[0:00:00] Given that we are running a little bit over scale and that this panel is what's standing in between you closing remarks and drinks, I'll try and be as...
[0:00:08] Sort of succinct and quick as possible. Lengthier introductions you can find in the booklet in front of you. So I won't sort of redo them. And I hope you'll bear with me. To start off this sort of spirit of interactivity of this final panel for the online viewers, please have a look at the poll that should come on online now. And we'll try and bring the answers of the poll into the
[0:00:30] And don't worry, for the people that are in the room, you'll get a chance to raise your hand in answer to the poll at a later stage. So don't panic there. As Nico has said, what we see is a really rapid escalating pace of AI development. And I think most of you will have seen this article in The Economist in June, which talked about how foundation models are really now powering progress in AI.
[0:00:56] I think many of us will have seen GPT-3 built by OpenAI, which was designed to basically predict the next word in a sentence. But it turned out it could also do translation. It could also write code. It could do basic arithmetic. And all these additional functions were discovered as the model came online. And if you follow Jack's newsletter, Import AI, can I see some hands who subscribes to Import AI?
[0:01:20] Oh, that's great. That's about half the room here. If you don't yet, Google Import AI. I can highly recommend it. I enjoy reading it every single week. But you can see in Jack's newsletter a new model coming out every single week, and you see this pace of accelerating progress. So maybe I'll get to the panel in one second. I'd like to sort of make one additional comment, one to the panelist, which is a request to be super brief.
[0:01:47] So two, three minutes for your answers so we can get to the audience as well. That would be excellent. And for the audience to please think about questions you have, if you're watching this online, please type the questions online and please address them to a single panelist because I can't afford questions to the whole panel because then we'll run out of time after one question. So please think about both your question and which panelist you want to ask it to. With that.
[0:02:14] Jack, can I ask you, maybe as sort of a representative of one of the leading AI companies now in the world, one of the highly capitalized startups of recent date, what is one AI development you're really excited about and what is one main risk that you see?
[0:02:31] Fantastic, thank you. I'm going to try a glass half full approach here, so less doomy than usual. What I'm very excited about is I truly think AI is going to revolutionize education. You can use these foundation models like GPT-3 and other things to...
[0:02:47] As didactic aids, they can help teach you things. I'm learning programming again using a model we have at Anthropic. It's much more patient than my previous teachers because it's a robot and doesn't care that I'm occasionally lazy. This stuff is incredibly exciting and useful. What is the risk is, well, I can teach myself anything with these models. And if we develop these models, you need to figure out ways so that they can teach people just the things that you think are good and they can't teach people to become some kind of like...
[0:03:16] Marvel supervillain. So the worries we have are how we create a regulatory enforcement infrastructure that guarantees everyone can benefit from these upsides, but these upsides don't allow people to sort of indulge their worst natures and do things that could harm others. Awesome. Compliments for keeping that super brief. One of the harms that have recently come up in the United States in particular is something that's been highlighted by House Representative Eshoo of Silicon Valley.
[0:03:49] Who talked about stable diffusion, which is an open source model that was used to create images of Asian women being abused of pornography and for which no one took liability because it was released as open source. I know the sort of Lucille Asioli, you're involved in the AI Act and sort of the minutia of current legislation. But as we sort of see this new development of open source models used in this way.
[0:04:18] Do you feel in future there might be a perspective on regulating open source or sort of how would the Commission look at analyzing that new emerging problem? Well, thanks a lot for the question. First of all, for us, we treat open source as we treat proprietary systems.
[0:04:39] In the current regulatory act on artificial intelligence, we don't have any exemption, let's say, for open source systems. So we do regulate those systems as well as we regulate the proprietary models. I think that the example you mentioned is something that would be actually captured by our regulatory framework.
[0:05:03] The way it is presented at the moment and with the modifications that have been added by the European Council most recently because what the European Council has added compared to
[0:05:14] The original Commission proposal is a clear requirement for general purpose artificial intelligence to be subject to certain rules. Now, these rules are not in the Act yet. We will have to make another Act, an implementing Act, which will describe how the general purpose technology artificial intelligence systems are actually regulated. But I think that through the risk management system, certainly a case like the one you have described would be kept.
