[0:00:17] All right, everyone, welcome. Thank you for coming and joining us also over livestream today. There was a lot of interest in this event today for, I think, good reasons. We had over 4,000 people trying to get into this event, which was fully booked within 30 minutes. So we are now full house here in the lecture hall in Audimax, and I'm very, very happy to see everyone here today.
[0:00:43] I'm very excited to welcome our special guests, our panelists, and all of you today here. We are bringing together, obviously, experts in leading AI to discuss the transformative power of AI, both in academia, industry, but also in society. So my name is Fatma Deniz. I am a professor in computer science at Intersection to Neuroscience and Artificial Intelligence.
[0:01:10] And also the Vice President of Digitalization and Sustainability here at the TU Berlin. So I think this event is basically also thematically a highlight for me. And I think also it's a fantastic way in some ways to wrap up and conclude our semester, which is going to be next week towards the end of the week.
[0:01:40] So today we will discuss the many potentials and challenges of AI. Among others, we will talk about how we can harness AI for the greater good and also while we are mitigating its risks. So with this very few introduction of words, let's drive into the conversation and please welcome our panelists, our guests to join me on stage. Sam Altman, Volker Makel and Nicole Bettner. Please come.
[0:02:22] So welcome. I would like to briefly introduce our guests. We have a true heavyweight in the world of AI and startups, to my direct left, Sam Altman, the man of the hour, basically. From his days at Y Combinator to leading OpenAI, Sam has spent his career in exciting and irritating, and as well as, I think, motivating and challenging. It's his colleagues, competitors,
[0:02:52] Scientists and the media alike. Please welcome Sam Altman. Next comes Volker Makl. I call him our data wizard, professor in computer science here at the TU Berlin, the chair of database systems and information management, chief scientist at German Research Center for AI, as well as the co-director at Beifold, one of the six national AI institutes here in Germany.
[0:03:28] Folka is a driving force behind some of the most advanced data processing and AI infrastructures out there. So please welcome Folka. I'm also thrilled to welcome a powerhouse in AI, economics and innovation, Nicole Bütner, is the CEO of Merantix. She is an economist, an entrepreneur, an AI expert, an investor. All of it actually while making cutting edge technology understandable.
[0:04:07] for businesses and also policymakers. Please welcome Nicole Bytnow. Okay, so I guess I'm wondering this question because maybe I catch myself doing this when I use chatbots. Are you actually kind to your chatbots? Do you ask questions like, or do you end your questions like please and thank you? I do. You do?
[0:04:41] How about you, Volker? I used to do it, but most recently I think I'm not doing it that much anymore. I also do it, so I don't have any rationale for it. Yeah, I was very ambivalent with this as well, but it's coming back and forth. Okay, so I guess, Sam, these are quite busy times for OpenAI and especially for you. All the more we are very happy to welcome you to TU.
[0:05:11] So let me ask you this. Chatbots like ChatGBT are used today for inspiration, for drafting texts, and also summarizing, right? Basically, it helps us to save some time, also, you know, cut corners. But what do you think does this mean for scientific and academic work? So what if basically the most insightful ideas come from when we take our time and not cut corners?
[0:05:39] We launched something on Monday called Deep Research. And Deep Research is, it's been sort of like almost a chat GPT moment in how people are reacting. And scientists are some of the people that are most excited about this. You can give Deep Research a query, and it can go off and do a bunch of research on your behalf. It can do tasks that would take you many hours, days, sometimes weeks. It's not going to...
[0:06:06] At least in its current form, it's not going to go generate a lot of deep new insights. But in terms of pulling information together to let you do what you can do as a scientist or researcher, it's really great. And that's kind of how I think this is supposed to work. AI does help us save time in some areas, and then we can work at a higher level. This is sort of the history of humans making tools. We let you do more by having the machine do things that you had to do before, and then you're freed up at a higher level of abstraction. And I think it'll all be great.
[0:06:36] Okay, thank you. Volker, how do you think could deep research, for example, or in general generative AI transform science and research? Yeah, so there is many different aspects on that, right? So the first one is, of course, transforming natural sciences and so on. So one of the important part is arriving at foundation models in that area, right?
[0:07:02] Foundation models there, because a lot of the data there is quite of different structure and modality, right? So the training and the integration requires still experts. So the question is, how can we do that better? But I think that we already see first steps towards that in areas like, for instance, pathology or earth observation or others where those foundation models are.
[0:07:30] In the making and can be used. So there's a transferable effect on the sciences in that respect. Of course, immediately, you know, like if you talk about computer science or so, right? So one of the things that a lot of people use it, as I said, to cut corners or to expedite things, right? Co-pilots, right? This is an important aspect here. So really using the tools in order to...
[0:07:53] Improve programming, right? So this is, for many scientists, you know, not having to think about complex Python logic or things like that and doing this in the copilot, I think, is a huge value add that really is changing the sciences. This is not so much true.
