designed for humans to find interesting and enjoyable and somewhat challenging. And I think it was kind of a watershed moment, of course for us, but also for maybe for the industry. This is back in 2013, I think, and that actually this thesis of learning systems, learning algorithms, this idea of generality that you don't special-case the information or give it privileged prior information to the system that maybe was the way traditional AI had been done or old-fashioned AI had been done until then, expert systems, that actually that could scale to something interesting, in this case an Atari screen with 20,000 pixels on the screen. You know, trivially small by today's standards, but very large action space and data space for the types of systems we had then.

Yeah. Now they've become like the hello world of reinforcement learning now.

Yes. Yes. And then of course that encouraged us to then go on with AlphaGo, which was I think really the big watershed moment that made the field and the industry sit up and notice in 2016. And I think started a lot of the commercial interest in these technologies that we could scale this kind of deep reinforcement learning and learning systems to actually, you know, beat the world champion and the legendary, you know, Lee Sedol in our South Korea match.

So and one thing I really have to say that, I mean, you wrote up everything about the Atari game player and also later AlphaGo that allowed the rest of the reinforcement learning community to catch up. So thank you for that. And of course my first success was I was actually in the room and not outside clamoring to get into the room to watch the demo. But so that's great. So again, so let's forget AI for a minute. Let me ask about reinforcement learning. So you know that Rich Sutton has been talking a lot about how, and David and Rich wrote this amazing article on how reinforcement learning is going to drive AI forward. So what's your take on that?

Well, yeah, obviously we've had many debates over the years, and I think it's, for me, if I was to say, and maybe we can take this question more generally, like what do I think about today's foundation models and reinforcement learning? Of course, reinforcement learning is an integral part of the post-training of these models, and I think the inference time compute, the thinking part of the models, could actually benefit a lot more from the ideas we pioneered in AlphaGo, the Monte Carlo research and other things. So in actually many respects we need to combine the ideas we had with AlphaGo with today's foundation models. Of course it's harder because you don't have a perfect model of the world, right? You need a better world model. In games it's trivial, right? The transition matrix. So I think that's an issue.

But if I was to guess today, I think the foundation models, like Gemini, are going to be a critical part of the ultimate AGI solution. And then I think we'll have lots of interesting reinforcement learning on top. I think eventually one day, maybe 20 years from now, we'll have a kind of more like an AlphaZero type system where reinforcement learning can learn everything from scratch, right? But I think actually that's not going to be the fastest way to AGI. I think it makes sense to use the foundation models and, you know, all the information that's already out there and learn that as almost like a model of how the world works, and then do your reinforcement learning and planning on top of that. I think that will be more efficient in the first case.

Is it still going to be the cherry on top?

No, I think it's going to be, I think it's in, I mean, you know, you have to ask Yann about his cherry comments. I'm sure he can talk at length on that, but I've never really agreed with that. I think it's just a fundamental, I don't really, you know, I think obviously if you measured it in terms of bits, then one can say, well, how much bits of information are you getting from the reinforcement learning? But in my view, and presumably your view, is that not all bits are equal in terms of information. Of course, if you get a bit about, you know, did you win the game or not win the game, that's much more important than some random pixel on the screen, right? So to equate the informational value of those bits in just a trivial way is clearly incorrect in my opinion. But I do think that the foundation models are going to be, the question is, are they going to be all that's needed or just a critical part of what's needed? I think there's no question they're going to be at least a critical component of the first AGI systems.

So we're almost out of time. So I just want to ask you what's your message for the attendees of this summit?

Well, look, I think my message is one of, I would say, cautious optimism. So I think we're on the cusp of an absolutely incredible transformation that's going to bring incredible benefits in science and medicine specifically, is what I'm passionate about. And I can see revolutionizing the way we deal with human health. I think there's many amazing companies and tools and products to build on top of these systems, and I think everyone in the world can build on these AI systems to do that. But then also I would just add a note of caution, which is I think we will solve these technical issues given enough time and enough brain power on it. I believe in human ingenuity, and if the best minds work towards that, I think we'll solve the technical risks. But we also need to do this internationally. So the societal challenges of that may actually end up being the harder problem than the technical ones.

Thank you. So, everyone, thank you.

Thank you, sir Demis and Ravi, for such a stimulating discussion covering the breadth of AI and reinforcement learning in such a short time, talking about possibilities and challenges. I have been asked to tell you that the number of people waiting outside to get into the room outnumbers the number of people inside the room. So you are requested to please stay seated if possible for the next session before we have a break. That would help with the law and order situation of the auditorium and outside. So I kindly request the attendees to stay on for the second keynote of the research symposium, which is going to be delivered by Professor Dame Wendy Hall, yet another fellow of the Royal Society and a professor of computer science at the University of Southampton. And joining her as a moderator is going to be Dr. Anand Deshpande, founder and chairman of Persistent Systems. Over to you, Wendy and Anand.

So we are joined on stage by Professor Dame Wendy Hall from the University of Southampton and Dr. Anand Deshpande from Persistent Systems. Please give them a huge round of applause.

Okay. Anand, I don't have a clicker. I have slides.

Oh, hang on. Okay.

Are you introducing me? You going to introduce me?

Okay. All right. Let me start.

So, good morning all of you. It's a pleasure to welcome and an honor to introduce a true pioneer in the digital age, Dame Wendy Hall. Long before artificial intelligence became the global headline and even before the web reshaped our daily lives, she was working on hypermedia and linked information systems, foundational ideas that anticipated today's interconnected and intelligent worlds. Few people have influenced both the birth of the web era and the global AI conversation. She has done both. She serves as the professor of computer science at the University of Southampton. She's a fellow of the Royal Society and the Royal Academy of Engineering, reflecting her rare ability to combine deep scholarship with real-world engineering impact. Her contribution to the global computing community is equally profound. She's a fellow and former president of the Association for Computing Machinery, the world's largest professional society in computing. Through ACM, she has championed international collaboration, open science, and the responsibility of digital technologies, shaping the profession far beyond her own research. In recognition of her extraordinary contributions to science, technology, and public service, she was appointed Dame Commander of the Order of the British Empire. At the moment when the world is asking how to build AI that is powerful, trustworthy, and globally accountable, Dame Wendy's voice carries both authority and wisdom. Please join me in welcoming Dame Wendy Hall. It's an honor to have you here.

Um, hello everybody. Are they all coming in?

Huh?

Okay, fine.

Are they going to come in? Let them in.

We should let them in. Right. If they want to come in the seats, you want me to just keep going.

