No — today's AI is not superintelligence. Current systems are ANI (artificial narrow intelligence), brilliant in some areas and useless in others; AGI has not been achieved, and ASI does not exist yet, though AI has repeatedly outdone predictions of its limits. The priority is building control, governance and safety now, before capability outstrips our ability to understand and correct it.
Searchable transcript of Is AI Really Superintelligence? ANI, AGI & ASI Explained — IBM Technology (13:26). Search for a phrase, then click its timestamp to jump straight to that moment in the video.
Captions sourced from the original video on YouTube, published by IBM Technology. The video, its captions and all related intellectual property remain the property of their respective owners; AINotes claims no ownership. Provided for research, accessibility and search — see the Transcript Notice and Copyright Policy.
00:00 Here's a question for you. Is artificial intelligence the same thing as superintelligence? Well, as the saying goes, everyone's entitled to their own opinion, but not their own facts. I could call this marker a banana, but I think you're going to disagree with me about that. And terminology really matters, especially when we're talking about technology.
00:21 So let's see what the people working in the field of AI say about this particular question. Have we achieved superintelligence yet? So first let's start with defining what we mean by intelligence. While there's no single definition as to what is and is not intelligent, there is general consensus around the idea that it's more than just knowing a lot of facts.
00:43 I mean, we have databases that can essentially memorize everything we put in them, but we don't really think of them as being smart. So intelligence involves the ability to perceive, to think, to learn and to act. For example, a traditional computer program of the sort that most of us write, can be very powerful, but not necessarily intelligent, because everything it does has to be programed into it.
01:11 We have to write precise, step by step instructions, telling it every detail in order to get the right result. People are intelligent. Well, most of them are. They aren't programed, they learn. And learning is different than programing. We can be given a request and then we can figure out what the steps are and when we fail, we learn not to do it that same way again.
01:34 Well, again, at least sometimes we learn not to do it again. We don't program a child, for instance. We teach them. And that's not the same thing. As an exasperated parent, you may wish you could just download new software into the kid's brain, but trust me, there are a lot more fun when we get to teach them, and see how they apply the lessons they learn as they grow into adults.
01:58 Learning involves cognition and the ability to infer reason and deduce, among a lot of other things. That's really hard to produce in a traditional program without writing tons and tons of code, and then having to rewrite every time we want to extend the capabilities of the system. Intelligence is much more adaptable. So let's take this idea of intelligence and break it down a little further.
02:18 Let's start with perception. So this is the part that deals with how an intelligence system takes in and interprets the world. So it's gonna involve things like sensation, which is receiving information through the senses. Or if it's an AI system, it's through sensors. It could involve pattern recognition because here we're identifying patterns and looking for anomalies.
02:46 And that's part of our intelligence as well. We learn through patterns. Language understanding is another, where we're interpreting words and their meanings. And that's what this video is in fact about. Object recognition is another, where we're identifying things in particular entities and being able to recall what they are. Spatial awareness. So being able to understand positions and relationships of things in three dimensional space.
03:12 Context understanding, so that we interpret information in a particular context and that gives it additional meaning. And then social perception, where we're recognizing emotions, intentions and behaviors. So that makes up the perception part of this. Next is thinking. This is how an intelligence system processes information and reaches conclusions through things like reasoning, where we're applying logic, or inference, where we're drawing conclusions from available information.
03:45 Causal reasoning, where we're understanding cause and effect. Problem solving, where we're anticipating particular types of solutions and finding those solutions. Prediction, where we actually do this a lot. We're predicting outcomes based upon particular behaviors. I see all of this. This is what's going to happen next. And intelligent systems can do that.
04:07 Decision making, which is something that basically involves selecting among a set of alternatives. Abstraction, the ability to think on a higher level and identify general concepts. Creativity is another, where we're generating novel ideas and solutions that did not exist before. Up next, learning. This is how an intelligence system acquires and improves its capabilities over time.
04:37 It involves things like memory, so that it knows what it's been doing, it's storing that information and able to retrieve it later. Learning from experience, which is another big part of this. Improving through feedback. Knowledge acquisition is another capability, where we're acquiring new information and skill acquisition, where we're acquiring new capabilities.
05:02 Adaptation, where we're adjusting to changing circumstances. Transfer, where we're applying knowledge in one context to a different and new situation. Self-correcting, so that we're able to recognize mistakes and correct for those. Then this idea of metacognition, which is where we're understanding our own knowledge and our own limitations, what we can and can't do.
05:31 And then finally, let's look at action. This is how an intelligence system turns understanding into decisions, and ultimately into results. So planning is a big part of this. We're determining how we're going to achieve a particular goal. Goal pursuit, going after that goal, working toward those particular objectives. Navigation, so we figure out how I get from here to there.
