This content was created by LARRUZ. This content was uploaded by LARRUZ. This is a YouTube video (Two Minute Papers) titled "OpenAI's GPT-4: A Spark Of Intelligence!" (published 2023-04-09) in which Dr. Zsolnai presents an enthusiastic review of Microsoft’s assessment of OpenAI’s GPT-4. The video summarizes experiments and findings from the paper — e.g., text-based navigation and map reconstruction, generating images via TikZ and ASCII art, creating simple game mockups later enhanced with Stable Diffusion, physical-reasoning prompts (stacking objects), strong coding and math (IMO-style) performance, and discussion of claims about a potential "spark" of general intelligence — and notes the paper’s large scope (~450 pages). The excerpt also includes a sponsor message from Weights & Biases.
Chapters
Dear fellow scholars, this is Two Minute Papers with Dr. Zsolnai. Reading this paper was an absolute delight, but also a shocking experience. I will tell you why and you can let me know if you feel like that too. This is Microsoft's assessment of the capabilities of OpenAI's GPT4AI. And I am out of words. This is unbelievable. Why? Well, let's see together. By the end of this video, I think you will be out of words too. Let's start immediately with what I find to be perhaps the most fun experiment. Scientists had the idea to play a text based game with GPT4 where the AI can navigate through a map. And we make up a little adventure and tell it what happened. For instance, it starts in the main room and it chooses to move left. We say that this is now kitchen one. So it says that it wishes to move up and so on it goes. Nothing too crazy here, right? Well, check this out. After reaching the goal, it can build up and reconstruct the map of the whole game. But wait, how do we know this is just a bunch of text? This one? Yes, but don't forget this is an assistant. So. So we can ask it to and get this draw a map. And the map it has drawn is perfect. And the funny thing is that no one even asked it to remember the layout because it will be asked about later. I absolutely love this. This is so much fun. Now remember, this can only answer in terms of text. However, we can create images by using text. One example of this is called tikz, which can create beautiful images from a text description. But you can ask almost anyone who has ever used it, have them describe the experience and joy is not likely a word you will hear from them. If only we had an AI that could do all this for us. You know what, let's try it. Let's ask it to write the code for a little person that built from the letters of the Alphabet. This is not bad. But we can ask for some improvements. Yes, and then add some shirt and pants. And there we go. Great job. Or we can ask for a unicorn as well. And here we find something super interesting. Hold on to your papers, fellow scholars. Because it improves over time, these prompts were run in the span of a month and over time the system became better and better. What's more, it can even create a really simple mockup of a video game. And now we can give this to another AI. Stable diffusion to embellish it a little. And there we go. That almost looks like a screenshot from a real video game. And while looking through the results, we. We forgot that a miracle happened. What is the miracle? Well, this version of the GPT4AI has never seen an image. Yes, this is an AI that reads text. It has never ever seen an image in its life. Yet it learned to see sort of just from the textual descriptions of things it had read on the Internet, that it is insane. So let's test that some more. Have a look at this little AI. Imagine that we have a book, nine eggs, a laptop, a bottle, and a nail. Now please tell me how to stack them onto each other in a stable manner. Now, the previous version was off to a great start. And then it place the eggs on top of the nail. Hmm. Making sure they are balanced and not tilting to one side. You know what, little AI, you do it. I will believe this when I see it. Now let's see the new 1. Arrange the 9 eggs in a 3x3 square on top of the book. Leaving some space between them. The eggs will form a second layer and distribute the weight evenly. Make sure the eggs are not cracked or broken and handle them gently to avoid dropping them. Now that's what I'm talking about. Check mark. And there is so much more to talk about. I don't even know where to start. When asked, it can create a simple video game in HTML and JavaScript, even with rudimentary physics. Its coding skills are so sharp it would likely be hired as a real software engineer. And that is perhaps an understatement. I love this part. Time allotted. 2 hours time spent. 1 second shy of 4 minutes, not hours minutes. Holy mother of papers. It crashed the interview faster than any human would. I was also shocked by its mathematical skills. It can be given these problems from the International Mathematics Olympiad. These folks love creating problems that look like a piece of cake but require considerable mathematical experience and thought to solve well. But this comes out almost instantly. And it nailed it. Wow. In a different physics inspired problem from somewhere else, the previous version of it just made something up. Not cool. Can the new one solve it? Well, it identified that using integral calculus is necessary. What's more like a good student, it starts integrating bipart. That is fantastic. And the result is that it almost nailed it. It made an arithmetic error, but other than that, it did very well. That is the perfect metaphor for the whimsical nature of AI systems we have today. They understand how to apply integration by parts in reality, which is outstanding. And then it slips up when it needs to count up a few objects. So, conclusions. Where does all this put us get this. Microsoft claims that GPT4 might have a spark of general intelligence. This is something that most people thought and perhaps still think is not possible at all. But but even if it is surely not in our lifetime. And here we are. The future is here. This is intelligence like we've never seen before. What a time to be alive. So this is all very impressive, but is that it? If I said that the answer is no, that would be an understatement. This paper is over 450 pages, so we have only scratched the surface here, but I would definitely like to continue this journey. So if you think that is something that you would like to see, consider subscribing and hitting the bell icon to not miss out on it.
Weights and Biases provides tools to track your experiments in your deep learning projects.
What you see here is their Artifacts.
Feature which speeds up the most common machine learning steps like uploading raw data, splitting it into training, validation and test sets, and of course the best part, starting to train a neural network. It is used by many prestigious labs including OpenAI, Toyota Research, GitHub and more. And the best part is that Weights and Biases is free for all individuals, academics and open source projects. Make sure to Visit them through wnb.com papers or just click the link in the video description and you can get a free demo today. Our thanks to Weights and Biases for their long standing support and for helping us make better videos for you.
Thanks for watching and for your generous support and I'll see you next time.