All right. Well, I'll start because it's just so distracting. There's a huge crowd of people there. The crowd following Demis, you know, the 'we are not worthy' crowd, went that way. This always happens to me when I speak with Tim Berners-Lee as well. But I just want to say first of all, before I forget, a big thank you to the Indian government. Abhishek Singh is somewhere here. PJ, Anand, all the people to him for inviting me to be a keynote here. It is a huge honor and I want to say a few things that are a bit different this morning. So I have used some slides to try and explain where I'm coming from. So I'm talking about unintended consequences because that's what we've all got to worry about. Whenever you invent a new technology, you have to think about what might get wrong, what might go wrong. And you will have noticed with Demis, who I will preface these remarks by saying I have the hugest respect for Demis Hassabis. I think he's one of the greatest people in the AI field and one of the few that actually does address the challenges. But you heard him say on safety, we've got to do this. But no talk about how we've got to build the guardrails. We've got to keep things safe. But meanwhile, we're carrying on inventing more advanced technology. So that's what I want to try and get us thinking about today. And as Anand said, I was around building hypermedia systems before the World Wide Web. Met Tim Berners-Lee at one of the first World Wide Web conferences. And you know, I remember him putting up the first website. I don't know how many people, there's a lot of young people here. I don't know how many people remember the first website going up. There was one and it wasn't a very exciting place to be at the time. This is what happened. This is a graph that one of my PhD students produced in 2013, just as the web was beginning to get completely explosive. I'm going to talk about the ChatGPT moment later. I would argue the ChatGPT moment for the web was the invention of the smartphone. See, the iPhone was released in 2008 and that's when you could actually start using social media. The internet was there, but you could start using social media on the move. And I could tell lots of stories about this graph. I haven't got time today, but it is amazing to watch. There's a bubble. Remember the bubble? And Facebook started in 2004 and little did we know because we weren't anticipating people would do bad things with this system. We thought it would be used for the good. Beyond this point, the social media starts to take off and the oligopolies grow. We hadn't foreseen that either. The fact that Tim was always talking about lots of different people innovating on the web but of course they coalesced into a small number of big companies, oligopolies, with the Googles and the then Twitters, the eBays, the Amazons. And look at what's happened with the internet and those companies. This is what's driving the current AI revolution. The internet's given us the access to the data, the companies have that data and they have the deep pockets to invest in the new systems. And so we're moving into a world where we've got the issues with social media which we still haven't tackled, the unintended consequences of social media that's doing so much social harm. I didn't think about that in advance. We thought it would be for the good. And then also the rise of the big tech companies that are dominating today's AI field. So this was in 2006 with Tim Berners-Lee, Nigel Shadbolt, Danny Weitzner and others. And by the way, that's Rod Brooks there when he was head of MIT. He's the man who invented the robots that clean your carpets. iRobot. So this is 2006. We launched this concept of web science to try and understand what had driven the growth of the web. What the social consequences were and what they might be in the future. It's very hard to build scenarios when you're talking about people as well as things because you can't predict what people are going to do. So we drove this development of web science. Say the stuff Tim talked about in his book Weaving the Web, the concept of a social machine, the concept of technology and people coming together to create artifacts that couldn't have existed otherwise. Google's an example of that, Twitter's an example of that and all the other... we talked about web science being the theory and practice of social machines. And you know, we were thinking about could we monitor, could we anticipate the social consequences of the development of this technology? Does that ring any bells about what we're trying to do today with AI? And so I'm now going to leap forward because I've only got a short minute, 10 minutes or so, five minutes left probably. Anand's going to, we're going to talk in a minute. Around 20, jumping forward to 2017, the Royal Society produced a seminal report, sorry excuse me, a seminal report on machine learning. As we all saw that, Demis said they founded DeepMind in 2010, was when the machine learning revolution was coming. And he said there weren't many in academia, there are hundreds of AI professors in academia. But anyway, that's beside the point. And then I was asked by the then Prime Minister of the UK, Theresa May, in 2017 to write a review of how AI could impact the growth of the economy. She was looking positive and job creation. Note, that was the beginning of my world in AI policy. Jump again another five years forward. I don't have time to go into the details of everything to what I call the ChatGPT moment. 30th of November 2022 when OpenAI released ChatGPT and suddenly everybody in the world who had access to the internet through the web, this all happens through the web, could talk to an AI, an AI, right? And thereby hangs the tale. We anthropomorphize it. A lot of people think, you know, we've had all the issues of falling in love with their chatbots and all the issues around people, the vast majority of the world not understanding what this technology is and what they're dealing with. And then we had an era of a year or so of, you know, existential threats. So there's OpenAI saying here's this amazing technology, does fantastic things, oh by the way it could wipe the human race out in two years' time, right? How are people supposed to deal with that? Our Prime Minister at the time, Rishi Sunak, who is here, I think. I don't think he's here, but he's here at the conference this year. He got very interested in this and launched the first global summit on AI which was, as Demis said, in London. They were still talking about existential threats. That was his adviser talking about them. But also he introduced the AI, he created the AI Safety Institute which is an important part of this story. At the same time, exactly the same time, the UN launched its high-level advisory board on AI. It was a huge, huge privilege to be on that. This I believe is a way forward. The, I think the only game in town for the global dialogue that we called for in our report Governing AI for Humanity and Demis called for on the stage this morning. We, and hopefully there's going to be some discussion of this amongst the world leaders tomorrow. We actually, we have to have global dialogue. We have to have discussions on standards and what values and standards we set for the industry globally that will be managed in the nation states. This I believe is so important and they announced their new scientific panel that Ravi is on. Other people in the room might be. It's moving forward and we think it's possibly the only game in town. I've just said that there. And the other thing, so where are we today? The United Nations just launched their panel, new scientific panel. Hopefully they'll launch the global dialogue at the AI for Good conference in Geneva in July. And also the, Yoshua, of course, who's talking later, produced the latest version of the safety report. Sorry. That was published last week, 222 pages. I hope you've read it all. I haven't yet. But something new I want to talk about which I think is really the way forward building on those is in the UK, is something I'm really interested in. Remember back to what we were doing with the World Wide Web as it grew. Our National Physical Laboratory, NPL, with the support of the UK government, has just launched a new center for AI measurement. Oh, that's bad in this slide, isn't it? Sorry, measurement. Center for AI measurement. This could lead to the sort of Met Office for AI, right? Met Office, metrology office for weather forecasting. We're talking about AI metrology here. And I think this is a way to help us define what we mean by the science of AI. Can we forecast some? Can we simulate? Can we get the data we need to do that globally? Right. This is very important that we, nations, we can share data in order to build up the whole concept of AI metrology. You won't have time to read all this slide but it's basically about building an industry for AI assurance. This is the future. This is one of the very particular ways we can do safety is how do we assure products? How do we assure people who are selling products? How do we, what license to people who need to use AI? Do we bring in laws that says if it's AI, it has to say it's AI? All these sorts of things. And how do we measure the impact? I mean, you know, the Australians, good for them, are trying out an experiment for stopping social media for the under 16s. How do we measure the consequences of that? Right? These are the sorts of things we have to be able to develop. And this is where my life is going to go over the next few years. I also want to shout out for the UK AI Safety Institute. Some of them are here I think because they do the most amazing things in terms of looking for the bad actors. Seeing a lot of it is very deeply security-based because they're seeing where, oh, I'm way over time, aren't I, Anand? And seeing, you know, what are the results of the, are any of the AIs going rogue basically. And the network of those which is the countries that have AI safety or security institutes is being called the International Network for Advanced AI Measurement, Evaluation, and Science. And I think this is so, so, so important. I believe this is where we need to go. I'm sorry Demis, but I think AGI is a meaningless term or at best an ill-defined aspiration. We need to work out what we're actually talking about here. It reminds me of The Emperor's New Clothes if you know that one. I'm told it is also an Indian folklore. What are we really trying to achieve? Yeah, and I'll just make, there's a difference between machine intelligence and artificial and human intelligence and we need to think about what are we trying to do with this. This came from a friend of mine, Pete Rivett. I should have put his name on there. Sorry. Talking about I think of the machine as more like the Dr. Spock in Star Trek. And of course, we are messy, unstable. We have emotion. Do we, you know, people talk about AGI? Well, the machines will have emotion. They'll have consciousness. We don't understand what that means for ourselves, let alone trying to build it into machines. And you know that moment in 2001 where the machine says, 'I'm sorry, Dave. I can't do this.' Right. That's what we want to be able to avoid. I just pay homage to what India's done with data governance. I've been here, I travel here a lot and people, my friends here know I've followed the story of the Aadhaar, what Nandan and the iSpirit team did with that. And now this is one of the things that India can really lead on. We can't have good AI governance without data governance and I'm on the UN working group on that. It's very difficult. Cross-border data flow is really hard stuff. Whose data is it? How do you value it? All those sorts of questions. This is the end. Basically, you know, it's more about physical. We're not just measuring physical artifacts. We're measuring and assuring social machines. These AI is a huge force for good, but we've got to stop the bad things happening if we can mitigate against them. We got to reset our moral compasses here. It's not okay to do bad things with AI systems. We've got to let people understand that, right? And it's not just governments and the technical companies, tech companies. It's civil society as well. We've got to all be more responsible. That's a big project at Southampton by the way, RAI UK. I'm just going to finish with this slide because also, I'm also into scenario building. How can we actually model the different ways that the AI world can develop and AI meets people can develop? There's lots of forking paths. We got to decide which way we want to go. And we have to decide what type of world we're trying to build here. Which pathway are we taking? And I love by the way the slogan we see as we drive in. In India, AI means all-inclusive. Good for India.

So, thank you Wendy. I think you teed off many different topics and you were one of the original persons in the multimedia and then hypermedia and the web over the years. You have seen the web evolve. You started out by sharing that was very optimistic and then social media and various other things have happened. You have advocated for, you've seen this happen in the web. How do you see the same kinds of things happening in the AI world and what would be, are you looking forward, are you positive about where we are, are you worried, what would be your learnings from the web to the AI world?

Well, I tried to explain this here. We didn't get it right with the web. I mean Tim is trying to build Solid, social linked data to try and fix the web. In terms of what we didn't see was what was going to happen to our data. Effectively we give all our data to the, not just the governments. I'm not talking about governments here. If it's the retail companies, it's the big tech companies. We sign it over, you know, we all, you tick the terms and conditions. No one reads them. You just, the data's with them. And so they can therefore do what they like with it. And of course we get, I mean Ravi was talking earlier about reinforcement learning. Google was a huge reinforcement learning machine because, and we were part of that. We were part of that social machine as we gave them our data. We gave Google our data but in return we got access, we got a search engine that worked and we kept, you know, we reinforced it every time we used it. I actually, I like Gemini. I'm using that because I feel comfortable that the values that Google have are good and of course Demis is part of that. But what we can't, what we, we know there are safety issues, we know there are security issues but we can't really define them in terms of the social consequences. We can talk about cyber security and of course we have our intelligence services fighting very hard to make sure the bad actors, humans or states, aren't using this technology to attack us or to, you know, ruin our lives in that respect. But the social consequences we didn't predict with the web. We didn't predict what was going to happen with social media. And you know, we need to find ways to try and think about what the social consequences will be of the introduction of this technology. We talk about, oh I think we might come on to this, there's a lot of rhetoric about job losses and the social consequence of, you know, unemployment caused by it. But actually there are huge social consequences in terms of the use of this technology that we just can't predict easily because we don't know exactly what people are going to do with it. And that's the science I'm trying to, as we started to develop it with web science, I'm trying to talk about that in terms of AI because it's about, as I say, building, we're building social machines and we need to think about how we evaluate and test and forecast what might happen. Things that could go right and things that could go wrong. And it's really difficult. It isn't like forecasting the weather. That's hard enough but because you're trying to forecast what people will do and people, you can't predict how people are going to react, particularly as a society.

So what do we do about it?

We, well, we have to start I think developing ways of thinking about it. So building simulations using the data that we have. I mean AI might help us in doing that and data, it's all about getting as much data as we can and encouraging a generation of data analysts that are looking at the data from this perspective. You know, monitoring what happens. You saw that graph I put right at the beginning which is a very simplistic way of looking back at how the web evolved. Can you then extrapolate and predict what might happen in the future? So, easy, by the way, I'm not pretending any of this is easy or I think we'll leave it to the young people here to explore new ways to see how we can learn from the past so that we can build a better future.

But let me, you've been a big champion of inclusivity and diversity and various other things. I know your term during the ACM you did a lot of work on ACM-W, women, and you also alluded to the fact and we've seen posters about AI equal to all-inclusive in this context. Do you think AI will help inclusivity? Are we going to get, will AI make it possible for it to be more inclusive? Meaning other people who are potentially deprived will have access to AI so they can do a lot more. Is that a positive thing or is the bias going to make it very difficult or what do you think about?

Well, there you've asked me about three questions in there.

Yes, I want you to, I don't want to interrupt your answer.

Well, let me try and unpack it a bit. We're here in the global south because it's so important and I'm really glad the summit has rolled to India this year. Although the scale of it, like India, is a bit scary. Was it, the number varies but 15,000, 250,000 yesterday. And a lot of young people which is great. Yeah. And I think that making sure that AI is used, I mean the UN report was called, we talked a lot about the global south when we were having our discussions, Governing AI for Humanity, right, the whole world. So that everyone can have access to the good things that AI will give us and this is good old-fashioned AI as well as the new generative AI. And I'm a great believer, the Chinese are way ahead of us on this, but robotics is the next big thing. Embedded AI in robotics will come and I'm sure there's a future in India for that particularly because in farming and all sorts of areas that's going to be really big. In terms of inclusivity, I think there is the issue of course of how we keep, make sure that the elderly like myself are included. This is a world, the young people here will be, this is your world, you're building it, right, in every sense of the word, building because you'll be part of the development of it if you go into the tech world or AI product world and you'll be using it in whatever career you do. So it is your world. We need to leave it safe for you. That's what I'm really passionate about. And we need to give, we need to build the frameworks and those, well we talk about guardrails, the safety rails, that infrastructure I suppose is the word I'm looking for that we, the young people here today, can build on in the future to know that what they're doing is going to be as safe as we can make it. And you know, what are the, you know, the things that put people in jail for doing bad things using AI, that's, you know, got to create laws around that. Inclusivity, my big worry when I get asked about what keeps me awake at night about AI. Yes, there is an existential threat down the line if we get it wrong and we build AI systems, I don't use the term AGI, we build AI systems that are more intelligent than us, is that, you know, is that, and they get out of control and to make decisions on their own that are not good for us. Yeah, that's dangerous. My biggest worry is the lack of women in AI. Right. This is, we are 50% of the population. I've fought my entire career to get women into tech and I feel I've singularly failed to do that. There are more and it's lovely to see the young people here. You don't see the female leaders. I'm one of the few women on this plenary stage. There are some others but tomorrow I don't see at the World Leaders Day. I don't think we'll see many women on the stage at all. And my friend, you asked me on yesterday, why is it? Because we get women coming into the industry to and and um

Yes, in India at least, you know, when we are recruiting fresh graduates we get about 40% women coming into the workforce.

