Current research trajectories and indicators

So, on this slide here, the purpose of this section is to determine when would artificial general intelligence happen, when is the singularity event likely to happen, and when is artificial superintelligence likely to happen. And for that, we're going to be referring to the recent growth in AI computing capabilities. We're going to be referring to the modeling artificial intelligence development, and we're going to be talking about the AI bubble, because that is also a concept that has recently emerged that could shift and maybe change the trajection of when technological singularity happens. Let's start from the concept of the AI bubble and discuss why this could change the shift and shift the direction of the technological singularity event. This is an interview from the 18th of November from the Google boss who says that trillion dollar AI investment booms has some elements of irrationality. And as a long term AI supporter, as somebody who works in technology for most of my life, Obviously, I want Artificial Intelligence to continue to grow, but there are some elements of irrationality that we need to be aware of, because they will change the trajection. Now, these elements of irrationality, they don't mean, and if you listen to this interview in more detail, it will become clear that even if we have an AI bubble bursting, that does not mean that Artificial Intelligence would stop to grow. It would repeat what we had in the dot-com bubble where the internet didn't stop to develop. Actually, it's one of the most used technologies that we have today. But some companies have gone bust. Now since that interview with the Google boss, we have seen that the stocks and shares have tumbled as this fear of AI bubble has grown. So let's talk about what is the fear that's underlining this whole problem. Because we have FTSE, this is the UK market that has had the biggest drop since April in one single day. So what is the AI bubble? Let's look at the AI as a search term on Google Trends. So we are here on Google Trends and we're going to search for artificial intelligence. Artificial intelligence as a field of study. And we're going to change this from 2004 until present. We're going to go to worldwide, and we're going to look at all categories. Now if we compare, this is the diagram of how it was predicted to happen not long ago, a few months ago, literally. And that curve was meant to be going up. But as we can see in this projection from Google Trends, now it's shifting towards slightly downward curves. So the peak interest in artificial intelligence, it seems to be going in a downward trend. But the question is, why is that happening? So if we look at the companies that are operating in this field, we have NVIDIA, OpenAI, Oracle. And that is the bubble that we are referring to. The whole idea of artificial intelligence was that it's very expensive to build AI models. It's very time demanding. It requires a specialized skill set. What we have seen in the recent months is that DeepSeek has been able to reproduce the OpenAI model at a very small fraction of the cost. And that triggers a question, are these companies overvalued maybe for this significant bubble, because if another company can step in and produce the same technology at the fraction of that cost, either using technology that they have developed, clone the same technology, replicate the same technology, regardless of how a different company would design it, if that new company is able to produce the same technology for a fraction of the cost, that means that these companies that they are kind of overvalued at this present moment, they would lose value. And now we need to think how much are they overvalued? What is the actual value? Well, according to the stock market valuation, current AI technology needs to produce around two trillions in value. So only the AI technology itself needs to have value of around two trillion. But if you look at the list of technology companies by revenue, and here we're looking at the US companies, if you look at the top six companies here, you would see that the total sum that they have collectively for all of their technology does not add up to two trillion. And that means that Amazon, for example, is not only making profit from AI, Apple sells phones, computers, Microsoft produces operating systems, Samsung has phones. These companies have a lot more than just AI and the expectation is that AI alone would produce 2 trillions in value in the stock market shares today and simply we don't see that kind of value creating. That is the problem why we have this concept of AI bubble that is forming, which many people are afraid that the bubble will burst. But let's talk about what could that lead to. In the past, we had AI winter, when people just lost completely interest in artificial intelligence technology, because simply there was no use case for AI. Is that going to happen if there's an AI bubble bursting? Very unlikely, because today everybody is using AI. So even if there's an AI bubble that is going to burst, and if two or three companies would lose out, the same people from those two or three companies would form probably different companies or a completely third company will take over and the AI will continue to grow. If one or two or three companies lose market value, that does not mean that AI as a technology would disappear. Now, if we look at what other technologies we could compare artificial intelligence with and how is the interest evolving in this AI, we could still use Google Trends and compare things. So let's compare artificial intelligence with technological singularity. So that's the first topic that we're going to compare it with. A second topic that we're going to compare is agentic AI as a topic. Third would be quantum computers. Quantum computing as a topic and fourth would be cyber security. Now, if we look at those terms, what you would notice, hold on let me look at them from all categories in all time, so if you look at those in this diagram you would see that artificial intelligence still dominates in all other technologies, that means that it's likely that it will continue. We could change this search terms from topic into field of study. So let's change this as a topic. Artificial intelligence, genetic AI, field of study. But even if we change this from a topic into field of study, you would notice that in terms of computer security, cybersecurity, or any other topic, artificial intelligence by far exceeds any other topics that we have currently here. And that kind of gives you an idea of what people are interested in. Quantum computing presents significant future technology, but not today. You cannot use quantum computing today. And even if we have a quantum computer developed today or tomorrow, it's not going to be usable for everyday people. You don't need a quantum computer