[0:05:43] And as I said at the beginning, it is not a question of open source. The systems are actually captured. That's really good news. And maybe relating back to the panel before, something where the US can take some inspiration from Europe. Turning to you, Deputy Minister Havarda, there is an increasing discussion around AI R&D and the effects that that has or could potentially have on the world.
[0:06:13] Especially when in March the Spitz Institute in Switzerland used an AI system that's usually used for design of medicines to design 40,000 new biochemical weapons in the R&D phase. That, of course, presents a whole host of questions. And I think it's these very, very rapid developments that suddenly pop up that really provide challenges to any solid AI governance framework.
[0:06:38] And my question to you is, therefore, do you see scope for regulating AI R&D? And how do you think AI governance can really adapt to all these rapidly changing developments? Thank you, Mark. It's a great pleasure to be here today. Well, I will do two things what I knew as a newer politician I would do. It's not to answer the question, one. And second, not to stick to the timeline that you gave me. Well, I think.
[0:07:09] And I will end up with the question. I think the AI is just a beginning. And as we discussed on the previous panels, and as you mentioned, especially with the GPS, they will see proliferation of new use cases in new domains completely. And this will require a completely new approach to much more agile approach to AI governance.
[0:07:35] This is why we do two things. So we proposed and we will have the general approach in coming days of the council. So agreement of the member states for the negotiations with the European Parliament is that first we put this flexibility in the hands of the commission. That's one thing. The other thing is we have explicitly included regulatory learning, especially for the sandboxes, but regulatory learning in order to create evidence.
[0:08:00] This is for evidence-based decision making into the act. And I think this is really crucial because otherwise we will not understand what the unintended consequences and what the negative impacts are in different use cases. And of course, then ultimately it comes to air safety as such. And let me step back and a little bit talk about human-centric for one minute, I think.
[0:08:29] Because we don't do it because of the sake of the regulation. We do it because we want AI technology, which is human centric. And let me add a little bit of, to share with you some ideas what human centric means, because we talk about it quite a lot, but it's very difficult to operationalize. And together with the colleagues, we put this statement in the plenipotentiary of the ITU just a couple months ago, which was not only signed by the 27, but almost by 60 countries.
[0:08:58] And when we were thinking what it is, is that I think the crucial is how we deal with the trade-offs which are inevitable in any deployment of AI. And I don't have to put this audience to be describing the trade-offs which are there between, for example, privacy and security and safety and performance. So in dealing with the trade-offs, I think if we want to be human centric, we need to take into the consideration
[0:09:24] Some key criteria. And these are just the criteria which we know today, which I will mention. Privacy, which is acknowledgement of the bias, which is in the historic data. Discrimination, obviously. But the most important is understanding of the impact in that very particular use case. And for that, I think we will just need to develop a completely new approach, how we gather the data and how we cooperate on that. I'm happy to talk about it later.
[0:09:51] So getting back to research, I think it's really difficult to be addressing research. What we need to do is to have a platform, as Paul was mentioning, basically, which would help us to gather the evidence on the concrete use cases. Because without it, we will not be able not only to have
[0:10:10] Rigorous impact assessment, which is crucial for the human centric and this needs to be based on the well-being approach basically and P700010 of IEEE might be very good basis for that. But we will not be able even to distinguish. So, without understanding the impacts, understanding high risk is really difficult. So, can I put you down for no?
[0:10:34] I'll put you down for no. Okay, so I'm going to ask the same question before I get to you, Sasha, to the audience. And if you could please raise your hand. So last March, an AI system designed for drug discovery was repurposed to develop 40,000 new biochemical weapons. Is this grounds for regulating AI R&D in future? Your options are yes, no, and maybe. Could I have everyone for yes, please? And if the panel wants to join in, please feel welcome.
[0:11:06] Okay, I'm going to repeat the question. So last March, an AI system designed for drug discovery was repurposed to develop 40,000 new biochemical weapons. Is this ground for regulating AI R&D in future? Yes? No? Maybe? Okay, I think we had this pass. I think if I can say that in the parliament.
[0:11:37] I hope that the MEPs present will take some inspiration, given that they usually only deal with things once they come onto the market. But I think it is a really interesting question that these types of developments now pose when it comes to AI development. Sasha, I'm really glad that you are here as policy lead for Amazon Web Services for not just Europe, but also the Middle East and Africa. Given that we've talked a lot about the AI Act and a bit about the US, and it's all been very Western-centric.