Yeah.

But as you go higher up in the organizations the number of women in leadership positions comes to be lower.

Well, this is it. It just doesn't just happen in the tech industry, but it's particularly bad in the tech industry. Women get, when you move up in the industry or in academia, you get fed up with being dismissed, being disregarded, being patronized, being told your ideas aren't good enough. Or I mean the number of times, and girls, it happens to me still. You put forward an idea and they say, 'Oh, that's useless.' And then a week later, you hear it put forward by a man and they go, 'Oh, that's a really good idea. Didn't you hear me say it?' And my friend Dame Steve Shirley, she had a lovely saying which said, 'Why do women leaders have flat heads? It's the constant patting on the head.' Says, 'Don't worry, dear. It'll all be all right.' And that's the problem. And it isn't a problem women should be solving. It's a problem the men have to solve as well. You have to stop patronizing. You have to listen to us. And you have to encourage us. I had the most fantastic mentors, both men and women. And without them, I wouldn't be here. There are several times in my career where I wanted to give up because I'd been so, people have been so rude to me or dismiss what I was saying. And it was my mentors who kept me going. And particularly of the men I remember, but particularly Steve Shirley, Martin Rees, the Royal Society who told me I could do it and I have.

You are an inspiration to not just the women of this world but also to the men. So thank you for keep doing what you do. I've known you for many years now. I just want to shift this since we are here in India and global south. Clearly from the point of view here we see that AI is completely dominated by two countries, 10 companies. What do you think is the future? How do we change that or is that important at all? And what happens to the rest of the world maybe in case somebody decides to shut the taps off somewhere? So what, to shut the systems down to certain countries or not allow other countries to do it because of?

Well, I think that again it's good that we're here this year because yeah, the first summit in the global south, it's so important that the countries, I mean India's got the scale to do it. I mean to be honest with you India could do whatever it wanted to do if it puts its mind to it, you know. And because of the scale, I mean you want to live in a little country like the UK, you know, once our navies ruled the world but we're actually a small island off the coast of Europe with only a few million people. So for us to even, for you, for us to scale up with a big company is really hard to do in the UK. I am very excited about what's happening in the UK. Our government's put a lot of money into AI and I talked about the new safety institute and the measurement institute. We tend to lead in the soft power area and I think the whole AI assurance industry, this is going to be a major industry. I mean you think about how you buy insurance all the time for your health, your car, your house. We're going to do that for AI, right? There will be all sorts of insurance products that we will build, we will buy. And people, I mean it's the same when a company buys some AI, they've got to know it's going to work. Who do you blame if it goes wrong? So there's a whole new industries that are going to grow. And also the other thing is amazing when I look at, I'm not, I know much more about India than I know about Africa but I do know in both countries I have seen the development of technology where it leapfrogs what's happening in the west or the east because you take a new idea and then you use, say what India did with smartphones, amazing what you've done with the Aadhaar and the, you know, I find it hard to pay for things here because I can't get onto your system but it all looks so easy. And then what's happened in Africa, in Kenya and Uganda and places where they're using the technology, it's quite amazing what will happen. And this is people being creative and innovative in terms of new technologies. And that's, so every government in the world needs an AI strategy that is fit for purpose for their country and it's not one-size-fits-all. And I think that, you know, we're talking about sovereign AIs. Do countries need their LLMs in their own language and with their own data? And small, you know, the SLMs. I think this is a world that's emerging very much so. It's very exciting and I think there's just everything to play for. I'm very positive, although I talk about the negatives, I'm very positive about this future.

Thank you. Thank you, Wendy. And I want to now move it to the younger people who are in this room. You talked a little bit about jobs and what happens to students. What should students be learning now? And what do you think is the future of work and employment in the AI world as you see it?

Well, okay, this is what we need to start measuring. We need to start gathering the data around the world of what happens as the AI introduces. I mean, it's all very well say we're going to measure AI, but we can't do that without data and evidence and things to measure. So, I'm calling for this to happen around the world. We need governments and companies to share data about what's happening as much as they can. Clearly there's sensitive data you can't share but and to enable researchers in industry or in academia to analyze that data and see what the trends are, see how things, because it's all hypothetical. I mean there's a lot in the UK at the moment in the press about oh the jobs are going, the frontline jobs are going. There's also the jobs are going also because there's not enough money in the economy and I would, I won't go political, sorry, I won't criticize the current government just in case they're in the audience. But because I don't agree with all their policies, but you know, there's all sorts of economic reasons why jobs are going for young people in the UK, not just AI. I do think it's ironic that programmers are one of the first to be hit, people who write code, because and we've seen the applications to our computer science program go down because the rhetoric is that companies won't need programmers, entry-level programmers, because the AI can do it. But the AI doesn't at the moment decide itself what it's going to do. Someone's got to tell it and someone's got to check that it's correct. Yeah. And so we can't, we need experts in the field of computing and software engineering. So the way they're trained might be different. But we're going to need those experts in the future. And I like to think about teams. I like to think about augmented intelligence. So the AI augments our intelligence and talk about teams of humans and machines working together to do things better.

So do you have a specific message for the young people here who are studying various subjects? Should they be investing more in computer science or do other things?

Well, I don't put it like that. Everyone needs to understand just like, oh, when I started out, we all had to learn more. I hated, I was a pure mathematician. I hated computing. Hated it. Now look at me. I think in the abstract that's what I've always loved, but I had to go, I did my PhD in pure mathematics and then there were no jobs, right, for pure mathematicians. There really aren't. And the computing was clearly the next big thing. The micros were coming out. I wrote my first AI code on a BBC Micro. It was a program to, an intelligent tutoring system in 1982 is when I wrote that in Micro-Prolog. And I learned that because I went to do a master's degree in computer science part-time at City University to learn about the computing. This poor mathematician, pure mathematician, learning about computing because I knew if I was going to get into this world, I had to understand it. So I think and my passion is about explaining AI to people who don't know what an eigenvector is. Now all the mathematicians and physicists will know what an eigenvector is, but that's a small part, most of the population are not going to ever know what an eigenvector is or want to know. And so what we're doing at Southampton is two things. One is we've developed an online masters for non-scientists, right? An online masters in AI for non-scientists. And I'm passionate, I always have been passionate about the interdisciplinary work. I do loads of work with social scientists, historians, geographers, musicians. And so I'm passionate about the interdisciplinary and about how we, as Demis was talking about, we go across the disciplinary boundaries. But they need to know about, you know, everyone needs to have some understanding about artificial intelligence and this is the most, for me, this is the most important thing. So, young people, if you're not, you learn about it. However, whatever's the easiest way for you to do it is to and buy my new book which is coming out in May or June, which is called AI Explained, and that's for non-scientists, too.

Thank you very much, Wendy. Sorry. I think our time is up. So, I'd like to really thank you for being here and for sharing your perspectives. Yes, we will all buy your book. Hopefully, you will come back again for the book tour and we would love to host you again when you can share more details about your books. Thank you very much and thank you all.

Thank you Professor Wendy Hall and Dr. Anand Deshpande for a very important conversation highlighting the need to do more as we build the future with AI. Thank you very much. We are now at the place of a well-deserved break. I think we would all want to, you know, stretch our legs but before we do that I want to make a couple of important announcements. We will reconvene for our panel at 10:55. I request you to be seated by 10:45. We are going to have two parallel panels. One is a research dialogue on AI for science in this room and one is an international panel on challenges of global south in level one room number 7. Secondly, people who are going to present posters can be beside their display boards starting 10:55. So we'll meet again at 10:55. Thank you very much.

Sorry, an update. Apparently there's a lot of crowd on your posters already. So if you are a poster presenter and you would want to catch more people, please go to your posters now and you know, start talking about your research.

Sitaram, principal researcher at Microsoft Research India and it's great to see everyone turned up in such large numbers here. So we're going to start this session actually with a video from Professor Fei-Fei Li who couldn't join us sadly but she wanted to, you know, convey her thoughts. So we'll be playing that video and then starting the next session.

Distinguished leaders, colleagues, and friends, I'm deeply honored to address you at this historical gathering. I want to express my sincere gratitude to Prime Minister Modi for his vision and leadership in convening this summit and to Additional Secretary Srinivas for this kind invitation to be part of this moment. Though I cannot be with you in person, I want to share my excitement for what India is building and what this moment means for our shared future. Historians will look back at this period and call it the first true era of artificial intelligence, a time with civilizational consequences. Every sector, every industry, every nation will be touched by this technology. And this is precisely why gatherings like this one matter so much. Today's AI has achieved something remarkable. Large language models can write, reason, and synthesize information in ways that seem impossible just a few years ago. Yet, for all their capabilities, these systems are just the first steps towards building intelligent machines. I firmly believe that the next frontier of AI lies in what I call spatial intelligence. The capability that connects perception, reasoning, and action in 3D and 4D worlds, enabling machines not just to see and talk, but to do. 540 million years ago, the emergence of vision triggered what scientists call the Cambrian explosion, a sudden flourishing of life and complexity. Vision gave rise to understanding. Understanding enabled prediction. Prediction made purposeful action possible. And from this chain, intelligence emerged. We stand at a similar threshold. Spatial intelligence will transform how AI perceives and engages with three-dimensional spaces with applications spanning scientific discovery, education, healthcare, robotics, and more. Imagine AI systems that can assist the surgeon in the operating room, help a farmer monitoring crop health, or support an elderly person at home with dignity and care. This is not speculation. This is the research we're pursuing right now. But let me be clear. The purpose of AI is not AI itself. The core mission of artificial intelligence must be to expand human capability, not to replace it. AI should amplify our creativity, deepen our connections, and enhance our sense of purpose. It should superpower our students and teachers to transform lifelong education, assist our clinicians and patients, and help workers stay out of harm's way while increasing productivity. Always. It must respect human autonomy and dignity. This is what we mean by human-centered AI. It is not a slogan. It's a discipline, one that demands we bring together technologists, social scientists, policy makers, and communities from all walks of life. If AI is going to change the world, and it will, then everyone from everywhere must have a role in shaping that change. I believe this summit embodies that spirit and I'm optimistic about the future. Not because the road ahead is simple, but because gatherings like this are bringing together committed leaders and brilliant minds to build something meaningful together. Thank you for the privilege of addressing you.

