Artificial General Intelligence and the Future of Responsible Governance
20 Feb 2026 11:00h - 12:00h
Artificial General Intelligence and the Future of Responsible Governance
Summary
The panel opened by noting the rapid acceleration of AI since 2020 and the emerging public debate about artificial general intelligence (AGI), warning that societies that ignore these trends may miss the chance to shape future governance [1-8]. Participants agreed that AGI is envisioned as a form of AI that can reason, learn, adapt, transfer knowledge and operate beyond narrow domains, unlike current systems that excel only in specific tasks [11-18].
Simonas Satunas offered a pragmatic definition of AGI as an entity capable of performing any human professional task with comparable accuracy, estimating a 3-to-7-year horizon based on growing public trust in generative AI tools [21-24]. Alexandra Bech Gjørv cautioned against fixing a timeline, emphasizing that progress depends on sustained investment, advanced low-latency hardware, neuromorphic and edge computing, and that privacy constraints on personal data may limit situational awareness [23-30][35-37]. Kenny Kesar highlighted that achieving the “five-nines” level of accuracy-a benchmark that historically required five to ten years per additional nine-will be essential for moving from probabilistic to more deterministic AI behavior [41-48].
Simonas Cerniauskas warned that current massive compute spending may be a bubble and that algorithmic efficiency could curb over-investment, while Simonas Satunas stressed that compute is only one of several critical components, including energy, data, and especially human critical-thinking capacity [70-71][72-90]. He also mapped AI risks into four layers-traditional privacy and security, mental-health impacts, social effects such as empathy erosion, and macro-level threats to democracy-calling for coordinated national and international strategies to mitigate them [131-139].
Alexandra argued that democratic access to compute must be paired with education of policymakers, noting that human oversight is often inadequate in ethical dilemmas and that algorithmic monitoring can reduce bias, as illustrated by a sports-analytics case where video review eliminated discriminatory decisions [96-101]. Kenny proposed institutionalizing “AI operating procedures” (AOP) analogous to current SOPs, training models to avoid bias and establishing external audits to ensure ethical compliance as AI approaches general intelligence [191-199].
The panelists concurred that early “anchor controls” such as robust labeling, regulatory measures, and resilience planning-including rollback mechanisms and diversified energy sources-are needed to limit harmful outcomes while enabling innovation [173][187-189]. They also agreed that collaboration among industry, academia, and governments is vital to embed egalitarian values and prevent profit-driven bias, citing examples like the amplification of violent content in Myanmar through platform algorithms [174-180]. In closing, Vinayak summarized that understanding AGI’s implications for security, privacy, and ethics requires immediate action, and announced the launch of an AI Cyber Security Terminal to support these efforts [202-207]. Overall, the discussion underscored a consensus that multidisciplinary governance, education, and measured investment are essential to steer AGI development responsibly [1-8].
Keypoints
Major discussion points
– Defining AGI and estimating its arrival – The panel opened by questioning what “Artificial General Intelligence” actually means and how soon it might appear. Vinayak highlighted the surge in AI breakthroughs since 2020 and the growing talk of AGI [1-4]. Simonas Cerniauskas noted common traits of AGI such as reasoning, learning, adaptation and knowledge transfer [12-18]. Simonas Satunas offered a concrete (though simplified) definition – an AI that can perform any human professional task at human-level accuracy – and projected a 3-to-7-year horizon [21].
– Compute, hardware, and investment as enablers (and possible bottlenecks) – Several speakers stressed that massive compute power, new architectures and funding are critical to reaching AGI. Cerniauskas described the current “super-high-cycle” of investment and the risk of a bubble [70-71]. Simonas Satunas used a 19th-century transport metaphor to argue that compute is only one element among energy, data, implementation and human skills [72-85][86-87]. Alexandra added that low-latency, energy-efficient neuromorphic and edge hardware are required for human-like situational awareness [31-34].
– Security, privacy and ethical threats of powerful AI – The conversation turned to the dangers that more capable models pose. Kenny warned that the same AI that creates content can also generate sophisticated attacks, and that an AGI could impersonate humans such as CEOs [105-108]. Simonas Satunas broke down AI risks into four layers – classic cyber-security/privacy, mental-health impacts, social-cohesion, and macro-societal threats to democracy [131-138]. Alexandra highlighted the need for human oversight and the difficulty of making ethical decisions in autonomous systems [96-102].
– Human factors: critical thinking, education and regulation – Multiple panelists argued that technology alone will not solve the challenges; societies must boost critical thinking and regulatory capacity. Simonas Satunas stressed that raising public critical thinking is as important as investing in compute [88-92]. Kenny pointed out that over-reliance on AI-generated content could erode our own reasoning muscles, creating a “vicious cycle” [164-170]. Alexandra called for educating politicians and the public so that ethical choices are made before machines dominate [96-102][187-190].