[0:12:03] But could you say a little bit about policy development challenges that you see for governance also outside of these two regions?
[0:12:11] Thank you and very happy to be here. I will, at the risk of sounding repetitive, repeat a little bit what Eva Medel and Eva Cady shared, which I think is also representative of the fact that the private and public sector are indeed working together in order to develop these kinds of generative conversations that Vilas talked about in the beginning, but also to translate policies into practice. And there are several common themes that we see at Amazon Web Services coming out as we explore and have these conversations with governments.
[0:12:41] in Europe, in the Middle East, and in Africa around AI and particularly regulating AI and deploying responsible AI. And the first thing that comes up is that it's not the Wild West. And I think this is a myth that needs to be debunked is that there are many laws that exist particularly around anti-discrimination and privacy that in fact cover a lot of issues and fundamental preoccupations of policymakers in the field of AI.
[0:13:08] The second thing that we see is the need to define artificial intelligence and machine learning accurately. And I know that this is a conversation that happens at the level of EU institutions, but also at the level of the OECD to say, okay, well, how can we draw on the definition of the OECD and adapt it to a fit for purpose, a regulatory definition so that it's clear what we're regulating and what we are not regulating. The third consideration that we see is the need to look at
[0:13:37] How we can deploy responsible AI in a way that is use case specific. And I love, again, something that Dilas Dara often says in his talks, which is the need to develop and deploy.
[0:13:48] Technology that is tech enabled, but fundamentally human centered. And this is where the use case specific approach that has come up by a lot of policymakers today, I think is absolutely critical that we are regulating the use cases, not the technology. Imagine if we had, for example, regulated the internet blanketly where we would be today.
[0:14:12] Another consideration that we see and that I think is absolutely fundamental that has been evoked by Eva Medel is the need to allocate responsibility for actors across the AI value chain. So what does a real shared responsibility model look like, including for, for example, general purpose AI systems, where it is allocated between the developer and the deployer and the end user in a way that is based on actual insights. And this is where I think private sector conversations with public policymakers.
[0:14:41] are absolutely essential. And then two last very quick points because I think that it came up on multiple occasions. The very crucial dimension of skills building. What we do not understand, we fear.
[0:14:55] And so we are over-regulating a technology that has enormous potential benefits for humankind. And so the aspects of understanding and building the skills and comprehension of AI, ML, of policymakers, but also of citizens, what the technology is, what responsible deployment means, and how we can scale and innovate for impact is absolutely essential. And then just the last point here that has been raised by many policymakers, so I won't dive deep into it, is the question of...
[0:15:24] Interoperability. So what does the future proofing of AI governance look like if, for example, we can introduce systems in Spain, but not in France, or sell them in the market in the United States, but not in Germany? And this is where we see a very important role for international standard and market interoperability. So there isn't a regulatory fragmentation that impacts the way we can innovate for good going forward.
[0:15:53] Excellent. With that and those introductions from the panel, I'd like to open it up to people online and also in the room if they have a question. If you have a question, could you raise your hand and also state to which panelist you're addressing the question? I have one question here for Sasha online, but I'll go to this gentleman over here first. Well, thank you very much. My name is Alexandros Piridonis from Erlan Tech. I want to raise my question to Sasha. How is it?
[0:16:20] If you are trying to regulate what you do as a business and as offering of services in Amazon, well, Amazon is international. So do you have a separate Amazon for Europe, a separate Amazon for the US, and a separate Amazon for the rest of the world? How do you manage in practice, since the service is the same for everyone and everyone wants the same from Amazon? How do you manage to settle all this between the different jurisdictions?
[0:16:50] And what would you like from the policymakers? What would you request from the policymakers from the business and the service point of view?
[0:17:04] So there are two parts to your question. The first is how we are structured, and the second is what we're looking for for policymakers. In terms of how we're structured, I'll focus in my response on compliance. And so what we do at Amazon Web Services is make sure that we are complying to all existing laws, and that means laws at the local and national level, but also at the regional and international level. And we have our Vice President for Public Policy, Michael Punk, tomorrow in his opening speech. We'll talk a little bit more.
[0:17:34] The second part is what we would ask for policymakers. And again, I really appreciated what Werner Stang said, is the need to understand who needs to comply with what.