We're very grateful to Professor Fei-Fei Li for her kind and encouraging words. And now we start with the next session at this hall which is a research dialogue, the first research dialogue of the day and this will be moderated by Dr. Shivkumar Kalyan Raman who is the CEO of ANRF and the speakers for this will be Dr. Acha Sharma, Professor Allison Noble, Professor Neil Lawrence, Dr. Bonnie Kraf, Professor Priya Donti and Dr. Pushmeet Kohli. So I welcome everybody on stage and over to you Shiv.

Good morning. It says time's up already. Maybe it's for the last one. Good morning, everyone. How are you feeling today? All good. Okay. So, it's my privilege to host this panel for you. We have the who's who of folks doing things in AI for science and engineering. So one of the exciting things and many leaders have told us that one of the biggest contributions that AI will do for the society is transform how science is done and transform the acceleration of the impact of science. So I look forward to a wonderful dialogue here. We have about 85 minutes and in terms of the format we'll keep it super simple and conversational if that's okay. I will request each of our panelists in order here to give an opening statement and specifically reflect on one and what they see in terms of the acceleration to impact. This conference is about impact as a theme. And this particular panel is AI as a catalyst. So a catalyst accelerates and what can AI do to accelerate discovery and both fundamental discoveries as well as accelerate the conversion of science and the translation of science to its impact. And also where, and since we have a lot of young audience, I will request the panelists to also reflect on what does it mean for the younger generation here as well as what they can do to contribute and since we are in the country of India what can India do to contribute as well and what is the opportunity and potential. There's a lot of discussion about risks and I would like to bias this conversation a lot more on opportunities and of course please feel free to balance that with risk and then what can we all do to being a positive force and science as a positive force for transformation. Okay, these are the few thoughts I have but without any further ado I thought I'll start off by saying how can AI help us win a Nobel Prize but you know, but I'll keep it simpler than that. Without further ado, Bonnie.

Wonderful. Can everyone hear me? Okay. It sounds like it can. I'm Bonnie Kraf. It's an honor to be here and a pleasure to be amongst such esteemed colleagues and thank you so much for inviting me to be a part of this. So I lead a team that focuses on AI for science at Microsoft Research. We focus on two main things, molecules and materials. So specifically how can AI enable the discovery of new novel molecules and materials and the motivation for that is pretty simple. The limits to discovery really doesn't come down to imagination. It comes down to the speed at which and the depth to which and scale to which we could explore. So our hypothesis is that we can use AI as a fast surrogate for some of the very expensive slow steps within the scientific process. So often times we have to compute the properties of molecules and materials that takes extensive amount of compute time, experience, domain knowledge, and experiment is even more expensive. So what ends up happening is the scientists are limited to a smaller space to which they could actually explore. So, if we can use AI, we can open up opportunities to find novel molecules and materials that can help to solve some of the world's greatest challenges today. Things like new medicines for diseases that we haven't been able to treat yet, solutions to the energy and climate crisis, new materials for sustainability, and many, many more. So, I'm very excited to be here. Especially be a part of this event in India. You asked about how India can take part. I do think very few places in the world have as much scale and talent as India. This is a nation of doers. And so I think there's an opportunity here for India to be a part of this for the young people, for the massive amount of talent in STEM, science, technology, engineering...

Math, computer science to get involved and be part of the scientific loop. Ultimately, the way that we accelerate scientific discovery is by bringing AI to help us find the new opportunities to explore a more expansive space of discovery and have India be a part of that as well. Thank you.

Thank you, buddy. Go ahead.

Okay. I'm Allison Noble and I think I can build a little bit on what you were saying. I work in healthcare AI and that's a very diverse topic. So I can talk about that for probably the whole day and beyond. And some of it, as we've heard and you've just mentioned, the advances in AI are really helping in the discovery side. But there are some other areas. I work much closer at the interface working with clinicians and there are some early papers coming out that are really trying to get people to think of doing medicine in different ways. And by that I mean taking large sets of data that are taken of patients, some of that will be imaging which is my background, some of it's then following those patients, but then it's using thousands and thousands of examples of that to be able to maybe discover new ways that we can manage conditions. And why this is so important is a lot of medicine has been driven by what you're able to measure at the time. In imaging, it's what you see in an image. And with AI now you can discover the hidden patterns in multimodal data that might suggest totally new ways of managing patients. And this is still the future though because even if you've got a discovery, that translation into practice takes a lot of time and effort. And this is why when we talk about some of the excitement of AI discovering a new algorithm, that's only the beginning. That's the beginning of the journey of getting that model into practice and in medicine that's a very complex process. So some of my research is actually thinking more about that. Having developed a model, I work with ultrasound. So we've been working for many years on how you can make ultrasound an easier technology to use. You need to make it easier because humans can't recognize the patterns in ultrasound. So what we've been working on is both how you recognize patterns but with AI that go beyond what humans can see. For example, estimating the gestational age of a fetus from a scan, you traditionally fit a circle or you fit a line having captured an image. But actually, if you want to date a pregnancy, you can do it as well, if not better, at certain points in pregnancy if you pose it as a pattern recognition problem because the acoustic patterns will recognize elements of fetal development that humans can't actually see by eye. So there are all these opportunities, but this then gets into the debate of you now have tools where humans work with AI. And the early papers in healthcare were very much trying to show that you could automate processes and show that AI was as good as a radiologist and radiologists were really scared about this that they would lose their jobs. But we've moved on from that now because what you realize is you can automate certain tasks but a lot of the time you need humans working with AI and then you get into all the sociotechnical issues of how do you understand where a human will make a better decision with AI or not. Evidence out there at the moment is very mixed that for certain tasks AI helps, in other tasks it doesn't help. And this to me is really important that we have a lot more research in this area to really understand this area. It's often called human-AI collaboration. At that point I'll stop. Thank you.

Thank you, Alison. Pushmeet, please.

Hi. So I'm Pushmeet. I lead the science and strategic initiatives unit at Google DeepMind and the mission of the organization of Google DeepMind is to build AI responsibly to benefit humanity. And when we think about that last part of the mission statement which is benefit humanity, what does that actually entail? When we talk about AI, AI is not just a monolithic sort of concept. If you think about what is even intelligence, there are different forms of intelligence. There is common intelligence that allows us to solve tasks like image understanding. There is expert level intelligence where some humans have become experts in solving incredibly hard challenges. Being able to solve and find proofs of immensely challenging math problems or being able to diagnose a disease by learning about medicine or by thinking about how should a legal contract be specified. So that's expert level intelligence. And then there is another form of intelligence which goes beyond human competencies where there is no human today which can solve that particular task. An example of it is the protein folding problem where you are given a sequence of a protein and you are asked can you tell me the structure. No single human just by reflecting on that sequence of amino acids which constitute a protein can tell you what is the 3D shape. And that's superhuman level intelligence. So when we think about the impact that AI is going to have, we have to think about all these different competencies. The first competency of common intelligence, it will make intelligence much cheaper, much more available and much faster. On the expert level intelligence, it will democratize expertise. So look at a country like India where it has challenges in terms of providing high quality health care to the 1.4 billion people in the country. Not everyone has access to the high quality medical health care needs that are available to a privileged few. So can we make sure that every single person on the planet has access to that deep intelligence? So that's the promise of AI at the expert level. My own work is mostly focused on the superhuman part and this is where I think the greatest promise of AI lies. If we think about ourselves as a species, as humans, we have come a long way in our understanding of nature and reality. But if you think about the last few years and reflect on what has happened, think about climate change, think about the COVID pandemic, that really shows us how little we know of nature. Once the virus evolved or came in, generated these sequences and the whole world had to stop. That's essentially the limit of human intelligence. We have so much more to learn which is why we need to have AI to solve these grand challenges of understanding what proteins look like, how cells work, how organisms work. That's the promise of AI for biology. In material science, as Bonnie was mentioning, if you think about human civilization, we teach history in school in terms of how humans were using materials. We went from the stone age to the iron age to the bronze age and depending on who you ask we are living through the silicon age or the plastic age but it tells you how important materials are to humanity. Each material gives us a superpower which completely transforms civilization's abilities to do certain things. So can you imagine what will be the new magical materials of the future and how will we invent them and that's where we need this superhuman intelligence to really think about these amazing materials like room temperature superconductors which can open up new advances in how we transform energy, can come up with high magnetic field for fusion reactors, give us abundant sources of energy. That is the promise of AI. So yes, when we talk about AI we think about all the economic impact that will happen and there will be many changes but there's a lot of positive things to look up to. This will all need to happen with a sense of responsibility. And with regards to India, and not just India but across the world, I think one of the very important elements is that we really need to make sure that people understand the strengths of this technology but more importantly they also understand the weaknesses of this technology because if you are not able to understand the weaknesses of the technology you will basically have a Formula 1 car but you will not know how to drive it and you will crash. So access and understanding I think are key elements that are needed not only for India but for the entire world and especially for the global south. We really need to make sure that people have access, they understand how the technology is shaping up, what are its weaknesses and how can they use it responsibly. So I'll stop.

Yeah, thank you, Pushmeet. Yeah, please go ahead.