[0:08:07] If you do system sciences, because there, the code is more, if you wish, in the long tail of the training, right? And it's usually not so much part of the foundation model. So if I want to, you know, like for the computer scientists among us, if we want to build a new database system, stream processing engine or whatnot, so that's obviously harder. But API integration and standard analysis tasks, it's a very powerful tool. So in this way, it's clearly transformatory. Also, if we look at, you know,
[0:08:37] Other types of sciences like humanities and whatnot. Training, or actually in this space, adapting the models with respect to texts, with respect to data that is specific to the domain, which is really what's happening, offers a lot of opportunities for...
[0:08:59] Question answering in areas and I mean we still haven't fully understood emergent behavior in this case right so there still may be some interesting things coming out of that as well. I think I mean with sort of the possibilities of like lab in the loop and just the acceleration of experimentation I think what's exciting me is that then these models can contribute also to knowledge discovery right and I think that's something I'm really looking forward to.
[0:09:27] I think we're within a couple of years now of a major discovery being done by AI autonomously. Yep. Cool. Totally. But maybe also from the business side, did you think that the ChatGPT type of chatbots basically would be such a game changer also in companies? In what? In companies, in what way? They were a game changer in this sense.
[0:09:56] I think it's an interesting question because we're seeing a lot of adoption. We know German businesses are actually very curious about this, using it a lot. And there are first experimentations, I would say. We see this a lot with rack systems, and I think that's great. We need to start this adoption. I'm personally more excited about what's coming and what's ahead when we're really looking into core business applications and getting that value started.
[0:10:22] And I think there will be a whole ecosystem that will be necessary to do this.
[0:10:28] There will be people using GPT or Copilot for this, but there's also another level of interaction with the technology needed. So I think we are just scratching the surface at the moment, and I think we need to work on things like data and capabilities and tech, also speed, maybe a little bit, to get this actually into business processes and away from just, say, rag systems. And I think that's really exciting.
[0:10:55] So maybe just to talk about that from the university perspective, right? So our customers, if you wish, are the students, right? And there's, of course, also interesting opportunities that one can explore in how to leverage large English models for students, right? Or in general, foundation models. And, of course, I mean, the obvious use cases that students use nowadays that are text summarization, and then we have to look at when, you know, some...
[0:11:24] Questions are answered by GPT. How do we ensure that the learning effect for the student is still there, right? However, what is really interesting is to have a chatbot as a tutor, right? So a really individualized tutor that teaches, like in our case, maybe some database systems content and you can interact with it based on your individual learning abilities on the one hand and also based on
[0:11:54] You know, the focus of the course. So I think there's lots of potential in that space which will eventually revolutionize teaching to some degree because it will relieve the scarce teachers and TAs and let them focus on the more difficult stuff. So I think this is a great opportunity that we have to seize and I think some people are seizing it already.
[0:12:13] Yeah, we will probably come back towards the end also about, you know, what is the risks in the society in these things? What are the fears in this aspect? But I would like to maybe go along some of the news that we had today between OpenAI and Bifold. You have just started basically a research partnership that will allow...
[0:12:37] OpenAI to provide 50K worth of API credits to facilitate the exploration of the O3 model in this case. And in tandem, you're looking into many new research opportunities together. So what are your expectations when you partner with such a foundational research organization? There are many reasons I'm excited by AI. We touched on this a little earlier. The most, personally, the single thing I'm most excited about.
[0:13:07] Is what this is going to do for scientific discovery. I believe that if we can accelerate scientific discovery, we can do 10 years of science in one year, and then someday 100 years of science in one year. What that'll do to quality of life, like solving our most pressing problems, addressing the climate, making life just better in all sorts of ways, curing disease. That will be an incredible gift. And I think AI is finally going to enable that. How many people feel smarter than GPT-4? Okay, how many of you think you're still going to be smarter than GPT-5?
[0:13:40] I was expecting more hands here. I don't think I'm going to be smarter than GPT-5, and I don't feel sad about it, because I think it just means that we'll be able to use it to do incredible things, and we want more science to get done. We want to enable researchers to do things they couldn't do before. This is the long history of humanity. It does feel a little different this time.
[0:14:10] Because of what this can enable. But if scientists can do things because they have like a crazy high IQ tool. And they can focus more on figuring out the right questions to ask. Address things quicker. Do their search space faster. That's just a win for all of us. So we're thrilled to get to enable that. And those two people that said smarter than GPT-5. I would like to hear back from you in a little bit of time. Yeah so from our perspective. So on the high level of course right.
[0:14:38] Collaborating with a company always grounds the research to some degree, which gives it more practicality and helps to focus on some actual real-world problems. And there's always a risk if you work on problems just from academia, that you are maybe researching in a direction that...
[0:14:57] might miss the core angle of practicality. So that's one aspect. The other one is, of course, access to resources, access to the models. So those are hugely valuable. And if you look at bifold specifically and the collaboration specifically, then there is, of course, we are doing research in applying models to the sciences. And we already talked about that before with Sam, you know, foundation models in pathology, foundation models in quantum chemistry and others, right?
