Ah, so great that you are still here. Um, so it's on super intelligence. Nine years ago I gave a talk in Berlin at a completely unknown >> [snorts] >> new conference called Rise of AI. And the title of my talk was super intelligence. Will machines be smarter than us? So, what has changed since then? This, by the way, was my starting slide nine years ago. And today I promise at the very end I will show you the succession of this, yeah. I will Okay. So, back then that's the last jump back. Back then was the time of narrow AI. There was no generative AI, no foundation models, no LLMs. And every AI system could only do one thing.
And everybody agreed this is the case because no AI system has all the knowledge. So, how can we get the knowledge in? By programming, knowledge engineering, we didn't know. There was no way, but everybody was clear, if you bring together the big data handling the AI training uh of the performance of the system plus the knowledge machine learning together, then we would have intelligence, yeah. So, that was But then in German there is the saying, but first things turn out different and second than you think. Uh things were a little different because uh 1 month after this talk uh the transformer paper appeared. And at this time in 2017, nobody, not even the authors of the paper, knew how much potential was in there.
So, it took again a couple of years, and then we saw knowledge is not the bottleneck anymore. So, what is it? So, we have to in in order to understand this, let's look for a moment in how people learn. Humans, how do we learn? We have two ways of learning. One is behavior, like routine skills, like walking, uh like dancing, things like this. Riding a bicycle, swimming. And then we also [clears throat] have another way of learning, and this is the type of learning where we learn most of our explicit knowledge. This is through language. We learn it in school, at the university, from television, from books, and so on. So, this those are the two ways of learning.
They're quite different. One difference is for instance, for the second one, it's enough that you are told something once. If it's important enough, you will remember. But if you learn the behavior way, once is not enough. You have to be told many times in order to let it be important enough for your system to be guided by this. So, how do we learn language then? Is it this way of learning, or is it that way of learning? What do you think? Well, it's both, but most of the stuff the child can already speak and use the full complexity of grammar before ever having entered the explicit learning stuff. That's the thing, yeah. So, we learn language actually by in situations, by talking to the parents and And what's the difference then between this and the LLMs?
How do LLMs learn language? They learn language in a way like humans as they learn language as behavior, not explicit way, but in a different way. So, the context that tells the machine the meaning is the textual context, and the human learns it together with a meaning in a situation. That's the important thing. So, human language learners cannot learn anything, no new concepts, no language without having the meaning at the same time. If you don't give it, it will not stick in memory. So, but in in the LLM, probabilities are learned layer over layer, millions, billions of times until everything falls into place. Completely different. There's lots of learning the first couple of millions of text, the system doesn't know anything.
And after some millions more, the system I mean, you get the feeling, right? The system starts to think, "Oh, I've seen this again. Now I finally I think I start understanding the world." But, of course, the system doesn't understand the world. It understands the world of text, the world only indirectly. So, that's important to know. So, AI today today is super has super knowledge. The knowledge is so much bigger than human knowledge. Not even 10,000, not even 50,000 humans, even well selected, can match the knowledge of one of the largest LLMs. That's for sure. Super knowledge and super eloquence. There's no human who can write uh job applications uh novels, poems, haikus everything in more than 100 languages.
It doesn't exist, yeah. So, so the system is much more eloquent than any human. So, but on the other hand so it then can translate and summarize and correct and you reformulate. It can deal with language quite a bit, but it still lacks autonomy and memory. It is a little bit like an Alzheimer patient. Yeah. Only knowing the old knowledge and everything new forgotten in a second again. And it's like a narcoleptic Alzheimer patient because you ask it something and after that falls asleep. You have to wake it up with a new question, yeah. So, it doesn't continue thinking. So, this lack of autonomy and this lack of memory is the biggest short come. How do we manage then to let it solve very important problems in industry and life by giving it agents, yeah, or by calling it many times if you have the time.
Uh but usually you give it an agent. The agent maybe incorporates a workflow and it calls this system, but the advantage of the agent is the agent can also call databases, do rag, it can also call the internet, it can call another language model, and we give it a memory but it kind of a context memory only of the context otherwise it would not remember what you have asked before. That's what we do. So that means >> [snorts] >> now we have super knowledge but outside is still thinking. We don't have thinking LLMs. We don't have planning because planning means thinking and we don't have learning through interaction because the system has Alzheimer's forgets everything so it doesn't learn through interaction.