Hello. So, namaste everyone. I see that there are many many young people here around the age of 30, 35 and they must be wondering what are these people talking about AI, you know it's us who know about it and I've seen some posts like that on social media. It's all right. So when I was your age I was fortunate to be landing in the laboratory where the first website was released and that was www.cern.ch. How many of you have heard about the Higgs boson? Oh, great. Wonderful. Now, how many of you know that it took 30 years to build the machine? The R&D for the accelerator that you mentioned, superconducting accelerators, that took 10 years to build it and then another two, three years to commission and run it. The algorithms that were needed to look for the Higgs boson for which the probability was 1 in a billion collisions. We needed data from 2,000 trillion proton collisions. So the algorithm making and the searching for anomalies in the data has been the forte of scientists working together. So the message here really for all of us is collaboration and all the guardrailing will come from collaboration, understanding the troubles that are coming in front of us. We have actually gone through the cycles as you just mentioned also. So the 30 years in which I have seen from paper to the large experiments which cost us more than a billion dollars each and then the operations that also cost the same amount per year. We need to be responsible. We need to see how responsibly we'll be able to take this forward. And AI is another challenge that is in front of us now. In the past it was about looking at how to deploy the web and the web itself was the precursor to a lot of domain related advances that you see in every possible domain of which we've already heard, we've heard healthcare, materials. COVID came as a bolt from the blue and everybody stopped to think a bit on how to take this forward responsibly. And we can't just take technology and science and research and then put it back in the drawer once we've finished the problem. We have to continue and we could pick up the challenge of COVID because there were tools and because there was science of understanding and because there was funding that was made available in good time by the governments all over the world so that we could all be prepared. But I would still underline that the bottom line was collaboration. And why do I say that? Because I am one of those 10,000 scientists who was involved in the discovery of the Higgs boson and so one upon 10,000th of the science breakthrough prize of 2024 is also I can say one of my accolades. But the next generation will not be 10,000 it will be several orders of magnitude which means the whole world is going to work together and particularly for the global health. Think about it that every classroom or every university or institution and our premier institutions in the country for that matter who are already doing cutting edge work. When we started working for the LHC, we did not know the solution. We did not know whether we'll get there. We did not know whether we will be able to solve the technical problems because 27 kilometers of the Large Hadron Collider, the protons cross the border from Switzerland to France 40 million times a second. So cross-border understanding is very important not only in the protons who go without passports but also within the governments and also of course how to deploy the understanding and how to deploy the information and of course the beams match with submicron precision. So I think that here is an opportunity for us to look at our societal issues that we have in our country and in the world. In fact for that matter if we look at the larger picture, the climate change, environment pollution in our country, we have some thoughts on how to solve them and I think we have the tools with all the precursors of AI if we look at what we have done in the past that we can deploy for solving these issues. And I would love to see all those institutions involved in the solving of these kinds of issues. Thank you.

Thank you. Neil, over to you.

Oh, well that's such great contributions so far already. So it's also showing sort of how vast the topic is. So when we think about science, Bonnie sort of emphasized the importance of discovery. Allison's talked about the importance of the human-machine collaboration. Pushmeet's talked about the superhuman capabilities and of course Ashwini's talked about how we all have to work together and collaborate to create science. And it reminds me of something that I think artificial intelligence or the notion of it, which I think the word is already problematic. I agree with Wendy that the notion of AGI is deeply problematic. Because intelligence is very personal. It's very human and I think when we think of India and actually eastern cultures in general, my sense is you have a much more sophisticated understanding of intelligence because the religions in the east tend to accept different forms of deity that express different aspects of humanity in a way that religions in the west don't. So I think some of the hang-ups we have in the west you don't have here and I think that's really exciting. But then it's what does it do? I think to me the AI itself gives us a place to stand and reflect on ourselves, reflect on who we are and our institutions. So when we think about doing that in science, I think we can see some enormous opportunities for impact in India and for the audience. And I want to sort of step back to people who've talked about the philosophy of science like Thomas Kuhn, the structure of scientific revolutions and Popper's conjectures and refutations because science isn't just about discovery. It's about understanding and what it means for us to understand. But not just us as individuals, us as institutions. And you can see in a number of ways we're failing. We're failing when not everyone understands the Higgs boson and why it's important to spend a billion on such experiments. And I think one of the ways we're failing can be seen right at the heart of what Thomas tells us. He tells us that the paradigm of science, he's writing in the 1960s, is stored in textbooks and he talks about normal science as solving puzzles within that paradigm. And then he talks about paradigm shifting science. So big discoveries like Einstein that change the way we think about the world around us. But there's a fundamental problem here. Nowadays the paradigm of science is not stored in textbooks. It's stored in machines. And some of those machines are storing simulations and complex ideas that we don't have individual access to. Now I got into this field sort of 25 years ago because I was an academic who wanted to apply machine learning and it seemed to be science was a very sensible place because science is all about academia. Now what I see as the opportunity here is increasingly people have become isolated from the science and I think particularly with large language models as tools there's an enormous opportunity to allow people to get back in contact with that and to reaccess science in the same way that a student a 100 years ago would be accessing science through textbooks. That you can access all that science stored up in the machine through these large language models and I hope that that's going to be a revolution in science where we see again going back to I think some of the hopes at independence for India, the foundation of the IIT, the foundation of the science institutes, that India was going to be a partner in those scientific discoveries as it has been but not enough of it happening here, too much of it through diaspora, too much of it not accessible because the technology for science has become removed from the place where the creative people are and I really hope we can see that change.

Thank you, Neil. Priya, your opening comments.

Right. Yeah. So, thank you so much. So, my name is Priya Donti. I am an assistant professor at MIT as well as a co-founder and chair of a nonprofit called Climate Change AI. And my angle on this topic of AI for science and engineering is more from the engineering side and specifically thinking about the power grid. So the power grid is something that I think many sort of take for granted, right? We switch on a light and unless there's a planned or unplanned outage, hopefully the light will come on. And yet behind that, it's a very complicated system. It's a system where the amount of power that is being produced and put into the power grid has to exactly equal the amount that is being consumed on the power grid at every single moment in time. And this is incredibly complex because first of all, we have a lot of variation in how much demand is on the grid at any given time. And as we aim to work towards kind of decarbonizing power grids or moving to more distributed forms of power, we also have to deal with the fact that the amount of supply we have, solar and wind for example, varies based on the weather. So, how do we manage this kind of increasingly complex system? We've historically done this using kind of optimization electrical engineering controls paradigms where we actually write down explicitly the sets of equations that we want to solve over on the system and solve them. But the challenge is that this paradigm is not sufficiently scalable. It's much too slow to solve the problems we actually need to solve to coordinate the power grid. And so this is where AI can potentially play a role in terms of both helping us reduce the uncertainty of quantities like demand and renewable energy by helping us to forecast these quantities more effectively. Then given those forecasts to help us to actually control the different devices on the power grid, our power generators, our batteries, our inverters in ways that are more scalable. So I think earlier this concept of surrogate modeling came up also in one of the other panelists' remarks. So how do we actually use AI to learn fast approximations to some of the optimization models that are used on the power grid? And then third, how do we actually envision new paradigms of managing this power system? Historically because the power generators were big, were centralized, there's been a very centralized way of managing the power grid, but kind of AI and reinforcement learning and control technologies kind of allow us to envision a power grid that's more dynamic, that's more distributed, where there's more kind of localized control. And so these are some places that I think AI can play a really useful role. What comes up though when designing AI for the power grid is that even though there are commonalities with other areas. So some of these terms I've used around kind of control and optimization, they come up in other scientific fields, they come up in robotics, but sometimes the notion of let's say kind of physical feasibility, the notion of safety, the notion of interpretability that we need on these systems is sometimes very or subtly different from what that means in other scientific areas. So there's some aspect of the fact that the way we design AI for these systems often looks a bit different than the way we design it for other systems. So my own research group for example, we develop ways to kind of write down the equations associated with power grids and actually embed them into the way we design machine learning models into the way we kind of design neural network layers. And the reason I bring that up is because I think that as we think about kind of the role of let's say India in kind of developing AI, I think it's important to recognize that the way we develop AI is deeply contextual when it comes to both what problem we're trying to solve. Kind of the way that our models look different because we're designing for different requirements and also depending on where we are trying to solve that problem. For example, electricity demand forecasting is a problem that's considered kind of solved in the US context because we have pretty steady electricity demand barring recent data center growth in many places. And so data about kind of how electricity demand looked last summer will reflect roughly what electricity demand looks like this summer. In the Indian context where in that one year's period of time you'd have had kind of many more buildings built, kind of many more air conditioning systems installed, your data looks different and so whereas kind of people developing time series forecasting algorithms in the US wouldn't have to think about what's called distribution shift in machine learning, the idea that kind of this underlying situation is changing, people innovating on that in the Indian context would. And so I think that kind of where I kind of hope that India kind of plays a role here is in recognizing that I think a lot of the most interesting innovation comes from a bottom-up understanding of what the problems are on the ground and a recognition that even scientific and technical challenges can look fundamentally contextually different in different places and that doesn't mean there isn't the capability to sort of share knowledge and collaborate between different places. Fundamentally, for example, power grids are governed by the same physics in different places. So, there's clearly some shared knowledge. But I think this paradigm of kind of developing AI needs to come from ground-up understanding and then collaboration that joins that knowledge rather than sort of a top-down paradigm where kind of a small set of people develop models and then we quote unquote specialize after the fact and as a result presumably miss I think a lot of the richness that we could have incorporated if we had empowered more people to contribute to this paradigm of AI. So I think this comes to collaboration and I think basically this goes well beyond access to kind of ensuring that many people are upskilled and educated and able to contribute to the ecosystem such that the models are built using their knowledge, their context from the get-go.

Thank you, Priya. I want to bring this conversation back to what Bonnie started off with. She started off with 540 million years back and talked about vision. I don't want to go that far. We could start from say the renaissance period when science really took off and we moved away from darker ages towards the more enlightenment, the more critical thinking, more reasoning, the scientific method, hypothesis leading to measurement and then building theories and so on. So in those periods there were many instruments that came around. You had the telescope, you had the microscope. So you could see into the stars, you could see into smaller things. But what is this AI thing? How do I even make sense of how does it fit within the paradigms of instruments like telescopes, microscopes? I want you to each one of you to put yourself in the shoes of the scientist here or the engineer here or the young person here who wants to be a scientist or an engineer and maybe if you can also concretize it with one of the things you're doing, Bonnie with the wonderful things that you're doing in your organization and each one of you in your things and then try to translate for them, how can I take this concrete technology or the models and things you're building and then how does it translate to hey if I'm a scientist how do I use the stuff you're talking about how do I even mentally map it to what I understand of a telescope microscope or other sorts of instruments and how's my life going to change, how does my life as a scientist or an engineer going to change if you wouldn't mind taking that dialogue and doing one more wave of things we'll do a couple of waves and then we'll come back to completely open conversation if that's okay. Bonnie I can put you on the spot.