[0:15:24] How can one leverage the systems in order to, for very interesting scientific disciplines, which is one of the core thing of Beifold, come up with novel insights, right? There's a huge promise for personalized medicine, there's a huge promise for improving care, and if we talk just about
[0:15:43] I'm saying that because Bifold is both T-Berlin and Charité, the university hospital, so there's a lot of opportunities in the healthcare domain, right? And we are just starting to touch those. On the other hand, Bifold, of course, we're also doing systems research, and I'm a data management person, right? So that is, you know...
[0:16:00] Data curation, data integration, just the preparation steps. Because if we look at how our model is trained, right? So we have all those different steps. How can we make those reproducible and documented? And with Sam, we just talked about how you can explain things and the features that are coming out there, right? So, I mean, there's a lot of, on the one hand, need and on the other hand, opportunity to look at the entire process that includes both systems to make things scalable. And this is one thing we are doing in BioWorld, conducting research on.
[0:16:29] More clever algorithms on systems architectures, on also programming models that help us with efficiently writing AI applications. And on the other hand, of course, intelligent algorithms, which is really the machine learning side. And as we all know, and probably we'll touch on that later, training is still quite expensive. So can we...
[0:16:51] Even while maintaining large data sets become more efficient because, of course, we can train on small data and do model composition, but training on large data, and I'm not sure how you see that, Sam, but I think the emergent behavior that we want to, which have, of course, promise and fear, I should argue, but that we still find very exciting, so those we probably will only observe by training on large data sets. At least that's my expectations that it's hard to get those out of these small data sets because then you kind of...
[0:17:20] Construct explicitly already what you want to achieve as opposed to getting this implicitly inducted behavior. Yeah, so I already hear a lot of research opportunities for the further partnership. That's amazing. So maybe moving on to AGI, Artificial General Intelligence. Your former colleague Sam, in this case, Dario Amadei, recently wrote an essay.
[0:17:46] Where he talks about his vision on next generation AI, basically AGI. And he mentions referring to AGI, he says it's basically a country of geniuses in a data center. So do you agree, first of all? And what does this mean for the next generation of scientists in general? I think at this point we're kind of close enough that the precise definition of AGI matters. You know, by the time you have...
[0:18:15] A world expert in every field, working together tirelessly. I think that's beyond what most people would consider AGI. So I think rather than talk about, you know, when we're gonna get there or whatever. Sorry, that was really inarticulate. I would say, I think we'll get to something in the next couple of years that many people will look at and say, I really didn't think a computer was gonna do that.
[0:18:43] And then I think we'll get to something that can do the scientific progress question. You know, 10 years of science in a year or whatever. That's further out, but that will be the moment where I think the world really kind of changes much more quickly and gets huge benefit. It does look like we're on a pretty steep trajectory. There was, you know, people love to be the doomsayers, so there was all this stuff last year about scaling's over, this is, you know, not gonna work.
[0:19:11] And we found a new paradigm, and we now have these reasoning models, and they are really smart. And that's going to scale for a while. And then I expect after that we'll find another paradigm again. So one thing I've learned in general is whenever you see these steep exponential curves in technology, you shouldn't bet against them. And I think you should all be very skeptical when people start saying this is about to run out, or that's about to run out, or we're going to hit this limit. It does look to us like we have this fundamental unlock of an algorithm that can really truly learn, and that's going to keep going.
[0:19:43] We will hit stumbling blocks. We'll have to figure out how to overcome them. But I think we will get to AGI and beyond, and it'll just be a pretty smooth scaling from here. Okay, maybe a follow-up question on that. Who would be, in your view, currently best equipped to independently vary claims of achievement of AGI? The only thing that matters is the amount of utility it provides to people. So if it's useful to you, if it's useful over some threshold, if it's useful to you, great.
[0:20:16] And if a bunch of experts want to debate if it's AGI or not, I don't think it matters at all. And again, I think it is going to be just this continuous exponential and this continuous rising utility. And the only question is how much is it helping people, how much is it helping the world?
[0:20:34] Would you agree? Well, I actually would agree. I think also it's probably a moving target, right? So we will see that over time as, you know, the technology progresses, right? AGI will move forward and forward, but indeed the usefulness metric that Sam just mentioned, I think to me is the key thing here. And I think that's also actually something that we need to deliver on more as the tech industry. And because we're debating a lot like about...
[0:21:01] What is happening? What are the newest developments? And obviously we're all curious about it. But I think for a lot of businesses, right? I mean, we want to put this, we're building this technology to put in the hands of teachers, nurses, business people to actually solve and help solve the challenge of the time labor shortage. Many, many services we need to deliver with increasing financial constraints, time constraints, like solving ailments and so on. So I think that's something that I'm also like...
[0:21:30] Even more interested in, when we get there, we need to be ready to bring it also to the people in a very useful manner. So this translation mechanism and this bridging, I think, is a really crucial part to definitely have fixed by the time we reach AGI. Basically, we can agree that the individuals who are using and getting the benefits out of it are the ones who could be the ones who can...
