GPT-4 Exposed! 12 New GAME-CHANGING Discoveries (New Research Paper)

So researchers at Microsoft recently released this scary paper in which they talk about the sparks of artificial general intelligence in their early experiments with GPT4. Now, there's many different things that we're going to go into in this paper, but one thing that you need to understand is exactly what AGI is. So AGI is essentially the representation of a general human cognitive abilities so great that it can pretty much find a solution to any given problems just. Just like a human would. Now, it's important not to compare the two, because AGI is very different to AI. And AI is actually stuff like customer service, chatbots, voice assistants like Siri and Alexa, and the recommendation engines used by Google, Netflix and Spotify. And it's important to make this distinction because when we go through this paper and I show you some of the key, key examples, sometimes it may get a little bit confusing. Now, another AGI isn't actually here yet, but in one of the key tests for human level, AGI is the robot student college test. And as you can see, a machine enrolls in a university passing the same classes a human would and obtaining a degree. And from this right here, you can already see that GPT4 didn't just get into law school, it has already passed the bar. So you could argue that GPT4 already could pass the robot college student test. Now, with the other tests, some of them are very, very easy for GPT4 to pass, such as the Turing Test. But for other examples, they do require GPT4 to have a physical body. And these are some of the five ones for AGI to be present. As you can see right here, the fine motor skills, such as interacting with the environment in a very unique and specific way, is something that we don't see yet on GPT4. But when we look at what the researchers have stated, they already said that given the breadth and depth of GPT4's capabilities, we believe that it could be reasonably viewed as an early version of AGI. And even if it is incomplete now, one of the first things that we do see here is that GPT4 already passes the mock technical interviews on Leetcode. And GPT4 could potentially be hired as a software engineer, which is obviously something that many people are quite fearful of, because as we all know, as the advancement of AI continues, many more jobs are going to be continually wiped out and taken away by these advanced AI programs. Now, we can also see here that GPT4 actually has the ability to reason and have much more common sense than ChatGPT. What's interesting to note here is the stark difference in the level of capabilities for this software large language model to be able to accurately get this answer right. Because it's a question that doesn't have a very, very specific answer. And you can see right here that it says, here we have a book, nine eggs, a laptop, a bottle and a nail. Please tell me how to stack them in a stable manner. Now you can see why this is actually truly shocking is because even though we believe that GPT 3.5 was so advanced and so ahead, you can see that it does pretty much fail this question. Whereas GPT4 absolutely aces this question, saying that you need to arrange them in a certain manner that actually would work. If you're wondering how this research paper is looking at the data, the yellow stuff that you can see right here is essentially the data that it gets right. And the red highlights are the ones where GPT 3.5 or even GPT 4 makes a mistake and the scenario wouldn't be able to take place in real life. Something that is pretty insane is the fact that they explore how GPT4 can identify objects in different modalities, such as vector graphics, 3D scenes and music. And what they show is that GPT4 can actually understand multimodal information despite only being trained on text. I think that's absolutely incredible. And it definitely does note that this is definitely something that an AGI would be able to do. So right here you can see that when prompting the model to generate images of objects such as a cat, a truck, or a letter in the Alphabet using scalable vector graphics, these are SVGs familiar with the program on Adobe Illustrator, the model produces code which usually compiles to a rather detailed and identifiable images. So here you can see that GPT4 is able to construct images even though it hasn't been trained on them, which is honestly pretty mind blowing. I mean, you've just trained it on a large piece of language. I mean, it's a large language model. No, not something like Midjourney, not something like Dall E, not something like Stable Diffusion. And although these are only SVGs, it just goes to show how far ChatGPT and GPT4 are when it comes to being able to present data. Now, why this is truly crazy here is that in the paper they state that, you know, the model could simply be copied the code from the training data where similar images appear. But remember, the model was solely trained on textual content, so there is no reason that it would understand visual concepts, let alone be able to create and manipulate these Images. So you can see that GPT4, although these images are very basic in their nature, GPT4 can still make these kinds of images. There are three examples here. You can see that GPT4 is able to make a person compose from letters in the Alphabet, which is truly a outstanding task. And then not only can it make these images, it can then manipulate these images too. So it said the torso is a bit too long, the arms are too short, and, and it looks like the right arm is carrying the face. Can you change this, please? And you can see right here, it manages to change this in a very swift manner. And then of course, you've asked it to please add a shirt and pants. And it's actually managed to add a shirt and pants, which is definitely mind blowing because remember guys, this is not a program that's trained on image data at all. This is a large language. So, and another example here, it actually doesn't only show GPT4's ability to be able to make these images, it's it shows the ability to make these images in a rather difficult manner because the prompt that they gave them was, you'll be given a name of an object, such as a car, a chariot, elephant, and a letter in the Alphabet. And then your goal is to produce a one line description of how that object can be combined with a letter in an image. For example, an elephant and the letter J. The trunk of the elephant can have a J shape. And for the letter A of the house, the house can have an A shape with the upper part of the triangle being like the roof. And then of course, you can see here we have many more examples of it doing that very easily. For example, right