[0:17:48] And we are a company that is dedicated to empowering and democratizing access to AI and ML services through the cloud. And so what that means is that in the conversations that we have with small scale startups around regulation is A, the need for clarity and B, the need for support.
[0:18:06] Compliance with certain regulations require an enormous amount of resources, both in terms of human resources, but also legal resources and financial resources. So the balance here is about ensuring the protection of human rights and also ensuring a future-proof regulation that promotes innovation and international cooperation. And here, our ask to policymakers would be, A,
[0:18:30] Where we are for these kinds of generative discussions and participatory public policy making approaches so that policies remain implementable. We are here for those kinds of open conversations to say what is technically feasible and what is not. For example, in the high risk requirements of the EUAI Act, what are specific to the AI as a technology and what perhaps is more broad and beyond AI and perhaps just about software in general. And so that for us is a very
[0:18:59] Thank you. I'll alternate with a question coming in online to Lucila about
[0:19:25] Other organizations beyond the EU and what role they can play to complement the EU's efforts in future-proofing AI governance.
[0:19:36] So there are many organizations that are actually working on artificial intelligence. There is one that is certainly helping us and will help us significantly in terms of governance is the OECD, because the OECD is very active in developing repositories, for example, metrics or tools that can be used to check compliance of artificial intelligence, high risk systems.
[0:20:04] Or simply providing information and guidance also to smaller companies about what can they do and what certain requirements actually mean. So we will be working very closely with the OECD in this.
[0:20:20] And I think other important organizations are, for example, UNESCO, because UNESCO has a recommendation on artificial intelligence, which is not binding. It's a recommendation. But still, it talks to developing countries and the countries that we normally maybe through OECD or dialogues that we have maybe speak less to. And so I think they play a very important role in terms of multiplying the messages in this.
[0:20:49] I think we, as the European Commission, have to think very close, very seriously about the governance of the artificial intelligence sector, exactly for the reasons that have just been explained. We're not here to create uncertainties to the world of startups. Actually, we have the opposite objective, which is to create more certainty and which is also to create more trust, because more trust
[0:21:16] These two more use of artificial intelligence systems. I mean, you will be surprised to hear that before we actually prepared the act, we ran a survey of enterprises and I think it was 65 or 75% of them said that they wouldn't use AI because they didn't trust it. And then there...
[0:21:35] Clients will not trust it. So, you know, you have to create trust if you want to also create a great use. And we will have to reflect a lot about the governance. And what we are reflecting about is to put in place maybe some kind of organizations that can really offer technical help. On the one hand, to the startups, and we are proposing regulatory sandboxes, for example, they are even in the legislation. And Spain at the moment is testing a regulatory sandbox.
[0:22:04] So we are very much preparing the ground also with the member states. And on the other hand, providing also resources to those who have to do market surveillance or have to do conformity assessments so that we can a little bit capitalize on skills and using synergies across European countries. Because these skills are not going to be easy to find. And we certainly don't want to take skills away from the private sector.
[0:22:33] To bring them in, you know, in government bodies that have to check how artificial intelligence develops. So we have to really have an intelligent approach to it. And we are reflecting very seriously about it. And we'll continue to do so with the Council and with the Parliament. I think it's clear that we need institutions to help sort of communicate to businesses and bridge that bridge that gap and make sure that this doesn't just stay in this room where we have all the
[0:23:01] Bubble working on AI. I see your question. I'll get to that in a second. Lucila, I realize your car is in a parking garage that will soon close. You're welcome to stay, but thank you very much for being with us. If you need to leave, could we have a round of applause for Lucila? Given that there's so many more people online, I'm going to prioritize one more online question before I get to you. Jack.
[0:23:37] One of the discussions that are really I think central to the national AI research resource in the United States is this.
[0:23:44] Concept of AI breakthroughs now being completely dominated by the private sector, whereas previously the academia had much more of a role in powering AI development. It's been, I think, a turning point about three, four years ago when companies took over all university work when it came to AI breakthroughs. How do you view that sort of development? And do you think there is a role for regulators, for governments in Europe, in the US, but also elsewhere to act upon that?