[0:21:58] Judge the system as well in this case for their general intelligence or not. Okay, switching gear a little bit towards investments in AI. There have been exciting times also in U.S. OpenAI has recently announced a big Stargate project. A new company that will invest around 500 million U.S. dollars in four years. I think to secure the... Maybe we'll get it in three. Okay. The American leadership as you call it in AI.
[0:22:28] And there are big players like Microsoft, Nvidia, and Oracle among them as well. But basically, apart from the big pile of money, what will Stargate, do you think, enable that OpenAI could not have done without, in this case, the joint venture? Yeah, there are two things. One, with a much bigger computer, we can train much better models.
[0:22:58] You know, I think the world doesn't want us to stop at GPT-5 or GPT-6, but would like to see that keep going. And I think Stargate will enable us to stay on the same curve of improvement we've been on for the last few years for the next few. And should get us to, I think it'll get us to extremely capable models. The other thing is people really want to use these systems a lot. As we make more progress, we continue to drive the cost down.
[0:23:28] Roughly speaking, every year we can take last year's intelligence and make it 10 times cheaper, and then 10 times cheaper again. Moore's Law changed the world, doubling every 18 months. This is a 10x every 12. It's a super powerful difference. But when we do that, the demand to use those services seems to go up way more than 10x. And also, as we train these truly capable models that can use a huge amount of computer inference time, you might really be willing to go use quite a lot of compute to cure cancer.
[0:23:58] And so the amount of compute we think the world is going to demand to run these services is going up pretty steeply. And if you look at our own inference consumption, it's going like that. So that's the other part of the project, is to enable people to use this. Hundreds of millions of people use OpenAI's products right now. It'll be billions soon. And they use them a lot, and they want to use them even more. Thank you. And Nicole, maybe looking at this target project, what do you think to European companies?
[0:24:29] Need in comparison. Yeah, I mean, we're talking a lot about regulation in Europe of AI and technologies. And I mean, I'm an entrepreneur and investor, so I want to talk about the opportunities and I want to see sort of more investment. And we see this that structurally, when we look at companies in Europe versus the US, for example, that they invest less in innovation and R&D structurally than in the US and also in different types of innovations more.
[0:24:59] Let's say middle tech than high tech. And I think this needs to change and it needs to change fast. So I'm looking forward to building alliances with more and more corporates to make this translation mechanism really effective and help on idea generation, how these investments can basically translate into business value really fast. So I think it's a really good.
[0:25:24] Very last wake-up call to say, hey, we're here, we have capital, we have great talent, we have very interesting data pools actually in Europe, very interesting industries here, and how can we use this unique advantage, right, to combine it with great technologies that have been built, and that, by the way, a lot of European research has contributed to, and really create a lot of jobs and growth.
[0:25:47] Business value and solve a lot of problems. So I hope this will give the impetus for this. And I'm, you know, that's what we're doing at Morantix, trying to also be that bridge for companies. Yeah. Yeah. So to me, it's all about the data. So as you already mentioned, right. So the key thing is there's a wealth of data available in European companies, particularly in the SMEs here in Germany that could be leveraged right from
[0:26:13] Production processes, the customer databases. There's lots of things that are currently not harvested at the same scale, right? And I think there's a huge opportunity here to do that if we talk about the applications. So, of course, there's regulation. Of course, there's the hardware scale and all that stuff. But I think a key aspect is really leveraging the data. And I think we have to get there. Germany is an incredible market for us. It's the biggest market.
[0:26:43] In Europe, it's top five globally. People are doing incredible work with the tool. I'm very convinced that most Europeans want AI to be used here, to happen here. They want it to revitalize economic growth. They want to drive science. They want European infrastructure, and so do we. We would love to do like a Stargate Europe. We need help, but we'd love to build something here for Europe to govern and operate someday. I think that would be great.
[0:27:09] We want to be able to deploy our products in Europe as quickly as we do in the rest of the world. I think a strong Europe is really important to the world. I think Europeans will do great stuff with the technology. I think the European people will have to decide how they want to set the rules for this technology. We'll obviously comply with whatever they are. But, and obviously I'm biased here, I think it is in Europe's interest to be able to adopt AI and not be behind the rest of the world.
[0:27:35] Well, that tackles me directly to maybe European AI Act in general. How do you see this for a US company looking towards Europe? We will comply with the law and respect the wishes of the European people. You all decide. We'll follow the rules. Nicole, do you? Sorry? Do you agree? But you do have to decide. If there are...
[0:28:11] There are benefits to different regulatory regimes and there are going to be economic impacts that will become societal impacts. I think we should have access as a continent to state-of-the-art technology. I don't see an advantage in not having this access, to be quite honest. For me, the only way to shape a technology is to have access to master it, to work with it.
[0:28:37] I'm a big proponent of not over-regulating. I mean, everybody here knows this. It's like, I think, and it's also very early to do so. So I feel that we want to keep this innovative capabilities and we want to encourage them. And I think it's not only about factual regulation, right? I mean, there's a whole other problem with like being clear about rules and just having like whatever the rules are just for them to be very clear. But I think it's a mindset, it's an attention.
