OpenAI’s ChatGPT Took An IQ Test!

A short YouTube explainer by Dr. Zsolnai (LARRUZ) about ChatGPT (a GPT-3 variant) showing five surprising uses: generating SVG/vector descriptions, writing pixel-shader code (Shadertoy), producing game-related code (mirror placement), finding/fixing security bugs, and taking standardized/linguistic IQ-style tests. The speaker highlights impressive capabilities (including an IQ claim of 147), cautions that results are informal (not peer-reviewed), and notes rapid user adoption and integrations (menu-bar apps, browser plugins) plus a sponsor mention of Weights & Biases.
Questions this video answers
January 18, 2023.
147.
One million users.
OpenAI, Toyota Research, GitHub (as cited in the video).
Chapters
Dear fellow scholars, this is two minute papers with Dr. Zsolnai. This is the amazing ChatGPT, a variant of the GPT3 language model AI that was unleashed to read the Internet and now we can use it as a conversational agent. It has a lot of superpowers and today we can even subject it to an IQ test. In a previous episode we saw that it could write new song lyrics or even give us ideas for designing a living room. However, you see, this is a language model, so it can only give us text, not images. But no one said that we cannot plug in these text prompts into a text to image AI. So these fellow scholars did just that. And thus we could even see these beautiful ideas come to life. This was one of my favorite examples and it turns out we can even get ChatGPT to give us some beautiful imagery with some other tricks, and it can do even more. Here are five more amazing examples that I absolutely adore and would love to share with you. One, we can ask it to create the description of an SVG file which, which is a vector image format. This describes in text what kinds of objects go where and Are you thinking what I am thinking? Oh yes. This is an excellent way of using text to create images. So it can create this red circle, blue triangle, or even a face. Now if you think that is all it can do, then have a look at this. For instance, can it create a vector version of the Twitter logo? Wow, it can. So cool. Okay, okay. But I am a computer graphics researcher by trade, so I would love to see something a little more advanced. So what other piece of text is there that can create absolutely beautiful images? Of course, computer code. And hold on to your papers because here it is. Two, it can write proper pixel shaders. What is that? This is a website called shadertoy where we can submit a piece of computer code that is executed by our graphics cards and produces an image. Oh yes, an even more powerful way of taking a piece of text and creating beautiful images, videos, or even a computer game with it. And ChatGPT can even write these cool little programs too. They compile and run so well. Wow. Now you may be thinking, are these just some cute little test programs or are they actually useful in any way? Well, three, check this out. This example shows that it can even write computer code to place a mirror in our computer game. It works like this. Yet another example of using a tool for something that it was not designed to do. I absolutely love these. For when we are talking about potential applications in the industry, here is another one. Whoa. ChatGPT can be asked to find and even fix security holes in in a piece of computer code. That is such an incredible thing because it requires knowledge on what happens if we execute a piece of code, and even better, what potential weaknesses may arise. That feels like talking to an intelligent being until we ask it to multiply two numbers together. And it then fails spectacularly. It is so close, but not quite correct, one of the amusing properties of a modern AI system. Loving it. 5. So actually, how reliable is it? Well, it is not always that unreliable. A fellow scholar had chatgpt take the SAT test and this, dear fellow scholars, is just about the SAT score for an average human. Wow. And it can also take a linguistic IQ test. And if you have a closer look, these questions require a considerable English vocabulary. So what does this test think ChatGPT's IQ is? It says 147. Holy mother of papers. Now before we get too excited, note that these are not peer reviewed experiments, so I am hesitant to give them too much credence and I hope you are too. So please don't take these too seriously. However, I feel that these are likely early indications of something amazing coming up real soon. I really hope that we will see a few peer reviewed papers on the academic capabilities of ChatGPT in more rigorous tests soon. So the original GPT was an AI that was meant to read the Internet, learn English and maybe be able to finish incomplete Amazon reviews. And now it knows computer graphics programming, Cybersecurity has lots of general human knowledge and more. Just imagine what this will be able to do a couple more papers down the line if you have some ideas and let me know in the comments below. And now if you have a Mac computer, you can even add it to your menu bar to access it quickly and for free. How cool is that? Or plugins for all major browsers now seem to be available too. So until the follow up papers appear, let the experiments begin. ChatGPT has inspired a lot of you fellow scholars all over the world and and has reached a million users in an incredibly small amount of time. That was one of the most explosive transitions from a paper to a real world product to show everyone that the papers are real. As real as it gets. Weights and Biases provides tools to track your experiments in your deep learning projects. Using their system you can create beautiful reports like this one to explain your findings to your colleagues better. 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.