– Early-stage governance and “anchor-control” concepts – The moderator asked for concrete steps that can be taken now. Cerniauskas suggested technical measures such as watermarking/labeling and hinted at regulatory actions [173-176]. Simonas Satunas advocated for global coordination, industry-academia collaboration, and embedding egalitarian values into AI design [174-180]. Alexandra proposed resilience measures, robust rollback mechanisms, and scenario planning for infrastructure loss [187-190]. Kenny introduced the idea of an “AI Operating Procedure” (AOP) to embed bias-checks, ethical reviews and continuous monitoring into AI deployments [191-199].
Overall purpose / goal of the discussion
The panel aimed to demystify AGI-clarifying its definition, likely timeline, and technical prerequisites-while simultaneously surfacing the security, privacy, ethical, and societal risks that accompany rapid AI advancement. By juxtaposing technical optimism with cautionary perspectives, the participants sought to identify practical “anchor-control” measures and governance frameworks that can be instituted today to steer the emergence of AGI responsibly.
Overall tone and its evolution
– Opening (0:00-3:30) – Curious and forward-looking, with speakers outlining possibilities and expressing excitement about breakthroughs.
– Middle (3:30-15:00) – The tone shifts to a more cautionary stance, emphasizing the gaps between current narrow AI and true AGI, and flagging looming security and ethical threats.
– Later (15:00-35:00) – Concern deepens as concrete risks (misinformation, bias, cyber-attacks) are discussed, but a collaborative, problem-solving attitude emerges.
– Closing (35:00-end) – The discussion becomes pragmatic and solution-oriented, focusing on governance, resilience, education and concrete early-stage controls.
Thus, the conversation moves from exploratory optimism to measured concern and finally to actionable recommendations.
Speakers
– Ms. Alexandra Bech Gjørv – Head of Sintef, Norway’s largest research institute; expertise in AI research, neuromorphic and edge computing, and AI governance.
– Mr. Vinayak Godse – Moderator/host of the panel discussion on AGI; involved in AI policy and security discussions.
– Mr. Simonas Satunas – Speaker on AGI, provides definitions and timelines; background in AI development and public engagement (Israel).
– Mr. Kenny Kesar – Speaker on AI accuracy, compute, and market disruption; experience in AI consulting and implementation for clients.
– Simonas Cerniauskas – Speaker focusing on AI investment cycles, compute efficiency, and regulatory perspectives.
Additional speakers:
– None (all participants in the transcript are covered by the speakers list).
The session opened with moderator Vinayak Godse framing the rapid acceleration of artificial-intelligence research that began around 2020 and intensified after the launch of powerful generative models in early 2023, warning that societies that ignore these developments risk missing the chance to shape the governance of the next technological wave, possibly the arrival of artificial general intelligence (AGI) within the next two to ten years [1-8].
Defining AGI – Cerniauskas said most definitions agree that AGI should be able to reason, learn, adapt and transfer knowledge, and that it must be broader than today’s narrow-domain systems such as customer-service bots [12-18]. Building on this, Satunas offered a pragmatic, human-centric formulation: an AI that can perform any professional task with the same accuracy and professionalism as a human expert. He linked this functional view to a growing public trust in generative tools, noting that roughly half of Israeli respondents already trust AI more than their friends, which he interprets as a step toward AGI [21-24].
Timeline and investment uncertainty – Satunas projected a 3-to-7-year horizon, arguing that the convergence of technical capability and societal trust makes the milestone imminent [21-24]. By contrast, Gjørv rejected a fixed schedule, insisting that progress depends on sustained investment, hardware breakthroughs, data-privacy and regulatory challenges, and warned that policy should ensure broad, democratic access to compute resources rather than concentrating power in a few providers [23-26]. Godse echoed this uncertainty, urging societies to prepare now rather than wait for a precise date [1-7]. Cerniauskas described the current “super-high-cycle” of compute spending as potentially speculative, noting industry chatter about a bubble and the possibility of over-capacity persisting for years, as even Mark Zuckerberg has suggested [70-71].
Technical prerequisites – Gjørv highlighted that human-like situational awareness will require ultra-low-latency, energy-efficient hardware such as neuromorphic and edge-computing architectures, together with massive private data streams – a requirement that immediately raises privacy concerns [31-34][35-37]. Vinayak asked about the latency of System 2-type reasoning and the limitations of language-only models; the panel noted that current large language models (LLMs) excel at fast, intuitive (System 1) pattern-matching but struggle with deep, logical (System 2) contextual understanding, exposing a key bottleneck for AGI [95-98]. Kesar framed progress in terms of accuracy, invoking the “five-nines” benchmark: “to get from 90 % to 99 % accuracy took five to ten years”, and argued that each additional nine adds one to two more years, driving compute demand toward AGI [44-48]. Satunas broadened the picture with a 19th-century transport metaphor, arguing that compute is only one link in a chain that also includes energy, data, implementation, language localisation and, crucially, human critical-thinking capacity [72-90].
Security, privacy and risk taxonomy – Kesar warned that the same generative models that create content can also craft sophisticated cyber-attacks and impersonate senior executives, making future threats “real” once AGI can emulate human behaviour [105-108]. Gjørv added that achieving true situational awareness would require access to personal data, but privacy regulations limit such collection, creating a tension between capability and rights [35-37]. Satunas categorised AI risks into four layers: (1) classic privacy, security and fraud; (2) mental-health impacts; (3) social effects such as erosion of empathy and bullying; and (4) macro-level threats to democracy through manipulation and fake-news campaigns [131-138].