[0:29:06] Do we want the elites of our country? Do we want all the tech talent that's sitting here? Do we want them focused on rules? Or do we want them focused on opportunities and how we can solve problems? And I think that's the question we need to ask ourselves. And I'm very clearly in Camp Opportunity. I can only concur to what you said. I think it's really crucial that we do not cut ourselves off from technologies and that we have the ability to
[0:29:42] Experiment to try out and explore. And of course, there has to be regulation to prevent harm in the end. But, you know, sometimes I think regulation comes prematurely and then there's a huge danger to that with cutting off the economy, cutting off markets and in general opportunities. Cool. That brings me towards the next topic that I had in my mind regarding open source and
[0:30:11] Basically, shortly just after Stargate project was announced, we had another big announcement from a Chinese company, in this case, who released their open source chatbot, DeepSeek R, in this case. And at the university, in general, in science, open source and open data is extremely important in relation to something we discussed earlier, explainability in AI, but also reproducibility.
[0:30:40] I guess after the release of DeepSeek R, you recently also said yourself, I think on Reddit, that OpenAI has been on the wrong side of history in regard to open source. So what does this release actually or in general open source mean for OpenAI in terms of competitiveness, but also in generative AI in general? Yeah, I think there's clearly a place for open source. And open source models are...
[0:31:10] Good to have in the world and will provide a lot of value to people. We can debate whether or not AGI should be open sourced. But I don't think anyone can debate that these models at the current level don't create value being open sourced. So I think it's good. We have a lot of priorities, which is why it's easier said than done. But I'm happy that there's good open source models out in the world. And are you currently also having discussions within OpenAI? We are. We don't have a decision yet, but yeah. Okay, cool.
[0:31:39] Volker, what do you think are the advantages and disadvantages on this? I think there's a couple of aspects. There's open source models, open sourcing the training data, open sourcing the training process. There's a couple of different aspects from a scientific perspective, of course. It's very valuable to have all three open because this allows further exploration.
[0:32:01] This allows for also improving our product scientifically. It's of course understood that companies that invest a large amount may not have an incentive to do that. So I understand that as well. But from a scientific perspective, of course, for us, this is crucial. And I think in the sciences, we will see this trend of open source also for other reasons like reproducibility. And it may help with explainability even, right? So I think this is, and this may be crucial for...
[0:32:28] Adoption in some areas. It's maybe also interesting to point out that very often, you know, the open source happened after the fact, right? So if we look at the history of computer science, there were the closed source systems first, and over time, when, you know, some things became more commoditized, they became open source. It's also sometimes a marketing strategy for effective followers and whatnot, right? There's also issues about that. But I would say that in the future, we'll probably see more
[0:32:57] More open source models, more hopefully open source even training processes and data in specific domains, in particular in the sciences. But don't you also agree that the fact that in case of large language models, but in general in AI, the fact that a lot of things have been open sourced early on has been leveraging and accelerating this development.
[0:33:26] Absolutely, yes. I mean, this is a key, right? So this is how science works. Science is about publishing, science is about open public processes, right? So this is totally understood in the scientific process. This is just good scientific process. Make things open, right? And this is what, in my opinion, should be and the most is the requirement for all scientists that are publicly funded, for sure. So, Nico, maybe a direct question.
[0:33:55] I'm wondering if you think also that this has kind of like made the field now more even basically in terms of also like companies, right? Because we have the possibility essentially with the open source chatbots as well. Well, I think, I mean, to me, it's like an innovation landscape and it's an ecosystem. So you need different parts for this to work really well. I mean, you need people who are pushing the boundaries.
[0:34:23] We know this is also a resource question, right? So access to these computer infrastructures is key and not everybody will be able to have it. We can debate whether that's right or wrong and how we want to do it as societies and economies. Right now, it is a bottleneck. And so I feel it's good that there is open source. There's more people contributing, pushing also, right? Some of the innovators. For me, it's not a question of either or.
[0:34:50] And we also know with open source technologies to make it usable on enterprise level, like that requires a whole nother layer because it's not like you can just like plug it in for safety, for security, for all these things. We will need those solutions. So for me, it's not either or. It's an ecosystem. It's all going to go hand in hand and hopefully accelerate and development, make it diverse and, you know, bring everybody on board to push the boundaries there.
[0:35:19] Okay, I would like to talk a little bit about something that, Volko, you touched upon, but also, Sam, you did earlier, data efficiency, but also, I guess, energy consumption. Something that I think a lot of also in the audience are thinking about. So, I guess, until recently, companies, AI companies, were more focused on training bigger on...
[0:35:50] Bigger and bigger models and bigger and bigger data sets. We have the discussion around data becoming a bottleneck even in the scientific communities. But we have now the shift, and again, we chatted about it earlier, towards more reasoning models, right? And to arrive basically at better answers, these models are using increasing computation also at inference time.
[0:36:19] I guess what capabilities do reasoning models offer that we were not possible before with traditional pre-training is my first question. Yeah, the models are getting incredibly efficient. If you think about the watts per token, they give a really high quality answer back. Probably we're more efficient than asking a human and the food they need to eat to run their own. They're getting super, super efficient.