[0:24:11] Yeah, I think that this is existential and crucial to the future of the human race. Power asymmetries tend to kill civilizations. We currently have a power asymmetry in AI where you have a huge amount of power and insight solely existing in private sector actors building these frontier models. Regulators and governments must look for ways to build capacity that creates shared experimental infrastructure so that the other sort of branches of society can participate in the development of these models. And also, most importantly,
[0:24:41] Have the leverage to critique the private sector. And I say that as someone in the private sector, just conscious of my own power and saying that this is a thing to fix. Very clear message. Gentleman over here. Hi. Yeah, I'm Tom Westcott from Tony Blair Institute for Global Change. Also one of Jack's many subscribers to the newsletter. And question is also for Jack, unfortunately.
[0:25:08] The staggering collapse of FTX and by sort of relation, the Future Fund has obviously had a big impact on the sustainability of the AI safety research community. And I was just wondering if we really want to future-proof AI governance, what do you see as the best way of future-proofing funding for AI safety? Well, the good news is... Oh, sorry.
[0:25:33] The good news is organizations like NSF and others are starting to fund AI safety generally, and you see a shift towards funding, you know, safety, interpretability, and transparency. I think one thing the EU AI Act gives us is a set of properties that AI systems need to meet, and I think we can translate that to funding agencies and say, well, the Act says this. If you want to, like, conform with this and have a thriving private sector, you probably need to fund research here.
[0:25:59] I also think it would be good to have a broader funding base for this type of research than just a narrow small set of factors. A question to you, Marek. We've heard from Sasha the need for skills, and I think it was also mentioned in a previous panel that we need regulators to also have the skills, and that that was the first thing that the National Security Commission on AI in the United States identified as one of their recommendations for the US federal government. Looking at it from a Czech perspective,
[0:26:30] What do you think is necessary to bring the Czech government up to speed in terms of having the skills and resources within the government to enforce current regulations that are under development, but also sort of make sure that AI governance is future proof going forward? I was actually really pleasantly surprised that we have so much skills in the house, in our government, in different parts of the government.
[0:26:59] I think we're not doing that bad, and I think we have proven it now working on the compromise test for the AI Act. I think what here might be very useful is the UNESCO recommendations which were mentioned, and they come up with an idea of so-called readiness index.
[0:27:19] And part of that Readiness Index should be exactly on that, on benchmarking your capabilities in the government. So what needs to come first is actually you need to understand what you need and then to see the gap, what you don't have. And this Readiness Index should help you in that. And then we will need mutual learning. This is how it's going to be. I think the position of the EU could be really crucial.
[0:27:46] Because it could serve as a place with a huge expertise, especially using the JRC, the Joint Research Center, and building on that. Then the cooperation with the OECD comes to the fore very clearly. We need to work as well with the Chinese closely on this. Maybe just...
[0:28:10] Use that opportunity and comment my vote on your question. I vote maybe. And the main reason is actually, and this is very related, is geopolitics. If we ban in democratic countries research and we don't manage to ban it, then I would like to learn. I spent with my team in Good AI years on thinking about this, how we could do it, how we could regulate research.
[0:28:37] If somebody can tell me how we can ban it in authoritarian countries like China, like Russia, but many other, North Korea, well, then maybe let's go for it and let's drag away somehow. But if you don't see the enforcement, that might be detrimental to the values we share. I'm going to controversially not pick up on that.
[0:29:00] Point about regulating the rest of the world, but maybe turn that upside down for a bit, because Sasha, given your experience, also having advised the African Union, there is, of course, the question of regulating dangerous research, but there is also a question of making sure that people all across the world can reap the benefits from artificial intelligence. What do you think can be done to ensure that capabilities outside of the West are developed and what opportunities do you see there?
[0:29:30] Great question. I just want to reflect very quickly on what my friend Mr. Havarda said about skills and just salute the work of the Czech government, both in terms of the EU presidency, but also in terms of their investment in skills building of the public sector and of civil society. And we at AWS are really excited to be working with them to expand the Elements of AI course, which is a course developed in Finland in the Czech Republic, in Czech, to upskill citizens and particularly...
[0:29:59] And that leads me to respond directly to your question around what's missing in terms of capabilities. And I'll start that off with what Lucia said, which is a lack of trust in AI. We do not.
[0:30:13] In general, trust AI because we do not understand AI. And so this is, again, where I think the skills approach is absolutely essential. And I would say one of the biggest barriers right now to adoption of AI at the national level in other regions, including Europe, is the question of lack of talent. There are so many jobs that will exist in 2030 that don't even exist yet.