Plus also the things below the line it doesn't have. It doesn't have interpersonal social skills because it didn't grow up in a family and it was not equipped with our innate capabilities. It has no fully understood world experience, no common sense in a broad sense. It has no instincts, no drives, no emotions. It has no embodiment, no sensory input. It has no consciousness. So it has almost nothing. These systems are so smart and they have almost nothing. So intelligence researchers in the psychometric psychology area they talk mainly that was the old Cattell system about two types or two main components of intelligence. One is the crystallized intelligence and the other one the fluid intelligence. The crystallized is all the knowledge that you have learned.
It's crystallized. Yeah, it's knowledge, it's wisdom. It is procedural knowledge. It is explicit and implicit knowledge and the fluid intelligence is what you make out of it. How do you deal with the knowledge and how you gain new knowledge, how you solve problems that you didn't have and so on. Yeah, How to So, this is all the And what is tested mainly in IQ tests? Not so much knowledge. It's mainly fluid intelligence. So, you get a high IQ test if you have very good fluid intelligence, not if you have so much crystallized. But if you apply for a job, the crystallized knowledge is very important. If you don't have it, yeah, it's very bad because you need it for your job.
That's the difference. And the crystallized intelligence, you decr- increases with age. You learn more and more. And the fluid intelligence decreases it, yeah? So, Nobel Prize winners in theoretical physics or mathematics usually don't do their contributions very late in age, but very early, yeah? So, that's a typical thing. So, now come to superintelligence. We are now at the point where we can tackle it. Uh one of the first um definitions came from Nick Bostrom, uh who said superintelligence is an intellect that is much smarter than the best human brain in practically every field, including scientific creativity, general wisdom, social skills. And I have to tell you this, in that sense of superintelligence, no way that we get it anytime soon.
But let's try a redefinition. Ah, let me first say it should exceed the human in quality and speed, and then it has this great potential. It can predict and optimize. It can even drive scientific uh innovation because it can feed itself. And it has recursive self-improvement. That's the thing also, yeah? It can improve itself, though. So, that's all that's Nick Bostrom's That's the idea about superintelligence. And in that sense, no way that we are going to see it fast. But, if we change it a little bit, if we say superintelligence could be an intellect that is much smarter than the best human brains in those dimensions that are essential for measuring intelligence. Measuring intelligence. So, what we think is important for getting be highly then I have to tell you that I believe and now in the rest of the talk I will explain why, that we are closer to this than we think.
That we are closer to that than we think, yeah. So, that will be the core of the talk. So, actually after the Cattell model of uh intelligence, which had the fluid and uh crystallized intelligence, uh other people like Horn and Cattell extended that. And this is right now the psychometric uh ruling theory, the CHC after Cattell, Horn, and Carroll uh uh theory of intelligence. It says there are many components. It starts with fluid and crystallized, but then there are these other points. And it turns out that uh if you go numerically, uh the first two account for in more than 90% of the intelligence. And uh the crystallized for maybe 80% of that. So, maybe 80 85, depending on the authors and the tests.
So, interesting. So, the fluid is the strongest part of intelligence. Then the the the the the knowledge place is very important. And then lots of other things contribute contribute to that, yeah. And now the question is simply how close are we now? And of course, there are people who have worked on that in the area of AGI because before super intelligence um people who search in this area have defined AGI the artificial general intelligence. General intelligence is like Nick Bostrom's definition, but not with that we exceed human intelligence, but that we reach it in almost every aspect. But then some people who tried now to measure how close are we, and they said uh the social intelligence, we cannot really test the system.
How should it have social intelligence? How should it Okay, not all of spatial reasoning. So, they took those elements that could be theoretically in a system. And then they measured how close how close are GPT-4 and GPT-5 to human measurements. And they they they just took intelligence tests, yeah, from the psychometric from the research tradition in psychology. So, then they took these and they said, "Look, this is how close we get." And the outer limit here, the outer limit of this kind of spider web is the 10, and that would mean 100% it would reach human intelligence. Now, look at it for a moment. Something you don't like? I have something that I don't like.
Why does GPT-5, I see in my slides here, I should look here. Why does GPT-5 not even get to human performance? Yeah, because they didn't check whether it exceeds human performance. They gave it some tasks that humans can solve. And they didn't give it tasks that humans cannot solve. And there were some parts in the especially in the social area and in the cultural area that the systems didn't know. So, they said, "Too bad. Uh GPT-5 is not as smart as not as knowledgeable as humans." But this is of course crazy, yeah, because GPT-5 is so much smarter in knowledge than any human. But this is how they measure. And you can see now that reasoning, there's a little bit of reasoning.