[0:36:48] I remember a long, long time ago when Google first came out, there was a sort of moral panic because people said, oh, you're like, you know, to ask that one question, you're using all this electricity in the data center, it's horrible. And I remember thinking, well, what that's replacing is someone like turning on their car engine, driving for 15 minutes to the library, looking something up, driving back.
[0:37:16] And even if they do it a hundred more times because the friction is lower, it's way less energy use than before. AI today uses a tiny amount of the world's energy. It's actually quite efficient per query. And the framing that I'm most interested in, like there is totally a world where we say we're gonna ban AI, it uses too much energy. We're gonna ban computers, we're gonna ban light bulbs while we're at it and we're gonna sit in the dark and we're not gonna burn any energy. That is a world we could choose.
[0:37:45] But there's another one where we say we're going to use AI, and even if we have to use hundreds of megawatts or gigawatts on this problem, if we can use AI to discover how to do efficient fusion, cheap fusion, and then very quickly replicate the thousands of gigawatts of generating capacity that's burning carbon around the world, that would be a huge win. And I think that is the promise of AI. So I could sit here and make arguments about efficiency, but what I would really say...
[0:38:12] We are totally screwed if we do not come up with new scientific solutions to address the climate challenge in front of us. We clearly have failed to do that fast enough on our own without AI. Let's try it with AI. Do you think even in terms of national power grids, etc., do you think renewable energy can fill the gap in energy needs that we have? I think fusion will be the way that most...
[0:38:47] Foka, maybe in this context, I would more be interested in your point in assessing the importance of high quality data and also efficient data pipelines for the use of AI. High quality of data and the efficient data pipelines. Efficient data processing. I agree with Sam, of course, we have to
[0:39:21] Look at how we use the data efficiently and the models. But there's one aspect. Indeed, it's probably not just about using the existing models. The question is, can we improve on the training process algorithmically? Because that might bring things down. There may be improvements in hardware that will help us as well. In addition to, of course, using what we have now. But this is, of course, the role of science to bring us forward here.
[0:39:47] And we can do this algorithmically. And we had some discussions about already overcoming the quadratic complexity of training. So this would be a huge win already for energy efficiency if one can bring that down to, let's say, n log n. And there might be other models than the attention mechanism and transformers in the future that will help us to improve on that. And this is also active research. This active research that we are...
[0:40:14] Conducting in our groups and many people around the world are looking into that. At the same time, of course, I'm not a hardware person, but improvements in hardware are another aspect where energy consumption will hopefully go down eventually.
[0:40:30] And so it's a holistic thing, and we have to look at this holistic process. But where I agree with Sam also is to look at energy consumption holistically. AI has a cost, but it also has a benefit, the traits of the energy consumption. And I think in this way, looking at this holistically, sometimes one might otherwise forget the benefits. But again, one has to be, of course, mindful, and we know our...
[0:40:56] We are in the era of climate change, we are in an era of problems, and we have to make sure that whatever AI does, it helps to solve it and not make it worse. So basically, we have to deal with it also in a very responsible way in this case, and keep all the risks with it in mind. So before we move on to Q&As, because we had a lot of interesting questions from audience as well.
[0:41:25] If we reference novel with sci-fi, because we have to do that, I think, and think about Space Odyssey's Hall, would you expect the first foundational models that exhibits AGI to be able to self-replicate, self-improve, self-modify, for example, based on an emergent need? And I guess maybe more importantly, even if you expect that, would OpenAI even be aware of this?
[0:41:54] Or companies be aware of this desire? And what is the responsibilities you see in companies like OpenAI have in reporting this and its potential social impact? Yeah, I think there are people who don't count it as an AGI until it's like a self-replicating, self-improving space probe or something. I will count it personally as AGI well before that milestone. And yes, of course, we have a responsibility to tell people if we get close to that. We will. We will tell people, I mean.
[0:42:23] Do you know how? How we'll tell people? No, how to communicate that and how to even see that desire. Oh, how we would see it in the system? I think it'll feel sort of gradual. People talk about self-improvement as this very binary thing, but really it'll be like it can help the scientist work a little faster and then a lot faster and now it can do more and now it can do more.
[0:42:52] And so I don't think it's like a clean binary, but we'll see something about the rate of the slope changing. Well, thank you for this part. We are moving on to the Q&A. As I mentioned earlier, we had a lot of interest.
[0:43:10] We had over 1,000 different questions. So what we did actually, we clustered these questions and selected them from topics as representative for these questions. And I will be reading some of these questions and asking this to the audience. And we have around, we will see how many of them we will get to. But I will do my best to keep the time. The first question comes from Oliver.
[0:43:39] And Oliver says, I'm curious to hear the panelists take on progressing society and economy with products that work like magic. That is, we don't understand what they work on. An analogy, how could we have succeeded with building cars that come out of a black box factory? So how can we basically trust also these models?
[0:44:06] That are all black box for us currently in terms of explainable AI. I use products every day that I don't understand how they work. If I understand that a particular input will give me a particular output, I can use incredibly complex systems. Maybe I understand how a car works. I certainly don't understand how an electron microscope works. I could sketch it out on a napkin, but I can't go make you one. But I could use one.