[0:30:39] And so the question is, how do we prepare young people, particularly in regions like Sub-Saharan Africa, which will have an exploding youth population, for the jobs of tomorrow? And that doesn't mean that everybody needs to be a coder. It does mean, though, that everybody needs to understand, for example, what's foreseen in the GDPR, when we are interacting with an AI system and what that means.
[0:31:02] It does mean that people need to understand what responsible use development and deployment of AI means. And it also means that we need to be preparing young people with the entrepreneurial skills and the business skills to innovate and scale. I loved, again, Vyles' example in his opening keynote speech because it is people who are in their local communities at the local level who have the most innovative solutions to local problems that can be scaled.
[0:31:30] The second challenge that I see is the lack of and also cost of developing purchasing hardware and keeping it up. And this is where the cloud plays a very important role in democratizing access to AIML services. And this is again why at Amazon Web Services, going back to the question of the local, we are supporting startups.
[0:31:51] In these kinds of innovations and the migration to the cloud, which is not only democratizing access, but also very importantly, reducing the carbon footprint in Europe. There are studies that have shown that migrating to the cloud reduces the carbon footprint by up to 80%. And this is something in the framework of particularly the EU twin digital and green transition that is particularly crucial.
[0:32:14] And then two last challenges. The third challenge that I would say is a lack of data. And Jack, I loved your articulation of this idea of shared experimental infrastructures. It sounds like a poem. But I think here that this is where building on pre-trained data is really important. And here again is where general purpose AI plays a very important role in reducing the carbon footprints of innovation worldwide. And then I would say the last challenge that we see here, again, looping back to my opening comment on the...
[0:32:43] The lack of trust is a lack of understanding of how to develop and deploy responsible AI. And organizations like the OECD and the IEEE have been talking about this. How do we translate principles of responsible AI like transparency, explainability, accountability, that we all agree are important?
[0:33:01] How do we translate that into practice as an engineer, as a philosopher, as an innovator, as business counterparts? And here that translation is absolutely critical. And we just launched actually earlier this week on our Machine Learning University platform, a course dedicated to responsible AI to say, okay, well, what do these principles mean in practice? And how do we translate these policies at all levels of the AI value chain?
[0:33:29] Thank you. I've been told there's a question in the back. I'll get to you in one second. I would, though, want to ask Jack at the moment. So if such as challenges have been overcome, right, and we have all this AI development all across the world and Marek's sort of demand stands that we need to regulate some of this dangerous research. How do we do that globally? Do you have any ideas? Thanks for the easy questions here.
[0:33:54] Well, it comes down to a lot of the emphasis of the EU AI Act is on these matters of measurement and standardization. You need an international set of standards and measures and things you can evaluate systems against. And you need capacity in governments to do that. It means nations need to spend money on technical capacity so that you can actually harmonize this system worldwide. And you can create, to my point about power asymmetries, just more kind of strength and leverage in government so we can have a better shared.
[0:34:23] And I think if you do that, we'll actually get pretty close to a good place. Thank you. Actually, that replied to my question. I'm Lida Alepi. I am a policy advisor dealing with development and human rights.
[0:34:43] With the external monitoring system of the international partnerships from the Commission. And actually one of the priorities of the strategy up until 2027 is harnessing the opportunities that new technologies can have. And one of the things that are difficult is to have actually indicators that measure how do you bring this result all over the globe. So the question which has been replied both by Sasha and the last
[0:35:12] comment is, in practice, when there is lack of infrastructure in these countries, has there been, from your side up until now, a result based on cooperation from the private sector to the public? So having the engagement from the private sector that actually has the capacities. Whoever wants to reply, I'm not playing this game. Thanks.
[0:35:45] So happy to give an answer to that and also share the answer with Mike.
[0:35:51] I think very concretely, since you are coming from the Commission, the European Innovation Council, for example, and what the External Action Service is doing around building platforms for public-private partnership through the European Innovation Council, through, for example, the Digital for Development Hubs, is absolutely crucial. And I do see as a way forward a very crucial role for public-private sector partnerships, especially in places where infrastructure is limited. But I would add here, of course, that where infrastructure is limited,
[0:36:20] Cloud plays an additionally important role because it reduced costs, but it also provides access to a lot.
[0:36:27] IML services that would otherwise not be available so so here there's again the migration to the cloud as an important vector and the cooperation between public and private sector and I would say here again not just public private sector cooperation but also by bringing on board academia and civil society and end users who are also in a position to identify opportunities and solutions that can begin at the local level and be scaled globally.