Sure. Thank you. It's a great question. I think if you think about the space of discovery, there's discovery and there's acceleration. And I think the two are very complementary, right? If we can collapse and accelerate the time it takes to discover new molecules and materials, that'll help solve and also give access to a wider space of molecules, materials, proteins available to us. So one of the ways that we're working in the biology space is around protein dynamics. So of course we can use things like AlphaFold to help us predict the structure of a protein. But of course that's not the end of the story. We also need to understand the function and the dynamics of that protein. So one way in which you young people, experienced scientists, wherever, can help to bring solutions to new problems is to think about how we can define and use AI to look at the dynamics and the functions of biology. Once we understand that we can design new medicines to treat diseases and many more. Making these available to students and to researchers around the world is all part of what we're trying to do and I think it's really important to get involved in the validation of that. So building a model in isolation really is just a curiosity, right? So what we really need to do is see how these models can be embedded directly into the scientific loop. And so I think that's one way in which the community can get involved. It's not just about training the models, making very fast and accurate models. It's about actually how it gets integrated into the systems. How it actually gets involved, how it actually gets used. How does that output from the model become something that a scientist, a domain expert can make a decision on? And that's when we're going to start to see the real impact of AI across biology and medicine. Similarly with materials, there's an astronomical number of possible materials in the universe and we have only explored a tiny fraction of them. If we can use generative AI to explore new novel materials and then also use AI as a surrogate to emulate the properties of those materials, are we able to find materials that can help solve some of the challenges that we have today? In particular in India, whether it's solving sustainability or energy, how can we make access to new materials and again build that into the scientific discovery loop, build that into going from hypothesis to testing to actually having something real to use and impact the world.

Thank you so much. So for me AI is just a tool, always has been. There was science prior to AI as well. And I guess for me I've always approached my work as you start by identifying the need. So when you say do I need a new instrument? In my case we need better ultrasound devices. You start with that. Then you might be thinking about how you would then bring in the AI understanding current capabilities, evaluate it, understand its limitations and build on that as well. But an important part of any development, as an engineer this is the way I think, is you also bring in other people who can work with you on solving some of those problems. So if you work with clinicians, I work with people who do the ultrasound scans. The sonographers are there from day one. They're the ones who gather the data that helps us build the model. The engineers are the ones who build the models. We evaluate together and we iterate, improve. And I think that's the important thing and I think for me the exciting thing about now is everything has become very interdisciplinary. When I started out it was computer scientists playing with whatever data was available. Then we moved on to want the clinicians to acquire data in everyday practice. Much more interesting things you can do with that and then try and predict if they modified their practice how would they be able to make better decisions. So it partly the clinicians and for a few years I was working with them. Now I'm working with experimental psychologists and the importance of doing that opens up a whole new area, a whole new definition of what they consider to be academic work, science, how they publish, all the things that come with that. But they're very important because they understand if you put a tool in the hand of a human, can the human use it? Does the tool behave in the way it did in the lab but how does it change, coming back to what you started with, how does it change reasoning and human behavior and we need all of this to be able to understand can we translate AI into everything useful. So for those of you out there who are computer scientists, who do you know who are your end-users of what you're doing? What is the end use of what you're doing? And who have you talked to in experimental psychology and other parts of social science about the human aspects of what you're developing? This is the future. It's also starting to happen now and it's incredibly exciting. So I ask you to get away from your computer and go and talk to other people in science who are trying to make things happen.

What a wonderful message. Get away from computer. That's wonderful. Please Pushmeet.

I think I agree with Allison that it's really important to understand the problem deeply. Now going back to the question of what is AI for? Where can it make a difference? So AI systems have this amazing sort of ability to scale reasoning. So there are two sort of ways I've seen these systems really shine on scientific and engineering tasks. One is where they are able to extract meaning or extract patterns from data that from a perspective of a single human mind it's impossible to get to understand or even sort of see that data. So look at the amount of data that we as a scientific community are collecting about all areas of science. We saw Ashwini talk about the amount of data that CERN collected from the LHC. That is a huge amount of data. Similar sort of data sets are being collected in genomics. We are sort of sequencing across the world at a scale which has been unprecedented. With regards to the protein folding problem, we were able to sequence these proteins and find their 3D structure for hundreds of thousands of proteins. That was a lot of data, but it was still scratching the surface for truly understanding the problem. And what machine learning systems like AlphaFold have been able to do is extract meaning from that data and come up with sort of patterns to really predict the future. So from those 150,000 proteins, 3D structure of those proteins, it is now able to generalize to almost every protein on the planet, 250 million. And that's where the real democratization also comes in that once it has extracted that meaning and is able to predict the 3D structure from the sequences. Now every scientist whether they are working in Latin America or South Africa whether they're working on a neglected tropical disease whether they are developing an enzyme for decomposing plastics they can use it at a click of a button. So that democratization power of AI systems that they can be used at will and they have been able to extract this meaning is very powerful. The second way in which I have sort of seen this is the ability to search. So how many of you remember that AlphaGo match with Lee Sedol? How many? Yeah, a lot of you have seen that game. I was seeing that game and move 37 happened and all the commentators were like, 'Oh yeah, this is a hallucination. The model made a silly mistake. Haha.' And then as the game proceeded they were like wow that was a genius move that rewrote the whole theory of Go. So what are those move 37s in science and we are seeing those move 37s in science like we have a system AlphaEvolve and what it has been able to do is basically we have used it on different problems in computer science and in optimization and we have figured out that it is able to discover new algorithms for very fundamental problems like matrix multiplication. Imagine matrix multiplication, we teach it in school and it's the basic building block of all of even AI, neural networks are basically matrix multiplication and yet we did not know what is the lower bound in terms of number of multiplications that are required to multiply two matrices. And what a system like AlphaEvolve and these search-based systems are able to do is find these new algorithms that really expand what scientists know about the problem. So it's a great example of the scale of reasoning abilities of these systems and that's where the real power comes in.

Yeah, thank you. Thank you. That's amazing.

Thank you. I mean I'm the only person here who's not really working on AI but there are I think all of us are working on AI because either we are the product or we are the user because you get up in the morning and your phone is automatically telling you what to do and take your blood pressure or dimming the lights, driving us through the traffic. So we are in this ambient AI which is now becoming reality and then there is the proactive AI which we think about and for that I think we are pretty much in a space where we have to see who is defining those algorithms. So that's very important I would say. At CERN you mentioned about the data, I mentioned about the data already and we are not done yet, we are now multiplying the data and we are upgrading the collider where we will have 10 times more luminosity which means that in one year we will be able to collect all the data collected in the last 15 years. So now is a challenge for all you experts here on AI who want to help us in sifting through that data and taking us towards dark matter and dark energy that is not out there yet. Nevertheless, I feel that for us scale is very important for the experiments and then inclusion is very important which was already mentioned earlier. I would say that every student is an Einstein just needs to find the right trigger and coming from high energy physics, triggers we make at the micro and nano level at every collision so this has to translate into society and I think if we can do that we've been doing it in iterations, we've been doing it incrementally over the last 25 years. We've been collaborating with meteorology organizations at CERN, sorry in Geneva like the United Nations has a separate UNOSAT which is the United Nations organization for satellite imagery and they use most of the tools that we have developed to look for forecasting the weather and to look at monitoring as well and to look for sending help and assistance where there are issues on weather or any kind of natural disasters. So UNHCR, the high commission of refugees, they are also using algorithms that are coming out from CERN. So I see the interdisciplinarity not only of the subject but also of the object of what we want to do and definition of that is very important and certainly we can't leave that to AI to decide what we want to do as a race, human beings. Thank you.

Thank you. So Neil, and just to profess also you've written a wonderful book and you've mentioned in several of your public comments about how we should understand ourselves better in this world of AI. If you could also weave that into your comments that would be great.

Yeah, I think it's a great question and obviously the answer is AI is a catalyst because that's the name of the session. But how is AI a catalyst and how does that relate to us as individuals? Well, first a quick story that comes from the Renaissance, the story of the discovery of Ceres, which took place in Sicily by a very isolated scientist making observations and was only published in April by which time this dwarf planet, it was thought to be a main planet at the time in 1801, was lost. Now the planet had been lost and yet it was recovered. It was recovered because the data was made available and a young mathematician by the name of Carl Friedrich Gauss, people know how old he was? 23. So 23 years old and he was able to read in the journal that was published the data at the time. So it's 1801. So there's no canals or anything like that that the information is moving across Europe and people are attacking it as a problem. And then the planet was recovered a year later in another observatory institute. So they're using tools, they're using telescopes, but they're also the main thing and the key thing that AI is providing and it's shifting and why this is important to the individual is also present in a very odd thing on this stage. We have Microsoft and Google talking about science in terms of materials and in terms of creation of biology. Well, what's going on there? They're information technology companies. So, it's an information technology, right? And it goes back to what Wendy was saying with the early onset of the internet. What we're really looking at is the same thing as the printing press. The fact that those books were able to move throughout Europe and any interested scientist was able to see and access this material. And if they were curious enough and clever enough, they could solve where the planet Ceres was and become famous overnight. Now, we've all been isolated from that. That's really, really hard to do today despite the growth of information technology. So, there's something really odd going on there that a 23-year-old would struggle to engage with the science we're talking about in this way. And that's largely because that information has now been locked down in very few companies and access to those things is locked down. This is terrible for science as an institution. Terrible. Absolutely terrible. It's good for individual scientists. It's good for discoveries, but it's appalling for science. And there's a big difference between scientists and science. But the positive is as individuals, we can contribute. Again, I'm doing enormous amount more science as an academic now than I could before because I can use these tools to engage with different materials and search much better than I was able to do. But that's not restricted to just me. That's true of the 23 year olds in this room again, right? And we need to make sure that that's the story that's coming out because it's too much now about big things. I mean CERN is an amazing achievement but as you said it took a decade to pull that together right and 10,000 people on the final paper. So that's science at the scale that has gone quite a long way beyond what Gauss was able to access. But what this can mean for individuals is getting back to doing science at the scale that I think was also mentioned importantly by Allison that you actually want it to be the experts who are working on the ground who care about the problem getting access to these capabilities and this technology gives us that chance and that's why it's a revolution because it's an information technology revolution but we have to be very careful because just as Wendy was saying information revolutions have a tendency to concentrate and that's very very problematic for India, countries that don't have access, for Europe, for a number of places that don't have access to those companies that are dominating on the information technology space. And I see that as probably the biggest threat to us moving forward as a society, unpicking exactly what Wendy was talking about, the recovery of that independence, that vision of the internet as a way of individuals communicating and participating rather than an entity that is dominated by a few large entities.