[0:44:38] And I think that's fine. I think we do need to have a high level of robustness and predictability and generally accepted safety for AI models. But also, users are sophisticated and they understand, like again, a lot of people use ChatGPT for important stuff. And on the whole, they understand super well when to trust it and when not to. And yeah, probably most of them couldn't go train it, but that's okay.
[0:45:03] Economic value exchange of the world. Some people build something they understand enough about how it works. We don't understand it perfectly, but we understand it surprisingly well. And the beauty in our job and what we get paid for is we want to make it so simple to use that you use it, it delivers great value for it, and you don't have to think about every detail of the training process. Yeah, so I mean there are safety considerations, of course, in systems, right? And we traditionally build engineering systems.
[0:45:33] We design them, we try to make things provably correct, look at many different aspects in order to ensure that we can use them. And so the way I see it with AI, we are not yet there, obviously. And so the question is for what kind of mission critical aspects can you use it in what way?
[0:45:58] So if we, you know, risk a nuclear meltdown in a power plant because of some margin of error there. So I think I would rather rely on a provably correct system that has been built without AI at this moment, right? At the same time, of course, there is systems where it's not that critical, where then you have to make a judgment, right? And I think it's always in society, it's a judgment call for which systems with which potential damage, which potential harm.
[0:46:26] You can use AI in this context. But personally, I would say there is still some areas where I would consider AI to be off limits at this moment. I mean, it's highly use case dependent, right? And I think when we talk about explainability, the question is also what...
[0:46:45] What do we mean and what do we actually want to understand? So do I need to understand which edge or things like an algorithm looked at in an image or do I actually want to understand did it look at the tumor and not at like something or is it looking at something behind it on the slide? So I feel we need, I'm not sure we need 100% explainability for stuff, I agree.
[0:47:11] I think we need to think about what is it we really need. And I see it's critical in some applications just for adoption because people are hesitant in some like more like safety health related things to use the output if they can't somehow understand how it came about. But I think it's, I mean, it's like a human. I think at the moment it's an illusion to have 100% explainability also for some decisions I take. Or everybody probably, I don't know.
[0:47:40] Okay, the next question comes from Veronique. Veronique asks, how can we ensure that AI only uses trustworthy data for training, congruent with the FAIR data principles, findable, now I'm missing my words, but the FAIR, accessible, findable, accessible, interoperable, and reusable. There we go. Well, that's a matter of data curation, right? So making sure that the data sources you use are of that.
[0:48:11] Nature, having some forms of validation. And there's, of course, also semi-automatic ways to do that, right? There's a lot of research in the areas of information extraction, information integration, some of which, again, using machine learning models to do so. And one has to apply those, right? And then it has to have checks and balances to validate. So, of course, I would imagine that, you know, in many cases, one does that. So, for instance, when one builds, like...
[0:48:40] Foundation models for specific corpora like pathology or whatever. You look at having those kinds of sources, right? So in this way, yeah, one has to do it according to those principles. I think what was interesting about this aspect is basically it sheds the light on quality of data, right? And how do we want to ensure quality of data? And there I think the human machine interaction will be interesting because we will need a lot of humans also to...
[0:49:07] Verify some of these things and make sure we're dealing with diverse teams doing this and basically like encouraging also humans to like, why would a radiologist, why would a high paid lawyer want to sort of give feedback to systems? So that's, I think, a matter of like designing the right interfaces and things that we want to be machine teachers, right? Also for some of these things. Okay, I will move on with the next question.
[0:49:33] Daniel is asking, what are the biggest technical difficulties to making AI more powerful, Sam? Better algorithms, bigger computers, harder data sets to train on, building better products that help us learn more from feedback with users about what's not working. But I cannot overstate how much progress I think we're going to make in the next two years. We know how to improve these models so, so much.
[0:50:08] And there's not an obvious roadblock in front of us. We gotta go build Stargate. We gotta go do a lot of great engineering. Our researchers need to keep doing stuff. But the progress that I would expect from February of 25 to February of 27, I think will feel more impressive than February of 23 to February of 25.
[0:50:31] Boga, would you like to add something? Yes, so just to add to what was said, I think it will be really interesting how we can incorporate multimodal data and some of the more sparse data from specific application domains in a meaningful way. So I think this will be quite interesting to make it more powerful for specific domains. Also, this is so exciting for you all. You are going to have... You are going to get to...
[0:51:00] Live through and sort of do impactful work in the steepest technology ramp probably that any of us have ever seen before. If we really are right about this rate of improvement, the things that you will be able to do in 2027, the things you can do right now that were just impossible a couple of years ago, the projects to start, the science to discover, the companies to build, it's unbelievable. I do agree with that. Nicole, do you have something?
[0:51:24] Yeah, I mean, we're so excited, right? We also run a fund and so we're very excited about the startups that are combining this with very specific domains and like tech bio and medical fields and like all sorts of fields. I think it's a fantastic time to invest and to see this unfold, right? It's like really, as an investor, very, very excited. So this links a little bit to, I think, if I pronounce this correctly, Dilshad's question.