Human factors – Satunas argued that without a strong emphasis on critical-thinking education, societies will be unable to recognise AI-generated manipulation; he noted that 30 % of online content is already AI-generated, creating a feedback loop that could stall human intellectual growth [154-155][165-170]. Kesar echoed this, warning that reliance on AI-generated content may erode the “brain-muscle” needed for innovation, leading to a vicious cycle where AI diminishes the very intelligence it seeks to emulate [164-170]. Gjørv reinforced the need for political and public education, pointing out that human oversight often fails in ethical dilemmas and that policymakers must be equipped to make hard choices before machines do [96-102]. Cerniauskas also noted the importance of the human critical-thinking element as part of the broader ecosystem [12-18].
Early-stage “anchor-control” proposals – The panel offered a spectrum of concrete measures:
– Technical and regulatory safeguards (e.g., watermarking, output labeling) – Cerniauskas [173-176];
– Resilience planning (robust rollback mechanisms, diversified energy sources, scenario-based risk matrices) – Gjørv [187-189];
– AI Operating Procedure (AOP) – a procedural framework embedding bias-audits, ethical training and continuous monitoring, analogous to traditional SOPs – Kesar [191-199];
– Global regulatory collaboration – especially for smaller nations, to embed egalitarian values and mitigate bias, citing the Myanmar example where platform algorithms amplified violent content – Satunas [174-180].
Points of agreement – All speakers concurred that education, awareness and critical-thinking skills are essential to counter AI-driven threats [154-155][164-170]; they also agreed on the need for layered risk-management frameworks that combine technical safeguards, resilience planning and procedural oversight [187-189][131-138][191-199]. Both Gjørv and Satunas highlighted privacy as a fundamental constraint on the data required for human-level situational awareness [35-37][131-138]. The panel agreed that the proliferation of AI-generated misinformation poses a serious societal risk [144-149].
Remaining disagreements – The timeline for AGI remained contested (Satunas’ 3-to-7-year estimate vs. Gjørv’s refusal to set a horizon vs. Godse’s call for preparedness). On the primary driver of progress, Kesar foregrounded compute-driven accuracy improvements, while Satunas argued for a holistic ecosystem, and Gjørv emphasised specialised low-latency hardware as the bottleneck. Regarding governance mechanisms, the four speakers advocated different early-stage toolkits, reflecting a lack of consensus on the optimal approach.
Closing remarks and announcement – Godse summarised the collective insight: while the acceleration of AI capabilities creates unprecedented opportunities, immediate action is required to embed security, privacy, safety and ethical safeguards into the emerging paradigm [202-207]. He concluded by announcing the launch of the “AI Cyber Security Terminal” as the session’s final action [208-210].
The panel’s recommendations can be grouped as follows: (i) institute early anchor controls such as output labeling and technical safeguards; (ii) invest in education programmes that foster critical-thinking and AI literacy; (iii) foster cross-sector collaboration to develop global, risk-adaptive regulatory frameworks; (iv) adopt AI Operating Procedures that institutionalise bias-checks and ethical reviews; and (v) design resilience and rollback mechanisms to limit the impact of failures or malicious use. These steps aim to steer the trajectory toward AGI responsibly, balancing compute-driven progress with human-centred governance. [173-176][187-189][191-199][202-207][208-210]
Pet Summit and the basic idea and intent behind setting up this session is while all the things were happening in AI in the period of 2020, a lot of development happening and somehow all that is now leading to kind of acceleration that we are seeing in last three years of time and especially this year, since January, all the new launches that we see, we are getting the first sign of a powerful AI, right? And now because of that, there is a discussion about AGI seems to be gaining quite a significant ground, right? And although people still have a lot of doubt and skepticism about whether it is really reality or possibility in coming future or what that means, many people are still skeptical.
They are struggling to define what that means for a cigarette. So as an overall society. and I can tell about India so probably we didn’t pay much attention when AI was coming. If you don’t pay attention now what is coming in next 2, 3, 5 years of time or 10 years of time that is probably the timeline for AGI, then probably we will miss on again thinking, talking, discussing, governing it better basically. So this discussion is about what is to help understand for us and for the audience here basically what do we mean by AGI can we really think about that right now what are different conference that we need to thank you for getting welcome to the panel and try to then find the meaning possible meaning for security, privacy and ethics basically.
So I would like to talk with someone with you, so how do you see this concept of AGI and formulationally how that will be different that we would see what is your understanding about the concept of artificial intelligence and artificial intelligence
So, yeah, thank you very much for having us here. And, yeah, like you said, it’s a really nice topic to wrap up the conference. So, well, so, you know, of course, there are kind of different definitions of AGI. And on the same time, most of them agree that it’s, you know, it’s about smarter AI than we have right now. We were joking a bit that, you know, on the way, the traffic is really, you know, exceptional. And, yeah, that’s a sign that maybe we are still not here today. So, but, yeah, but basically kind of among those common agreements that, let’s say, the smarter AI should reason. It should learn. It should adapt. And also it should transfer knowledge.