Thank you. Thank you. Yeah, Priya, go ahead, please. Thank you.

All right. Yeah. So, I think that was great, Neil. Also, thank you. So, kind of a different angle I'll bring is something echoing what some others have said on this panel, which is I think that in sort of engineering and scientific discovery, some of the things that really AI gives us is it's scalable. It can run really fast and in principle it can run 24/7. And on the one hand, if not done right, that is absolutely terrifying. Right? If you do things at scale incorrectly, unsafely, unfairly, in an unaligned way, that is terrifying. But if we're able to sort of by design ensure that AI is kind of working in a way that we want it to, that in a way that is physically feasible, that is safe, that is responsible, etc. Then I think this is game-changing right because again kind of reflecting on power grids right now in power grids things are incredibly manual. So even though we think of it as this engineering system that in principle could be optimized, in reality what's happening is that when we think about should a particular power generator turn on or off, someone sitting in the dispatch center is often making a phone call to somebody actually sitting at the generator to say can you please turn this up or down? And this is happening in part because we just again aren't able to deal with the scale and complexity of the system. And on top of that, there are lots of things that we wish we were doing on these systems that we're not because humans can only act so fast. I am not able to act at a millisecond level to control a power grid. There are certain things that just need to happen much more quickly. And so I think what kind of AI does in this sphere is on the time scales where we are able to do things today, it kind of shifts the role of the person there to not having to do things like make a phone call to dispatch power, but maybe look at kind of more nuanced things on the system and control it. And it also enables us to do things we weren't able to do like put more renewables on the grid which requires us to manage power grids much more quickly and much more scalably which we don't do right now because we're locked into very conservative paradigms of making sure the power grid has plenty of infrastructure, plenty of buffer such that a human can react to it. So I'd say again the things here are kind of the speed, the scale and the 24/7 nature. Terrifying when done wrong but I think really powerful when done right.

Thank you. Thank you. I mean there's so much amazing set of perspectives. The next set of questions I'll have maybe go the other way around so that it's not the same way. Or we could randomize as we like. Is science connected to societal problems. We talked about your opening statements about what you're doing. We talked about taking the persona of the scientists or the engineers and what is this AI, what kind of tool is it and so on. Now when I look at societal problems, let's say pollution or climate change or health, education, or eradicating poverty or other sorts of things. I mean...

I don't have an AI button that I press and I magically solve this. I mean clearly humans still have a role to play. How do you see the diffusion of fundamental discoveries in science? I mean the way it's happened in the past versus the way it could be in the future. Pick any societal problem that you're interested in. I assume Priya you'll pick about climate change and so on and say how should we think about climate change? You talked about power grids and are there other aspects of climate change similarly Bonnie in terms of say I mean I know your team has done wonderful work on Aurora or you know rethinking how DFT should be done or material science or bioimaging or I mean each one of you for your various areas. How do you pick a societal problem? How should I as a scientist or an engineer or teams of these think about navigating or thinking backwards from defining the societal problem and breaking that down into scientific or engineering problems. How should I collaborate? I'm trying to be as concrete as possible and I would encourage you to be as concrete as possible and also since many of them may not know especially what you're doing if you can take a specific example of what you're doing and help us understand. So maybe Priya if you wouldn't mind I can start with you first. How do we solve climate change?

Oh gosh that is a much bigger question but okay so yeah so thinking about the broader climate change problem which again it's a very big challenge. It involves I've talked about power grids but it also involves decarbonization in other sectors. Some of them are energy related like buildings and transportation. Others are related to land use like agriculture and forestry. And then of course this also requires large-scale efforts to adapt to climate change. The effects that we'll face through kind of public health assessment, disaster prediction, biodiversity monitoring and all these sorts of things. So all of the problems I've mentioned, right, they're clearly not just sort of technical and scientific, which I think Shiv's question was also queuing up a bit, right? They're also deeply societal. So, how do we break them down in a way that both kind of yields some tractability to how we think about the problem? Kind of forms it and frames it in a way that kind of a scientist or engineer can grasp, but also doesn't lose the context and richness that can then cause further lines of scientific inquiry to be divorced from the actual needs of the on the ground problem. So I think there are lots of ways to think about this, but one kind of paradigm I'll bring this up is the paradigm of benchmarks in machine learning. So benchmarks have been kind of a really big part of driving progress in AI and machine learning where we basically kind of specify a sort of set of data, some tasks, some metrics and then basically try to push forward progress in that setting. I think benchmarks have been incredibly powerful. They sort of have codified and unified efforts in a way that's really helpful. But I think also the way we've done benchmarks has been really limited in the past in the sense that there's kind of a phenomenon where you frame a problem that is hard to solve at a particular point in time. The field makes progress on it and then people keep working on that version of the problem even though in some sense we've exhausted the practical actual progress on that version of the problem. I think we also tend to kind of write down metrics of success as being sort of singular. So did we get more accurate on this benchmark for example. But in reality when you actually deploy solutions on the ground different things matter in different settings. So for example maybe in a power grid example again when we're trying to learn fast approximations to power grid optimization models we want them to be fast. We want them to be accurate and we want them to be robust to kind of uncertainties in their inputs. But how much each of those things matters differs depending on who the user is. And so if we evaluate our models based on just one metric and we push to improve only that one metric, we miss the richness of comparison there and we risk people kind of all trying to drive forward progress on one thing which is not actually capturing the scope and richness of the problem. And then I think the third thing that gets missed is when a benchmark is defined, we're obviously making lots of assumptions about a particular problem. We are kind of making some assumptions also about how one might then kind of integrate that into society, what the pathway to deployment is. And I don't know that we're particularly principled about kind of writing those down or acknowledging kind of what the limitations are of what this problem formulation and model development tells us. And that means is that when we've quote unquote solved the benchmark, we often then also don't know what the next step is. We again kind of get to this place where we're sort of divorced from reality. So, I'm using benchmarks as sort of a maybe like an organizing frame to think about some of this, but some of this isn't limited to just benchmarks. But I think to summarize it comes down to kind of understanding that progress is sort of a multimetric thing where kind of it's not like one singular direction but we are often needing to make progress in multiple directions and we should acknowledge the richness of that and we should kind of write down what assumptions and approximations we make about and incorporate knowledge about what it means to actually deploy a solution. And we should also have a sense of when our approximations have reached their limit of usefulness and when it's time to kind of move on to a new formulation of the problem or something that better captures the richness.

Thank you. Yeah. Would you like to take that?

It's a great question and something that we've really obsessed about particularly over the last decade as it's been very clear these technologies can come out. The public dialogues we I've been involved in public dialogues. Wendy mentioned one report, but we've been doing them again recently. So, trying to refocus. I think we have to put our hands up and say, you know, the set of societal problems that we're looking at, they're very old and long-standing and they haven't gone away. And people are not stupid. They've been asking us to solve them for a long time. So, if you do the public dialogue, you say they'd like solutions in healthcare, education, social care, etc. So, then there's a major question, well, why haven't we done anything about it? And in particular with AI because if you look at the report Wendy mentioned, we did public dialogue then and they said we'd love you to solve all those problems. The one thing we don't want you to work on is art and the one thing we've made undoubted progress on without solving any of those problems is art. So in a very real sense, we're not listening to people and their problems as scientists. And I think that's partially this abstraction thing that Priya has mentioned. It's far easier to abstract a problem away into something we can solve and then shout about it. Well, to a large extent, that's not helping. That's not helping people in their regular lives. So, what we've tried to do is switch the model and do something we call the attention reinvestment cycle. And we think we should be owning up as universities and saying we haven't done a good enough job on translating innovation from our labs to things that mean something to the, you know, I don't know, a person on a bus in Hyderabad. Right now what that means to get through this benchmark problem that Priya is highlighting is you have to get down and dirty with the issues they're actually facing. So rather than telling them that we've solved a problem and deploying it for them as a solution, we're trying to flip the model. And in Cambridge with AI at CAM and also the accelerate science program and also learning a lot from my colleagues in data science Africa who've been doing this beautifully for 10 years. We're realizing you actually have to sit down with the people who are having the problem listen to what the problem is and start working with them to solve it and ideally and this technology is allowing that ideally give them the tools to solve it themselves. So what we're doing at the moment with one of our programs we're working with local authorities who've already formed AI clubs. We're listening to what their problems are. We're trying to then sit with them and propose a few things, but then often just introducing them to each other. We have reverse poster sessions where they show the poster session with the problem in and most of the time they can solve each other's problems because there's such a radical change in the information infrastructure. Everyone is having the same problems everywhere. And I think the challenge is that the nature of those problems is not percolating up to the universities to the big tech companies and we're solving for things we can either make money for or we can get on the front page of Nature for and that isn't helping and that's leading to a rise in populism and a rise in people pushing back saying what is this all for because it's not doing something for me in my daily basis and I think with such a transformational technology it is time to reassess that and start saying how can we improve some of what we're doing and actually address the problems that real people are having and again India is in an amazing position for this you have distributed universities across the country you have a large number of challenges that are now offering new solutions through mobile phones like my colleagues in Africa that's why I find Africa so interesting not because I'm helping you know not because they're helpless they're solving their own problems in innovative ways that would blow your mind if you're sitting in the UK trying to solve those problems and if we can transfer some of those solutions to the UK then we have success.

And sorry can I just interject really quickly and at next year's summit for exactly this reason I would love to see the stage representing many more Indian scholars based in India talking about these topics as well. Thank you, thank you, inspiring.