[0:51:54] If you were a student today, how would you prepare for the future? At least I'll let you start. Do you want to go first? I think continuing to think, critically think, I think will be quite important. I also think I like this word liminality. I discovered it recently. It's the status between what has been and what will be. It's the intersection of it. And I think you're...
[0:52:25] We're all maybe feeling some of it like the past is like we're running like it doesn't give us more options but the future like we can we can feel what's happening but we don't know exactly what's happening and so to be able to deal with this state of liminality and ambiguity and still remain like in a constructive action driven outcome driven way so embrace this embrace that things are changing there's it means it's a lot of opportunity and I think that's also a mindset question.
[0:52:53] The obvious tactical thing is to get really good at using the tools. When I was in college, the thing to do was learn how to program, and that was great. But programming at the beginning of 2025 and at the end of 2025 are gonna be completely different things. And you wanna be on the good side of that trend. So learning how to use AI tools, learning how to
[0:53:20] Just sort of like do way more than you could before. That's the obvious tactical thing that makes sense. I think the sort of less obvious things are adaptability and resilience. These are gonna be hugely important skills. These are super learnable and super practical. They get much better with practice. The skill of figuring out what other people want. I think this is also like a learnable thing and when AI can kind of do anything, create anything, then deciding what to do and what people value.
[0:53:50] It's gonna be really important, and humans are so hardwired to care about other people and what other people create, and I think that's gonna keep going. Some people say it's figuring out what questions to ask, but I really think the answer is figuring out what other people want, what's gonna create the value, how you're gonna direct the AI. I used to be a startup investor at this firm called Y Combinator, and one of my biggest takeaways from that is the degree to which people can learn this skill.
[0:54:19] It's much harder than I would have thought and something that I think we should teach. So I'll let you all answer first because this is a very common question and my answer is always the same. And the answer is, you know, in the end, you know, there was this term that was shaped about the T-shaped student, right? That I think many of you have heard. And a T-shaped student really means you should have knowledge in the breadth and knowledge in the depth. And I think it's very important to still protest both. And now you can, of course, discuss in your specific domain and specific subject area what that means.
[0:54:49] So for instance, in computer scientists, as Samurai said, programming, but also a specific specialization area, be it machine learning algorithms or data systems or robotics or whatever. So you should have one specialization area that you know in depth. And even if some of the knowledge that you learn there in depth is eventually outdated, you will still be able to have learned how to adapt, how to carry it over, which was the adaptability that was just discussed about.
[0:55:17] And this was, you know, an interesting aspect. So we have depth knowledge and breadth knowledge. And when we had Eric Schmidt here and we had exactly the same question, right, then in the end it was dubbed, you know, that don't study marketing. This was essentially what came out of that from Eric Schmidt. The reason was because that's breadth knowledge. It's, of course, not true. Of course, you might also want to have some knowledge there. But you also need to have some technical depth, right, in combination to that. And that's why I would say always follow.
[0:55:46] This guideline of the T-shaped student. That was a great, I guess, lesson for the students in this case. But maybe the last question for everyone. Just if you could briefly tell me or tell us what you are most excited about. And you hinted towards that already for AI in 2025. Yeah, I'll stick with my scientific progress answer.
[0:56:12] Yeah, I mean, there's lots of things to be excited about, right? So, I mean, we're really seeing a time of drastic changes in the scientific process, support through AI systems, so how that can be leveraged. So I'm really excited about that. I'm also excited about new systems developments. And so I think it's an exciting time in general to be a researcher in the field of AI and data because there's lots of opportunities, opportunities for...
[0:56:39] Entrepreneurship, right? So there's some options that we have there and I'm sure you will talk to that as well. And there's lots of options for deep foundational research because it's still, even though it's evolving, it's still a very open field in many areas and they just can encourage many of you to join this endeavor either in academia or as entrepreneurs or in industry because it's where the future is, right? And that's what's exciting. Yeah, I mean, I'm very excited about, for example,
[0:57:08] We're looking a lot into tech bio and medical use cases. I would say that's really amazing, understanding the body and how we can create remedies better. I think that's very exciting. I think combining logistics, manufacturing, there are so many cool things. Basically, it's about going deeper into the core of...
[0:57:31] Business, new business creation, new business model creation and like accelerating things at the core where there's really an intersection between the tech and the core expertise. I think that's what we're going to see more coming from more like, let's say, knowledge retrieval systems to like more to the heart of the matter. And I think that's always very exciting. Thank you very much for this very exciting future look, I guess, in this case. And we are almost at the end. I'm saying almost because we would like to take a
[0:58:01] I would like the audience to just to be seated and follow my instructions. I will ask the panelists now to stand up with me and then we will just turn around. I think you can all wave as well and bring your hands up or are you done?
[0:58:38] Okay, we are done, I guess. Okay, with that, I would like to really warmly thank you all for this great conversation. Thanks to the audience. And thank you. Thank you. Thank you, everyone.