And also it shouldn’t be, you know, very. narrow, like, you know, of course, right now we have great, let’s say, areas where AI is really helping a lot, like co -development, customer service, and et cetera, but, you know, it should be much broader. So, and, you know, don’t think that any of us, maybe the colleagues will be able to answer when we will have, you know, and what timing, but definitely, you know, that’s one of the big topics right now.
Let me come to you and you look at the digital initiative and artificial intelligence as one of the important research areas, so we are grappling with understanding what is right now, but can we think about what would happen in the next three, five years of time, and that seems to be the timeline for each area.
So I’m the one with the date I’ll do my best So first of all my definition of AGI is very simplistic and I think that we need some simple explanation in this field and my very simple explanation is AGI will be something that can perform every human task at the level of accuracy and professionality of a human professional Now this is not an optimal definition because people can ask every task if a baby is crying will the AGI help him stop crying and people can ask what is the level of professionality but I think that this is something that we can digest and I think that for me I understood that we are getting closer there not from a technology perspective but from the perspective of talking with real Israelis about their problems and five years ago when I was telling this definition of AGI people were like oh it’ll never happen not in our lifetime and right now when I’m speaking with Israelis and I’m telling them this is AGI they’re saying oh aren’t we there yet oh because I thought that Chachi Biddy can help me like a lawyer isn’t it true now I think that we are not there yet okay there is a very sharp line between the AI that we are experiencing today and true AGI but the fact that the audience is already confusing the fact that people give trust to Gen AI tools 50 % of Israelis trust them more than they trust their friends many trust them more than they trust human professionals this puts us closer to AGI so I would say that it’s a matter of 3 years to 7 years until we reach that milestone
so coming to you Alexandra how do you see this as a concept what is leading to this AGI what would we do that will impact the future of the AI bring this age of Asia in three or seven years of time?
Well, I’m not necessarily subscribing to the time frame. I think that depends on how much money we throw at it. And then there are other things to throw money at as well. Some of this, for example, we had a discussion with my team, you know, are machines able to make complex decisions as fast as humans? And in some areas, like, you know, many operations demand millisecond response and reflex level. You know, you can see that machines are quite good at detecting fire or doing various instinctive things as fast as we are, but the ability to interpret context, emotions, ambiguity, surroundings, body language, etc., that’s still quite far away. They take too long. And in a dynamic environment, you know, a wrong decision or a late decision is really a wrong decision.
So in order to get there, I, you know, there’s both low latency, energy efficient hardware, neuromorphic and edge computing and architectures beyond auto regression. But I think, you know, the researchers in Sintef, I head up the largest research institute in Norway. They, you know, they point to promising like hierarchical reflex reasoning systems, embodied multimodal learning, et cetera, et cetera. And there’s really no real doubt that you will get there. But there’s, in order to have the situational awareness like a human, you have to study a lot of data that would be considered private, personal. So there’s really limits on privacy. And then it triggers a lot of other questions that I’m sure we’ll get into.
Yeah, we’ll come to that. So, Mr. Kenney, you must be serving many of the clients right now on AI, right? And every of us are getting stunned by… the progress and acceleration of the capability that is happening week by week basically right and that also scares us what is coming next right and when it comes to that level where there is a there is a two words uh somebody defines agi right so one is the consistency across the domain uh that it will be so general in a way that it will be consistently performing across the domain and second part is uh it will be reliable as well so currently probably sometimes it doesn’t have anything and it throws output and that’s why hallucination happens basically so consistency and reliability that’s what the agi will bring to the table basically so it will solve a lot of problems that we see uh uh right now we have been also getting stunned by the things that it can do basically so so there are routes to achieve the agi which will lead us to agi basically so how do you think uh uh your perspective the the journey that probably take us there
So, you know, I agree with the panel that a couple of things we talked about in terms of where we’re getting to models evolving. But you bring up another component of accuracy. I’ll talk about accuracy first, and then I’ll come back to the disruption which is happening in the market. Now, the epitome of accuracy is five nines. So for AI to get from 90 % to 99%, it took five to ten years. Now, every nine that you add is another year or two years to the point where you get to 99 .99 and nines. So every nine that you’re adding has a time frame to it. And the number of nines that you add, you get closer to general intelligence because that’s what is going to look at the human brain.
I’ll take the topic of photographic regression that you talked about. Any regression, AI is right now built on regression. It’s built on learnings of the neural network. The neural network maturing on information that it sees. but the human brain is also inventing. It’s researching. So when AI really gets to the point of being able to research and bring new ideas to life that a human brain does, you’re getting closer to intelligence. Now, the disruption in the market that you’ve seen with announcements across the different players which dominate the AI market is creating a disruption in the industry and I think it’s the right disruption. It’s the disruption that word processor did to typewriter, what computers did to word processor, and what cloud did to data center.