Why? Because the system can already in one step do some reasoning. And secondly, because you can call it several times, but actually there's not very good reasoning, yeah. And the memory is missing. That's why you get all the memory stuff, the storage, the retrieval very very bad in memory, because there's no memory in there. No no no no problem. It's clear. So, I I reformulated this it's the same results in a slightly different way and I show you in a moment why I did it, yeah. But before I go to that, uh let's briefly look at some new let's go to the main labs in the world, Google, Meta, the Chinese labs, and see what is the next generation of systems.
Because it is so obvious that the memory part is missing. This is very clear. It's the memory part. The working memory is the most defining part of human intelligence. And this is not there. So, we look we find two new lines of research. One comes from the Google people, and not surprisingly what they do is they take the transformer and they add memory. But not memory outside like uh OpenAI does with ChatGPT, but they added really as a neural network that learns while doing. Not just is updated, not just a static memory of context, okay? But it really learns. So, that is what they do at Google research. And then there's another line of research mainly at Stanford and Nvidia the TTT um test time training areas.
They That's models that work partially with recurrent neural nets and learn from data partially while they are testing. So, the testing and learning is fully integrated. Or the Moonshot AI the Kimi one of the smartest systems in China is also using one version of it. The linear the Kimi Delta attention is using recurrent neural networks for learning continuously. Or LS X LSTM You remember Schmidhuber and Hochreiter LSTMs very important, yeah? So, Sepp Hochreiter has now further developed it and has put some of the power of the transformer in there by saying we had it all the time. We didn't need the transformer because Schmidhuber and Hochreiter have done everything already as you know before everybody else.
And so >> >> they're known for that and but it's an interesting model and he said, "I simply don't have enough I don't have enough money. If I had more money, I could show that it blows all the others out of the water, yeah?" And who knows? Nobody can see right now who of these things is win winning. My guess would be that the Google people are a little closer because they don't need to uh redefine the whole thing. They have the power of the transformer plus the additional memory. They may come first, but in the long run, I would bet on recur on recurrent neural networks because that's like the human brain works, yeah.
Uh like the human for activation. But this is very important to know. So, this will come. I will not go in there, but in detail, but for instance, the the Google system So, from the past data stream, you build a long-term memory. Then you have the current context. And then you learn, you use still the attention mechanism, but you have the feedback. The feedback is not shown here, but from it gets a feedback into the long-term memory, and then you get the output prediction, yeah. And the nested uh learning is even more interesting because in the nested learning uh and it's a terrific, it's a brilliant system uh in the whole architecture. There's this architecture that they could show, not yet at the same size as GPT or Claude or Gemini in smaller systems, but they have already systems that have not only a memory, but that can even update the old knowledge, the existing knowledge, continual learning.
Why is that such a difficulty, by the way? Uh one of the difficulties is because it learn this system learned by thousands of occurrences, of instances. And if you now give it something only once, yeah, you tell them, for instance, uh Mario Adolf is dead, yeah. So, you tell it. But before it had so much other information. Now the system should say, "Why don't you tell me a thousand time, and then I start believing it?" You you do that. So, the thing is what do you do with new knowledge? Do you destroy old knowledge or not? If you don't if you then it will not change anything. So, you need to give it very high weight, but then also lots of noise comes in and you get a phenomenon which is called catastrophic forgetting or model collapse.
The new data bring in and destroy all the old data, yeah? So, this is a well-known thing. This makes it so hard to do continual learning because the machine doesn't have explicit knowledge. It needs to learn the behavior by training, yeah? You cannot just tell it how it's different, yeah? So, that's the thing. So, now let's imagine you have all this new memory. Then I would predict you we would get something like this. In the middle you see what the human can do because right now already the knowledge is far exceeding and the memory I mean all all psychologists agree that the only reason why we cannot program much better, why we cannot solve physics problems and math problems is mainly because of our small memory.
But here we can build huge memory compared to people, yeah? So, the advantage and and if I had more time I could talk about this, but I cannot. So, still not modeled and still not modeled is interpersonal social skills, instincts, embodiment, consciousness. Question is of course will explicit knowledge emerge? Or will functional counterparts to consciousness emerge? That's an open philosophical question. Will that come? Is it avoidable or will it not come? We don't know. So, now important questions. How dangerous is it? Each of these questions would be an own talk. Very dangerous. I think I really believe I'm I'm not I I usually don't like if people warn too much or create panic, but I think it's very dangerous.