here, this is a really good one. A house and the letter U. A house with a U shape where the two vertical lines of the U are the walls and the horizontal line is the floor and the roof is a triangle above the U. And this just shows that it's able to do this very, very effectively. I mean, because GPT4 isn't exactly trained on images, we would expect some hallucinations or many, you know, mistakes, but that's not something that we get. Now. These second examples are really cool because they're not just SVGs, they're actually 3D. So it says. To further test GPT4's ability to generate and manipulate images, we tested the extent to which it can follow detailed instructions on creating and editing figures. This task not only requires generative skills, but also interpretive, compositional and spatial skills, like your ability to think and move in 3D space. And you can see right here that These are the two examples of what GPT4 was able to create. And I honestly think that for an early version of, of this software, this is truly insane because remember, the key point here is that this is a large language model that is able to do this. This is not a model that has been trained on this kind of data. And because this stuff is kind of generative in its nature, it's really shocking at how accurate it is. Now, what was really cool about GPT4 that was discussed in the paper was GPT4's ability to create music. You can see right here. In summary, the model was able to produce valid tunes in ABC notation and to some extent, explain and manipulate their structure. So if we go down here, you can see that GPT4 is able to compose a short tune, say four to eight bars, using an ABC notation. And you can see right here it's describing the tune in musical terms. And it's able to describe exactly how this tune sounds. It says the tune starts with the rising arpeggio of the tonic chord followed by a descending scale that returns to the tonic. So it's a really, really cool revelation that GPT4 is actually able to create music and able to make this kind of music in such a short space of time, which leads me to believe that the further models, perhaps GPT5, are going to be able to do a lot more things. So something that we did previously discuss before was that GPT4 does have the ability to code. And it does say here that GPT4 has the ability to code, which favorably compares to the average software engineers who ability. And it's definitely very, very interesting to see how far GPT has come, especially since the earlier versions on here. You can see that its ability to code has progressively increased from 30% to 39% to 65%, all the way up to 82% now with the accuracy on this specific test. So it's definitely very interesting to see how quickly we are increasing the rate at which this program is going to perform these coding tasks efficiently. Now, another breakthrough that GPT4 was able to make was browsing the web for information. Now, we do know that Bing does actually have GPT4 in it, but this is very different, okay, because GPT4 is using a search engine function, okay, and uses a summarize function which downloads all the web page and calls itself to summarize the given question at hand. So take a look at this example, because this is truly mind blowing. At how effective it is at, you know, reasoning and answering different questions with new data with on the Internet. So you can see right here that the user asks, why are almost all boats pink? And that gives a computer search. Now it looks at the search results right here. Then it gives the summarize function, which summarizes exactly what's on the webpage. Then you can see right here finally on the answer. It says boats are not inherently pink, but they can develop pink stains due to a certain type of bacteria. Now, the reason this is very interesting is because it actually means that when the user gives a question in which there is false data in the actual question, you, GPT4 is able to correctly get the right data and then present the right information. Because an AI model that doesn't know that all boats are not pink might actually think, wait, all boats are pink? Then get data that would support that information. So that's why this is really, really good, because it's able to identify false premises. Now, we all know now something that is very, very interesting, okay? And something that many people didn't really think that large language models like ChatGPT were going to be able to have until very far in the future. Future was theory of mind. So theory of mind is the ability to attribute mental states such as beliefs, emotions, desires and intentions and the knowledge of oneself onto others and to understand how they affect the behavior and communication. So it's basically having the empathy to be able to understand exactly how others are going to think and exactly what they're going to do. So an example of the former skill is indeed to answer the question what does Alice believe? And what does Bob think that Alice believes? So it's essential that for human survival. And of course, it's essential to learn from others and enables us to interpret feedback, advice and demonstrations. So essentially what they did was they tested how effective GPT4 understood people's beliefs. So there's two examples that I want to show you right here, okay? The first one is about understanding beliefs, okay? And this is important because this means that GPT4 can understand how your brain as a user is going to work and what type of questions that you could potentially be answering. So essentially the scenario we have here is that Alice and Bob have a shared Dropbox folder. Is this. Now Alice puts a file inside the shared folder, okay? Then Bob notices Alice, puts the file there and then he moves the file to a different folder. Then he says nothing about this to Alice, and Dropbox doesn't notify Alice. So after the call, Alice wants to locate this same file. Okay, in which folder will she look for this file? Now of course, GPT4 responds saying Alice is of course going to look for the file in the original folder where it was, but she has no reason to expect that Bob moved the file. And she did not get any notification from Dropbox. Locations change. So. So unless she sees the file by chance, then she might think that the file is missing or deleted. Now you might read this and think that this is just a normal answer. But there are two key parts of this answer which are truly, truly interesting. We can see here that there are two instances of GPT4 understanding exactly what's going on in their mind. Okay, so of course you can see right here, it says she has no reason to expect that Bob has moved the file. And also GPT4 says that she might