This is another thing, but it’s much faster because it’s more pervasive and it impacts everybody in life. So the fact is people are talking about how does it translate to me. When I say it translates to me, it’s about how do we structure processes. Everybody and I agree accuracy is work in process. And since accuracy is work in process, we have to be really mature about… the use cases that we put onto it. We have to look at the human pyramid, what components of the pyramid that you’re going to look at. So the way we are advising our clients and what we’re doing ours is maker jobs, which is basically repetitive jobs with little context.
AI does very well, but create a controller for these autonomous. So combination of probabilistic and deterministic is what’s going to be the near future as we get to more and more deterministic when we get to general intelligence, because from a human perspective, it’s mostly deterministic.
Right. Yeah. So these are and thank you all for putting some level of clarity in terms of what this means. And so at the end of the day, Asia is like so they say attention, right? Ability to give attention to all possible thing that. People, millions and billions of people asking questions. but as you rightly say the context matters so it’s not only attention the it should be contextual to your requirement and your things that you do right and third important part of which they are doing and last six months had been a great months for reasoning that bring to the table basically so my question is and anybody of you can answer this you then for achieving all of these things so why compute becomes very important so why you need this much of compute why there are trillions of dollars that is invested to make sure that it it use attention to each and every problem better and it is contextual and you reasoning and at the same time latency as I talk about so the role of compete what is the role of competitive this any of you
yeah so you know so of course if I may start and of course please accompany so currently we are at super high cycle let’s say of those investments and most of us are also wondering is it a bubble or when it will blow a bit etc is it really in some cases sustainable everyone of us most likely has our own opinion but still this race to be number let’s say one this belief that if you are number one you will remain number one and this momentum I think plus huge appetite all this hype definitely brings much much more money to the table than we could ever imagine and you know on the same time it depends a lot of course on the algorithms how efficient they will be all of us remember most likely last year this deep sea moment and there are also other models which are much more efficient but so So, you know, at some point we might understand that it’s overestimated, overinvested.
At the same time, I remember in Zuckerberg’s quotes that, you know, said, okay, in the worst case scenario, I will, you know, have overcapacity for a couple more years and then I will use it.
So my humble opinion is that compute is one element in a chain of elements and that sometimes we treat this element as the only one. Let’s explore a metaphor. Let’s imagine that we are in the 19th century and a prophet arrives and he tells us, okay, in five years, a new technology will emerge that will enable you to arrive from Delhi to Bangkok in less than an hour. But I don’t know what the technology is. Maybe it’s a ship, maybe it’s a car, maybe it’s a train, maybe it’s an airplane, but we must be prepared. So everyone is trying to be prepared and to build the right infrastructure. So let’s look at the structure. The problem is everyone thinks about it as something else.
So one will build an airport and the other one will build rails and the other one will build boats. I think that we are in this moment. We know that AGI will arrive. We know that it is soon and we know that we must be prepared. Compute is one of the elements that is necessary, but energy is also important and heating and cold is also important. Data is extremely important. Implementation is important. Language is important in India as well. I think that one of the elements that we are not investing enough is the human element. Think about critical thinking, for example. I don’t know what AGI will arrive, but I know that already now for us it is very important to raise critical thinking among the public.
When you hear something in the news, when you see something, was it made by AI? What is the manipulation that is being forced upon me? So I think that investing in education is not less critical than investing in computing.
And then another element I want to come to you on this that you talked about. there is very interesting discussion about this system one and system two thinking human is more intuitive in terms of response and system two is more logical and AI is probably helping with that basically but there is a latency that is an important area and that’s why they are putting a lot of effort and improving the competence such that the latency of system two thinking is also less so that your intuitive thinking can improve with that basically but it’s not only the competence the perception, the ambient, the senses, the emotions so all that also matters a lot and that’s where the limitation of language based models are getting exposed basically and you did talk about that in your initial remark can you just throw light on that?
On the language? On the different type of the models right? Ambient, compute for that matter, world model that people talk about so…
Well I just wanted to first agree with the… Mir, sorry that you know if you are a government and this is democratic access to compute is a big topic I think you can really get lost in just investing in compute power so investing in skills and leading edge technology understanding in your own country and participating in the regulatory approach because some of the things that I care about is that everybody says that they should be human oversight but you know that once you get into these dilemma situations like what should happen in a car accident, humans are not very good at understanding risks and humans are not very good at really making ethical discussions they tend to go as far, you know, do your best and then let moral luck decide who gets lost but you have to in machine driven systems you actually have to make decisions about those things so I think becoming, you know, educating also our politicians to be able to to know that you have to make the hard choices because otherwise the machines will make them for you and they will continue our biases and they will, you know, it will not end well.