[0:36:59] I've been just thinking a little bit more on the PPP, on the public-private partnerships. And I think this is a really difficult one to get right. And just risk and transfer of the risk from private to public or otherwise, whereas the public has not been very good at it. So to really have good structures, we still will need to work on this. But it is crucial that we provide it to the global south.
[0:37:28] Because otherwise others will provide it. And what is crucial is that it is provided in line with our values in multiple criteria, multiple dimensions from vendors, reliable vendors as the first thing to the services which are run in line with our values. And this takes me to the second point, which I wanted to make. It's really crucial we are able to measure the impact. So the positive, as you were describing, but as well, the negative.
[0:37:55] And for this, we will need to really develop new metrics. We don't have it at the moment. And I just give one example. We've been losing, at least myself, the more I'm using some of the maps, the more I'm using the ability to navigate in the space. And this might be really important for some communities. I hope that was a great answer. You even got two replies, which was illegal.
[0:38:23] That is better than you gambled for. I see a question from, I think it was Simeon in the back. Yes, thank you. So I'm Simeon Campos, founder of Safer AI, an AI auditing organization focused on general purpose AI systems. And my question is on the coding AIs that are coming. It's very likely that in the next couple of years, there will be highly competent coding AIs.
[0:38:53] And that has tremendous consequences. For instance, it could cause large scale hacking, things like that. And if you want to know more about that, there is one of the most prominent cybersecurity experts called Bruce Schneier, who published The Coming AI Hackers. And he's highly alarmist on that issue. And I'd be curious.
[0:39:17] To have your view on what it would take to govern such models, in particular, Jack, given that you are aware of the most capable models because you work in such an organization and you're in the governance space. Thank you. So this is really challenging. This illustrates the dual use problem of AI. It can do amazing things like we will revolutionize cyber defense. AI is going to actually improve.
[0:39:49] Like cyber defense amazingly through these tools. And then you have the problem of how to deal with cyber offense. A lot of it is about using the technology to harden systems, right? And try and create a defense dominant environment. But we'll also need to conduct some experiments to explore how effective these offensive capabilities are. And then if it turns out that they're legitimately state level important, then you need to involve states in a conversation about them. And I think it's going to be messy.
[0:40:16] And complicated, but I agree that is something that may be a few years out, but an important issue to raise. I want to take one final question from the room and one final online question. I'll start with the online question, which is similar to the one Jack just answered, but then to Sasha. How would you feel general purpose AI systems ought to be regulated from Amazon Web Services perspective?
[0:40:43] So I'll give an answer in two parts, building on the previous question that I answered from the audience, is that general purpose AI in building on pre-trained models plays a really important role in the twin digital and green transition. And in international fora, including in, for example, the 2019 OECD AI principles, but also from the European Parliament in 2021 that talks about...
[0:41:09] Specifically, incentivizing sustainable development and applications of AI. We see an increasing interest in looking at how to reduce the carbon footprint in Europe. And I know that IEEE and Clara Kneppel is here today is looking at how to develop even standards in this field that looks at the intersections of innovation and sustainability. And to that extent, when we can build on general AI systems and pre-trained models, it should be encouraged.
[0:41:38] General purpose AI. There are lots of different conversations going on, particularly at the EU level, where the Commission is discussing their approach. The Council is discussing their approach. The European Parliament is discussing another approach. One of the things that we see the need for, and which is also the challenge of general purpose AI, and I know that Jack shares this perspective, is that there is no intended purpose of general purpose AI. So how do you regulate a technology that has no intended purpose?
[0:42:08] And here again, I think there are lessons to be learned and was really excited to hear Paul earlier today from the work of the GDPR and Article 28, where there is a possibility to, for example, enter into contractual agreements between developers and deployers based on actual insights of actors in the AI value chain. And again, and Werner said it very clearly, that we need to understand what is being regulated and how and why and what is needed for us to comply.
[0:42:37] And in order to do that, that needs to be based on real insights. And so we do see the need to have tailored requirements in the framework of the EUAI Act and beyond with regards to how we will approach general purpose AI based on feasibility and also based on value chain insights. Thank you. And then I see a number, I saw a number of hands, but the final question goes to the gentleman here in the middle at the back. Yes. Thank you.