05:56 Motor control, the ability to control physical movement and coordination, where we're able to coordinate multiple actions or body parts, or things that are interacting with the physical world. Manipulation, which is physically interacting with those objects. Communication is another, where we're conveying information or intent, and then adaptation in action, where we're adjusting behavior in real time.
06:26 So that's not even an exhaustive list, although I do feel pretty exhausted after going through the whole list. The point is, intelligence is complex and multifaceted, and its definition is as well. So that's intelligence. But what is AI or artificial intelligence? Well, facetiously, I could just say it's the opposite of NS, or natural stupidity, but that's a topic for another video.
06:50 The simplest definition I know of for AI is that it's the branch of computer science that deals with matching or exceeding human intelligence in a computer. In other words, doing all of that stuff in a machine. But here's an important distinction. Being capable of an intelligent task, doesn't necessarily mean possessing intelligence in the same way a human does.
07:15 An AI system might be extraordinarily good at one particular capability, while being completely useless at another. And that's where we get three terms that you need to know. Those three terms are ANI, AGI and ASI. First let's talk about ANI. So artificial narrow intelligence, this is what AI designed today is able to perform particular classes of tasks and do them very well.
07:50 For instance, a chess engine can play chess really well. Our recommendation system can predict what you might want to watch next. A speech recognition system can turn speech into text. A modern large language model can perform astonishingly broad range of tasks, but it still has limitations. The key characteristic of ANI is that its capabilities are still bounded by what the system has been designed and trained to do.
08:17 This is what we have today. A genius in some areas and an idiot in others. The same system that can do one intelligent task has no real clue how to do the rest. Let's take a look at what I mean by that. So if we graph out capabilities, then we'll set the baseline as being what humans can do. So, this is capability here, and this is what a human is able to do.
08:39 Now, an ANI system might be not so good at this, but really good at this particular task. Or maybe it's only about as good as a person here, but all the rest of it it's really not good at. That's artificial narrow intelligence. AGI, artificial general intelligence is a much higher bar because this refers to an AI system with broad general purpose capabilities, the ability to learn and perform across many different domains, rather than being limited to just a single task or 2 or 3.
09:21 Here we're talking about a system that basically matches humans across the board, so there may be some areas it's a little better, but it's mostly us. And that's what artificial general intelligence is. We don't have that yet. Let's be clear about that. And some people think we never will. Well, as I said in a previous video about the limitations of AI, don't bet against AI, because it seems that most of the things people have said over the years that AI could never do, well it's now doing that stuff.
09:51 So stay tuned on that one. Now, if we get to the really interesting one, that's this artificial superintelligence. Here we're talking about a system whose intellectual capabilities are substantially exceeding those of a human. Here we go with ASI, and it's doing everything across all the cognitive domains better than a person could. This is superhuman intelligence in all the areas that make up what we consider to be intelligence.
10:22 So is today's AI superintelligent? The short answer is no. Today's systems have remarkable capabilities, but they also have significant limitations. They can make factual errors, we call those hallucinations. They can struggle with reasoning in certain situations. They can lack reliable understanding about the physical world around them, and their abilities can be uneven, extremely impressive in one particular context, and surprisingly poor in another.
10:54 We are definitely not at the AI side level yet, but here's a really important question. Can we control something that's actually smarter than us? The answer is that we'd better learn how and I think we can and will. How do we make sure that increasing intelligence remains aligned with human interests? Just remember this equation. Capability minus control equals risk.
11:21 We put guardrails around powerful technologies because capability without control creates risk and AI is no different. We need technical safeguards. We need security. We need testing and evaluation. We need governance. We need humans to remain in the loop and in control of everything that we're doing with regards to these decisions. What we don't want is for AI to be driving the car with us, locked in the trunk.
11:50 AI doesn't have to be evil for that to happen. It just has to become more capable than our ability to understand, control and correct it. We've been spending a lot of effort on teaching AI to be smart, but that's not nearly the same thing as teaching it how to behave, how to be ethical, how to do what's right. We need to teach it not only the facts, figures, and reasoning, but also give it an ethical framework to operate within.
12:17 You learned that growing up from your parents, teachers and friends. AI is still going through its teenage years now, and moments of brilliance and others, well, not brilliance. Speaking as a former teenager, I know. We need to make sure it has some adult supervision, maybe even a curfew. We need a kill switch with a human in the loop to flip that when it's necessary.
12:43 That way we end up with an AI that augments human intelligence rather than detracting from it. So is AI really superintelligence? Is AI really SI? Well, the simple answer is no. At least not today. But we need to be prepared to govern it. And when we ever do get to that point, if we ever do, the good news is we can do this because we've got a lot of work to do, but we've also got a lot of tools to help us get there.