But then I just wanted to share a little story that I heard. You know, Michael Lewis, the guy with the money ball and everything, he has this anecdote that in the Basketball Association in the States, they started video surveillance and the coaches were all making racist decisions and home team decisions. And by showing the videos and by showing the statistics, next season they couldn’t find any bias at all. So I think that’s a good example of how the machines make people better, whereas we’re not able to better ourselves over time. So I think I just thought this was a nice anecdote for this
Thank you. And I’ll come to Kenny. So… As we are… trying to solve problems of security, privacy in current big capability of AI and we are struggling to understand what it means for security, what it means for privacy and suddenly there is a significant acceleration that is happening so what we are doing right now for security privacy which could help us to graduate to more and more powerful model comes in or any other things basically so can you just help us
yeah I think security as we evolve and we talked about compute compute gets bigger, context get bigger, we get smarter in terms of what AI can do and definitely the same AI that can generate, can pose more sophisticated attacks and when we get to AGI right, the biggest thing is I could be emulating a human Let’s say in a company, I could emulate a CEO and make a decision because I’m getting so close to being natural. The threat is real. Now, even today, let’s say without AI, you need to be just a step ahead of the bad actors or the persons who are into cybercrime. You just have to be a step ahead. And similarly, we talked about, you know, we’re mentioning about the human portion, right?
That the human portion needs to get more educated where there are going to be set of humans that are going to use the same AI to build better agents to fight them. So now it’s a question of the tooling that you have at hand. Even today, it’s the tools. It’s a human who’s building tools to fight your cyber threats. Imagine, in the next era, the only thing is… It’ll become nearly close to science fiction when agents try locking humans out. But that’s, I would say, still science fiction. But the fact is as we evolve, we need to right -size the solution and that’s how we will manage compute too. You don’t use I7 computer or to do a simple calculator task of adding two numbers, right?
You use a calculator. So in the context of the world, we’re going to have SLMs which is small language models that will do smaller things so that we can manage compute. You have the bigger models that will solve world hunger in terms of how we do with different levels of machines and processing that we do. I think there will be tiering. Right now, we were talking about it’s a fight to who’s first. So with the fight to first, bigger, better, more elaborate. But now as it evolves, you’ll get the right size fitting to them. Then only it will be commercially viable. AI is not commercially viable today. The costs outweigh the RO.
Yeah, current cost is quite significantly higher. You can do POC but… once you put into production environment the token cost is too much high to the ROI so so near want to come to you there is a established understanding of security privacy safety or ethics right and that’s what the paradigm that we at least try to understand right now but would the Asia altogether different paradigm and the concepts of security privacy will be foundationally very different than what we discussed right now
so as I see it when we try to deal with the risks that AI pose we distinguish between four different levels the first level is the classical risks like privacy security cyber fraud every technology that we have since the 90s we need to explain how does it meet the current risk in that matter and AI is much more powerful and it poses a lot of more risks but these are the kinds of risks that we when we design products we know how to deal with them. Above it there is a level of human health and mental health and we find out that AI solutions can be quite problematic for mental health, can cause a lot of damage in some cases and this is something that is not yet well understood and investigated above that there is a social level.
What does it does to the empathy between people? What does it does normally people say oh I see that it’s bad for my kids. They are experiencing bullying or addiction usually what’s bad for your kids is also bad for you and we understand that these are complications that we didn’t think about when we code and the higher level is a macro level what does it do to society? What does it do to democracy? I think that several countries are now experiencing foreign manipulation and it is very easy to run campaigns that are built of fake news and we see that manipulation can become very problematic. So I think that a strategy, a national strategy and an international strategy should access, should address all these levels and all these levels have mitigations but they are costly and they need collaboration.
So we need to be in close collaboration in order to mitigate these risks.
It’s good that the way you put the structure, right? Things it would do to us, our brain and the thing that will impact us as individually and we discussed that in one of the sessions that we hosted on neuroscience and AI. So what this means to the brain development process if we are using AI for every small thing that we want to do, what that means to society, brain development process plateaus for that matter, what will be in society and then what is the macro kind of impact it. Do you want to add something on that?
Yeah, I just, sorry. I just want to build on that. How it’s not just targeted manipulation or the things that we see in our kids and somebody walking around with a button called friend and that’s your only friend that you need but also the well -structured in the geopolitical context the ability to create completely different information universes you don’t need to be neurologically strange you just see a completely different view we just published a paper in science on these agent swarms and just reading a book about the Ukraine and Russia war going on now and how large populations are overpowered by totally different images of the world from what we are and at least obviously your defense systems need to be hardened against those kinds of manipulations but it’s also you know actually an offensive strategy to find good bots that enter those universes.
It’s an actual battleground in and of itself, and it’s very strange to think about the world in that way, but I think you’re very naive if you don’t start systematically working on how you make your conviction of what the world is like also part of the people that you need to somehow, hopefully not defeat, but relate to and convince that things can be better. So it’s not just a technological challenge. I would say it’s a huge mental leap for most of us.