[0:43:05] So we heard that example of AI that code and start cyber attacks. I was actually one that said no, that makes also no sense to regulate on that bio weapons, simply because the ones who are doing these are the ones that don't care about regulation. So even if there's regulation, they would still attack and they would still develop these weapons.
[0:43:31] Now in the AI act for example we have that rule of CE marking and so kind of highlighting that something is using AI. Now do we create a false trust through that that we say well the ones that actually comply with AI they are kind of
[0:43:58] I have that signature. The ones that are intentionally trying to fake and attack us don't use it. So we create a trust while we make it even more open for those. And maybe to Marek, how do we think about that from the government perspective? How do we deal with those that intentionally do not want to follow regulation? This is a great question.
[0:44:24] And just Nico. Yes, with Nico, I'm the global partnership on AI. I am the managing director of Applied AI, a large national initiative in Germany focusing on kind of how to stay globally competitive as Europe. Well, this is a great question. And really, I'm afraid to be headed in that direction in many of the initiatives. And I don't want to be no naming. But I think what is really crucial is this more agile approach to regulation in order to understand if
[0:44:54] First of all, the impacts it has and then for that we need metrics and we as well need to establish the causality between the impacts and which we observe and the system which we think is actually causing these impacts. So we need the normal scientific approach, full scientific approach, large studies, many people involved. Second, we need to understand that the regulation we put in place actually delivers. So we need to, this is why I started, this is why
[0:45:25] I made a step back in my introductory remark and I went back to the human centric and human rights based approach because we need to understand why we regulate. So it's not just the checking the boxes exercise or kind of a little bit more sophisticated checking the boxes exercise using maybe some sophisticated tools provided by some very smart people, but exactly leading to the situation. And maybe I just use this. I have the mic.
[0:45:54] I'm even more afraid not only about the cybersecurity, but I'm really afraid of using AI to manipulate people. I think this is a core for me. This is a core theme for me. And what we need to do is, and this is as well the answer how to regulate the general purpose AI systems. We really need to see concrete use cases, but then we need to generate maybe two classes of use cases, which we either regulate or ban.
[0:46:24] And this is a little bit where the AI goes, for example, with the social scoring. So it is about understanding where the manipulation happens. We need to have, again, metrics for that and understanding of that impact. And then we need agile regulation, which will allow us to put it on the list and make sure we enforce it. With that, I want to...
[0:46:47] Thank you for some excellent answers. Marek's asked if he can have one sort of closing remark. So I'm going to grant you all a closing remark. But I've also seen that sign which was held up said that I had like two minutes and 20 seconds remaining. So if you could please do that in two sentences, that would be awesome. And Jack, could you would you be willing to take the floor first? We can have a truly fantastic future with AI and it depends on us.
[0:47:18] Actually making investments across society in academia and in government and in the private sector so that we can all work collectively to get to a good place. And I'm optimistic we can do that.
[0:47:33] I'm going to adopt the EU's digital compass approach and propose the four-pronged approach for this digital compass that I hope will continue to guide the way in which we develop jointly together responsible AI. The first word, I would say participation. And again, I really like this idea that Vila said around generative conversations, that we need participatory public policymaking processes, that we need everybody around the table, and that we need everybody who is...
[0:48:02] The second part of that compass I would say would be cooperation. We need cooperation between the public and private sector. We need that in policy, but we also need that in AI adoption.
[0:48:19] Third part of that compass, collaboration. We need more collaboration with academia. We need to be working closely with research. We need to be working backwards from what our goal is, not developing technology for technology's sake, but developing technology for impact. And then lastly, trust. We need trust, and that takes skills. I think Sasha is first, then the politicians actually. First of all,
[0:48:50] Big thanks for a great discussion. I think this was amazing and at first I think we need more discussions like this. I'm talking about not only this panel but all the panels. And second is I think we really need to develop more centralized approach to gathering the impressions or perceptions of the impact of AI and then using scientific methods really proving that it leads to concrete use cases and then maybe concrete even AI systems.
[0:49:20] Thank you. We covered a lot here from AI R&D to open source to capabilities to skills. And I know that's way more than you can fit into 55 minutes. So please use the drinks that will be following in a while to sort of continue this conversation. You made this panel sweat. I think you asked them a lot of difficult questions. So please give them a very warm round of applause.