Very dangerous to a point that it will be only AI that can control it. Humans, if that comes, we have no chance anymore. We need to have the smartest AI to control it. The second, do we need to master it? Of course. I mean, >> [clears throat] >> it's the smartest technology whatsoever. It's the most dangerous and most powerful. If we It has nothing to do with political sovereignty. We are crazy and dumb if we don't want to master it ourselves. If we leave it to other parts of the world. How can it be educated? Only by a massive interdisciplinary effort because educating such a thing Everything will depend on the education of the thing.
And then the last one, isn't it the wrong technology because of the ecological concerns? Because it gets even worse in energy consumption. But now I will tell you something. If you think that if you think that LLMs use up much much electricity, much energy, I have to tell you they don't. They don't. It's an illusion. The illusion comes because we concentrate all the energy consumption. We have, let's say, 300 million, now GPT-5, 800 million users. We concentrate all the stuff for them at one point, one power station in training and in also providing providing the knowledge or inferencing, called, yeah. We concentrate it and we concentrate millions of texts or billions of texts in there.
That's why the energy is so great. If you now do books, if you only give to 300 million people who were the users of GPT-4 or today with 800 will be bigger, only send them print one book for each of them, only one book, you already get much more energy consumption. So, our old technologies, books, were much more energy consuming than this one. So, please, if you get argumentation with those people who tell you that it uses much energy, it doesn't. No, it simply doesn't. And the same is true with water. I calculated this also with water. It's the same thing. So, [snorts] it's only even if we provide every user of GPT-5 with one egg each, it will be much more energy than the whole training of the system.
That Yeah. Okay. >> >> So, what about Europe? Any hope for AI leadership? Of course, we have lots of hope for B2B application, lots of hope for AI uptake in industry. Modest hopes. We heard today about public administration. But little hope for catching up with USA and China. Why? There are good things, and I wish I could talk about them, but I need to finish about the next frontier AI initiative of the EU and say a little bit whether I think and and the uh SPRIND next frontier next frontier AI initiative of SPRIND and the the frontier AI initiative. Wonderful programs, especially the one I have to say here is coming something from Germany. Unfortunately, it will not come overnight.
It is a 24-month thing, but it could be really it could lead to something. And please look into that. It's it's worthwhile. It's really worthwhile. Um Did we miss the train in GenAI? Yes, some trains have left the station, but as Antonio Krüger, who talked today, said, in AI, there are always new trains departing, and it's true. It depends on whether you can take them. So, deep seek nested learning, all that stuff is new that came in the last few months. What we What would it take to catch up? And it's not as much as you think. It's only one or few, maybe two, three labs that have a fair chance. That's all. Because if we have the talent in Europe, and we can get it.
If we can pay the people a decent salary, they don't even need to get the highest salary, but we cannot right now. If we had one or two, three labs that would have the compute power of the labs of Google and Meta, where people can also experiment and get into that, we would not even have to worry. We would not have to give them tasks or to supervise them or watch them by the government. The It It would happen. We would be up in the frontrunner group. Instead, we are talking and talking and talking and talking. Yeah, so this [snorts] is the thing. So, we cannot rely on big industry because they may change their plans.
SAP may say one day, we need our own model, and then maybe a year later, they say, "No, no, maybe we just go and do a strategic partnership with somebody else." Small industries, they will be bought by the Americans before they can scale. So, we need to do a social effort, a big social effort. So, and all it takes is top talent, locally grown, repatriated, newly attracted. And the compute infrastructure that is not enough and that is a mistake the politicians make. Not just enough or that the HPC people make. Not just enough to train the next model, but to let people experiment. And otherwise, we don't get the talent that's good enough to try the new things.
If that is not there for experimentation, just giving them like the EU they are calculating. Oh, can we afford it? Uh how much is needed for the final training of GPT-4? If we give exactly that to the people, will be No, it won't be. It's not it Maybe maybe they get a copy out of GPT or maybe they fail like well-known companies in Europe have. Uh and that's it. That's the end of it. Yeah, there's nothing coming up. Okay. Yeah. That's all. So, what about then superhuman performance? Now, let's look. Oh, what's happening? >> I let him win. He thinks I'm a dumb human. >> Okay, thank you for your attention.