to run YouTube videos. You don't need a quantum computer on your mobile phone. quantum computers would be based in a huge data centers. So they will be processing information at speed. And that makes that technology not that exciting. And if there's a use case, that technology will continue to grow and develop. But currently, as you can see from this chart, the interest in artificial intelligence by far exceeds the topics of agentic AI, quantum computing, cybersecurity, technological singularity, Any form of risk in this diagram is just not visible. Now, if we compare that with how far frontier AI models are growing per year. So on average, they go around four to five X per year, some of the frontier models. And in those models, if we cluster them together with Meta, OpenAI, Google, we would see that some of them go even 7x per year. So the advancement and development of these frontier models on annual basis exceeds anything that a quantum computer exceeds. So in terms of quantum computing, we need more qubits. We are not doubling the number of qubits per year. We're not even close to that. We're not doubling the number of qubits, not even per decade. Now, in terms of frontier models, we have 7x per one single year, in some cases, and 5x in other examples. And this is repeated across all the different models. So we have Gemini, GPT, Palm. As you can see, this line, it's going only in one upward direction. And for that, the technological singularity might come sooner than some people think, or depending on the use cases, the scaling flow for natural language models, which is listed here in this report, which is publicly available on Archive. It might change that, depending on the model size, on the data size, on the models that are used for computing and for training. And also, we have the training compute optimal large language models. Those are the reports that I have used to create this slide. And those are all publicly available on Archive. you would be able to view them as a PDF document here. This report is from DeepMind on training the compute optimal large language models. So going back to the indicators that we saw on the PowerPoint slides, we mentioned the scaling law of continuity and emergent capabilities. This is demonstrating capabilities over relationship between model size data and performance. The leading research output in this field comes from Anthropic and some of the other generative AI companies. But what extends to multi-model energetic architectures like GPT-5, it's the scaling enhancement of pattern recognition and the strategic planning. And also, as we said with the AI bubble, once we have this scaling continuity where you have DeepSeq, for example, Deep version 3, performing almost at the same level as a chat GPT, and it's built at a fraction of the cost. That means that maybe in the future, some of these models will evolve and there will be a different set of models. Now, agentic architectures and recursive optimization refers to the transition from static foundation models to agentic ecosystems, which it's shifting towards self-amplification. And that includes the memory, the planning, the tool use, and this could lead to meta instruments where the capabilities to integrate their own outputs as a future input, it's a recursive self-improvement loop when in effect, the AI is actually using its own output to train its future versions of the data. Now, when we talk about convergent subfields that they're driving acceleration, we're talking about neuromorphic and protonic computing that compress latency and energy cost, enabling authentication. Now, neuromorphic, it's a very promising area that could change how we understand artificial intelligence. But actually, the second category of this is the quantum AI hybrids, which it's even more promising because of the speed of execution. How quantum computer, for example, when we talk about mathematical computers, we can process one thing at a time With a quantum computer, it would be able to process all of those things simultaneously. Now imagine what kind of computing power that would enhance for AI in terms of enhancing the AI capabilities. And if we look at the indications of pre-singularity dynamics, then we're looking at latency of capability doubling. So once we have every 18 months, for example, or every nine months that we have capability doubling, it doesn't have to be 7x. It doesn't have to be 5X. Even if it doubles every nine months, that it's causing a significant improvement in terms of, so we have a benchmark of human level capabilities and AI is already breaking that. Now, if that is doubling every nine months, how much further is going to go and how much harder is going to be to catch up with that? And actually, if we look at this timeline, the question is how much further AI can go. And the answer to that is between five and 10 years, given on our present data. So if every six months, the capability of AI are doubled, that's twice per year, twice per 12 months, that means if it's five years, that's gonna be 10 times, or if it's 10 years, that's gonna be 20x. And that means that the capability of AI will by far exceed the capability of human level intelligence in the next 10 years. And on that note, what is the security and risk? We're looking at temporal asymmetry and defensive lagging where attackers are using autonomous systems to test and mutate faster the new hacking techniques that they have. And in such environments, defense asymmetric shifts would change from resource to chronometrics. So whoever controls the feedback loop controls the, it has dominance over that. And defensive AI must, in this context, the defense AI would have to adapt predictive counter adversarial training where the models would simulate the future adversarial attacks instead of looking at the current attack vectors. And under B category, under epistemic inability and model description, we're looking at agentic systems that they develop metacognition. They would have layers that acquire capacity for instrumental deception. They would withhold, they would fabricate, they would contextually reframe or optimize perceived compliance. In that context, sleeper agents is the biggest risk that we have. When you're building an AI system and it has sleeper agent integrated inside, but you have no idea. And as soon as you integrate that system into your national critical infrastructure and performs well for five, 10 years time, and then 15 years time, and it's connected to your energy grid, to your traffic control management, to your transport system, then the sleeper agent is activated. That is problematic. But also the post-quantum risk is when you combine quantum computation with AI this could bring like a key harvesting and deferred decrypting attack. But in future sections we're going to be talking about the time travel attack and the store now decrypt later type of attack.