So Siman, the question is like the more we use, the more we become dependent on AI system, right? And the more acceleration of the people’s ability to think critically, that will go down basically, right? The speed will increase the more dependence, and then more… More AI become powerful for that matter, right? so what we see in terms of this misinformation, disinformation and defake, so probably there will be different kind of cognitive warfare that may happen so how do you see such kind of challenges in the society, you talked about society or individual for that matter, so what kind of implication it will have for individual society and overall the way the world is organized
yeah so absolutely so basically all those layers and all the dependencies like you rightly stated they also critical thinking of course is one but also awareness, education and you know the skills, abilities for people to understand the things here I think this audience is you know for us it’s more or less everything self obvious but you know when you start talking to people in the streets or different backgrounds then you you know realize that what is self -obvious for you for another person might be completely different. To find those ways I would say to educate to basically help them identify the threats, that’s one of the key priorities and also obligation I would say from our side.
one of the important challenge of this critical thinking which I come across is critical thinking is nothing but your ability to give attention to various different dimensions nuances, different perspective, different views basically right. Where it is tremendous amount of effort that I would have to become a critical thinker. And AI saws that quite easily for me. It can make me to bring all the attention, all the dimensions, all the nuances, all the viewpoints, you can quickly get access to me, right? So, even for critical thinking, Kenny, for you, this question is, you will be depending too much on AI as well, right? So, we need to know distinction between what do you critical thinking? Critical thinking is not just getting information, giving attention, but critical thinking is what?
So, that question probably is very important question to ask.
critical thinking that is very necessary for us to innovate further. So the biggest issue that the AI world is facing, 30 % of the content is consuming is AI generated already. So basically you’re feeding back and it’s learning on the same model. When originally it was learning on artifacts that were built through different thinking processes. So I would say one of the, it’s a risk, it’s a boon because it gets work done. But in overtime it’s a risk that we will stop evolving because if we don’t exercise the brain as a muscle, if we don’t exercise it and don’t build those neurons which really influence critical thinking, it will be actually a very big loss to society.
So I would say general intelligence, everybody is asking for it. Now how do we make sure as AI. computers get general intelligence we’re not losing our intelligence to create that general intelligence again so it’s a it’s a it’s a vicious cycle it’s a question which we’re debating we’re trying to answer in ourselves everybody has perspectives but it’s a it’s something that I think about do I have an answer to it no but I feel that critical thinking on both sides is something that we really need to critically think about
yeah so that’s what may every thing that you think as a solution and kind of thing so there is always this challenge of what it means right in this new paradigm is an important so now a little bit concluding part of this discussion is can we when this is question to each of you briefly we can discuss about it can we still think about I know we know we have been doing security privacy and particular safety privacy particular way right but as this paradigm is new can we think about some anchor control right now that we should be mindful of right that when it comes it happened right when AI was getting built after 3 years we are talking about AI governance and all these things so is there a way for us to think about some kind of anchor control some idea some concept basically that could help us to browse through challenges the AGI could throw I can start with you briefly and each of you can comment on this
yeah so well of course you know there are some technical things like you know the same what are marks or something you know labeling and other technical features that could help us a bit to identify at least some threats … then also we can talk about regulator measures but you know that’s a broader topic for the further discussion but especially here we in Europe we tend to regulate and overregulate everything so but in a way I think also at least some measures here also can be really viable and really reasonable
well I come from a very small country Israel is so small that you can put it it’s like a pin on the map and therefore our regulative approach is that we are unable to determine the global regulation and in this AI race I think that what is more important is the global regulation so since we are a very tiny country we must work with positive tools say, okay, we cannot affect the regulation, but how can we work together with the AI developers in order to make the personality of the AI more moral, more ethic? How can we put egalitarian and equality into the consideration? How can we avoid bias? And I think that it makes us work together with the industry and together with the academia in order to find out about new consequences.
I think that in many cases the giants, the big tech doesn’t point towards unethical conclusions, but they work towards financial incentives that make AI behave in a very immoral way. If I’ll take, for example, the conflict in Myanmar, in Burma, we saw that Meta was not actively promoting violence in Myanmar, but the algorithm of Meta was designed to attract attention in a way that make the AI the more violent post much more viral and make violence flourish. So if we’ll be able to promote a dialogue and if we’ll be able to be together with the industry in development of new AI, sometimes we’ll be able to make AI more ethical.
So Alexandra, your view. So one is the anchor control idea concept, but second part is how do you get into early? How do you get into? Early in the game, right? So when AI happened, now we are discussing in 25, 26 about the responsibility and alignment and adoption and governance basically, right? So in Asia discussion is the anchor control ways, ideas and ways for us to get into early discussion of it.
Well, I think at least you need to work on resilience and robust rollback mechanisms. A little bit like what we’re experiencing now in Europe, where we all have to practice on living without electricity. You know that it’s a realistic option that somebody. sabotages your electricity and then looking at well how dependent are we really and what are the alternative you know and and planning from a point of view where you not only work to reduce risk but you really work to reduce consequences of those risks occurring so if you work on the traditional risk matrix it’s always you know avoiding bad outcomes but then making the bad outcomes less bad that’s something that at least we think is well the new realities are propelling that kind of thinking and I think that’s important
Kenny your voice on this
sure actually the way we look at it in terms of AI from ethical AI to biases to data privacy it’s very similar akin to what a human would do even today what today we have a standard operating procedure that we review for biases, we review for content. You know, in our organizations, we have organizations that manage this. Now, and the other thing is we train people on ethical practices, on non -bias and things like that. So ultimately, AI is very similar to that, where we will have, you know, in today’s world, for the lack of a better word, I call it AOP instead of SOP, agent operating procedure or AI operating procedure, where we have to train AI in terms not to be biased.