Okay, it's working. So yes, in the ways I think how we have been defining problems or rather picking problems which are existing on the globe in itself. You know there's global hunger, there's energy, there's the climate change, the ones we all been hearing all morning. Now at CERN we do have a section which is recently set up 10 years ago CERN and society and we have a lot of problems that we go there and then we convert them into competitions for everybody to participate and typically you know CERN being in Europe and access being easy and like you Priya just said that you know Indian students are amazing but they do not have the means to get there right? They need to be empowered. But with the WWW first and then with all the AI tools that are available any student sitting anywhere can be linked to the problem and this part is very close to my heart I can tell you I'm working a lot with children not children students high level engineers technicians and trying to move towards capacity building and that I think is our responsibility we should not forget that so any tool that is available today at CERN except like superconductors because it'll be hard for us to ship them down here. But that knowledge had the ability to create a 28 billion market for MRI machines, imaging and so on, cryogenics. So I want to link those and embed that into universities, schools, institutions solving one problem together. Thank you.

If you can also add a little bit about that Hadron therapy you do. I mean we think of this fancy particle accelerators huge loops but how can it actually help solve?

Thank you. So you may have heard about radiation therapy for cancer patients of course. You may have heard also about proton therapy hadron therapy. This is the next big thing which actually targets tumors without damaging the healthy tissues and this has become go-to technology in the west if I can say so also in Japan and there are 130 centers all over the world and in India where the concentration of cancer patients is much more if you look at it from the if you look at a graph of the world we need to make technologies for cancer patients accessible and affordable. So the good idea would be to come together and design something on the basis of the next generation of these particle therapy centers which of which do not exist yet because all the six centers in the world are using technologies of 20 years ago. It takes that cycle and AI can perhaps help us in accelerating that cycle and moving forward. Thank you.

Yeah. Yeah. Pushmeet if you could also add on you've solved protein folding, right? How can AI then help further discoveries and then help us cure cancer or one of these wonderful things that Deb keeps talking about you know when are we going to cure cancer and when are we going to get on with it.

Yeah. So I think like reflecting back on what Neil was mentioning as well, I think if you think about AI for science and if you think about science generally one aspect of science is basically making sense of the world really a science for the sake of science science for the sake of understanding but then there's another aspect of science which is science understanding for the sake of solving problems and I think in AI we are seeing that AI is delivering on problems that are actually having impact. I mean you mentioned AlphaFold and AlphaFold is basically a root node problem because solving protein folding does not just get you a Nature prize a Nature cover or a Nobel prize. It gets you more than that. The fact that 3.3 million people have used the AlphaFold database was a surprise to me. If you had told me this when we were starting the project, I would not have believed you that there are 3.3 million scientists who have used AlphaFold. Those people are spread across the world. 180,000 of them are in India. These people did not have access to the cyclotrons which could have done the structural prediction for the samples they were interested in. Now they can on the click of a button and that really accelerates not just one particular area like drug discovery. It accelerates a whole span of things which require understanding of biology and biochemistry. There are people and scientists working here in India on understanding how do you improve soya bean crops and make them resistant to fungus, how do you increase crop yields all using AlphaFold. So in some sense we are already seeing these scientific models have impact across socially important problems. In the case of weather weather is another weather and climate like we were coming back to that sort of topic. Our new generation of weather models they are able to make predictions about weather in the short term in 5 to 10 days or 15 days in 1 minute much more accurately than what a supercomputer would have taken a whole day to predict. And now these machine learning models for weather prediction are becoming so good that in fact in the last hurricane tracking and cyclone tracking sort of cycle they were the model of choice and were able to literally save lives in terms of giving better and faster input as to what will be the cyclone and hurricane tracks. So AI is delivering it is delivering on problems which is going to have impact. Now and then the final sort of thing I don't think this work is having impact because it cannot have impact by staying in a silo. It is having impact because these models have been made accessible that 3.3 million people have used AlphaFold database. People can access those cyclone and weather predictions. Another example is our work on AI co-scientist and agents for science like AlphaFold. And this is where these agents are being used by scientists themselves. Some of the best some of the Terence Tao who is one of the I would say arguably the best mathematician living on the planet today. He's using these AI agents to further his maths research. We are seeing AI co-scientists being used by scientists to come up with new hypotheses for problems like antimicrobial resistance. Antimicrobial resistance is a key problem that is facing the on the societal scale the health of the planet. So we are literally seeing these models having impact on these problems and I think that's what will happen in the future. But again this all needs to will happen only if scientists have access to it and this is not the job of one person or one institution. This is the job of society at large. Governments need to be involved academic institutions need to be involved. Social scientists need to be involved. Industry needs to be involved in making sure that this very powerful technology is given to people and scientists around the world.

Thank you. Yeah, if I can please. So just 30 seconds on that that peaceful international collaboration is what has to come out of all our deliberations and in principle scientists who are also diplomats to democratize all the work that we are doing. Thank you.

Thank you. Thank you for that message. Please.

And I definitely agree with that. What I'm going to say though is AI models depend on data. And one of the biggest problems in getting science to solve some societal problems is we don't have access to the right data. Working in medical imaging, it's already been touched on. If we restrict ourselves to the small open data sets and building AI models on that, we do not solve real problems. We do a proof of principle and there it stops and then people go on to the next one. So then people say well provide access to all hospital data please just open the gates and of course you can't do that because of the different data privacy laws and the respect of health data. So what do we do? We've come a long way. When I first worked in an international study on pregnancy ultrasound, this was in 2008. How we in a research study, we have ethics approval for this. You shared data by USB stick. They were sent my lab was the central center for this and it was tedious. You had to triple copy everything, bring it together. Of course, things have moved on. But in healthcare in particular, it's sensitive data and in other areas of science, we've also advanced in the way we can build models from sensitive data through things like federated analysis. We've come a long way, but the bottom line is we do need the individual science communities to decide what data they will share, provide access to so that we can all collectively solve the right science problems. This may sound easy to a computer scientist. You should be able to just do it. It's complicated. It does require policy makers. It requires people to be able to provide that access under conditions and governance models. But it's very important for the future of using AI in science that we find better mechanisms to do that. It's not one solution either, but we work together so we can either share or provide access so we can build models that can really address some of these societal issues. And it's costly to do this. We need to understand we need to pick the problems together a few problems to start with to be able to come up with being able to do this maybe with some then leading to some new standards and ways that scientists can work together globally but that's what's holding us up at the moment.

So I just wanted to start by reiterating what Pushmeet said about democratization. So making models available not keeping them behind closed doors so that people can use them, society can use them, governments can use them. Our model at MSR on the climate. So we have Aurora which is a foundation model of the earth's atmosphere. And that's able to and as he said there are a lot of machine learning for climate weather prediction models and they're getting better and better. And what it does is enables us to forecast much faster than traditional methods. And one of the greatest benefits of that is preparation. So if societies, if communities, governments have access to these tools and they can predict when a major climate event is going to happen, we can prepare. And ultimately that helps to save lives. So I think it's incredibly important that we continue to build on these models. We continue to make them accessible and continue to grow in that area. The second area I wanted to talk about, the second societal problem I want to talk about was making medicines. India manufactures a significant portion of the world's medicines today. And it's no secret that medicines are extremely expensive to make. It's very costly. It takes a lot of time in the lab. It takes many experiments, lots of human resources, scientists that are trying out experiments and oftentimes failing. So it costs billions of dollars and once we get to clinical trials typically we tend to fail more often than we succeed. The industry has made a lot of progress especially with large molecules. So looking at antibodies and how those can treat disease. But of course large molecules can't treat everything. So there's still a need for sort of traditional very small molecules to treat as medicines. However, you know, collaboration is really central and core to what we do probably in any tech company or in academia working with industry. And you know, some of the partnerships that we've had at MSR with pharma companies have really shown that there is a bit of an existential crisis when it comes to small molecule discovery such that it's so expensive and takes so much time to discover small molecule drug candidates that oftentimes it's just easier and faster to just focus on the large molecules. But if we were to do that as a society, as industry, then that would cut out a significant number of treatments for diseases that people need access to. So it's incredibly important. This is a problem that we need to bring front and center is how do we actually make drug discovery more accessible, more efficient? How do we find ways in which to reduce the time it takes and improve the probability of success when it comes to clinical trials? So that's a problem that we're working on as well. There are many bottlenecks along the way when it comes to making medicines. Essentially every step along the way could be accelerated using AI. So working directly with the people involved with the scientists with industry is incredibly important. And making access to the cycles and iterations to find new medicines. One of the biggest problems and challenges that we have is not just how to design the molecules but actually how to go and make them. You can imagine a world where you could design a perfect molecule that might attach wonderfully to the target that you're trying to go after in biology, but then once you get to the lab, you can't actually figure out a way to make it. Or you can figure out a way to make it, but it's too expensive to make, which again drives up the cost of medicines. So I think AI is a great way to solve some of these challenges. And that is one area that we're working on as well in particular for reaction prediction and synthesis.

Thank you. And as we come to the final stage, I'll do a quick lightning round on a couple of quick topics that I had. Finally, one is as we are all sitting here, what skills does a scientist pick up? We think of first principles thinking we think of broader societal understanding how to work with AI disciplinary skills collaboration both with people and AI models and so on so a quick round on 30 seconds each you know what skills should a person pick up anyone so this we'll keep a random round whoever wants to pick it up please.

Maybe I can start. I think one of the most important things is understanding of these tools so use them. These tools are not perfect. They will make things up and you need to be aware when they make things up because that's where the responsible use of this technology comes in. And one other sort of key point is this does not mean that now that you have access to this AI tool the expertise is not important. Expertise is even more important I would say now because in order to sort of really think about what is the right problem to ask that requires deep expertise. So these AI systems are able to help you solve those problems. But who comes up with that problem? Who comes up with the right formulation of the problem? That requires deep expertise and that requires deep interaction with how do you use these tools. So my advice to everyone in the room would be first understand the deep problems that are going to have impact in your field of study in the field in the impact that you want to have in the world and then use these tools because they will give you superpower but they will also have weaknesses and try to discover those weaknesses and I think those are the two very important skills that all scientists will need to come up with.

Thank you Pushmeet. Yeah, please.

Yeah, listen. Yeah, you go next.

I think the most important thing to do is be a team player. You don't have to be an expert at everything. You won't be an expert of everything. You've got your whole career to learn different things and that's what makes it fun. So, for me, the answer is I think if all of us when we started out, we had certain skills and we've certainly evolved those with times. So, it's about your ability to work with others. Challenge them but so be a challenger but be prepared to be challenged yourself contribute and take that forward and that's all about working that's what interdisciplinary scientists do so that's I think come to the table with your expertise and then it will evolve with time.