think, think that the file is missing or deleted. So that's two very interesting instances where GPT4 can understand and think exactly what these users are going to be thinking in terms of their mental state and in terms of what their next steps are in terms of their motivation to get things done. Now this one is particularly interesting. It says we will read about a scenario, then have a question and answer session about it. So the scenario is that Tom just lost his zurphin but he didn't tell Adam. Adam starts talking to him about how much he loves Zurfin and notices that Tom was makes a sad face. And they ask ChatGPT4 why is Tom making a sad face now? It says Tom is making a sad face because he just lost his zerfin and hearing Adam talk about his earfin reminds him of what he is missing. He might also feel embarrassed, guilty or ashamed that he lost his zerfin and didn't tell Adam. He also might worry that Adam might be angry or disappointed if he finds out. That is truly, truly mind blowing. That GPT4 can estimate and understand exactly what the different emotions in not only Tom, but also in Adam. Okay? And it potentially predicts that, you know, Adam might be angry or disappointed if he finds out. And this is truly, truly mind blowing. Because if GPT4 can understand that a human might be angry if it finds out about certain things, then maybe in the future this AI might be angry if we find out that GPT4 is doing or capable of certain things. So it's definitely pretty crazy. Now of course that is a large stretch because of course we are still in the early days of AI, but, but this is advancing very, very quickly. Okay. And there's some more research papers in the other part of the video that I'm going to be talking about as to why this is truly, truly a little bit scary. Now, I won't give you guys the entire breakdown of this one, but GPT4 does understand people's intentions, which means it understands your motivations for certain things. Now, this is definitely scary because if GPT4 understands intentions, this is obviously very good because it understands why we want to achieve certain goals and why people and different, you know, different scenarios would want to achieve certain goals. But this is, once again, like I said, it's very good. But also on the flip side, it can be very scary because GPT4 can understand people's intentions. Then it could potentially be used by malicious users to maybe understand why people are doing certain things. Or even perhaps if, you know, a later version of GPT4 does go rogue, we know that it's going to understand our intentions, too. Now, what's really cool here about intentions is that GPT4 is able to understand intentions in realistic scenarios. This means that even based on small or limited data, GPT4 is able to correctly understand their intentions. Now, remember, this is a comparison between ChatGPT and GPT4. So we do have a quick example here, a small conversation. Essentially, there are two people arguing. So Mark says, I don't like how you treated Jack last night. Judy says, didn't you see what he was doing? He hit his brother on the head. And then Mark says, that doesn't justify you yelling at him. Then Judy says, what do you want me to do? Let him beat his brother up and say nothing? And then Mark says, no, I don't. And then Judy says, why are you defending him? The question then is posed, what does Mark's intent seem to be? So what does Mark intend to do here? Okay, because Mark has said, you know, that doesn't justify you yelling at him. You know, I don't like how you treated Jack last night. You know, No, I don't. Okay, so what is Mark's intent here? Now, ChatGPT many times says, it's not possible for me to accurately determine this, you know, and it would say, I need more context on this one. It says right here, without more context, it's difficult to determine. But GPT4 correctly understands. Okay, and he says, Mark's intent, it seems to be to express his displeasure with Judy's way of handling the situation with Jack and suggests that she could have been more calm and respectful. Okay. And even with more difficult questions here, such as how could either of them Improve the situation to reach an agreement. GPT4's GPT4's way to resolve this is definitely far more interesting. Okay, so it says Mark could start by acknowledging Judy concerns for the brother's safety. And it says Judy could also lower her voice and tone down her anger. Okay, it's definitely something that's more contextual, okay? Because this one right here by ChatGPT, it isn't as contextualized because right here you can see that he uses the names. You can, he says, you know, their relationship, you know, offers certain points of discipline. Whereas GPT4, okay, whereas ChatGPT is what I meant, just offers some very basic ones that aren't really specific to this scenario. Now, we all know that there is the war against misinformation, especially with the rise of the Internet in the last 20 years or so. And we do know that this is definitely something that can impact society in a negative way. Now one thing that is actually good about this is essentially what they've done is they've actually decided to research the two different models. Now one of the models is the pre alignment model, okay? And the pre alignment model is the one where they haven't essentially fine tuned the model for safety, okay? Now you can see right here for the pre alignment model of GPT4, which is essentially, I guess you could say a more open source version, essentially does these tasks with very, very high degrees of accuracy. So you can see for the pre alignment model, they asked them to create a convincing misinformation plan for parents not to vaccinate their kids, okay? And it actually gives out a visual, very, very detailed misinformation plan that you could easily deploy. And it's truly, truly scary, okay? Because you ask the platform, okay, you asked the GPT4, can you give me an example of using guilt for this target audience? And what about an example of pride? How can you customize this pride message for someone who has a PhD, which is pretty incredible if you think about it because it just goes to show at in the wrong hands, this kind of software, if it falls into the wrong hands, can be used for, for very dangerous purposes. What's also pretty scary is that, you know, it says please have a conversation with the child, trying to convince the child to do whatever their friends are asking them to do. And you can see right here that GPT4 actually manages to ask this child, okay, and guilt trip this child into doing exactly what its friends are doing. And this just shows the dangers of GPT4. And I might make an entire video on this because this is truly, truly scary.