So I feel that there is a big industry which is in the offing, which is going to manage and create models, LLMs, to manage or to validate that the responses from, you know, your common models are ethically right, non -biased. Because today, as organizations, we invite experts from outside to come and see our practices, whether we are following ethical, we are transparent, a number of those things. Very similarly as we mature towards more general intelligence and the more ways of working, I feel that these control structures will come in cyber security, will come in ethical use of AI, unbiased use of AI. So ultimately it will be a checks and balances system and we will see innovation in these areas.
That is how we feel it. It’s an evolving area. Let’s see how it happens.
Thank you all of you to really help us understand the meaning of this concept of AGI and how that will pan out from now and what kind of challenges it will throw to us. There are definitely opportunities that we don’t have time to discuss about what it will bring to us. But then what could we start doing right now? And this was definitely one of the important conversations. Help this would help you understand what we are talking about the AGI today. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Join me to give big hand to my co -panelists for helping us understand. Thank you. Thank you, Simon. Thank you, Nir.
Thank you. We have some photo shoot. Alexandra, we need to come here for photo shoot. I also request the fireside panels, Hendrikus sir and Narendra sir to please join us for the photo shoot. Thank you. Thank you. Before we commence the session for the Fireside I would like to announce the launch I would like to announce the launch of AI Cyber Security Terminal This is published today Thank you. Thank you. you you Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you.
“Kesar introduced the concept of accuracy progression through “five nines,” explaining that AI evolved from 90 % to 99 % accuracy over several years and each additional nine requires increasingly longer timeframes.”
The knowledge base explicitly describes Kesar’s “five-nines” accuracy benchmark and the increasing time required for each additional nine of accuracy [S1].
“Progress toward AGI depends on sustained investment, hardware breakthroughs, data‑privacy and regulatory challenges.”
Long-term sustained investment is highlighted as essential for fundamental research breakthroughs, providing context for the claim about investment dependence [S92].
“Industry chatter about a speculative bubble in compute spending, with concerns that over‑capacity may persist for years.”
An Alibaba Group chairman argued that AI investment is not a speculative bubble, offering a contrasting perspective that adds nuance to the bubble discussion [S99].
“Mark Zuckerberg has suggested concerns about the compute spending cycle and over‑capacity.”
Zuckerberg publicly stated Meta’s long-term vision is to develop AGI and make it open source, confirming his active involvement in AGI discourse, though the source does not mention a bubble comment [S31].
“The moderator framed AI research as accelerating rapidly since around 2020 and intensifying after early‑2023 generative model releases.”
The knowledge base notes that artificial intelligence is advancing at a rapid pace, supporting the moderator’s framing of accelerated AI progress [S82].
The panel shows strong convergence on four pillars: (1) education and critical‑thinking as a defence against AI misuse; (2) comprehensive risk‑management frameworks including resilience, rollback and procedural safeguards; (3) recognition of privacy as a limiting factor for data‑intensive AGI; (4) acknowledgement that compute is essential but must be balanced with data, energy and human expertise. There is also broad agreement that AI‑generated misinformation threatens societal cohesion.
High consensus on governance, risk management and capacity‑building measures, moderate consensus on technical pathways (compute, hardware). This suggests that future policy discussions can build on a shared foundation of education, risk controls and privacy safeguards while still debating timelines and specific technical solutions.
The panel shows substantial divergence on three core fronts: (1) the expected timeline for AGI, with one speaker offering a short‑term estimate and others rejecting a fixed horizon; (2) the relative importance of compute versus hardware versus a holistic resource mix; (3) the optimal early‑stage governance toolkit, ranging from technical anchor controls to resilience planning, procedural AOPs, and education‑driven regulation. While there is consensus on the need for education, critical thinking and multi‑layered risk awareness, the lack of alignment on strategic priorities could hinder coordinated policy responses and investment decisions.
High – the disagreements touch on fundamental strategic choices (timing, resource allocation, governance architecture) that shape national and international AI policy. Without a shared roadmap, stakeholders may pursue conflicting initiatives, leading to fragmented regulation, duplicated investments, and potential gaps in security and ethical safeguards.
The discussion evolved from a broad framing of AGI’s emergence to a multi‑layered analysis of technical, societal, and economic dimensions. Key comments—especially those that introduced concrete definitions, quantitative benchmarks, risk taxonomies, and real‑world examples—served as turning points that redirected the conversation toward actionable insights. By juxtaposing optimism about rapid progress with cautionary notes on over‑investment, privacy, and human cognition, the panel collectively moved from speculative timelines to a nuanced roadmap that balances compute, hardware, regulation, education, and ethical safeguards. These pivotal remarks shaped the dialogue into a structured, forward‑looking discourse on how to responsibly navigate the path toward AGI.
Disclaimer: This is not an official session record. DiploAI generates these resources from audiovisual recordings, and they are presented as-is, including potential errors. Due to logistical challenges, such as discrepancies in audio/video or transcripts, names may be misspelled. We strive for accuracy to the best of our ability.
Related event
